[WebAssembly] Add passes for GEP lowering

[samparker]

Summary

At optimization level -O2 and above, add SeparateConstOffsetFromGEP, EarlyCSE, InstCombine and LICM to the addIRPasses phase.

These passes will also enable D139415.

[sbc100]

llvm/test/CodeGen/WebAssembly/fpclamptosat.ll

184

What happened here?

[samparker]

llvm/test/CodeGen/WebAssembly/fpclamptosat.ll

184

Yeah... I need to look at these clamping files. I'm assuming CSE is responsible, but the reasons for the changes are non-obvious to me.

[samparker]

llvm/test/CodeGen/WebAssembly/fpclamptosat.ll

184

Looks like DAGCombiner is now failing to produce fp_to_uint_sat, I will investigate more tomorrow.

[samparker]

Hopefully the FP clamping and min/max idiom misses have now been fixed in D142093, D142481, D142535, D143106 and D143256.

With my local benchmark suite, using wasi-sdk and running on node, this patch affects slightly less than half of the benchmarks and gives a geomean improvement of ~5%.

[tlively]

Impressive performance wins! How is code size affected?

[samparker]

How is code size affected?

Using wasi-sdk-19, the total code size of my suite is 7985 KB. With this patch it's reduced slightly to 7982 KB.

[dschuff]

Do those figures include using wasm-opt after linking?

[samparker]

Do those figures include using wasm-opt after linking?

I don't know... I just use wasi-sdk and I assumed it just uses clang, llvm and lld.

Though, after reporting the (small) code size changes, the reported performance numbers (large) didn't make sense. I'm still investigating what's going on.

[tlively]

The clang driver will automatically call wasm-opt if it finds it in your PATH, so it's worth checking whether wasm-opt is being run or not.

[junparser]

llvm/lib/Target/WebAssembly/WebAssemblyTargetMachine.cpp

466

have you ever test with createSeparateConstOffsetFromGEPPass(true)

[samparker]

Sorry for the delay in replying.

I found it impossible to reproduce my previously quoted numbers. I'm juggling an evolving benchmark suite, while modifying wasi-sdk and NodeJS, so I'm guessing I got something out-of-sync!

Below are my revised numbers, with the varying configuration options for completeness. I'm running on an Ampere Altra, with a modified version of NodeJS which uses trap handlers for OOB checks (the patch is currently in review for V8), and I'm only using the TurboFan compiler to reduce noise. wasm-opt is not on my PATH. The main takeaways are:

• LICM is no longer included.
• Running InstCombine increases code size a little, but I haven't included any of the patches to mitigate the known problems with running it late.
• Lowering GEPs, in SeparateConstOffsetFromGEP, increases code size and execution time.
• OptimizeGEPs still gives a useful improvement in size and execution time.

Pass Config	Total code size (KB)	Speedup (%)
vanilla wasi-sdk-16	7809.192	-
SeparateConstOffsetFromGEP(false), EarlyCSE	7805.771	0.1
SeparateConstOffsetFromGEP(false), EarlyCSE, InstCombine	7810.446	0.1
SeparateConstOffsetFromGEP(true), EarlyCSE	7819.303	-0.22
SeparateConstOffsetFromGEP(true), EarlyCSE, InstCombine	7819.988	-0.24
SeparateConstOffsetFromGEP(false), EarlyCSE, OptimizeGEPs	7798.139	1.78
SeparateConstOffsetFromGEP(false), EarlyCSE, OptimizeGEPs, InstCombine	7802.854	1.7

llvm/lib/Target/WebAssembly/CMakeLists.txt

add_llvm_target(WebAssemblyCodeGen
   intrinsics_gen
 
   LINK_COMPONENTS
   Analysis
   AsmPrinter
   BinaryFormat
   CodeGen
   Core
+  InstCombine
   MC
   Scalar
   SelectionDAG
   Support
   Target
   TargetParser
   TransformUtils
   WebAssemblyDesc

llvm/lib/Target/WebAssembly/WebAssemblyTargetMachine.cpp

 #include "llvm/CodeGen/Passes.h"
 #include "llvm/CodeGen/RegAllocRegistry.h"
 #include "llvm/CodeGen/TargetPassConfig.h"
 #include "llvm/IR/Function.h"
 #include "llvm/InitializePasses.h"
 #include "llvm/MC/MCAsmInfo.h"
 #include "llvm/MC/TargetRegistry.h"
 #include "llvm/Target/TargetOptions.h"
+#include "llvm/Transforms/InstCombine/InstCombine.h"
 #include "llvm/Transforms/Scalar.h"
 #include "llvm/Transforms/Scalar/LowerAtomicPass.h"
 #include "llvm/Transforms/Utils.h"
 #include <optional>
 using namespace llvm;
 
 #define DEBUG_TYPE "wasm"
 
void WebAssemblyPassConfig::addIRPasses() {
   // transformation algorithms with Emscripten SjLj, so we run
   // LowerEmscriptenEHSjLj pass also when Wasm SjLj is enabled.
   if (WasmEnableEmEH || WasmEnableEmSjLj || WasmEnableSjLj)
     addPass(createWebAssemblyLowerEmscriptenEHSjLj());
 
   // Expand indirectbr instructions to switches.
   addPass(createIndirectBrExpandPass());
 
+  if (getOptLevel() >= CodeGenOpt::Default) {
+    addPass(createSeparateConstOffsetFromGEPPass());
+    addPass(createEarlyCSEPass());
+    addPass(createInstructionCombiningPass());
+    addPass(createLICMPass());
+  }
+
   TargetPassConfig::addIRPasses();
 }
 
 void WebAssemblyPassConfig::addISelPrepare() {
   WebAssemblyTargetMachine *WasmTM = static_cast<WebAssemblyTargetMachine*>(TM);
   const WebAssemblySubtarget *Subtarget = WasmTM
             ->getSubtargetImpl(std::string(WasmTM->getTargetCPU()),
                                std::string(WasmTM->getTargetFeatureString()));

llvm/test/CodeGen/WebAssembly/PR41149.ll

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
-; RUN: llc < %s -mtriple=wasm32-unknown-unknown | FileCheck %s
+; RUN: llc < %s -O1 -mtriple=wasm32-unknown-unknown | FileCheck %s
 
 ; Regression test for PR41149.
 
 define void @mod() {
 ; CHECK-LABEL: mod:
 ; CHECK:         .functype mod () -> ()
 ; CHECK-NEXT:    .local i32
 ; CHECK-NEXT:  # %bb.0:

llvm/test/CodeGen/WebAssembly/address-offsets.ll

 ; TODO: load_test1 - load_test8 are disabled because folding GA+reg is disabled
 ; (there are cases where the value in the reg can be negative).
 ; Likewise for stores.
 
 define i32 @load_test1(i32 %n) {
 ; CHECK-LABEL: load_test1:
 ; CHECK:         .functype load_test1 (i32) -> (i32)
 ; CHECK-NEXT:  # %bb.0:
+; CHECK-NEXT:    global.get $push2=, g@GOT
 ; CHECK-NEXT:    i32.const $push0=, 2
 ; CHECK-NEXT:    i32.shl $push1=, $0, $pop0
-; CHECK-NEXT:    global.get $push2=, g@GOT
-; CHECK-NEXT:    i32.add $push3=, $pop1, $pop2
-; CHECK-NEXT:    i32.const $push4=, 40
-; CHECK-NEXT:    i32.add $push5=, $pop3, $pop4
-; CHECK-NEXT:    i32.load $push6=, 0($pop5)
-; CHECK-NEXT:    return $pop6
+; CHECK-NEXT:    i32.add $push3=, $pop2, $pop1
+; CHECK-NEXT:    i32.load $push4=, 40($pop3)
+; CHECK-NEXT:    return $pop4
   %add = add nsw i32 %n, 10
   %arrayidx = getelementptr inbounds [0 x i32], ptr @g, i32 0, i32 %add
   %t = load i32, ptr %arrayidx, align 4
   ret i32 %t
 }
 
 define i32 @load_test2(i32 %n) {
 ; CHECK-LABEL: load_test2:
 ; CHECK:         .functype load_test2 (i32) -> (i32)
 ; CHECK-NEXT:  # %bb.0:
+; CHECK-NEXT:    global.get $push2=, g@GOT
 ; CHECK-NEXT:    i32.const $push0=, 2
 ; CHECK-NEXT:    i32.shl $push1=, $0, $pop0
-; CHECK-NEXT:    global.get $push2=, g@GOT
-; CHECK-NEXT:    i32.add $push3=, $pop1, $pop2
-; CHECK-NEXT:    i32.const $push4=, 40
-; CHECK-NEXT:    i32.add $push5=, $pop3, $pop4
-; CHECK-NEXT:    i32.load $push6=, 0($pop5)
-; CHECK-NEXT:    return $pop6
+; CHECK-NEXT:    i32.add $push3=, $pop2, $pop1
+; CHECK-NEXT:    i32.load $push4=, 40($pop3)
+; CHECK-NEXT:    return $pop4
   %add = add nsw i32 10, %n
   %arrayidx = getelementptr inbounds [0 x i32], ptr @g, i32 0, i32 %add
   %t = load i32, ptr %arrayidx, align 4
   ret i32 %t
 }
 
 define i32 @load_test3(i32 %n) {
 ; CHECK-LABEL: load_test3:
; CHECK-NEXT:    return $pop6
   %t = load i32, ptr %add.ptr, align 4
   ret i32 %t
 }
 
 define i32 @load_test6(i32 %n) {
 ; CHECK-LABEL: load_test6:
 ; CHECK:         .functype load_test6 (i32) -> (i32)
 ; CHECK-NEXT:  # %bb.0:
+; CHECK-NEXT:    global.get $push2=, g@GOT
 ; CHECK-NEXT:    i32.const $push0=, 2
 ; CHECK-NEXT:    i32.shl $push1=, $0, $pop0
-; CHECK-NEXT:    global.get $push2=, g@GOT
-; CHECK-NEXT:    i32.add $push3=, $pop1, $pop2
-; CHECK-NEXT:    i32.const $push4=, 40
-; CHECK-NEXT:    i32.add $push5=, $pop3, $pop4
-; CHECK-NEXT:    i32.load $push6=, 0($pop5)
-; CHECK-NEXT:    return $pop6
+; CHECK-NEXT:    i32.add $push3=, $pop2, $pop1
+; CHECK-NEXT:    i32.load $push4=, 40($pop3)
+; CHECK-NEXT:    return $pop4
   %add = add nsw i32 %n, 10
   %add.ptr = getelementptr inbounds [0 x i32], ptr @g, i32 0, i32 %add
   %t = load i32, ptr %add.ptr, align 4
   ret i32 %t
 }
 
 define i32 @load_test7(i32 %n) {
 ; CHECK-LABEL: load_test7:
; CHECK-NEXT:    return $pop4
   %t = load i32, ptr %add.ptr1, align 4
   ret i32 %t
 }
 
 define i32 @load_test8(i32 %n) {
 ; CHECK-LABEL: load_test8:
 ; CHECK:         .functype load_test8 (i32) -> (i32)
 ; CHECK-NEXT:  # %bb.0:
+; CHECK-NEXT:    global.get $push2=, g@GOT
 ; CHECK-NEXT:    i32.const $push0=, 2
 ; CHECK-NEXT:    i32.shl $push1=, $0, $pop0
-; CHECK-NEXT:    global.get $push2=, g@GOT
-; CHECK-NEXT:    i32.add $push3=, $pop1, $pop2
-; CHECK-NEXT:    i32.const $push4=, 40
-; CHECK-NEXT:    i32.add $push5=, $pop3, $pop4
-; CHECK-NEXT:    i32.load $push6=, 0($pop5)
-; CHECK-NEXT:    return $pop6
+; CHECK-NEXT:    i32.add $push3=, $pop2, $pop1
+; CHECK-NEXT:    i32.load $push4=, 40($pop3)
+; CHECK-NEXT:    return $pop4
   %add = add nsw i32 10, %n
   %add.ptr = getelementptr inbounds [0 x i32], ptr @g, i32 0, i32 %add
   %t = load i32, ptr %add.ptr, align 4
   ret i32 %t
 }
 
 define i32 @load_test9() {
 ; CHECK-LABEL: load_test9:

 }
 
 define i32 @load_test12(ptr %p, i32 %n) {
 ; CHECK-LABEL: load_test12:
 ; CHECK:         .functype load_test12 (i32, i32) -> (i32)
 ; CHECK-NEXT:  # %bb.0:
 ; CHECK-NEXT:    i32.const $push0=, 2
 ; CHECK-NEXT:    i32.shl $push1=, $1, $pop0
-; CHECK-NEXT:    i32.add $push2=, $pop1, $0
-; CHECK-NEXT:    i32.const $push3=, 40
-; CHECK-NEXT:    i32.add $push4=, $pop2, $pop3
-; CHECK-NEXT:    i32.load $push5=, 0($pop4)
-; CHECK-NEXT:    return $pop5
+; CHECK-NEXT:    i32.add $push2=, $0, $pop1
+; CHECK-NEXT:    i32.load $push3=, 40($pop2)
+; CHECK-NEXT:    return $pop3
   %add = add nsw i32 %n, 10
   %arrayidx = getelementptr inbounds i32, ptr %p, i32 %add
   %t = load i32, ptr %arrayidx, align 4
   ret i32 %t
 }
 
 define i32 @load_test13(ptr %p, i32 %n) {
 ; CHECK-LABEL: load_test13:
 ; CHECK:         .functype load_test13 (i32, i32) -> (i32)
 ; CHECK-NEXT:  # %bb.0:
 ; CHECK-NEXT:    i32.const $push0=, 2
 ; CHECK-NEXT:    i32.shl $push1=, $1, $pop0
-; CHECK-NEXT:    i32.add $push2=, $pop1, $0
-; CHECK-NEXT:    i32.const $push3=, 40
-; CHECK-NEXT:    i32.add $push4=, $pop2, $pop3
-; CHECK-NEXT:    i32.load $push5=, 0($pop4)
-; CHECK-NEXT:    return $pop5
+; CHECK-NEXT:    i32.add $push2=, $0, $pop1
+; CHECK-NEXT:    i32.load $push3=, 40($pop2)
+; CHECK-NEXT:    return $pop3
   %add = add nsw i32 10, %n
   %arrayidx = getelementptr inbounds i32, ptr %p, i32 %add
   %t = load i32, ptr %arrayidx, align 4
   ret i32 %t
 }
 
 define i32 @load_test14(ptr %p, i32 %n) {
 ; CHECK-LABEL: load_test14:

 }
 
 define i32 @load_test17(ptr %p, i32 %n) {
 ; CHECK-LABEL: load_test17:
 ; CHECK:         .functype load_test17 (i32, i32) -> (i32)
 ; CHECK-NEXT:  # %bb.0:
 ; CHECK-NEXT:    i32.const $push0=, 2
 ; CHECK-NEXT:    i32.shl $push1=, $1, $pop0
-; CHECK-NEXT:    i32.add $push2=, $pop1, $0
-; CHECK-NEXT:    i32.const $push3=, 40
-; CHECK-NEXT:    i32.add $push4=, $pop2, $pop3
-; CHECK-NEXT:    i32.load $push5=, 0($pop4)
-; CHECK-NEXT:    return $pop5
+; CHECK-NEXT:    i32.add $push2=, $0, $pop1
+; CHECK-NEXT:    i32.load $push3=, 40($pop2)
+; CHECK-NEXT:    return $pop3
   %add = add nsw i32 %n, 10
   %add.ptr = getelementptr inbounds i32, ptr %p, i32 %add
   %t = load i32, ptr %add.ptr, align 4
   ret i32 %t
 }
 
 define i32 @load_test18(ptr %p, i32 %n) {
 ; CHECK-LABEL: load_test18:

 }
 
 define i32 @load_test19(ptr %p, i32 %n) {
 ; CHECK-LABEL: load_test19:
 ; CHECK:         .functype load_test19 (i32, i32) -> (i32)
 ; CHECK-NEXT:  # %bb.0:
 ; CHECK-NEXT:    i32.const $push0=, 2
 ; CHECK-NEXT:    i32.shl $push1=, $1, $pop0
-; CHECK-NEXT:    i32.add $push2=, $pop1, $0
-; CHECK-NEXT:    i32.const $push3=, 40
-; CHECK-NEXT:    i32.add $push4=, $pop2, $pop3
-; CHECK-NEXT:    i32.load $push5=, 0($pop4)
-; CHECK-NEXT:    return $pop5
+; CHECK-NEXT:    i32.add $push2=, $0, $pop1
+; CHECK-NEXT:    i32.load $push3=, 40($pop2)
+; CHECK-NEXT:    return $pop3
   %add = add nsw i32 10, %n
   %add.ptr = getelementptr inbounds i32, ptr %p, i32 %add
   %t = load i32, ptr %add.ptr, align 4
   ret i32 %t
 }
 
 define i32 @load_test20(ptr %p) {
 ; CHECK-LABEL: load_test20:
; CHECK-NEXT:    return
   store i32 %i, ptr getelementptr ([0 x i32], ptr @g, i32 0, i32 10), align 4
   ret void
 }
 
 define void @store_test1(i32 %n, i32 %i) {
 ; CHECK-LABEL: store_test1:
 ; CHECK:         .functype store_test1 (i32, i32) -> ()
 ; CHECK-NEXT:  # %bb.0:
+; CHECK-NEXT:    global.get $push2=, g@GOT
 ; CHECK-NEXT:    i32.const $push0=, 2
 ; CHECK-NEXT:    i32.shl $push1=, $0, $pop0
-; CHECK-NEXT:    global.get $push2=, g@GOT
-; CHECK-NEXT:    i32.add $push3=, $pop1, $pop2
-; CHECK-NEXT:    i32.const $push4=, 40
-; CHECK-NEXT:    i32.add $push5=, $pop3, $pop4
-; CHECK-NEXT:    i32.store 0($pop5), $1
+; CHECK-NEXT:    i32.add $push3=, $pop2, $pop1
+; CHECK-NEXT:    i32.store 40($pop3), $1
 ; CHECK-NEXT:    return
   %add = add nsw i32 %n, 10
   %arrayidx = getelementptr inbounds [0 x i32], ptr @g, i32 0, i32 %add
   store i32 %i, ptr %arrayidx, align 4
   ret void
 }
 
 define void @store_test2(i32 %n, i32 %i) {
 ; CHECK-LABEL: store_test2:
 ; CHECK:         .functype store_test2 (i32, i32) -> ()
 ; CHECK-NEXT:  # %bb.0:
+; CHECK-NEXT:    global.get $push2=, g@GOT
 ; CHECK-NEXT:    i32.const $push0=, 2
 ; CHECK-NEXT:    i32.shl $push1=, $0, $pop0
-; CHECK-NEXT:    global.get $push2=, g@GOT
-; CHECK-NEXT:    i32.add $push3=, $pop1, $pop2
-; CHECK-NEXT:    i32.const $push4=, 40
-; CHECK-NEXT:    i32.add $push5=, $pop3, $pop4
-; CHECK-NEXT:    i32.store 0($pop5), $1
+; CHECK-NEXT:    i32.add $push3=, $pop2, $pop1
+; CHECK-NEXT:    i32.store 40($pop3), $1
 ; CHECK-NEXT:    return
   %add = add nsw i32 10, %n
   %arrayidx = getelementptr inbounds [0 x i32], ptr @g, i32 0, i32 %add
   store i32 %i, ptr %arrayidx, align 4
   ret void
 }
 
 define void @store_test3(i32 %n, i32 %i) {
; CHECK-NEXT:    return
   store i32 %i, ptr %add.ptr, align 4
   ret void
 }
 
 define void @store_test6(i32 %n, i32 %i) {
 ; CHECK-LABEL: store_test6:
 ; CHECK:         .functype store_test6 (i32, i32) -> ()
 ; CHECK-NEXT:  # %bb.0:
+; CHECK-NEXT:    global.get $push2=, g@GOT
 ; CHECK-NEXT:    i32.const $push0=, 2
 ; CHECK-NEXT:    i32.shl $push1=, $0, $pop0
-; CHECK-NEXT:    global.get $push2=, g@GOT
-; CHECK-NEXT:    i32.add $push3=, $pop1, $pop2
-; CHECK-NEXT:    i32.const $push4=, 40
-; CHECK-NEXT:    i32.add $push5=, $pop3, $pop4
-; CHECK-NEXT:    i32.store 0($pop5), $1
+; CHECK-NEXT:    i32.add $push3=, $pop2, $pop1
+; CHECK-NEXT:    i32.store 40($pop3), $1
 ; CHECK-NEXT:    return
   %add = add nsw i32 %n, 10
   %add.ptr = getelementptr inbounds [0 x i32], ptr @g, i32 0, i32 %add
   store i32 %i, ptr %add.ptr, align 4
   ret void
 }
 
 define void @store_test7(i32 %n, i32 %i) {
; CHECK-NEXT:    return
   store i32 %i, ptr %add.ptr1, align 4
   ret void
 }
 
 define void @store_test8(i32 %n, i32 %i) {
 ; CHECK-LABEL: store_test8:
 ; CHECK:         .functype store_test8 (i32, i32) -> ()
 ; CHECK-NEXT:  # %bb.0:
+; CHECK-NEXT:    global.get $push2=, g@GOT
 ; CHECK-NEXT:    i32.const $push0=, 2
 ; CHECK-NEXT:    i32.shl $push1=, $0, $pop0
-; CHECK-NEXT:    global.get $push2=, g@GOT
-; CHECK-NEXT:    i32.add $push3=, $pop1, $pop2
-; CHECK-NEXT:    i32.const $push4=, 40
-; CHECK-NEXT:    i32.add $push5=, $pop3, $pop4
-; CHECK-NEXT:    i32.store 0($pop5), $1
+; CHECK-NEXT:    i32.add $push3=, $pop2, $pop1
+; CHECK-NEXT:    i32.store 40($pop3), $1
 ; CHECK-NEXT:    return
   %add = add nsw i32 10, %n
   %add.ptr = getelementptr inbounds [0 x i32], ptr @g, i32 0, i32 %add
   store i32 %i, ptr %add.ptr, align 4
   ret void
 }
 
 define void @store_test9(i32 %i) {

 }
 
 define void @store_test12(ptr %p, i32 %n, i32 %i) {
 ; CHECK-LABEL: store_test12:
 ; CHECK:         .functype store_test12 (i32, i32, i32) -> ()
 ; CHECK-NEXT:  # %bb.0:
 ; CHECK-NEXT:    i32.const $push0=, 2
 ; CHECK-NEXT:    i32.shl $push1=, $1, $pop0
-; CHECK-NEXT:    i32.add $push2=, $pop1, $0
-; CHECK-NEXT:    i32.const $push3=, 40
-; CHECK-NEXT:    i32.add $push4=, $pop2, $pop3
-; CHECK-NEXT:    i32.store 0($pop4), $2
+; CHECK-NEXT:    i32.add $push2=, $0, $pop1
+; CHECK-NEXT:    i32.store 40($pop2), $2
 ; CHECK-NEXT:    return
   %add = add nsw i32 %n, 10
   %arrayidx = getelementptr inbounds i32, ptr %p, i32 %add
   store i32 %i, ptr %arrayidx, align 4
   ret void
 }
 
 define void @store_test13(ptr %p, i32 %n, i32 %i) {
 ; CHECK-LABEL: store_test13:
 ; CHECK:         .functype store_test13 (i32, i32, i32) -> ()
 ; CHECK-NEXT:  # %bb.0:
 ; CHECK-NEXT:    i32.const $push0=, 2
 ; CHECK-NEXT:    i32.shl $push1=, $1, $pop0
-; CHECK-NEXT:    i32.add $push2=, $pop1, $0
-; CHECK-NEXT:    i32.const $push3=, 40
-; CHECK-NEXT:    i32.add $push4=, $pop2, $pop3
-; CHECK-NEXT:    i32.store 0($pop4), $2
+; CHECK-NEXT:    i32.add $push2=, $0, $pop1
+; CHECK-NEXT:    i32.store 40($pop2), $2
 ; CHECK-NEXT:    return
   %add = add nsw i32 10, %n
   %arrayidx = getelementptr inbounds i32, ptr %p, i32 %add
   store i32 %i, ptr %arrayidx, align 4
   ret void
 }
 
 define void @store_test14(ptr %p, i32 %n, i32 %i) {

 }
 
 define void @store_test17(ptr %p, i32 %n, i32 %i) {
 ; CHECK-LABEL: store_test17:
 ; CHECK:         .functype store_test17 (i32, i32, i32) -> ()
 ; CHECK-NEXT:  # %bb.0:
 ; CHECK-NEXT:    i32.const $push0=, 2
 ; CHECK-NEXT:    i32.shl $push1=, $1, $pop0
-; CHECK-NEXT:    i32.add $push2=, $pop1, $0
-; CHECK-NEXT:    i32.const $push3=, 40
-; CHECK-NEXT:    i32.add $push4=, $pop2, $pop3
-; CHECK-NEXT:    i32.store 0($pop4), $2
+; CHECK-NEXT:    i32.add $push2=, $0, $pop1
+; CHECK-NEXT:    i32.store 40($pop2), $2
 ; CHECK-NEXT:    return
   %add = add nsw i32 %n, 10
   %add.ptr = getelementptr inbounds i32, ptr %p, i32 %add
   store i32 %i, ptr %add.ptr, align 4
   ret void
 }
 
 define void @store_test18(ptr %p, i32 %n, i32 %i) {

 }
 
 define void @store_test19(ptr %p, i32 %n, i32 %i) {
 ; CHECK-LABEL: store_test19:
 ; CHECK:         .functype store_test19 (i32, i32, i32) -> ()
 ; CHECK-NEXT:  # %bb.0:
 ; CHECK-NEXT:    i32.const $push0=, 2
 ; CHECK-NEXT:    i32.shl $push1=, $1, $pop0
-; CHECK-NEXT:    i32.add $push2=, $pop1, $0
-; CHECK-NEXT:    i32.const $push3=, 40
-; CHECK-NEXT:    i32.add $push4=, $pop2, $pop3
-; CHECK-NEXT:    i32.store 0($pop4), $2
+; CHECK-NEXT:    i32.add $push2=, $0, $pop1
+; CHECK-NEXT:    i32.store 40($pop2), $2
 ; CHECK-NEXT:    return
   %add = add nsw i32 10, %n
   %add.ptr = getelementptr inbounds i32, ptr %p, i32 %add
   store i32 %i, ptr %add.ptr, align 4
   ret void
 }
 
 define void @store_test20(ptr %p, i32 %i) {

llvm/test/CodeGen/WebAssembly/atomic-fence.ll

-; RUN: llc < %s | FileCheck %s --check-prefix NOATOMIC
-; RUN: llc < %s -asm-verbose=false -wasm-disable-explicit-locals -wasm-keep-registers -mattr=+atomics | FileCheck %s
+; RUN: llc < %s -O1 | FileCheck %s --check-prefix NOATOMIC
+; RUN: llc < %s -O1 -asm-verbose=false -wasm-disable-explicit-locals -wasm-keep-registers -mattr=+atomics | FileCheck %s
 
 target triple = "wasm32-unknown-unknown"
 
 ; A multithread fence is lowered to an atomic.fence instruction.
 ; CHECK-LABEL: multithread_fence:
 ; CHECK:  atomic.fence
 ; NOATOMIC-NOT: i32.atomic.rmw.or
 define void @multithread_fence() {

llvm/test/CodeGen/WebAssembly/bulk-memory.ll

-; RUN: llc < %s -asm-verbose=false -verify-machineinstrs -disable-wasm-fallthrough-return-opt -wasm-disable-explicit-locals -wasm-keep-registers -mcpu=mvp -mattr=+bulk-memory | FileCheck %s --check-prefixes CHECK,BULK-MEM
-; RUN: llc < %s -asm-verbose=false -verify-machineinstrs -disable-wasm-fallthrough-return-opt -wasm-disable-explicit-locals -wasm-keep-registers -mcpu=mvp -mattr=-bulk-memory | FileCheck %s --check-prefixes CHECK,NO-BULK-MEM
+; RUN: llc < %s -O1 -asm-verbose=false -verify-machineinstrs -disable-wasm-fallthrough-return-opt -wasm-disable-explicit-locals -wasm-keep-registers -mcpu=mvp -mattr=+bulk-memory | FileCheck %s --check-prefixes CHECK,BULK-MEM
+; RUN: llc < %s -O1 -asm-verbose=false -verify-machineinstrs -disable-wasm-fallthrough-return-opt -wasm-disable-explicit-locals -wasm-keep-registers -mcpu=mvp -mattr=-bulk-memory | FileCheck %s --check-prefixes CHECK,NO-BULK-MEM
 
 ; Test that basic bulk memory codegen works correctly
 
 target triple = "wasm32-unknown-unknown"
 
 declare void @llvm.memcpy.p0.p0.i8(ptr, ptr, i8, i1)
 declare void @llvm.memcpy.p0.p0.i32(ptr, ptr, i32, i1)
 

llvm/test/CodeGen/WebAssembly/bulk-memory64.ll

-; RUN: llc < %s -asm-verbose=false -verify-machineinstrs -disable-wasm-fallthrough-return-opt -wasm-disable-explicit-locals -wasm-keep-registers -mcpu=mvp -mattr=+bulk-memory | FileCheck %s --check-prefixes CHECK,BULK-MEM
-; RUN: llc < %s -asm-verbose=false -verify-machineinstrs -disable-wasm-fallthrough-return-opt -wasm-disable-explicit-locals -wasm-keep-registers -mcpu=mvp -mattr=-bulk-memory | FileCheck %s --check-prefixes CHECK,NO-BULK-MEM
+; RUN: llc < %s -O1 -asm-verbose=false -verify-machineinstrs -disable-wasm-fallthrough-return-opt -wasm-disable-explicit-locals -wasm-keep-registers -mcpu=mvp -mattr=+bulk-memory | FileCheck %s --check-prefixes CHECK,BULK-MEM
+; RUN: llc < %s -O1 -asm-verbose=false -verify-machineinstrs -disable-wasm-fallthrough-return-opt -wasm-disable-explicit-locals -wasm-keep-registers -mcpu=mvp -mattr=-bulk-memory | FileCheck %s --check-prefixes CHECK,NO-BULK-MEM
 
 ; Test that basic bulk memory codegen works correctly
 
 target triple = "wasm64-unknown-unknown"
 
 declare void @llvm.memcpy.p0.p0.i8(ptr, ptr, i8, i1)
 declare void @llvm.memcpy.p0.p0.i64(ptr, ptr, i64, i1)
 

llvm/test/CodeGen/WebAssembly/cfg-stackify-dbg-skip.ll

-; RUN: llc %s -stop-after wasm-cfg-stackify -o - | FileCheck %s
+; RUN: llc %s -O1 -stop-after wasm-cfg-stackify -o - | FileCheck %s
 
 ; The test ensures "block" instruction is not inserted in the middle of a group
 ; of instructions that form a stackified expression when DBG_VALUE is present
 ; among them.
 
 ; CHECK: body:
 ; CHECK: BLOCK
 ;                       <-- Stackified expression starts

llvm/test/CodeGen/WebAssembly/cfg-stackify.ll

-; RUN: llc < %s -asm-verbose=false -disable-wasm-fallthrough-return-opt -wasm-disable-explicit-locals -wasm-keep-registers -disable-block-placement -verify-machineinstrs -fast-isel=false -machine-sink-split-probability-threshold=0 -cgp-freq-ratio-to-skip-merge=1000 | FileCheck %s
+; RUN: llc < %s -start-after=licm -asm-verbose=false -disable-wasm-fallthrough-return-opt -wasm-disable-explicit-locals -wasm-keep-registers -disable-block-placement -verify-machineinstrs -fast-isel=false -machine-sink-split-probability-threshold=0 -cgp-freq-ratio-to-skip-merge=1000 | FileCheck %s
 
 ; Test the CFG stackifier pass.
 
 ; Explicitly disable fast-isel, since it gets implicitly enabled in the
 ; optnone test.
 
 target triple = "wasm32-unknown-unknown"
 

llvm/test/CodeGen/WebAssembly/comparisons-f32.ll

 define void @xor_add_switch(float %a, float %b) {
 ; CHECK-LABEL: xor_add_switch:
 ; CHECK:         .functype xor_add_switch (f32, f32) -> ()
 ; CHECK-NEXT:  # %bb.0: # %entry
 ; CHECK-NEXT:    block
 ; CHECK-NEXT:    block
 ; CHECK-NEXT:    block
 ; CHECK-NEXT:    block
-; CHECK-NEXT:    local.get $push8=, 0
-; CHECK-NEXT:    local.get $push7=, 1
-; CHECK-NEXT:    f32.ge $push1=, $pop8, $pop7
-; CHECK-NEXT:    i32.const $push2=, 1
-; CHECK-NEXT:    i32.xor $push3=, $pop1, $pop2
-; CHECK-NEXT:    i32.const $push6=, 1
-; CHECK-NEXT:    i32.add $push4=, $pop3, $pop6
-; CHECK-NEXT:    i32.const $push5=, 1
-; CHECK-NEXT:    i32.xor $push0=, $pop4, $pop5
+; CHECK-NEXT:    i32.const $push3=, 0
+; CHECK-NEXT:    i32.const $push2=, 3
+; CHECK-NEXT:    local.get $push5=, 0
+; CHECK-NEXT:    local.get $push4=, 1
+; CHECK-NEXT:    f32.ge $push1=, $pop5, $pop4
+; CHECK-NEXT:    i32.select $push0=, $pop3, $pop2, $pop1
 ; CHECK-NEXT:    br_table $pop0, 0, 1, 2, 3 # 0: down to label12
 ; CHECK-NEXT:    # 1: down to label11
 ; CHECK-NEXT:    # 2: down to label10
 ; CHECK-NEXT:    # 3: down to label9
 ; CHECK-NEXT:  .LBB21_1: # %sw.bb.1
 ; CHECK-NEXT:    end_block # label12:
 ; CHECK-NEXT:    call foo1
 ; CHECK-NEXT:    return

llvm/test/CodeGen/WebAssembly/comparisons-f64.ll

 define void @xor_add_switch(double %a, double %b) {
 ; CHECK-LABEL: xor_add_switch:
 ; CHECK:         .functype xor_add_switch (f64, f64) -> ()
 ; CHECK-NEXT:  # %bb.0: # %entry
 ; CHECK-NEXT:    block
 ; CHECK-NEXT:    block
 ; CHECK-NEXT:    block
 ; CHECK-NEXT:    block
-; CHECK-NEXT:    local.get $push8=, 0
-; CHECK-NEXT:    local.get $push7=, 1
-; CHECK-NEXT:    f64.ge $push1=, $pop8, $pop7
-; CHECK-NEXT:    i32.const $push2=, 1
-; CHECK-NEXT:    i32.xor $push3=, $pop1, $pop2
-; CHECK-NEXT:    i32.const $push6=, 1
-; CHECK-NEXT:    i32.add $push4=, $pop3, $pop6
-; CHECK-NEXT:    i32.const $push5=, 1
-; CHECK-NEXT:    i32.xor $push0=, $pop4, $pop5
+; CHECK-NEXT:    i32.const $push3=, 0
+; CHECK-NEXT:    i32.const $push2=, 3
+; CHECK-NEXT:    local.get $push5=, 0
+; CHECK-NEXT:    local.get $push4=, 1
+; CHECK-NEXT:    f64.ge $push1=, $pop5, $pop4
+; CHECK-NEXT:    i32.select $push0=, $pop3, $pop2, $pop1
 ; CHECK-NEXT:    br_table $pop0, 0, 1, 2, 3 # 0: down to label12
 ; CHECK-NEXT:    # 1: down to label11
 ; CHECK-NEXT:    # 2: down to label10
 ; CHECK-NEXT:    # 3: down to label9
 ; CHECK-NEXT:  .LBB21_1: # %sw.bb.1
 ; CHECK-NEXT:    end_block # label12:
 ; CHECK-NEXT:    call foo1
 ; CHECK-NEXT:    return

llvm/test/CodeGen/WebAssembly/debugtrap.ll

-; RUN: llc < %s -asm-verbose=false -verify-machineinstrs | FileCheck %s
+; RUN: llc < %s -O1 -asm-verbose=false -verify-machineinstrs | FileCheck %s
 
 ; Test lowering of __builtin_debugtrap in cases where lowering it via
 ; the normal UNREACHABLE instruction would yield invalid
 ; MachineFunctions.
 
 target triple = "wasm32"
 
 declare void @llvm.debugtrap()

llvm/test/CodeGen/WebAssembly/externref-ptrtoint.ll

-; RUN: llc < %s --mtriple=wasm32-unknown-unknown -asm-verbose=false -mattr=+reference-types 2>&1 | FileCheck %s
+; RUN: llc < %s -O1 --mtriple=wasm32-unknown-unknown -asm-verbose=false -mattr=+reference-types 2>&1 | FileCheck %s
 
 %externref = type ptr addrspace(10)
 
 define i32 @externref_to_int(%externref %ref) {
   %i = ptrtoint %externref %ref to i32
   ret i32 %i
 }
 
 ; CHECK-LABEL: externref_to_int:
 ; CHECK-NEXT: .functype       externref_to_int (externref) -> (i32)
 ; CHECK-NEXT: .local i32
 ; CHECK-NEXT: unreachable
 ; CHECK-NEXT: local.get 1
 ; CHECK-NEXT: end_function

llvm/test/CodeGen/WebAssembly/externref-tableget.ll

-; RUN: llc < %s --mtriple=wasm32-unknown-unknown -asm-verbose=false -mattr=+reference-types | FileCheck %s
+; RUN: llc < %s -O1 --mtriple=wasm32-unknown-unknown -asm-verbose=false -mattr=+reference-types | FileCheck %s
 
 %externref = type ptr addrspace(10) ;; addrspace 10 is nonintegral
 
 @externref_table = local_unnamed_addr addrspace(1) global [0 x %externref] undef
 
 declare %externref @llvm.wasm.table.get.externref(ptr addrspace(1), i32) nounwind
 
 define %externref @get_externref_from_table(i32 %i) {

llvm/test/CodeGen/WebAssembly/externref-tableset.ll

-; RUN: llc --mtriple=wasm32-unknown-unknown -asm-verbose=false -mattr=+reference-types < %s | FileCheck %s
+; RUN: llc -O1 --mtriple=wasm32-unknown-unknown -asm-verbose=false -mattr=+reference-types < %s | FileCheck %s
 
 %externref = type ptr addrspace(10) ;; addrspace 10 is nonintegral
 
 @externref_table = local_unnamed_addr addrspace(1) global [0 x %externref] undef
 
 declare void @llvm.wasm.table.set.externref(ptr addrspace(1), i32, %externref) nounwind
 
 define void @set_externref_table(%externref %g, i32 %i) {

llvm/test/CodeGen/WebAssembly/f32.ll

; CHECK-NEXT:    return $pop0
   %a = call float @llvm.rint.f32(float %x)
   ret float %a
 }
 
 define float @fmin32(float %x) {
 ; CHECK-LABEL: fmin32:
 ; CHECK:         .functype fmin32 (f32) -> (f32)
 ; CHECK-NEXT:  # %bb.0:
-; CHECK-NEXT:    local.get $push2=, 0
 ; CHECK-NEXT:    f32.const $push0=, 0x0p0
-; CHECK-NEXT:    f32.min $push1=, $pop2, $pop0
-; CHECK-NEXT:    return $pop1
+; CHECK-NEXT:    local.get $push5=, 0
+; CHECK-NEXT:    local.get $push4=, 0
+; CHECK-NEXT:    f32.const $push3=, 0x0p0
+; CHECK-NEXT:    f32.ge $push1=, $pop4, $pop3
+; CHECK-NEXT:    f32.select $push2=, $pop0, $pop5, $pop1
+; CHECK-NEXT:    return $pop2
   %a = fcmp ult float %x, 0.0
   %b = select i1 %a, float %x, float 0.0
   ret float %b
 }
 
 define float @fmax32(float %x) {
 ; CHECK-LABEL: fmax32:
 ; CHECK:         .functype fmax32 (f32) -> (f32)
 ; CHECK-NEXT:  # %bb.0:
-; CHECK-NEXT:    local.get $push2=, 0
 ; CHECK-NEXT:    f32.const $push0=, 0x0p0
-; CHECK-NEXT:    f32.max $push1=, $pop2, $pop0
-; CHECK-NEXT:    return $pop1
+; CHECK-NEXT:    local.get $push5=, 0
+; CHECK-NEXT:    local.get $push4=, 0
+; CHECK-NEXT:    f32.const $push3=, 0x0p0
+; CHECK-NEXT:    f32.le $push1=, $pop4, $pop3
+; CHECK-NEXT:    f32.select $push2=, $pop0, $pop5, $pop1
+; CHECK-NEXT:    return $pop2
   %a = fcmp ugt float %x, 0.0
   %b = select i1 %a, float %x, float 0.0
   ret float %b
 }
 
 declare float @llvm.minimum.f32(float, float)
 define float @fmin32_intrinsic(float %x, float %y) {
 ; CHECK-LABEL: fmin32_intrinsic:

llvm/test/CodeGen/WebAssembly/f64.ll

; CHECK-NEXT:    return $pop0
   %a = call double @llvm.rint.f64(double %x)
   ret double %a
 }
 
 define double @fmin64(double %x) {
 ; CHECK-LABEL: fmin64:
 ; CHECK:         .functype fmin64 (f64) -> (f64)
 ; CHECK-NEXT:  # %bb.0:
-; CHECK-NEXT:    local.get $push2=, 0
 ; CHECK-NEXT:    f64.const $push0=, 0x0p0
-; CHECK-NEXT:    f64.min $push1=, $pop2, $pop0
-; CHECK-NEXT:    return $pop1
+; CHECK-NEXT:    local.get $push5=, 0
+; CHECK-NEXT:    local.get $push4=, 0
+; CHECK-NEXT:    f64.const $push3=, 0x0p0
+; CHECK-NEXT:    f64.ge $push1=, $pop4, $pop3
+; CHECK-NEXT:    f64.select $push2=, $pop0, $pop5, $pop1
+; CHECK-NEXT:    return $pop2
   %a = fcmp ult double %x, 0.0
   %b = select i1 %a, double %x, double 0.0
   ret double %b
 }
 
 define double @fmax64(double %x) {
 ; CHECK-LABEL: fmax64:
 ; CHECK:         .functype fmax64 (f64) -> (f64)
 ; CHECK-NEXT:  # %bb.0:
-; CHECK-NEXT:    local.get $push2=, 0
 ; CHECK-NEXT:    f64.const $push0=, 0x0p0
-; CHECK-NEXT:    f64.max $push1=, $pop2, $pop0
-; CHECK-NEXT:    return $pop1
+; CHECK-NEXT:    local.get $push5=, 0
+; CHECK-NEXT:    local.get $push4=, 0
+; CHECK-NEXT:    f64.const $push3=, 0x0p0
+; CHECK-NEXT:    f64.le $push1=, $pop4, $pop3
+; CHECK-NEXT:    f64.select $push2=, $pop0, $pop5, $pop1
+; CHECK-NEXT:    return $pop2
   %a = fcmp ugt double %x, 0.0
   %b = select i1 %a, double %x, double 0.0
   ret double %b
 }
 
 declare double @llvm.minimum.f64(double, double)
 define double @fmin64_intrinsic(double %x, double %y) {
 ; CHECK-LABEL: fmin64_intrinsic:

llvm/test/CodeGen/WebAssembly/fast-isel-br-i1.ll

-; RUN: llc < %s -fast-isel -asm-verbose=false -wasm-keep-registers | FileCheck %s
+; RUN: llc < %s -O1 -fast-isel -asm-verbose=false -wasm-keep-registers | FileCheck %s
 
 target triple = "wasm32-unknown-unknown"
 
 ; Fast-isel uses a 32-bit xor with -1 to negate i1 values, because it doesn't
 ; make any guarantees about the contents of the high bits of a register holding
 ; an i1 value. Test that when we do a `br_if` or `br_unless` with what what an
 ; i1 value in LLVM IR, that we only test the low bit.
 

llvm/test/CodeGen/WebAssembly/fpclamptosat.ll

 
 define i32 @stest_f16i32_cse(half %x) {
 ; CHECK-LABEL: stest_f16i32_cse:
 ; CHECK:         .functype stest_f16i32_cse (f32) -> (i32)
 ; CHECK-NEXT:  # %bb.0: # %entry
 ; CHECK-NEXT:    local.get 0
 ; CHECK-NEXT:    call __truncsfhf2
 ; CHECK-NEXT:    call __extendhfsf2
-; CHECK-NEXT:    i64.trunc_sat_f32_s
-; CHECK-NEXT:    i32.wrap_i64
+; CHECK-NEXT:    i32.trunc_sat_f32_s
 ; CHECK-NEXT:    # fallthrough-return
 entry:
   %conv = fptosi half %x to i64
   %conv6 = trunc i64 %conv to i32
   ret i32 %conv6
 }
 
 define i32 @utesth_f16i32(half %x) {

 
 define i32 @utesth_f16i32_cse(half %x) {
 ; CHECK-LABEL: utesth_f16i32_cse:
 ; CHECK:         .functype utesth_f16i32_cse (f32) -> (i32)
 ; CHECK-NEXT:  # %bb.0: # %entry
 ; CHECK-NEXT:    local.get 0
 ; CHECK-NEXT:    call __truncsfhf2
 ; CHECK-NEXT:    call __extendhfsf2
-; CHECK-NEXT:    i64.trunc_sat_f32_u
-; CHECK-NEXT:    i32.wrap_i64
+; CHECK-NEXT:    i32.trunc_sat_f32_u
 ; CHECK-NEXT:    # fallthrough-return
 entry:
   %conv = fptoui half %x to i64
   %conv6 = trunc i64 %conv to i32
   ret i32 %conv6
 }
 
 define i32 @ustest_f16i32(half %x) {
 ; CHECK-LABEL: ustest_f16i32:
 ; CHECK:         .functype ustest_f16i32 (f32) -> (i32)
 ; CHECK-NEXT:  # %bb.0: # %entry
 ; CHECK-NEXT:    local.get 0
 ; CHECK-NEXT:    call __truncsfhf2
 ; CHECK-NEXT:    call __extendhfsf2
 ; CHECK-NEXT:    i32.trunc_sat_f32_u
 ; CHECK-NEXT:    # fallthrough-return
 entry:
   %conv = fptosi half %x to i64
   %0 = icmp slt i64 %conv, 4294967295
   %spec.store.select = select i1 %0, i64 %conv, i64 4294967295
   %1 = icmp sgt i64 %spec.store.select, 0
   %spec.store.select7 = select i1 %1, i64 %spec.store.select, i64 0
   %conv6 = trunc i64 %spec.store.select7 to i32
   ret i32 %conv6
 }
 
 define i32 @ustest_f16i32_cse(half %x) {
 ; CHECK-LABEL: ustest_f16i32_cse:
 ; CHECK:         .functype ustest_f16i32_cse (f32) -> (i32)
 ; CHECK-NEXT:  # %bb.0: # %entry
entry:
   %spec.store.select7 = select i1 %1, i32 %spec.store.select, i32 -32768
   %conv6 = trunc i32 %spec.store.select7 to i16
   ret i16 %conv6
 }
 
 define i16 @utesth_f16i16(half %x) {
 ; CHECK-LABEL: utesth_f16i16:
 ; CHECK:         .functype utesth_f16i16 (f32) -> (i32)
-; CHECK-NEXT:    .local i32
 ; CHECK-NEXT:  # %bb.0: # %entry
 ; CHECK-NEXT:    local.get 0
 ; CHECK-NEXT:    call __truncsfhf2
 ; CHECK-NEXT:    call __extendhfsf2
 ; CHECK-NEXT:    i32.trunc_sat_f32_u
-; CHECK-NEXT:    local.tee 1
-; CHECK-NEXT:    i32.const 65535
-; CHECK-NEXT:    local.get 1
-; CHECK-NEXT:    i32.const 65535
-; CHECK-NEXT:    i32.lt_u
-; CHECK-NEXT:    i32.select
 ; CHECK-NEXT:    # fallthrough-return
 entry:
   %conv = fptoui half %x to i32
   %0 = icmp ult i32 %conv, 65535
   %spec.store.select = select i1 %0, i32 %conv, i32 65535
   %conv6 = trunc i32 %spec.store.select to i16
   ret i16 %conv6
 }
entry:
   %conv = fptoui half %x to i32
   %conv6 = trunc i32 %conv to i16
   ret i16 %conv6
 }
 
 define i16 @ustest_f16i16(half %x) {
 ; CHECK-LABEL: ustest_f16i16:
 ; CHECK:         .functype ustest_f16i16 (f32) -> (i32)
-; CHECK-NEXT:    .local i32
 ; CHECK-NEXT:  # %bb.0: # %entry
 ; CHECK-NEXT:    local.get 0
 ; CHECK-NEXT:    call __truncsfhf2
 ; CHECK-NEXT:    call __extendhfsf2
-; CHECK-NEXT:    i32.trunc_sat_f32_s
-; CHECK-NEXT:    local.tee 1
-; CHECK-NEXT:    i32.const 65535
-; CHECK-NEXT:    local.get 1
-; CHECK-NEXT:    i32.const 65535
-; CHECK-NEXT:    i32.lt_s
-; CHECK-NEXT:    i32.select
-; CHECK-NEXT:    local.tee 1
-; CHECK-NEXT:    i32.const 0
-; CHECK-NEXT:    local.get 1
-; CHECK-NEXT:    i32.const 0
-; CHECK-NEXT:    i32.gt_s
-; CHECK-NEXT:    i32.select
+; CHECK-NEXT:    i32.trunc_sat_f32_u
 ; CHECK-NEXT:    # fallthrough-return
 entry:
   %conv = fptosi half %x to i32
   %0 = icmp slt i32 %conv, 65535
   %spec.store.select = select i1 %0, i32 %conv, i32 65535
   %1 = icmp sgt i32 %spec.store.select, 0
   %spec.store.select7 = select i1 %1, i32 %spec.store.select, i32 0
   %conv6 = trunc i32 %spec.store.select7 to i16

 ; CHECK-NEXT:    local.get 3
 ; CHECK-NEXT:    i64.const 0
 ; CHECK-NEXT:    local.get 2
 ; CHECK-NEXT:    i64.const 1
 ; CHECK-NEXT:    i64.lt_s
 ; CHECK-NEXT:    local.tee 1
 ; CHECK-NEXT:    i64.select
 ; CHECK-NEXT:    local.tee 3
-; CHECK-NEXT:    i64.const 0
 ; CHECK-NEXT:    local.get 3
 ; CHECK-NEXT:    i64.const 0
-; CHECK-NEXT:    i64.ne
 ; CHECK-NEXT:    local.get 2
 ; CHECK-NEXT:    i64.const 1
 ; CHECK-NEXT:    local.get 1
 ; CHECK-NEXT:    i64.select
 ; CHECK-NEXT:    local.tee 2
 ; CHECK-NEXT:    i64.const 0
 ; CHECK-NEXT:    i64.gt_s
+; CHECK-NEXT:    i64.select
 ; CHECK-NEXT:    local.get 2
 ; CHECK-NEXT:    i64.eqz
-; CHECK-NEXT:    i32.select
 ; CHECK-NEXT:    i64.select
 ; CHECK-NEXT:    # fallthrough-return
 entry:
   %conv = fptosi double %x to i128
   %0 = icmp slt i128 %conv, 18446744073709551616
   %spec.store.select = select i1 %0, i128 %conv, i128 18446744073709551616
   %1 = icmp sgt i128 %spec.store.select, 0
   %spec.store.select7 = select i1 %1, i128 %spec.store.select, i128 0

 ; CHECK-NEXT:    local.get 3
 ; CHECK-NEXT:    i64.const 0
 ; CHECK-NEXT:    local.get 2
 ; CHECK-NEXT:    i64.const 1
 ; CHECK-NEXT:    i64.lt_s
 ; CHECK-NEXT:    local.tee 1
 ; CHECK-NEXT:    i64.select
 ; CHECK-NEXT:    local.tee 3
-; CHECK-NEXT:    i64.const 0
 ; CHECK-NEXT:    local.get 3
 ; CHECK-NEXT:    i64.const 0
-; CHECK-NEXT:    i64.ne
 ; CHECK-NEXT:    local.get 2
 ; CHECK-NEXT:    i64.const 1
 ; CHECK-NEXT:    local.get 1
 ; CHECK-NEXT:    i64.select
 ; CHECK-NEXT:    local.tee 2
 ; CHECK-NEXT:    i64.const 0
 ; CHECK-NEXT:    i64.gt_s
+; CHECK-NEXT:    i64.select
 ; CHECK-NEXT:    local.get 2
 ; CHECK-NEXT:    i64.eqz
-; CHECK-NEXT:    i32.select
 ; CHECK-NEXT:    i64.select
 ; CHECK-NEXT:    # fallthrough-return
 entry:
   %conv = fptosi float %x to i128
   %0 = icmp slt i128 %conv, 18446744073709551616
   %spec.store.select = select i1 %0, i128 %conv, i128 18446744073709551616
   %1 = icmp sgt i128 %spec.store.select, 0
   %spec.store.select7 = select i1 %1, i128 %spec.store.select, i128 0
entry:
   %spec.store.select7 = select i1 %1, i128 %spec.store.select, i128 -9223372036854775808
   %conv6 = trunc i128 %spec.store.select7 to i64
   ret i64 %conv6
 }
 
 define i64 @utesth_f16i64(half %x) {
 ; CHECK-LABEL: utesth_f16i64:
 ; CHECK:         .functype utesth_f16i64 (f32) -> (i64)
-; CHECK-NEXT:    .local i32, i64, i64
 ; CHECK-NEXT:  # %bb.0: # %entry
-; CHECK-NEXT:    global.get __stack_pointer
-; CHECK-NEXT:    i32.const 16
-; CHECK-NEXT:    i32.sub
-; CHECK-NEXT:    local.tee 1
-; CHECK-NEXT:    global.set __stack_pointer
-; CHECK-NEXT:    local.get 1
 ; CHECK-NEXT:    local.get 0
 ; CHECK-NEXT:    call __truncsfhf2
 ; CHECK-NEXT:    call __extendhfsf2
-; CHECK-NEXT:    call __fixunssfti
-; CHECK-NEXT:    local.get 1
-; CHECK-NEXT:    i32.const 8
-; CHECK-NEXT:    i32.add
-; CHECK-NEXT:    i64.load 0
-; CHECK-NEXT:    local.set 2
-; CHECK-NEXT:    local.get 1
-; CHECK-NEXT:    i64.load 0
-; CHECK-NEXT:    local.set 3
-; CHECK-NEXT:    local.get 1
-; CHECK-NEXT:    i32.const 16
-; CHECK-NEXT:    i32.add
-; CHECK-NEXT:    global.set __stack_pointer
-; CHECK-NEXT:    local.get 3
-; CHECK-NEXT:    i64.const 0
-; CHECK-NEXT:    local.get 2
-; CHECK-NEXT:    i64.eqz
-; CHECK-NEXT:    i64.select
+; CHECK-NEXT:    i64.trunc_sat_f32_u
 ; CHECK-NEXT:    # fallthrough-return
 entry:
   %conv = fptoui half %x to i128
   %0 = icmp ult i128 %conv, 18446744073709551616
   %spec.store.select = select i1 %0, i128 %conv, i128 18446744073709551616
   %conv6 = trunc i128 %spec.store.select to i64
   ret i64 %conv6
 }
 
 define i64 @utesth_f16i64_cse(half %x) {
 ; CHECK-LABEL: utesth_f16i64_cse:
 ; CHECK:         .functype utesth_f16i64_cse (f32) -> (i64)
-; CHECK-NEXT:    .local i32, i64
 ; CHECK-NEXT:  # %bb.0: # %entry
-; CHECK-NEXT:    global.get __stack_pointer
-; CHECK-NEXT:    i32.const 16
-; CHECK-NEXT:    i32.sub
-; CHECK-NEXT:    local.tee 1
-; CHECK-NEXT:    global.set __stack_pointer
-; CHECK-NEXT:    local.get 1
 ; CHECK-NEXT:    local.get 0
 ; CHECK-NEXT:    call __truncsfhf2
 ; CHECK-NEXT:    call __extendhfsf2
-; CHECK-NEXT:    call __fixunssfti
-; CHECK-NEXT:    local.get 1
-; CHECK-NEXT:    i64.load 0
-; CHECK-NEXT:    local.set 2
-; CHECK-NEXT:    local.get 1
-; CHECK-NEXT:    i32.const 16
-; CHECK-NEXT:    i32.add
-; CHECK-NEXT:    global.set __stack_pointer
-; CHECK-NEXT:    local.get 2
+; CHECK-NEXT:    i64.trunc_sat_f32_u
 ; CHECK-NEXT:    # fallthrough-return
 entry:
   %conv = fptoui half %x to i128
   %conv6 = trunc i128 %conv to i64
   ret i64 %conv6
 }
 
 define i64 @ustest_f16i64(half %x) {

 ; CHECK-NEXT:    local.get 3
 ; CHECK-NEXT:    i64.const 0
 ; CHECK-NEXT:    local.get 2
 ; CHECK-NEXT:    i64.const 1
 ; CHECK-NEXT:    i64.lt_s
 ; CHECK-NEXT:    local.tee 1
 ; CHECK-NEXT:    i64.select
 ; CHECK-NEXT:    local.tee 3
-; CHECK-NEXT:    i64.const 0
 ; CHECK-NEXT:    local.get 3
 ; CHECK-NEXT:    i64.const 0
-; CHECK-NEXT:    i64.ne
 ; CHECK-NEXT:    local.get 2
 ; CHECK-NEXT:    i64.const 1
 ; CHECK-NEXT:    local.get 1
 ; CHECK-NEXT:    i64.select
 ; CHECK-NEXT:    local.tee 2
 ; CHECK-NEXT:    i64.const 0
 ; CHECK-NEXT:    i64.gt_s
+; CHECK-NEXT:    i64.select
 ; CHECK-NEXT:    local.get 2
 ; CHECK-NEXT:    i64.eqz
-; CHECK-NEXT:    i32.select
 ; CHECK-NEXT:    i64.select
 ; CHECK-NEXT:    # fallthrough-return
 entry:
   %conv = fptosi half %x to i128
   %0 = icmp slt i128 %conv, 18446744073709551616
   %spec.store.select = select i1 %0, i128 %conv, i128 18446744073709551616
   %1 = icmp sgt i128 %spec.store.select, 0
   %spec.store.select7 = select i1 %1, i128 %spec.store.select, i128 0
entry:
   %spec.store.select7 = call i32 @llvm.smax.i32(i32 %spec.store.select, i32 -32768)
   %conv6 = trunc i32 %spec.store.select7 to i16
   ret i16 %conv6
 }
 
 define i16 @utesth_f16i16_mm(half %x) {
 ; CHECK-LABEL: utesth_f16i16_mm:
 ; CHECK:         .functype utesth_f16i16_mm (f32) -> (i32)
-; CHECK-NEXT:    .local i32
 ; CHECK-NEXT:  # %bb.0: # %entry
 ; CHECK-NEXT:    local.get 0
 ; CHECK-NEXT:    call __truncsfhf2
 ; CHECK-NEXT:    call __extendhfsf2
 ; CHECK-NEXT:    i32.trunc_sat_f32_u
-; CHECK-NEXT:    local.tee 1
-; CHECK-NEXT:    i32.const 65535
-; CHECK-NEXT:    local.get 1
-; CHECK-NEXT:    i32.const 65535
-; CHECK-NEXT:    i32.lt_u
-; CHECK-NEXT:    i32.select
 ; CHECK-NEXT:    # fallthrough-return
 entry:
   %conv = fptoui half %x to i32
   %spec.store.select = call i32 @llvm.umin.i32(i32 %conv, i32 65535)
   %conv6 = trunc i32 %spec.store.select to i16
   ret i16 %conv6
 }
 
 define i16 @ustest_f16i16_mm(half %x) {
 ; CHECK-LABEL: ustest_f16i16_mm:
 ; CHECK:         .functype ustest_f16i16_mm (f32) -> (i32)
-; CHECK-NEXT:    .local i32
 ; CHECK-NEXT:  # %bb.0: # %entry
 ; CHECK-NEXT:    local.get 0
 ; CHECK-NEXT:    call __truncsfhf2
 ; CHECK-NEXT:    call __extendhfsf2
-; CHECK-NEXT:    i32.trunc_sat_f32_s
-; CHECK-NEXT:    local.tee 1
-; CHECK-NEXT:    i32.const 65535
-; CHECK-NEXT:    local.get 1
-; CHECK-NEXT:    i32.const 65535
-; CHECK-NEXT:    i32.lt_s
-; CHECK-NEXT:    i32.select
-; CHECK-NEXT:    local.tee 1
-; CHECK-NEXT:    i32.const 0
-; CHECK-NEXT:    local.get 1
-; CHECK-NEXT:    i32.const 0
-; CHECK-NEXT:    i32.gt_s
-; CHECK-NEXT:    i32.select
+; CHECK-NEXT:    i32.trunc_sat_f32_u
 ; CHECK-NEXT:    # fallthrough-return
 entry:
   %conv = fptosi half %x to i32
   %spec.store.select = call i32 @llvm.smin.i32(i32 %conv, i32 65535)
   %spec.store.select7 = call i32 @llvm.smax.i32(i32 %spec.store.select, i32 0)
   %conv6 = trunc i32 %spec.store.select7 to i16
   ret i16 %conv6
 }
entry:
   %spec.store.select7 = call i128 @llvm.smax.i128(i128 %spec.store.select, i128 -9223372036854775808)
   %conv6 = trunc i128 %spec.store.select7 to i64
   ret i64 %conv6
 }
 
 define i64 @utesth_f16i64_mm(half %x) {
 ; CHECK-LABEL: utesth_f16i64_mm:
 ; CHECK:         .functype utesth_f16i64_mm (f32) -> (i64)
-; CHECK-NEXT:    .local i32, i64, i64
 ; CHECK-NEXT:  # %bb.0: # %entry
-; CHECK-NEXT:    global.get __stack_pointer
-; CHECK-NEXT:    i32.const 16
-; CHECK-NEXT:    i32.sub
-; CHECK-NEXT:    local.tee 1
-; CHECK-NEXT:    global.set __stack_pointer
-; CHECK-NEXT:    local.get 1
 ; CHECK-NEXT:    local.get 0
 ; CHECK-NEXT:    call __truncsfhf2
 ; CHECK-NEXT:    call __extendhfsf2
-; CHECK-NEXT:    call __fixunssfti
-; CHECK-NEXT:    local.get 1
-; CHECK-NEXT:    i32.const 8
-; CHECK-NEXT:    i32.add
-; CHECK-NEXT:    i64.load 0
-; CHECK-NEXT:    local.set 2
-; CHECK-NEXT:    local.get 1
-; CHECK-NEXT:    i64.load 0
-; CHECK-NEXT:    local.set 3
-; CHECK-NEXT:    local.get 1
-; CHECK-NEXT:    i32.const 16
-; CHECK-NEXT:    i32.add
-; CHECK-NEXT:    global.set __stack_pointer
-; CHECK-NEXT:    local.get 3
-; CHECK-NEXT:    i64.const 0
-; CHECK-NEXT:    local.get 2
-; CHECK-NEXT:    i64.eqz
-; CHECK-NEXT:    i64.select
+; CHECK-NEXT:    i64.trunc_sat_f32_u
 ; CHECK-NEXT:    # fallthrough-return
 entry:
   %conv = fptoui half %x to i128
   %spec.store.select = call i128 @llvm.umin.i128(i128 %conv, i128 18446744073709551616)
   %conv6 = trunc i128 %spec.store.select to i64
   ret i64 %conv6
 }
 
entry:
   %spec.store.select7 = call i128 @llvm.smax.i128(i128 %spec.store.select, i128 0)
   %conv6 = trunc i128 %spec.store.select7 to i64
   ret i64 %conv6
 }
 
 define i64 @utesth_f16i64_mm_cse(half %x) {
 ; CHECK-LABEL: utesth_f16i64_mm_cse:
 ; CHECK:         .functype utesth_f16i64_mm_cse (f32) -> (i64)
-; CHECK-NEXT:    .local i32, i64
 ; CHECK-NEXT:  # %bb.0: # %entry
-; CHECK-NEXT:    global.get __stack_pointer
-; CHECK-NEXT:    i32.const 16
-; CHECK-NEXT:    i32.sub
-; CHECK-NEXT:    local.tee 1
-; CHECK-NEXT:    global.set __stack_pointer
-; CHECK-NEXT:    local.get 1
 ; CHECK-NEXT:    local.get 0
 ; CHECK-NEXT:    call __truncsfhf2
 ; CHECK-NEXT:    call __extendhfsf2
-; CHECK-NEXT:    call __fixunssfti
-; CHECK-NEXT:    local.get 1
-; CHECK-NEXT:    i64.load 0
-; CHECK-NEXT:    local.set 2
-; CHECK-NEXT:    local.get 1
-; CHECK-NEXT:    i32.const 16
-; CHECK-NEXT:    i32.add
-; CHECK-NEXT:    global.set __stack_pointer
-; CHECK-NEXT:    local.get 2
+; CHECK-NEXT:    i64.trunc_sat_f32_u
 ; CHECK-NEXT:    # fallthrough-return
 entry:
   %conv = fptoui half %x to i128
   %conv6 = trunc i128 %conv to i64
   ret i64 %conv6
 }
 
 declare i32 @llvm.smin.i32(i32, i32)
 declare i32 @llvm.smax.i32(i32, i32)
 declare i32 @llvm.umin.i32(i32, i32)
 declare i64 @llvm.smin.i64(i64, i64)
 declare i64 @llvm.smax.i64(i64, i64)
 declare i64 @llvm.umin.i64(i64, i64)
 declare i128 @llvm.smin.i128(i128, i128)
 declare i128 @llvm.smax.i128(i128, i128)
 declare i128 @llvm.umin.i128(i128, i128)

llvm/test/CodeGen/WebAssembly/fpclamptosat_vec.ll

entry:
   %spec.store.select7 = select <8 x i1> %1, <8 x i32> %spec.store.select, <8 x i32> <i32 -32768, i32 -32768, i32 -32768, i32 -32768, i32 -32768, i32 -32768, i32 -32768, i32 -32768>
   %conv6 = trunc <8 x i32> %spec.store.select7 to <8 x i16>
   ret <8 x i16> %conv6
 }
 
 define <8 x i16> @utesth_f16i16(<8 x half> %x) {
 ; CHECK-LABEL: utesth_f16i16:
 ; CHECK:         .functype utesth_f16i16 (f32, f32, f32, f32, f32, f32, f32, f32) -> (v128)
-; CHECK-NEXT:    .local v128
 ; CHECK-NEXT:  # %bb.0: # %entry
-; CHECK-NEXT:    local.get 5
-; CHECK-NEXT:    call __truncsfhf2
-; CHECK-NEXT:    call __extendhfsf2
-; CHECK-NEXT:    local.set 5
-; CHECK-NEXT:    local.get 4
-; CHECK-NEXT:    call __truncsfhf2
-; CHECK-NEXT:    call __extendhfsf2
-; CHECK-NEXT:    local.set 4
-; CHECK-NEXT:    local.get 6
-; CHECK-NEXT:    call __truncsfhf2
-; CHECK-NEXT:    call __extendhfsf2
-; CHECK-NEXT:    local.set 6
-; CHECK-NEXT:    local.get 7
-; CHECK-NEXT:    call __truncsfhf2
-; CHECK-NEXT:    call __extendhfsf2
-; CHECK-NEXT:    local.set 7
 ; CHECK-NEXT:    local.get 1
 ; CHECK-NEXT:    call __truncsfhf2
 ; CHECK-NEXT:    call __extendhfsf2
 ; CHECK-NEXT:    local.set 1
 ; CHECK-NEXT:    local.get 0
 ; CHECK-NEXT:    call __truncsfhf2
 ; CHECK-NEXT:    call __extendhfsf2
 ; CHECK-NEXT:    i32.trunc_sat_f32_u
-; CHECK-NEXT:    i32x4.splat
+; CHECK-NEXT:    i16x8.splat
 ; CHECK-NEXT:    local.get 1
 ; CHECK-NEXT:    i32.trunc_sat_f32_u
-; CHECK-NEXT:    i32x4.replace_lane 1
+; CHECK-NEXT:    i16x8.replace_lane 1
 ; CHECK-NEXT:    local.get 2
 ; CHECK-NEXT:    call __truncsfhf2
 ; CHECK-NEXT:    call __extendhfsf2
 ; CHECK-NEXT:    i32.trunc_sat_f32_u
-; CHECK-NEXT:    i32x4.replace_lane 2
+; CHECK-NEXT:    i16x8.replace_lane 2
 ; CHECK-NEXT:    local.get 3
 ; CHECK-NEXT:    call __truncsfhf2
 ; CHECK-NEXT:    call __extendhfsf2
 ; CHECK-NEXT:    i32.trunc_sat_f32_u
-; CHECK-NEXT:    i32x4.replace_lane 3
-; CHECK-NEXT:    v128.const 65535, 65535, 65535, 65535
-; CHECK-NEXT:    local.tee 8
-; CHECK-NEXT:    i32x4.min_u
+; CHECK-NEXT:    i16x8.replace_lane 3
 ; CHECK-NEXT:    local.get 4
+; CHECK-NEXT:    call __truncsfhf2
+; CHECK-NEXT:    call __extendhfsf2
 ; CHECK-NEXT:    i32.trunc_sat_f32_u
-; CHECK-NEXT:    i32x4.splat
+; CHECK-NEXT:    i16x8.replace_lane 4
 ; CHECK-NEXT:    local.get 5
+; CHECK-NEXT:    call __truncsfhf2
+; CHECK-NEXT:    call __extendhfsf2
 ; CHECK-NEXT:    i32.trunc_sat_f32_u
-; CHECK-NEXT:    i32x4.replace_lane 1
+; CHECK-NEXT:    i16x8.replace_lane 5
 ; CHECK-NEXT:    local.get 6
+; CHECK-NEXT:    call __truncsfhf2
+; CHECK-NEXT:    call __extendhfsf2
 ; CHECK-NEXT:    i32.trunc_sat_f32_u
-; CHECK-NEXT:    i32x4.replace_lane 2
+; CHECK-NEXT:    i16x8.replace_lane 6
 ; CHECK-NEXT:    local.get 7
+; CHECK-NEXT:    call __truncsfhf2
+; CHECK-NEXT:    call __extendhfsf2
 ; CHECK-NEXT:    i32.trunc_sat_f32_u
-; CHECK-NEXT:    i32x4.replace_lane 3
-; CHECK-NEXT:    local.get 8
-; CHECK-NEXT:    i32x4.min_u
-; CHECK-NEXT:    i16x8.narrow_i32x4_u
+; CHECK-NEXT:    i16x8.replace_lane 7
 ; CHECK-NEXT:    # fallthrough-return
 entry:
   %conv = fptoui <8 x half> %x to <8 x i32>
   %0 = icmp ult <8 x i32> %conv, <i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535>
   %spec.store.select = select <8 x i1> %0, <8 x i32> %conv, <8 x i32> <i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535>
   %conv6 = trunc <8 x i32> %spec.store.select to <8 x i16>
   ret <8 x i16> %conv6
 }

 ; CHECK-NEXT:    call __extendhfsf2
 ; CHECK-NEXT:    i32.trunc_sat_f32_s
 ; CHECK-NEXT:    i32x4.replace_lane 2
 ; CHECK-NEXT:    local.get 3
 ; CHECK-NEXT:    call __truncsfhf2
 ; CHECK-NEXT:    call __extendhfsf2
 ; CHECK-NEXT:    i32.trunc_sat_f32_s
 ; CHECK-NEXT:    i32x4.replace_lane 3
-; CHECK-NEXT:    v128.const 65535, 65535, 65535, 65535
-; CHECK-NEXT:    local.tee 8
-; CHECK-NEXT:    i32x4.min_s
 ; CHECK-NEXT:    v128.const 0, 0, 0, 0
-; CHECK-NEXT:    local.tee 9
+; CHECK-NEXT:    local.tee 8
 ; CHECK-NEXT:    i32x4.max_s
+; CHECK-NEXT:    v128.const 65535, 65535, 65535, 65535
+; CHECK-NEXT:    local.tee 9
+; CHECK-NEXT:    v128.and
 ; CHECK-NEXT:    local.get 4
 ; CHECK-NEXT:    i32.trunc_sat_f32_s
 ; CHECK-NEXT:    i32x4.splat
 ; CHECK-NEXT:    local.get 5
 ; CHECK-NEXT:    i32.trunc_sat_f32_s
 ; CHECK-NEXT:    i32x4.replace_lane 1
 ; CHECK-NEXT:    local.get 6
 ; CHECK-NEXT:    i32.trunc_sat_f32_s
 ; CHECK-NEXT:    i32x4.replace_lane 2
 ; CHECK-NEXT:    local.get 7
 ; CHECK-NEXT:    i32.trunc_sat_f32_s
 ; CHECK-NEXT:    i32x4.replace_lane 3
 ; CHECK-NEXT:    local.get 8
-; CHECK-NEXT:    i32x4.min_s
-; CHECK-NEXT:    local.get 9
 ; CHECK-NEXT:    i32x4.max_s
+; CHECK-NEXT:    local.get 9
+; CHECK-NEXT:    v128.and
 ; CHECK-NEXT:    i16x8.narrow_i32x4_u
 ; CHECK-NEXT:    # fallthrough-return
 entry:
   %conv = fptosi <8 x half> %x to <8 x i32>
   %0 = icmp slt <8 x i32> %conv, <i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535>
   %spec.store.select = select <8 x i1> %0, <8 x i32> %conv, <8 x i32> <i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535>
   %1 = icmp sgt <8 x i32> %spec.store.select, zeroinitializer
   %spec.store.select7 = select <8 x i1> %1, <8 x i32> %spec.store.select, <8 x i32> zeroinitializer

 ; CHECK-NEXT:    i32.const 32
 ; CHECK-NEXT:    i32.add
 ; CHECK-NEXT:    global.set __stack_pointer
 ; CHECK-NEXT:    local.get 5
 ; CHECK-NEXT:    i64.const 9223372036854775807
 ; CHECK-NEXT:    local.get 5
 ; CHECK-NEXT:    i64.const 9223372036854775807
 ; CHECK-NEXT:    i64.lt_u
+; CHECK-NEXT:    i64.select
+; CHECK-NEXT:    local.get 5
+; CHECK-NEXT:    i64.const 9223372036854775807
 ; CHECK-NEXT:    local.get 4
 ; CHECK-NEXT:    i64.const 0
 ; CHECK-NEXT:    i64.lt_s
+; CHECK-NEXT:    i64.select
 ; CHECK-NEXT:    local.get 4
 ; CHECK-NEXT:    i64.eqz
-; CHECK-NEXT:    i32.select
-; CHECK-NEXT:    local.tee 1
 ; CHECK-NEXT:    i64.select
 ; CHECK-NEXT:    local.tee 5
 ; CHECK-NEXT:    i64.const -9223372036854775808
 ; CHECK-NEXT:    local.get 5
 ; CHECK-NEXT:    i64.const -9223372036854775808
 ; CHECK-NEXT:    i64.gt_u
-; CHECK-NEXT:    local.get 4
-; CHECK-NEXT:    i64.const 0
-; CHECK-NEXT:    local.get 1
 ; CHECK-NEXT:    i64.select
+; CHECK-NEXT:    local.get 5
+; CHECK-NEXT:    i64.const -9223372036854775808
+; CHECK-NEXT:    local.get 4
+; CHECK-NEXT:    local.get 4
+; CHECK-NEXT:    i64.const 63
+; CHECK-NEXT:    i64.shr_s
+; CHECK-NEXT:    i64.and
 ; CHECK-NEXT:    local.tee 4
 ; CHECK-NEXT:    i64.const -1
 ; CHECK-NEXT:    i64.gt_s
+; CHECK-NEXT:    i64.select
 ; CHECK-NEXT:    local.get 4
 ; CHECK-NEXT:    i64.const -1
 ; CHECK-NEXT:    i64.eq
-; CHECK-NEXT:    i32.select
 ; CHECK-NEXT:    i64.select
 ; CHECK-NEXT:    i64x2.splat
 ; CHECK-NEXT:    local.get 3
 ; CHECK-NEXT:    i64.const 9223372036854775807
 ; CHECK-NEXT:    local.get 3
 ; CHECK-NEXT:    i64.const 9223372036854775807
 ; CHECK-NEXT:    i64.lt_u
+; CHECK-NEXT:    i64.select
+; CHECK-NEXT:    local.get 3
+; CHECK-NEXT:    i64.const 9223372036854775807
 ; CHECK-NEXT:    local.get 2
 ; CHECK-NEXT:    i64.const 0
 ; CHECK-NEXT:    i64.lt_s
+; CHECK-NEXT:    i64.select
 ; CHECK-NEXT:    local.get 2
 ; CHECK-NEXT:    i64.eqz
-; CHECK-NEXT:    i32.select
-; CHECK-NEXT:    local.tee 1
 ; CHECK-NEXT:    i64.select
 ; CHECK-NEXT:    local.tee 4
 ; CHECK-NEXT:    i64.const -9223372036854775808
 ; CHECK-NEXT:    local.get 4
 ; CHECK-NEXT:    i64.const -9223372036854775808
 ; CHECK-NEXT:    i64.gt_u
-; CHECK-NEXT:    local.get 2
-; CHECK-NEXT:    i64.const 0
-; CHECK-NEXT:    local.get 1
 ; CHECK-NEXT:    i64.select
+; CHECK-NEXT:    local.get 4
+; CHECK-NEXT:    i64.const -9223372036854775808
+; CHECK-NEXT:    local.get 2
+; CHECK-NEXT:    local.get 2
+; CHECK-NEXT:    i64.const 63
+; CHECK-NEXT:    i64.shr_s
+; CHECK-NEXT:    i64.and
 ; CHECK-NEXT:    local.tee 2
 ; CHECK-NEXT:    i64.const -1
 ; CHECK-NEXT:    i64.gt_s
+; CHECK-NEXT:    i64.select
 ; CHECK-NEXT:    local.get 2
 ; CHECK-NEXT:    i64.const -1
 ; CHECK-NEXT:    i64.eq
-; CHECK-NEXT:    i32.select
 ; CHECK-NEXT:    i64.select
 ; CHECK-NEXT:    i64x2.replace_lane 1
 ; CHECK-NEXT:    # fallthrough-return
 entry:
   %conv = fptosi <2 x double> %x to <2 x i128>
   %0 = icmp slt <2 x i128> %conv, <i128 9223372036854775807, i128 9223372036854775807>
   %spec.store.select = select <2 x i1> %0, <2 x i128> %conv, <2 x i128> <i128 9223372036854775807, i128 9223372036854775807>
   %1 = icmp sgt <2 x i128> %spec.store.select, <i128 -9223372036854775808, i128 -9223372036854775808>

 ; CHECK-NEXT:    local.get 5
 ; CHECK-NEXT:    i64.const 0
 ; CHECK-NEXT:    local.get 4
 ; CHECK-NEXT:    i64.const 1
 ; CHECK-NEXT:    i64.lt_s
 ; CHECK-NEXT:    local.tee 1
 ; CHECK-NEXT:    i64.select
 ; CHECK-NEXT:    local.tee 5
-; CHECK-NEXT:    i64.const 0
 ; CHECK-NEXT:    local.get 5
 ; CHECK-NEXT:    i64.const 0
-; CHECK-NEXT:    i64.ne
 ; CHECK-NEXT:    local.get 4
 ; CHECK-NEXT:    i64.const 1
 ; CHECK-NEXT:    local.get 1
 ; CHECK-NEXT:    i64.select
 ; CHECK-NEXT:    local.tee 4
 ; CHECK-NEXT:    i64.const 0
 ; CHECK-NEXT:    i64.gt_s
+; CHECK-NEXT:    i64.select
 ; CHECK-NEXT:    local.get 4
 ; CHECK-NEXT:    i64.eqz
-; CHECK-NEXT:    i32.select
 ; CHECK-NEXT:    i64.select
 ; CHECK-NEXT:    i64x2.splat
 ; CHECK-NEXT:    local.get 3
 ; CHECK-NEXT:    i64.const 0
 ; CHECK-NEXT:    local.get 2
 ; CHECK-NEXT:    i64.const 1
 ; CHECK-NEXT:    i64.lt_s
 ; CHECK-NEXT:    local.tee 1
 ; CHECK-NEXT:    i64.select
 ; CHECK-NEXT:    local.tee 4
-; CHECK-NEXT:    i64.const 0
 ; CHECK-NEXT:    local.get 4
 ; CHECK-NEXT:    i64.const 0
-; CHECK-NEXT:    i64.ne
 ; CHECK-NEXT:    local.get 2
 ; CHECK-NEXT:    i64.const 1
 ; CHECK-NEXT:    local.get 1
 ; CHECK-NEXT:    i64.select
 ; CHECK-NEXT:    local.tee 2
 ; CHECK-NEXT:    i64.const 0
 ; CHECK-NEXT:    i64.gt_s
+; CHECK-NEXT:    i64.select
 ; CHECK-NEXT:    local.get 2
 ; CHECK-NEXT:    i64.eqz
-; CHECK-NEXT:    i32.select
 ; CHECK-NEXT:    i64.select
 ; CHECK-NEXT:    i64x2.replace_lane 1
 ; CHECK-NEXT:    # fallthrough-return
 entry:
   %conv = fptosi <2 x double> %x to <2 x i128>
   %0 = icmp slt <2 x i128> %conv, <i128 18446744073709551616, i128 18446744073709551616>
   %spec.store.select = select <2 x i1> %0, <2 x i128> %conv, <2 x i128> <i128 18446744073709551616, i128 18446744073709551616>
   %1 = icmp sgt <2 x i128> %spec.store.select, zeroinitializer

 ; CHECK-NEXT:    i32.const 32
 ; CHECK-NEXT:    i32.add
 ; CHECK-NEXT:    global.set __stack_pointer
 ; CHECK-NEXT:    local.get 5
 ; CHECK-NEXT:    i64.const 9223372036854775807
 ; CHECK-NEXT:    local.get 5
 ; CHECK-NEXT:    i64.const 9223372036854775807
 ; CHECK-NEXT:    i64.lt_u
+; CHECK-NEXT:    i64.select
+; CHECK-NEXT:    local.get 5
+; CHECK-NEXT:    i64.const 9223372036854775807
 ; CHECK-NEXT:    local.get 4
 ; CHECK-NEXT:    i64.const 0
 ; CHECK-NEXT:    i64.lt_s
+; CHECK-NEXT:    i64.select
 ; CHECK-NEXT:    local.get 4
 ; CHECK-NEXT:    i64.eqz
-; CHECK-NEXT:    i32.select
-; CHECK-NEXT:    local.tee 1
 ; CHECK-NEXT:    i64.select
 ; CHECK-NEXT:    local.tee 5
 ; CHECK-NEXT:    i64.const -9223372036854775808
 ; CHECK-NEXT:    local.get 5
 ; CHECK-NEXT:    i64.const -9223372036854775808
 ; CHECK-NEXT:    i64.gt_u
-; CHECK-NEXT:    local.get 4
-; CHECK-NEXT:    i64.const 0
-; CHECK-NEXT:    local.get 1
 ; CHECK-NEXT:    i64.select
+; CHECK-NEXT:    local.get 5
+; CHECK-NEXT:    i64.const -9223372036854775808
+; CHECK-NEXT:    local.get 4
+; CHECK-NEXT:    local.get 4
+; CHECK-NEXT:    i64.const 63
+; CHECK-NEXT:    i64.shr_s
+; CHECK-NEXT:    i64.and
 ; CHECK-NEXT:    local.tee 4
 ; CHECK-NEXT:    i64.const -1
 ; CHECK-NEXT:    i64.gt_s
+; CHECK-NEXT:    i64.select
 ; CHECK-NEXT:    local.get 4
 ; CHECK-NEXT:    i64.const -1
 ; CHECK-NEXT:    i64.eq
-; CHECK-NEXT:    i32.select
 ; CHECK-NEXT:    i64.select
 ; CHECK-NEXT:    i64x2.splat
 ; CHECK-NEXT:    local.get 3
 ; CHECK-NEXT:    i64.const 9223372036854775807
 ; CHECK-NEXT:    local.get 3
 ; CHECK-NEXT:    i64.const 9223372036854775807
 ; CHECK-NEXT:    i64.lt_u
+; CHECK-NEXT:    i64.select
+; CHECK-NEXT:    local.get 3
+; CHECK-NEXT:    i64.const 9223372036854775807
 ; CHECK-NEXT:    local.get 2
 ; CHECK-NEXT:    i64.const 0
 ; CHECK-NEXT:    i64.lt_s
+; CHECK-NEXT:    i64.select
 ; CHECK-NEXT:    local.get 2
 ; CHECK-NEXT:    i64.eqz
-; CHECK-NEXT:    i32.select
-; CHECK-NEXT:    local.tee 1
 ; CHECK-NEXT:    i64.select
 ; CHECK-NEXT:    local.tee 4
 ; CHECK-NEXT:    i64.const -9223372036854775808
 ; CHECK-NEXT:    local.get 4
 ; CHECK-NEXT:    i64.const -9223372036854775808
 ; CHECK-NEXT:    i64.gt_u
-; CHECK-NEXT:    local.get 2
-; CHECK-NEXT:    i64.const 0
-; CHECK-NEXT:    local.get 1
 ; CHECK-NEXT:    i64.select
+; CHECK-NEXT:    local.get 4
+; CHECK-NEXT:    i64.const -9223372036854775808
+; CHECK-NEXT:    local.get 2
+; CHECK-NEXT:    local.get 2
+; CHECK-NEXT:    i64.const 63
+; CHECK-NEXT:    i64.shr_s
+; CHECK-NEXT:    i64.and
 ; CHECK-NEXT:    local.tee 2
 ; CHECK-NEXT:    i64.const -1
 ; CHECK-NEXT:    i64.gt_s
+; CHECK-NEXT:    i64.select
 ; CHECK-NEXT:    local.get 2
 ; CHECK-NEXT:    i64.const -1
 ; CHECK-NEXT:    i64.eq
-; CHECK-NEXT:    i32.select
 ; CHECK-NEXT:    i64.select
 ; CHECK-NEXT:    i64x2.replace_lane 1
 ; CHECK-NEXT:    # fallthrough-return
 entry:
   %conv = fptosi <2 x float> %x to <2 x i128>
   %0 = icmp slt <2 x i128> %conv, <i128 9223372036854775807, i128 9223372036854775807>
   %spec.store.select = select <2 x i1> %0, <2 x i128> %conv, <2 x i128> <i128 9223372036854775807, i128 9223372036854775807>
   %1 = icmp sgt <2 x i128> %spec.store.select, <i128 -9223372036854775808, i128 -9223372036854775808>

 ; CHECK-NEXT:    local.get 5
 ; CHECK-NEXT:    i64.const 0
 ; CHECK-NEXT:    local.get 4
 ; CHECK-NEXT:    i64.const 1
 ; CHECK-NEXT:    i64.lt_s
 ; CHECK-NEXT:    local.tee 1
 ; CHECK-NEXT:    i64.select
 ; CHECK-NEXT:    local.tee 5
-; CHECK-NEXT:    i64.const 0
 ; CHECK-NEXT:    local.get 5
 ; CHECK-NEXT:    i64.const 0
-; CHECK-NEXT:    i64.ne
 ; CHECK-NEXT:    local.get 4
 ; CHECK-NEXT:    i64.const 1
 ; CHECK-NEXT:    local.get 1
 ; CHECK-NEXT:    i64.select
 ; CHECK-NEXT:    local.tee 4
 ; CHECK-NEXT:    i64.const 0
 ; CHECK-NEXT:    i64.gt_s
+; CHECK-NEXT:    i64.select
 ; CHECK-NEXT:    local.get 4
 ; CHECK-NEXT:    i64.eqz
-; CHECK-NEXT:    i32.select
 ; CHECK-NEXT:    i64.select
 ; CHECK-NEXT:    i64x2.splat
 ; CHECK-NEXT:    local.get 3
 ; CHECK-NEXT:    i64.const 0
 ; CHECK-NEXT:    local.get 2
 ; CHECK-NEXT:    i64.const 1
 ; CHECK-NEXT:    i64.lt_s
 ; CHECK-NEXT:    local.tee 1
 ; CHECK-NEXT:    i64.select
 ; CHECK-NEXT:    local.tee 4
-; CHECK-NEXT:    i64.const 0
 ; CHECK-NEXT:    local.get 4
 ; CHECK-NEXT:    i64.const 0
-; CHECK-NEXT:    i64.ne
 ; CHECK-NEXT:    local.get 2
 ; CHECK-NEXT:    i64.const 1
 ; CHECK-NEXT:    local.get 1
 ; CHECK-NEXT:    i64.select
 ; CHECK-NEXT:    local.tee 2
 ; CHECK-NEXT:    i64.const 0
 ; CHECK-NEXT:    i64.gt_s
+; CHECK-NEXT:    i64.select
 ; CHECK-NEXT:    local.get 2
 ; CHECK-NEXT:    i64.eqz
-; CHECK-NEXT:    i32.select
 ; CHECK-NEXT:    i64.select
 ; CHECK-NEXT:    i64x2.replace_lane 1
 ; CHECK-NEXT:    # fallthrough-return
 entry:
   %conv = fptosi <2 x float> %x to <2 x i128>
   %0 = icmp slt <2 x i128> %conv, <i128 18446744073709551616, i128 18446744073709551616>
   %spec.store.select = select <2 x i1> %0, <2 x i128> %conv, <2 x i128> <i128 18446744073709551616, i128 18446744073709551616>
   %1 = icmp sgt <2 x i128> %spec.store.select, zeroinitializer

 ; CHECK-NEXT:    i32.const 32
 ; CHECK-NEXT:    i32.add
 ; CHECK-NEXT:    global.set __stack_pointer
 ; CHECK-NEXT:    local.get 6
 ; CHECK-NEXT:    i64.const 9223372036854775807
 ; CHECK-NEXT:    local.get 6
 ; CHECK-NEXT:    i64.const 9223372036854775807
 ; CHECK-NEXT:    i64.lt_u
+; CHECK-NEXT:    i64.select
+; CHECK-NEXT:    local.get 6
+; CHECK-NEXT:    i64.const 9223372036854775807
 ; CHECK-NEXT:    local.get 5
 ; CHECK-NEXT:    i64.const 0
 ; CHECK-NEXT:    i64.lt_s
+; CHECK-NEXT:    i64.select
 ; CHECK-NEXT:    local.get 5
 ; CHECK-NEXT:    i64.eqz
-; CHECK-NEXT:    i32.select
-; CHECK-NEXT:    local.tee 2
 ; CHECK-NEXT:    i64.select
 ; CHECK-NEXT:    local.tee 6
 ; CHECK-NEXT:    i64.const -9223372036854775808
 ; CHECK-NEXT:    local.get 6
 ; CHECK-NEXT:    i64.const -9223372036854775808
 ; CHECK-NEXT:    i64.gt_u
-; CHECK-NEXT:    local.get 5
-; CHECK-NEXT:    i64.const 0
-; CHECK-NEXT:    local.get 2
 ; CHECK-NEXT:    i64.select
+; CHECK-NEXT:    local.get 6
+; CHECK-NEXT:    i64.const -9223372036854775808
+; CHECK-NEXT:    local.get 5
+; CHECK-NEXT:    local.get 5
+; CHECK-NEXT:    i64.const 63
+; CHECK-NEXT:    i64.shr_s
+; CHECK-NEXT:    i64.and
 ; CHECK-NEXT:    local.tee 5
 ; CHECK-NEXT:    i64.const -1
 ; CHECK-NEXT:    i64.gt_s
+; CHECK-NEXT:    i64.select
 ; CHECK-NEXT:    local.get 5
 ; CHECK-NEXT:    i64.const -1
 ; CHECK-NEXT:    i64.eq
-; CHECK-NEXT:    i32.select
 ; CHECK-NEXT:    i64.select
 ; CHECK-NEXT:    i64x2.splat
 ; CHECK-NEXT:    local.get 4
 ; CHECK-NEXT:    i64.const 9223372036854775807
 ; CHECK-NEXT:    local.get 4
 ; CHECK-NEXT:    i64.const 9223372036854775807
 ; CHECK-NEXT:    i64.lt_u
+; CHECK-NEXT:    i64.select
+; CHECK-NEXT:    local.get 4
+; CHECK-NEXT:    i64.const 9223372036854775807
 ; CHECK-NEXT:    local.get 3
 ; CHECK-NEXT:    i64.const 0
 ; CHECK-NEXT:    i64.lt_s
+; CHECK-NEXT:    i64.select
 ; CHECK-NEXT:    local.get 3
 ; CHECK-NEXT:    i64.eqz
-; CHECK-NEXT:    i32.select
-; CHECK-NEXT:    local.tee 2
 ; CHECK-NEXT:    i64.select
 ; CHECK-NEXT:    local.tee 5
 ; CHECK-NEXT:    i64.const -9223372036854775808
 ; CHECK-NEXT:    local.get 5
 ; CHECK-NEXT:    i64.const -9223372036854775808
 ; CHECK-NEXT:    i64.gt_u
-; CHECK-NEXT:    local.get 3
-; CHECK-NEXT:    i64.const 0
-; CHECK-NEXT:    local.get 2
 ; CHECK-NEXT:    i64.select
+; CHECK-NEXT:    local.get 5
+; CHECK-NEXT:    i64.const -9223372036854775808
+; CHECK-NEXT:    local.get 3
+; CHECK-NEXT:    local.get 3
+; CHECK-NEXT:    i64.const 63
+; CHECK-NEXT:    i64.shr_s
+; CHECK-NEXT:    i64.and
 ; CHECK-NEXT:    local.tee 3
 ; CHECK-NEXT:    i64.const -1
 ; CHECK-NEXT:    i64.gt_s
+; CHECK-NEXT:    i64.select
 ; CHECK-NEXT:    local.get 3
 ; CHECK-NEXT:    i64.const -1
 ; CHECK-NEXT:    i64.eq
-; CHECK-NEXT:    i32.select
 ; CHECK-NEXT:    i64.select
 ; CHECK-NEXT:    i64x2.replace_lane 1
 ; CHECK-NEXT:    # fallthrough-return
 entry:
   %conv = fptosi <2 x half> %x to <2 x i128>
   %0 = icmp slt <2 x i128> %conv, <i128 9223372036854775807, i128 9223372036854775807>
   %spec.store.select = select <2 x i1> %0, <2 x i128> %conv, <2 x i128> <i128 9223372036854775807, i128 9223372036854775807>
   %1 = icmp sgt <2 x i128> %spec.store.select, <i128 -9223372036854775808, i128 -9223372036854775808>
   %spec.store.select7 = select <2 x i1> %1, <2 x i128> %spec.store.select, <2 x i128> <i128 -9223372036854775808, i128 -9223372036854775808>
   %conv6 = trunc <2 x i128> %spec.store.select7 to <2 x i64>
   ret <2 x i64> %conv6
 }
 
 define <2 x i64> @utesth_f16i64(<2 x half> %x) {
 ; CHECK-LABEL: utesth_f16i64:
 ; CHECK:         .functype utesth_f16i64 (f32, f32) -> (v128)
-; CHECK-NEXT:    .local i32, i64, i64, i64, i64
 ; CHECK-NEXT:  # %bb.0: # %entry
-; CHECK-NEXT:    global.get __stack_pointer
-; CHECK-NEXT:    i32.const 32
-; CHECK-NEXT:    i32.sub
-; CHECK-NEXT:    local.tee 2
-; CHECK-NEXT:    global.set __stack_pointer
-; CHECK-NEXT:    local.get 2
-; CHECK-NEXT:    i32.const 16
-; CHECK-NEXT:    i32.add
 ; CHECK-NEXT:    local.get 1
 ; CHECK-NEXT:    call __truncsfhf2
 ; CHECK-NEXT:    call __extendhfsf2
-; CHECK-NEXT:    call __fixunssfti
-; CHECK-NEXT:    local.get 2
+; CHECK-NEXT:    local.set 1
 ; CHECK-NEXT:    local.get 0
 ; CHECK-NEXT:    call __truncsfhf2
 ; CHECK-NEXT:    call __extendhfsf2
-; CHECK-NEXT:    call __fixunssfti
-; CHECK-NEXT:    local.get 2
-; CHECK-NEXT:    i32.const 16
-; CHECK-NEXT:    i32.add
-; CHECK-NEXT:    i32.const 8
-; CHECK-NEXT:    i32.add
-; CHECK-NEXT:    i64.load 0
-; CHECK-NEXT:    local.set 3
-; CHECK-NEXT:    local.get 2
-; CHECK-NEXT:    i64.load 16
-; CHECK-NEXT:    local.set 4
-; CHECK-NEXT:    local.get 2
-; CHECK-NEXT:    i32.const 8
-; CHECK-NEXT:    i32.add
-; CHECK-NEXT:    i64.load 0
-; CHECK-NEXT:    local.set 5
-; CHECK-NEXT:    local.get 2
-; CHECK-NEXT:    i64.load 0
-; CHECK-NEXT:    local.set 6
-; CHECK-NEXT:    local.get 2
-; CHECK-NEXT:    i32.const 32
-; CHECK-NEXT:    i32.add
-; CHECK-NEXT:    global.set __stack_pointer
-; CHECK-NEXT:    local.get 6
-; CHECK-NEXT:    i64.const 0
-; CHECK-NEXT:    local.get 5
-; CHECK-NEXT:    i64.eqz
-; CHECK-NEXT:    i64.select
+; CHECK-NEXT:    i64.trunc_sat_f32_u
 ; CHECK-NEXT:    i64x2.splat
-; CHECK-NEXT:    local.get 4
-; CHECK-NEXT:    i64.const 0
-; CHECK-NEXT:    local.get 3
-; CHECK-NEXT:    i64.eqz
-; CHECK-NEXT:    i64.select
+; CHECK-NEXT:    local.get 1
+; CHECK-NEXT:    i64.trunc_sat_f32_u
 ; CHECK-NEXT:    i64x2.replace_lane 1
 ; CHECK-NEXT:    # fallthrough-return
 entry:
   %conv = fptoui <2 x half> %x to <2 x i128>
   %0 = icmp ult <2 x i128> %conv, <i128 18446744073709551616, i128 18446744073709551616>
   %spec.store.select = select <2 x i1> %0, <2 x i128> %conv, <2 x i128> <i128 18446744073709551616, i128 18446744073709551616>
   %conv6 = trunc <2 x i128> %spec.store.select to <2 x i64>
   ret <2 x i64> %conv6

 ; CHECK-NEXT:    local.get 6
 ; CHECK-NEXT:    i64.const 0
 ; CHECK-NEXT:    local.get 5
 ; CHECK-NEXT:    i64.const 1
 ; CHECK-NEXT:    i64.lt_s
 ; CHECK-NEXT:    local.tee 2
 ; CHECK-NEXT:    i64.select
 ; CHECK-NEXT:    local.tee 6
-; CHECK-NEXT:    i64.const 0
 ; CHECK-NEXT:    local.get 6
 ; CHECK-NEXT:    i64.const 0
-; CHECK-NEXT:    i64.ne
 ; CHECK-NEXT:    local.get 5
 ; CHECK-NEXT:    i64.const 1
 ; CHECK-NEXT:    local.get 2
 ; CHECK-NEXT:    i64.select
 ; CHECK-NEXT:    local.tee 5
 ; CHECK-NEXT:    i64.const 0
 ; CHECK-NEXT:    i64.gt_s
+; CHECK-NEXT:    i64.select
 ; CHECK-NEXT:    local.get 5
 ; CHECK-NEXT:    i64.eqz
-; CHECK-NEXT:    i32.select
 ; CHECK-NEXT:    i64.select
 ; CHECK-NEXT:    i64x2.splat
 ; CHECK-NEXT:    local.get 4
 ; CHECK-NEXT:    i64.const 0
 ; CHECK-NEXT:    local.get 3
 ; CHECK-NEXT:    i64.const 1
 ; CHECK-NEXT:    i64.lt_s
 ; CHECK-NEXT:    local.tee 2
 ; CHECK-NEXT:    i64.select
 ; CHECK-NEXT:    local.tee 5
-; CHECK-NEXT:    i64.const 0
 ; CHECK-NEXT:    local.get 5
 ; CHECK-NEXT:    i64.const 0
-; CHECK-NEXT:    i64.ne
 ; CHECK-NEXT:    local.get 3
 ; CHECK-NEXT:    i64.const 1
 ; CHECK-NEXT:    local.get 2
 ; CHECK-NEXT:    i64.select
 ; CHECK-NEXT:    local.tee 3
 ; CHECK-NEXT:    i64.const 0
 ; CHECK-NEXT:    i64.gt_s
+; CHECK-NEXT:    i64.select
 ; CHECK-NEXT:    local.get 3
 ; CHECK-NEXT:    i64.eqz
-; CHECK-NEXT:    i32.select
 ; CHECK-NEXT:    i64.select
 ; CHECK-NEXT:    i64x2.replace_lane 1
 ; CHECK-NEXT:    # fallthrough-return
 entry:
   %conv = fptosi <2 x half> %x to <2 x i128>
   %0 = icmp slt <2 x i128> %conv, <i128 18446744073709551616, i128 18446744073709551616>
   %spec.store.select = select <2 x i1> %0, <2 x i128> %conv, <2 x i128> <i128 18446744073709551616, i128 18446744073709551616>
   %1 = icmp sgt <2 x i128> %spec.store.select, zeroinitializer
entry:
   %spec.store.select7 = call <8 x i32> @llvm.smax.v8i32(<8 x i32> %spec.store.select, <8 x i32> <i32 -32768, i32 -32768, i32 -32768, i32 -32768, i32 -32768, i32 -32768, i32 -32768, i32 -32768>)
   %conv6 = trunc <8 x i32> %spec.store.select7 to <8 x i16>
   ret <8 x i16> %conv6
 }
 
 define <8 x i16> @utesth_f16i16_mm(<8 x half> %x) {
 ; CHECK-LABEL: utesth_f16i16_mm:
 ; CHECK:         .functype utesth_f16i16_mm (f32, f32, f32, f32, f32, f32, f32, f32) -> (v128)
-; CHECK-NEXT:    .local v128
 ; CHECK-NEXT:  # %bb.0: # %entry
-; CHECK-NEXT:    local.get 5
-; CHECK-NEXT:    call __truncsfhf2
-; CHECK-NEXT:    call __extendhfsf2
-; CHECK-NEXT:    local.set 5
-; CHECK-NEXT:    local.get 4
-; CHECK-NEXT:    call __truncsfhf2
-; CHECK-NEXT:    call __extendhfsf2
-; CHECK-NEXT:    local.set 4
-; CHECK-NEXT:    local.get 6
-; CHECK-NEXT:    call __truncsfhf2
-; CHECK-NEXT:    call __extendhfsf2
-; CHECK-NEXT:    local.set 6
-; CHECK-NEXT:    local.get 7
-; CHECK-NEXT:    call __truncsfhf2
-; CHECK-NEXT:    call __extendhfsf2
-; CHECK-NEXT:    local.set 7
 ; CHECK-NEXT:    local.get 1
 ; CHECK-NEXT:    call __truncsfhf2
 ; CHECK-NEXT:    call __extendhfsf2
 ; CHECK-NEXT:    local.set 1
 ; CHECK-NEXT:    local.get 0
 ; CHECK-NEXT:    call __truncsfhf2
 ; CHECK-NEXT:    call __extendhfsf2
 ; CHECK-NEXT:    i32.trunc_sat_f32_u
-; CHECK-NEXT:    i32x4.splat
+; CHECK-NEXT:    i16x8.splat
 ; CHECK-NEXT:    local.get 1
 ; CHECK-NEXT:    i32.trunc_sat_f32_u
-; CHECK-NEXT:    i32x4.replace_lane 1
+; CHECK-NEXT:    i16x8.replace_lane 1
 ; CHECK-NEXT:    local.get 2
 ; CHECK-NEXT:    call __truncsfhf2
 ; CHECK-NEXT:    call __extendhfsf2
 ; CHECK-NEXT:    i32.trunc_sat_f32_u
-; CHECK-NEXT:    i32x4.replace_lane 2
+; CHECK-NEXT:    i16x8.replace_lane 2
 ; CHECK-NEXT:    local.get 3
 ; CHECK-NEXT:    call __truncsfhf2
 ; CHECK-NEXT:    call __extendhfsf2
 ; CHECK-NEXT:    i32.trunc_sat_f32_u
-; CHECK-NEXT:    i32x4.replace_lane 3
-; CHECK-NEXT:    v128.const 65535, 65535, 65535, 65535
-; CHECK-NEXT:    local.tee 8
-; CHECK-NEXT:    i32x4.min_u
+; CHECK-NEXT:    i16x8.replace_lane 3
 ; CHECK-NEXT:    local.get 4
+; CHECK-NEXT:    call __truncsfhf2
+; CHECK-NEXT:    call __extendhfsf2
 ; CHECK-NEXT:    i32.trunc_sat_f32_u
-; CHECK-NEXT:    i32x4.splat
+; CHECK-NEXT:    i16x8.replace_lane 4
 ; CHECK-NEXT:    local.get 5
+; CHECK-NEXT:    call __truncsfhf2
+; CHECK-NEXT:    call __extendhfsf2
 ; CHECK-NEXT:    i32.trunc_sat_f32_u
-; CHECK-NEXT:    i32x4.replace_lane 1
+; CHECK-NEXT:    i16x8.replace_lane 5
 ; CHECK-NEXT:    local.get 6
+; CHECK-NEXT:    call __truncsfhf2
+; CHECK-NEXT:    call __extendhfsf2
 ; CHECK-NEXT:    i32.trunc_sat_f32_u
-; CHECK-NEXT:    i32x4.replace_lane 2
+; CHECK-NEXT:    i16x8.replace_lane 6
 ; CHECK-NEXT:    local.get 7
+; CHECK-NEXT:    call __truncsfhf2
+; CHECK-NEXT:    call __extendhfsf2
 ; CHECK-NEXT:    i32.trunc_sat_f32_u
-; CHECK-NEXT:    i32x4.replace_lane 3
-; CHECK-NEXT:    local.get 8
-; CHECK-NEXT:    i32x4.min_u
-; CHECK-NEXT:    i16x8.narrow_i32x4_u
+; CHECK-NEXT:    i16x8.replace_lane 7
 ; CHECK-NEXT:    # fallthrough-return
 entry:
   %conv = fptoui <8 x half> %x to <8 x i32>
   %spec.store.select = call <8 x i32> @llvm.umin.v8i32(<8 x i32> %conv, <8 x i32> <i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535>)
   %conv6 = trunc <8 x i32> %spec.store.select to <8 x i16>
   ret <8 x i16> %conv6
 }
 

 ; CHECK-NEXT:    call __extendhfsf2
 ; CHECK-NEXT:    i32.trunc_sat_f32_s
 ; CHECK-NEXT:    i32x4.replace_lane 2
 ; CHECK-NEXT:    local.get 3
 ; CHECK-NEXT:    call __truncsfhf2
 ; CHECK-NEXT:    call __extendhfsf2
 ; CHECK-NEXT:    i32.trunc_sat_f32_s
 ; CHECK-NEXT:    i32x4.replace_lane 3
-; CHECK-NEXT:    v128.const 65535, 65535, 65535, 65535
-; CHECK-NEXT:    local.tee 8
-; CHECK-NEXT:    i32x4.min_s
 ; CHECK-NEXT:    v128.const 0, 0, 0, 0
-; CHECK-NEXT:    local.tee 9
+; CHECK-NEXT:    local.tee 8
 ; CHECK-NEXT:    i32x4.max_s
+; CHECK-NEXT:    v128.const 65535, 65535, 65535, 65535
+; CHECK-NEXT:    local.tee 9
+; CHECK-NEXT:    v128.and
 ; CHECK-NEXT:    local.get 4
 ; CHECK-NEXT:    i32.trunc_sat_f32_s
 ; CHECK-NEXT:    i32x4.splat
 ; CHECK-NEXT:    local.get 5
 ; CHECK-NEXT:    i32.trunc_sat_f32_s
 ; CHECK-NEXT:    i32x4.replace_lane 1
 ; CHECK-NEXT:    local.get 6
 ; CHECK-NEXT:    i32.trunc_sat_f32_s
 ; CHECK-NEXT:    i32x4.replace_lane 2
 ; CHECK-NEXT:    local.get 7
 ; CHECK-NEXT:    i32.trunc_sat_f32_s
 ; CHECK-NEXT:    i32x4.replace_lane 3
 ; CHECK-NEXT:    local.get 8
-; CHECK-NEXT:    i32x4.min_s
-; CHECK-NEXT:    local.get 9
 ; CHECK-NEXT:    i32x4.max_s
+; CHECK-NEXT:    local.get 9
+; CHECK-NEXT:    v128.and
 ; CHECK-NEXT:    i16x8.narrow_i32x4_u
 ; CHECK-NEXT:    # fallthrough-return
 entry:
   %conv = fptosi <8 x half> %x to <8 x i32>
   %spec.store.select = call <8 x i32> @llvm.smin.v8i32(<8 x i32> %conv, <8 x i32> <i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535, i32 65535>)
   %spec.store.select7 = call <8 x i32> @llvm.smax.v8i32(<8 x i32> %spec.store.select, <8 x i32> zeroinitializer)
   %conv6 = trunc <8 x i32> %spec.store.select7 to <8 x i16>
   ret <8 x i16> %conv6
entry:
   %spec.store.select7 = call <2 x i128> @llvm.smax.v2i128(<2 x i128> %spec.store.select, <2 x i128> <i128 -9223372036854775808, i128 -9223372036854775808>)
   %conv6 = trunc <2 x i128> %spec.store.select7 to <2 x i64>
   ret <2 x i64> %conv6
 }
 
 define <2 x i64> @utesth_f16i64_mm(<2 x half> %x) {
 ; CHECK-LABEL: utesth_f16i64_mm:
 ; CHECK:         .functype utesth_f16i64_mm (f32, f32) -> (v128)
-; CHECK-NEXT:    .local i32, i64, i64, i64, i64
 ; CHECK-NEXT:  # %bb.0: # %entry
-; CHECK-NEXT:    global.get __stack_pointer
-; CHECK-NEXT:    i32.const 32
-; CHECK-NEXT:    i32.sub
-; CHECK-NEXT:    local.tee 2
-; CHECK-NEXT:    global.set __stack_pointer
-; CHECK-NEXT:    local.get 2
-; CHECK-NEXT:    i32.const 16
-; CHECK-NEXT:    i32.add
 ; CHECK-NEXT:    local.get 1
 ; CHECK-NEXT:    call __truncsfhf2
 ; CHECK-NEXT:    call __extendhfsf2
-; CHECK-NEXT:    call __fixunssfti
-; CHECK-NEXT:    local.get 2
+; CHECK-NEXT:    local.set 1
 ; CHECK-NEXT:    local.get 0
 ; CHECK-NEXT:    call __truncsfhf2
 ; CHECK-NEXT:    call __extendhfsf2
-; CHECK-NEXT:    call __fixunssfti
-; CHECK-NEXT:    local.get 2
-; CHECK-NEXT:    i32.const 16
-; CHECK-NEXT:    i32.add
-; CHECK-NEXT:    i32.const 8
-; CHECK-NEXT:    i32.add
-; CHECK-NEXT:    i64.load 0
-; CHECK-NEXT:    local.set 3
-; CHECK-NEXT:    local.get 2
-; CHECK-NEXT:    i64.load 16
-; CHECK-NEXT:    local.set 4
-; CHECK-NEXT:    local.get 2
-; CHECK-NEXT:    i32.const 8
-; CHECK-NEXT:    i32.add
-; CHECK-NEXT:    i64.load 0
-; CHECK-NEXT:    local.set 5
-; CHECK-NEXT:    local.get 2
-; CHECK-NEXT:    i64.load 0
-; CHECK-NEXT:    local.set 6
-; CHECK-NEXT:    local.get 2
-; CHECK-NEXT:    i32.const 32
-; CHECK-NEXT:    i32.add
-; CHECK-NEXT:    global.set __stack_pointer
-; CHECK-NEXT:    local.get 6
-; CHECK-NEXT:    i64.const 0
-; CHECK-NEXT:    local.get 5
-; CHECK-NEXT:    i64.eqz
-; CHECK-NEXT:    i64.select
+; CHECK-NEXT:    i64.trunc_sat_f32_u
 ; CHECK-NEXT:    i64x2.splat
-; CHECK-NEXT:    local.get 4
-; CHECK-NEXT:    i64.const 0
-; CHECK-NEXT:    local.get 3
-; CHECK-NEXT:    i64.eqz
-; CHECK-NEXT:    i64.select
+; CHECK-NEXT:    local.get 1
+; CHECK-NEXT:    i64.trunc_sat_f32_u
 ; CHECK-NEXT:    i64x2.replace_lane 1
 ; CHECK-NEXT:    # fallthrough-return
 entry:
   %conv = fptoui <2 x half> %x to <2 x i128>
   %spec.store.select = call <2 x i128> @llvm.umin.v2i128(<2 x i128> %conv, <2 x i128> <i128 18446744073709551616, i128 18446744073709551616>)
   %conv6 = trunc <2 x i128> %spec.store.select to <2 x i64>
   ret <2 x i64> %conv6
 }

llvm/test/CodeGen/WebAssembly/funcref-tableget.ll

-; RUN: llc < %s --mtriple=wasm32-unknown-unknown -asm-verbose=false -mattr=+reference-types | FileCheck %s
+; RUN: llc < %s -O1 --mtriple=wasm32-unknown-unknown -asm-verbose=false -mattr=+reference-types | FileCheck %s
 
 %funcref = type ptr addrspace(20) ;; addrspace 20 is nonintegral
 
 @funcref_table = local_unnamed_addr addrspace(1) global [0 x %funcref] undef
 
 declare %funcref @llvm.wasm.table.get.funcref(ptr addrspace(1), i32) nounwind
 
 define %funcref @get_funcref_from_table(i32 %i) {

llvm/test/CodeGen/WebAssembly/funcref-tableset.ll

-; RUN: llc --mtriple=wasm32-unknown-unknown -asm-verbose=false -mattr=+reference-types < %s | FileCheck %s
+; RUN: llc -O1 --mtriple=wasm32-unknown-unknown -asm-verbose=false -mattr=+reference-types < %s | FileCheck %s
 
 %funcref = type ptr addrspace(20) ;; addrspace 20 is nonintegral
 
 @funcref_table = local_unnamed_addr addrspace(1) global [0 x %funcref] undef
 
 declare void @llvm.wasm.table.set.funcref(ptr addrspace(1), i32, %funcref) nounwind
 
 define void @set_funcref_table(%funcref %g, i32 %i) {

llvm/test/CodeGen/WebAssembly/function-addr-offset.ll

-; RUN: llc < %s -verify-machineinstrs | FileCheck %s
+; RUN: llc < %s -O1 -verify-machineinstrs | FileCheck %s
 
 ; Wasm does not currently support function addresses with offsets, so we
 ; shouldn't try to create a folded SDNode like (function + offset). This is a
 ; regression test for the folding bug and this should not crash in MCInstLower.
 
 target triple = "wasm32-unknown-unknown"
 
 ; 'hidden' here should be present to reproduce the bug

llvm/test/CodeGen/WebAssembly/global-get-unlowerable.ll

-; RUN: not llc < %s --mtriple=wasm32 2>&1 | FileCheck %s
+; RUN: not llc < %s -O1 --mtriple=wasm32 2>&1 | FileCheck %s
 
 ; Demonstrates a code pattern that could be encountered even with frontend
 ; restrictions on creating new pointers to globals. In the absence of a better
 ; fix, the backend should produce a comprehensible message for why it can't
 ; continue.
 
 ; CHECK: LVM ERROR: Encountered an unlowerable load from the wasm_var address space
 
 @g1 = external addrspace(1) global i32
 @g2 = external addrspace(1) global i32
 
 define i32 @global_get_phi(i1 zeroext %bool) {
   %sel = select i1 %bool, ptr addrspace(1) @g1, ptr addrspace(1) @g2
   %gval = load i32, ptr addrspace(1) %sel
   ret i32 %gval
 }

llvm/test/CodeGen/WebAssembly/i128.ll

; CHECK-NEXT:    return
   %c = lshr i128 %x, %z
   %d = or i128 %b, %c
   ret i128 %d
 }
 
 define i128 @masked_rotl(i128 %x, i128 %y) {
 ; CHECK-LABEL: masked_rotl:
 ; CHECK:         .functype masked_rotl (i32, i64, i64, i64, i64) -> ()
-; CHECK-NEXT:    .local i32, i32
+; CHECK-NEXT:    .local i32, i64
 ; CHECK-NEXT:  # %bb.0:
-; CHECK-NEXT:    global.get $push13=, __stack_pointer
-; CHECK-NEXT:    i32.const $push14=, 32
-; CHECK-NEXT:    i32.sub $push25=, $pop13, $pop14
-; CHECK-NEXT:    local.tee $push24=, 5, $pop25
-; CHECK-NEXT:    global.set __stack_pointer, $pop24
-; CHECK-NEXT:    local.get $push26=, 5
-; CHECK-NEXT:    i32.const $push17=, 16
-; CHECK-NEXT:    i32.add $push18=, $pop26, $pop17
-; CHECK-NEXT:    local.get $push29=, 1
-; CHECK-NEXT:    local.get $push28=, 2
-; CHECK-NEXT:    local.get $push27=, 3
-; CHECK-NEXT:    i32.wrap_i64 $push0=, $pop27
-; CHECK-NEXT:    i32.const $push1=, 127
-; CHECK-NEXT:    i32.and $push23=, $pop0, $pop1
-; CHECK-NEXT:    local.tee $push22=, 6, $pop23
-; CHECK-NEXT:    call __ashlti3, $pop18, $pop29, $pop28, $pop22
-; CHECK-NEXT:    local.get $push33=, 5
-; CHECK-NEXT:    local.get $push32=, 1
-; CHECK-NEXT:    local.get $push31=, 2
-; CHECK-NEXT:    i32.const $push2=, 128
-; CHECK-NEXT:    local.get $push30=, 6
-; CHECK-NEXT:    i32.sub $push3=, $pop2, $pop30
-; CHECK-NEXT:    call __lshrti3, $pop33, $pop32, $pop31, $pop3
+; CHECK-NEXT:    local.get $push31=, 0
+; CHECK-NEXT:    local.get $push25=, 1
+; CHECK-NEXT:    local.get $push24=, 2
+; CHECK-NEXT:    local.get $push23=, 3
+; CHECK-NEXT:    i64.const $push1=, 64
+; CHECK-NEXT:    i64.and $push2=, $pop23, $pop1
+; CHECK-NEXT:    i64.const $push3=, 6
+; CHECK-NEXT:    i64.shr_u $push4=, $pop2, $pop3
+; CHECK-NEXT:    i32.wrap_i64 $push22=, $pop4
+; CHECK-NEXT:    local.tee $push21=, 5, $pop22
+; CHECK-NEXT:    i64.select $push20=, $pop25, $pop24, $pop21
+; CHECK-NEXT:    local.tee $push19=, 6, $pop20
+; CHECK-NEXT:    local.get $push26=, 3
+; CHECK-NEXT:    i64.shl $push8=, $pop19, $pop26
+; CHECK-NEXT:    local.get $push29=, 2
+; CHECK-NEXT:    local.get $push28=, 1
+; CHECK-NEXT:    local.get $push27=, 5
+; CHECK-NEXT:    i64.select $push18=, $pop29, $pop28, $pop27
+; CHECK-NEXT:    local.tee $push17=, 1, $pop18
+; CHECK-NEXT:    i64.const $push5=, 1
+; CHECK-NEXT:    i64.shr_u $push6=, $pop17, $pop5
+; CHECK-NEXT:    local.get $push30=, 3
+; CHECK-NEXT:    i64.const $push0=, -1
+; CHECK-NEXT:    i64.xor $push16=, $pop30, $pop0
+; CHECK-NEXT:    local.tee $push15=, 2, $pop16
+; CHECK-NEXT:    i64.shr_u $push7=, $pop6, $pop15
+; CHECK-NEXT:    i64.or $push9=, $pop8, $pop7
+; CHECK-NEXT:    i64.store 8($pop31), $pop9
 ; CHECK-NEXT:    local.get $push36=, 0
-; CHECK-NEXT:    local.get $push34=, 5
-; CHECK-NEXT:    i32.const $push19=, 16
-; CHECK-NEXT:    i32.add $push20=, $pop34, $pop19
-; CHECK-NEXT:    i32.const $push4=, 8
-; CHECK-NEXT:    i32.add $push5=, $pop20, $pop4
-; CHECK-NEXT:    i64.load $push6=, 0($pop5)
-; CHECK-NEXT:    local.get $push35=, 5
-; CHECK-NEXT:    i32.const $push21=, 8
-; CHECK-NEXT:    i32.add $push7=, $pop35, $pop21
-; CHECK-NEXT:    i64.load $push8=, 0($pop7)
-; CHECK-NEXT:    i64.or $push9=, $pop6, $pop8
-; CHECK-NEXT:    i64.store 8($pop36), $pop9
-; CHECK-NEXT:    local.get $push39=, 0
-; CHECK-NEXT:    local.get $push37=, 5
-; CHECK-NEXT:    i64.load $push10=, 16($pop37)
-; CHECK-NEXT:    local.get $push38=, 5
-; CHECK-NEXT:    i64.load $push11=, 0($pop38)
-; CHECK-NEXT:    i64.or $push12=, $pop10, $pop11
-; CHECK-NEXT:    i64.store 0($pop39), $pop12
-; CHECK-NEXT:    local.get $push40=, 5
-; CHECK-NEXT:    i32.const $push15=, 32
-; CHECK-NEXT:    i32.add $push16=, $pop40, $pop15
-; CHECK-NEXT:    global.set __stack_pointer, $pop16
+; CHECK-NEXT:    local.get $push33=, 1
+; CHECK-NEXT:    local.get $push32=, 3
+; CHECK-NEXT:    i64.shl $push12=, $pop33, $pop32
+; CHECK-NEXT:    local.get $push34=, 6
+; CHECK-NEXT:    i64.const $push14=, 1
+; CHECK-NEXT:    i64.shr_u $push10=, $pop34, $pop14
+; CHECK-NEXT:    local.get $push35=, 2
+; CHECK-NEXT:    i64.shr_u $push11=, $pop10, $pop35
+; CHECK-NEXT:    i64.or $push13=, $pop12, $pop11
+; CHECK-NEXT:    i64.store 0($pop36), $pop13
 ; CHECK-NEXT:    return
   %a = and i128 %y, 127
   %z = sub i128 128, %a
   %b = shl i128 %x, %a
   %c = lshr i128 %x, %z
   %d = or i128 %b, %c
   ret i128 %d
 }
; CHECK-NEXT:    return
   %c = shl i128 %x, %z
   %d = or i128 %b, %c
   ret i128 %d
 }
 
 define i128 @masked_rotr(i128 %x, i128 %y) {
 ; CHECK-LABEL: masked_rotr:
 ; CHECK:         .functype masked_rotr (i32, i64, i64, i64, i64) -> ()
-; CHECK-NEXT:    .local i32, i32
+; CHECK-NEXT:    .local i32, i64, i64
 ; CHECK-NEXT:  # %bb.0:
-; CHECK-NEXT:    global.get $push13=, __stack_pointer
-; CHECK-NEXT:    i32.const $push14=, 32
-; CHECK-NEXT:    i32.sub $push25=, $pop13, $pop14
-; CHECK-NEXT:    local.tee $push24=, 5, $pop25
-; CHECK-NEXT:    global.set __stack_pointer, $pop24
-; CHECK-NEXT:    local.get $push26=, 5
-; CHECK-NEXT:    i32.const $push17=, 16
-; CHECK-NEXT:    i32.add $push18=, $pop26, $pop17
-; CHECK-NEXT:    local.get $push29=, 1
-; CHECK-NEXT:    local.get $push28=, 2
-; CHECK-NEXT:    local.get $push27=, 3
-; CHECK-NEXT:    i32.wrap_i64 $push0=, $pop27
-; CHECK-NEXT:    i32.const $push1=, 127
-; CHECK-NEXT:    i32.and $push23=, $pop0, $pop1
-; CHECK-NEXT:    local.tee $push22=, 6, $pop23
-; CHECK-NEXT:    call __lshrti3, $pop18, $pop29, $pop28, $pop22
-; CHECK-NEXT:    local.get $push33=, 5
-; CHECK-NEXT:    local.get $push32=, 1
-; CHECK-NEXT:    local.get $push31=, 2
-; CHECK-NEXT:    i32.const $push2=, 128
-; CHECK-NEXT:    local.get $push30=, 6
-; CHECK-NEXT:    i32.sub $push3=, $pop2, $pop30
-; CHECK-NEXT:    call __ashlti3, $pop33, $pop32, $pop31, $pop3
-; CHECK-NEXT:    local.get $push36=, 0
-; CHECK-NEXT:    local.get $push34=, 5
-; CHECK-NEXT:    i32.const $push19=, 16
-; CHECK-NEXT:    i32.add $push20=, $pop34, $pop19
-; CHECK-NEXT:    i32.const $push4=, 8
-; CHECK-NEXT:    i32.add $push5=, $pop20, $pop4
-; CHECK-NEXT:    i64.load $push6=, 0($pop5)
-; CHECK-NEXT:    local.get $push35=, 5
-; CHECK-NEXT:    i32.const $push21=, 8
-; CHECK-NEXT:    i32.add $push7=, $pop35, $pop21
-; CHECK-NEXT:    i64.load $push8=, 0($pop7)
-; CHECK-NEXT:    i64.or $push9=, $pop6, $pop8
-; CHECK-NEXT:    i64.store 8($pop36), $pop9
-; CHECK-NEXT:    local.get $push39=, 0
-; CHECK-NEXT:    local.get $push37=, 5
-; CHECK-NEXT:    i64.load $push10=, 16($pop37)
-; CHECK-NEXT:    local.get $push38=, 5
-; CHECK-NEXT:    i64.load $push11=, 0($pop38)
-; CHECK-NEXT:    i64.or $push12=, $pop10, $pop11
-; CHECK-NEXT:    i64.store 0($pop39), $pop12
-; CHECK-NEXT:    local.get $push40=, 5
-; CHECK-NEXT:    i32.const $push15=, 32
-; CHECK-NEXT:    i32.add $push16=, $pop40, $pop15
-; CHECK-NEXT:    global.set __stack_pointer, $pop16
+; CHECK-NEXT:    local.get $push29=, 0
+; CHECK-NEXT:    local.get $push23=, 1
+; CHECK-NEXT:    local.get $push22=, 2
+; CHECK-NEXT:    local.get $push21=, 3
+; CHECK-NEXT:    i64.const $push1=, 64
+; CHECK-NEXT:    i64.and $push2=, $pop21, $pop1
+; CHECK-NEXT:    i64.eqz $push20=, $pop2
+; CHECK-NEXT:    local.tee $push19=, 5, $pop20
+; CHECK-NEXT:    i64.select $push18=, $pop23, $pop22, $pop19
+; CHECK-NEXT:    local.tee $push17=, 6, $pop18
+; CHECK-NEXT:    i64.const $push3=, 1
+; CHECK-NEXT:    i64.shl $push4=, $pop17, $pop3
+; CHECK-NEXT:    local.get $push24=, 3
+; CHECK-NEXT:    i64.const $push0=, -1
+; CHECK-NEXT:    i64.xor $push16=, $pop24, $pop0
+; CHECK-NEXT:    local.tee $push15=, 7, $pop16
+; CHECK-NEXT:    i64.shl $push5=, $pop4, $pop15
+; CHECK-NEXT:    local.get $push27=, 2
+; CHECK-NEXT:    local.get $push26=, 1
+; CHECK-NEXT:    local.get $push25=, 5
+; CHECK-NEXT:    i64.select $push14=, $pop27, $pop26, $pop25
+; CHECK-NEXT:    local.tee $push13=, 2, $pop14
+; CHECK-NEXT:    local.get $push28=, 3
+; CHECK-NEXT:    i64.shr_u $push6=, $pop13, $pop28
+; CHECK-NEXT:    i64.or $push7=, $pop5, $pop6
+; CHECK-NEXT:    i64.store 8($pop29), $pop7
+; CHECK-NEXT:    local.get $push34=, 0
+; CHECK-NEXT:    local.get $push30=, 2
+; CHECK-NEXT:    i64.const $push12=, 1
+; CHECK-NEXT:    i64.shl $push8=, $pop30, $pop12
+; CHECK-NEXT:    local.get $push31=, 7
+; CHECK-NEXT:    i64.shl $push9=, $pop8, $pop31
+; CHECK-NEXT:    local.get $push33=, 6
+; CHECK-NEXT:    local.get $push32=, 3
+; CHECK-NEXT:    i64.shr_u $push10=, $pop33, $pop32
+; CHECK-NEXT:    i64.or $push11=, $pop9, $pop10
+; CHECK-NEXT:    i64.store 0($pop34), $pop11
 ; CHECK-NEXT:    return
   %a = and i128 %y, 127
   %z = sub i128 128, %a
   %b = lshr i128 %x, %a
   %c = shl i128 %x, %z
   %d = or i128 %b, %c
   ret i128 %d
 }

llvm/test/CodeGen/WebAssembly/implicit-def.ll

-; RUN: llc -o - %s -asm-verbose=false -wasm-keep-registers -disable-wasm-fallthrough-return-opt -mattr=+simd128 | FileCheck %s
+; RUN: llc -o - %s -O1 -asm-verbose=false -wasm-keep-registers -disable-wasm-fallthrough-return-opt -mattr=+simd128 | FileCheck %s
 
 target triple = "wasm32-unknown-unknown"
 
 ; Test that stackified IMPLICIT_DEF instructions are converted into
 ; CONST_XXX instructions to provide an explicit push.
 
 ; CHECK-LABEL: implicit_def_i32:
 ; CHECK:      i32.const $push{{[0-9]+}}=, 0{{$}}

llvm/test/CodeGen/WebAssembly/indirect-import.ll

-; RUN: llc < %s -asm-verbose=false -disable-wasm-fallthrough-return-opt -wasm-keep-registers -verify-machineinstrs | FileCheck %s
-; RUN: llc < %s -asm-verbose=false -disable-wasm-fallthrough-return-opt -wasm-keep-registers -verify-machineinstrs -fast-isel | FileCheck %s
+; RUN: llc < %s -O1 -asm-verbose=false -disable-wasm-fallthrough-return-opt -wasm-keep-registers -verify-machineinstrs | FileCheck %s
+; RUN: llc < %s -O1 -asm-verbose=false -disable-wasm-fallthrough-return-opt -wasm-keep-registers -verify-machineinstrs -fast-isel | FileCheck %s
 
 ; ModuleID = 'test/dot_s/indirect-import.c'
 source_filename = "test/dot_s/indirect-import.c"
 target triple = "wasm32"
 
 ; CHECK: .functype extern_fd (f64) -> (f32)
 ; CHECK: .functype extern_vj (i64) -> ()
 ; CHECK: .functype extern_v () -> ()

llvm/test/CodeGen/WebAssembly/indirectbr.ll

-; RUN: llc < %s -asm-verbose=false -disable-wasm-fallthrough-return-opt -wasm-disable-explicit-locals -wasm-keep-registers | FileCheck %s
+; RUN: llc < %s -O0 -asm-verbose=false -disable-wasm-fallthrough-return-opt -wasm-disable-explicit-locals -wasm-keep-registers | FileCheck %s
 
 ; This tests that indirectbr instructions are lowered to switches. Currently we
 ; just re-use the IndirectBrExpand Pass; it has its own IR-level test.
 ; So this test just ensures that the pass gets run and we can lower indirectbr
 
 target triple = "wasm32"
 
 @test1.targets = constant [4 x ptr] [ptr blockaddress(@test1, %bb0),

llvm/test/CodeGen/WebAssembly/ir-locals-stackid.ll

-; RUN: llc -mtriple=wasm32-unknown-unknown -asm-verbose=false < %s | FileCheck %s --check-prefix=CHECKCG
-; RUN: llc -mtriple=wasm32-unknown-unknown -stop-after=finalize-isel < %s | FileCheck %s --check-prefix=CHECKISEL
+; RUN: llc -O1 -mtriple=wasm32-unknown-unknown -asm-verbose=false < %s | FileCheck %s --check-prefix=CHECKCG
+; RUN: llc -O1 -mtriple=wasm32-unknown-unknown -stop-after=finalize-isel < %s | FileCheck %s --check-prefix=CHECKISEL
 
 %f32_cell = type ptr addrspace(1)
 
 ; CHECKISEL-LABEL: name: ir_local_f32
 ; CHECKISEL:       stack:
 ; CHECKISEL:       id: 0, name: retval, type: default, offset: 1, size: 1, alignment: 4,
 ; CHECKISEL-NEXT:  stack-id: wasm-local
 

llvm/test/CodeGen/WebAssembly/ir-locals.ll

-; RUN: llc < %s --mtriple=wasm32-unknown-unknown -asm-verbose=false | FileCheck %s
+; RUN: llc < %s -O1 --mtriple=wasm32-unknown-unknown -asm-verbose=false | FileCheck %s
 
 %i32_cell = type ptr addrspace(1)
 %i64_cell = type ptr addrspace(1)
 %f32_cell = type ptr addrspace(1)
 %f64_cell = type ptr addrspace(1)
 
 ; We have a set of tests in which we set a local and then reload the
 ; local.  If the load immediately follows the set, the DAG combiner will

llvm/test/CodeGen/WebAssembly/libcalls.ll

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
-; RUN: llc < %s -disable-wasm-fallthrough-return-opt -wasm-keep-registers | FileCheck %s
+; RUN: llc < %s -O1  -asm-verbose=false -disable-wasm-fallthrough-return-opt -wasm-keep-registers | FileCheck %s
 
 ; Test a subset of compiler-rt/libm libcalls expected to be emitted by the wasm backend
 
 target triple = "wasm32-unknown-unknown"
 
 declare fp128 @llvm.sqrt.f128(fp128)
 declare fp128 @llvm.floor.f128(fp128)
 declare fp128 @llvm.trunc.f128(fp128)
 declare fp128 @llvm.nearbyint.f128(fp128)
 declare fp128 @llvm.pow.f128(fp128, fp128)
 declare fp128 @llvm.powi.f128.i32(fp128, i32)
 
 declare double @llvm.cos.f64(double)
 declare double @llvm.log10.f64(double)
 declare double @llvm.pow.f64(double, double)
 declare double @llvm.powi.f64.i32(double, i32)
 declare double @llvm.log.f64(double)
 declare double @llvm.exp.f64(double)
 declare i32 @llvm.lround(double)
 
 
 
 define fp128 @fp128libcalls(fp128 %x, fp128 %y, i32 %z) {
   ; compiler-rt call
-; CHECK-LABEL: fp128libcalls:
-; CHECK:         .functype fp128libcalls (i32, i64, i64, i64, i64, i32) -> ()
-; CHECK-NEXT:    .local i32
-; CHECK-NEXT:  # %bb.0:
-; CHECK-NEXT:    global.get $push28=, __stack_pointer
-; CHECK-NEXT:    i32.const $push29=, 144
-; CHECK-NEXT:    i32.sub $push73=, $pop28, $pop29
-; CHECK-NEXT:    local.tee $push72=, 6, $pop73
-; CHECK-NEXT:    global.set __stack_pointer, $pop72
-; CHECK-NEXT:    local.get $push74=, 6
-; CHECK-NEXT:    i32.const $push32=, 128
-; CHECK-NEXT:    i32.add $push33=, $pop74, $pop32
-; CHECK-NEXT:    local.get $push78=, 1
-; CHECK-NEXT:    local.get $push77=, 2
-; CHECK-NEXT:    local.get $push76=, 3
-; CHECK-NEXT:    local.get $push75=, 4
-; CHECK-NEXT:    call __addtf3, $pop33, $pop78, $pop77, $pop76, $pop75
-; CHECK-NEXT:    local.get $push79=, 6
-; CHECK-NEXT:    i32.const $push36=, 112
-; CHECK-NEXT:    i32.add $push37=, $pop79, $pop36
-; CHECK-NEXT:    local.get $push80=, 6
-; CHECK-NEXT:    i64.load $push3=, 128($pop80)
-; CHECK-NEXT:    local.get $push81=, 6
-; CHECK-NEXT:    i32.const $push34=, 128
-; CHECK-NEXT:    i32.add $push35=, $pop81, $pop34
-; CHECK-NEXT:    i32.const $push0=, 8
-; CHECK-NEXT:    i32.add $push1=, $pop35, $pop0
-; CHECK-NEXT:    i64.load $push2=, 0($pop1)
-; CHECK-NEXT:    local.get $push83=, 3
-; CHECK-NEXT:    local.get $push82=, 4
-; CHECK-NEXT:    call __multf3, $pop37, $pop3, $pop2, $pop83, $pop82
-; CHECK-NEXT:    local.get $push84=, 6
-; CHECK-NEXT:    i32.const $push40=, 96
-; CHECK-NEXT:    i32.add $push41=, $pop84, $pop40
-; CHECK-NEXT:    local.get $push85=, 6
-; CHECK-NEXT:    i64.load $push6=, 112($pop85)
-; CHECK-NEXT:    local.get $push86=, 6
-; CHECK-NEXT:    i32.const $push38=, 112
-; CHECK-NEXT:    i32.add $push39=, $pop86, $pop38
-; CHECK-NEXT:    i32.const $push71=, 8
-; CHECK-NEXT:    i32.add $push4=, $pop39, $pop71
-; CHECK-NEXT:    i64.load $push5=, 0($pop4)
-; CHECK-NEXT:    local.get $push88=, 3
-; CHECK-NEXT:    local.get $push87=, 4
-; CHECK-NEXT:    call __divtf3, $pop41, $pop6, $pop5, $pop88, $pop87
-; CHECK-NEXT:    local.get $push89=, 6
-; CHECK-NEXT:    i32.const $push44=, 80
-; CHECK-NEXT:    i32.add $push45=, $pop89, $pop44
-; CHECK-NEXT:    local.get $push90=, 6
-; CHECK-NEXT:    i64.load $push9=, 96($pop90)
-; CHECK-NEXT:    local.get $push91=, 6
-; CHECK-NEXT:    i32.const $push42=, 96
-; CHECK-NEXT:    i32.add $push43=, $pop91, $pop42
-; CHECK-NEXT:    i32.const $push70=, 8
-; CHECK-NEXT:    i32.add $push7=, $pop43, $pop70
-; CHECK-NEXT:    i64.load $push8=, 0($pop7)
-; CHECK-NEXT:    call sqrtl, $pop45, $pop9, $pop8
-; CHECK-NEXT:    local.get $push92=, 6
-; CHECK-NEXT:    i32.const $push48=, 64
-; CHECK-NEXT:    i32.add $push49=, $pop92, $pop48
-; CHECK-NEXT:    local.get $push93=, 6
-; CHECK-NEXT:    i64.load $push12=, 80($pop93)
-; CHECK-NEXT:    local.get $push94=, 6
-; CHECK-NEXT:    i32.const $push46=, 80
-; CHECK-NEXT:    i32.add $push47=, $pop94, $pop46
-; CHECK-NEXT:    i32.const $push69=, 8
-; CHECK-NEXT:    i32.add $push10=, $pop47, $pop69
-; CHECK-NEXT:    i64.load $push11=, 0($pop10)
-; CHECK-NEXT:    call floorl, $pop49, $pop12, $pop11
-; CHECK-NEXT:    local.get $push95=, 6
-; CHECK-NEXT:    i32.const $push52=, 48
-; CHECK-NEXT:    i32.add $push53=, $pop95, $pop52
-; CHECK-NEXT:    local.get $push96=, 6
-; CHECK-NEXT:    i64.load $push15=, 64($pop96)
-; CHECK-NEXT:    local.get $push97=, 6
-; CHECK-NEXT:    i32.const $push50=, 64
-; CHECK-NEXT:    i32.add $push51=, $pop97, $pop50
-; CHECK-NEXT:    i32.const $push68=, 8
-; CHECK-NEXT:    i32.add $push13=, $pop51, $pop68
-; CHECK-NEXT:    i64.load $push14=, 0($pop13)
-; CHECK-NEXT:    local.get $push99=, 3
-; CHECK-NEXT:    local.get $push98=, 4
-; CHECK-NEXT:    call powl, $pop53, $pop15, $pop14, $pop99, $pop98
-; CHECK-NEXT:    local.get $push100=, 6
-; CHECK-NEXT:    i32.const $push56=, 32
-; CHECK-NEXT:    i32.add $push57=, $pop100, $pop56
-; CHECK-NEXT:    local.get $push101=, 6
-; CHECK-NEXT:    i64.load $push18=, 48($pop101)
-; CHECK-NEXT:    local.get $push102=, 6
-; CHECK-NEXT:    i32.const $push54=, 48
-; CHECK-NEXT:    i32.add $push55=, $pop102, $pop54
-; CHECK-NEXT:    i32.const $push67=, 8
-; CHECK-NEXT:    i32.add $push16=, $pop55, $pop67
-; CHECK-NEXT:    i64.load $push17=, 0($pop16)
-; CHECK-NEXT:    local.get $push103=, 5
-; CHECK-NEXT:    call __powitf2, $pop57, $pop18, $pop17, $pop103
-; CHECK-NEXT:    local.get $push104=, 6
-; CHECK-NEXT:    i32.const $push60=, 16
-; CHECK-NEXT:    i32.add $push61=, $pop104, $pop60
-; CHECK-NEXT:    local.get $push105=, 6
-; CHECK-NEXT:    i64.load $push21=, 32($pop105)
-; CHECK-NEXT:    local.get $push106=, 6
-; CHECK-NEXT:    i32.const $push58=, 32
-; CHECK-NEXT:    i32.add $push59=, $pop106, $pop58
-; CHECK-NEXT:    i32.const $push66=, 8
-; CHECK-NEXT:    i32.add $push19=, $pop59, $pop66
-; CHECK-NEXT:    i64.load $push20=, 0($pop19)
-; CHECK-NEXT:    call truncl, $pop61, $pop21, $pop20
-; CHECK-NEXT:    local.get $push109=, 6
-; CHECK-NEXT:    local.get $push107=, 6
-; CHECK-NEXT:    i64.load $push24=, 16($pop107)
-; CHECK-NEXT:    local.get $push108=, 6
-; CHECK-NEXT:    i32.const $push62=, 16
-; CHECK-NEXT:    i32.add $push63=, $pop108, $pop62
-; CHECK-NEXT:    i32.const $push65=, 8
-; CHECK-NEXT:    i32.add $push22=, $pop63, $pop65
-; CHECK-NEXT:    i64.load $push23=, 0($pop22)
-; CHECK-NEXT:    call nearbyintl, $pop109, $pop24, $pop23
-; CHECK-NEXT:    local.get $push111=, 0
-; CHECK-NEXT:    local.get $push110=, 6
-; CHECK-NEXT:    i32.const $push64=, 8
-; CHECK-NEXT:    i32.add $push25=, $pop110, $pop64
-; CHECK-NEXT:    i64.load $push26=, 0($pop25)
-; CHECK-NEXT:    i64.store 8($pop111), $pop26
-; CHECK-NEXT:    local.get $push113=, 0
-; CHECK-NEXT:    local.get $push112=, 6
-; CHECK-NEXT:    i64.load $push27=, 0($pop112)
-; CHECK-NEXT:    i64.store 0($pop113), $pop27
-; CHECK-NEXT:    local.get $push114=, 6
-; CHECK-NEXT:    i32.const $push30=, 144
-; CHECK-NEXT:    i32.add $push31=, $pop114, $pop30
-; CHECK-NEXT:    global.set __stack_pointer, $pop31
-; CHECK-NEXT:    return
   %a = fadd fp128 %x, %y
   %b = fmul fp128 %a, %y
   %c = fdiv fp128 %b, %y
   ; libm calls
   %d = call fp128 @llvm.sqrt.f128(fp128 %c)
   %e = call fp128 @llvm.floor.f128(fp128 %d)
   %f = call fp128 @llvm.pow.f128(fp128 %e, fp128 %y)
   %g = call fp128 @llvm.powi.f128.i32(fp128 %f, i32 %z)
   %h = call fp128 @llvm.trunc.f128(fp128 %g)
   %i = call fp128 @llvm.nearbyint.f128(fp128 %h)
   ret fp128 %i
 }
 
 define i128 @i128libcalls(i128 %x, i128 %y) {
   ; Basic ops should be expanded
-; CHECK-LABEL: i128libcalls:
-; CHECK:         .functype i128libcalls (i32, i64, i64, i64, i64) -> ()
-; CHECK-NEXT:    .local i32, i64
-; CHECK-NEXT:  # %bb.0:
-; CHECK-NEXT:    global.get $push11=, __stack_pointer
-; CHECK-NEXT:    i32.const $push12=, 32
-; CHECK-NEXT:    i32.sub $push23=, $pop11, $pop12
-; CHECK-NEXT:    local.tee $push22=, 5, $pop23
-; CHECK-NEXT:    global.set __stack_pointer, $pop22
-; CHECK-NEXT:    local.get $push24=, 5
-; CHECK-NEXT:    i32.const $push15=, 16
-; CHECK-NEXT:    i32.add $push16=, $pop24, $pop15
-; CHECK-NEXT:    local.get $push26=, 1
-; CHECK-NEXT:    local.get $push25=, 3
-; CHECK-NEXT:    i64.add $push21=, $pop26, $pop25
-; CHECK-NEXT:    local.tee $push20=, 6, $pop21
-; CHECK-NEXT:    local.get $push28=, 2
-; CHECK-NEXT:    local.get $push27=, 4
-; CHECK-NEXT:    i64.add $push0=, $pop28, $pop27
-; CHECK-NEXT:    local.get $push30=, 6
-; CHECK-NEXT:    local.get $push29=, 1
-; CHECK-NEXT:    i64.lt_u $push1=, $pop30, $pop29
-; CHECK-NEXT:    i64.extend_i32_u $push2=, $pop1
-; CHECK-NEXT:    i64.add $push3=, $pop0, $pop2
-; CHECK-NEXT:    local.get $push32=, 3
-; CHECK-NEXT:    local.get $push31=, 4
-; CHECK-NEXT:    call __multi3, $pop16, $pop20, $pop3, $pop32, $pop31
-; CHECK-NEXT:    local.get $push37=, 5
-; CHECK-NEXT:    local.get $push33=, 5
-; CHECK-NEXT:    i64.load $push7=, 16($pop33)
-; CHECK-NEXT:    local.get $push34=, 5
-; CHECK-NEXT:    i32.const $push17=, 16
-; CHECK-NEXT:    i32.add $push18=, $pop34, $pop17
-; CHECK-NEXT:    i32.const $push4=, 8
-; CHECK-NEXT:    i32.add $push5=, $pop18, $pop4
-; CHECK-NEXT:    i64.load $push6=, 0($pop5)
-; CHECK-NEXT:    local.get $push36=, 3
-; CHECK-NEXT:    local.get $push35=, 4
-; CHECK-NEXT:    call __umodti3, $pop37, $pop7, $pop6, $pop36, $pop35
-; CHECK-NEXT:    local.get $push39=, 0
-; CHECK-NEXT:    local.get $push38=, 5
-; CHECK-NEXT:    i32.const $push19=, 8
-; CHECK-NEXT:    i32.add $push8=, $pop38, $pop19
-; CHECK-NEXT:    i64.load $push9=, 0($pop8)
-; CHECK-NEXT:    i64.store 8($pop39), $pop9
-; CHECK-NEXT:    local.get $push41=, 0
-; CHECK-NEXT:    local.get $push40=, 5
-; CHECK-NEXT:    i64.load $push10=, 0($pop40)
-; CHECK-NEXT:    i64.store 0($pop41), $pop10
-; CHECK-NEXT:    local.get $push42=, 5
-; CHECK-NEXT:    i32.const $push13=, 32
-; CHECK-NEXT:    i32.add $push14=, $pop42, $pop13
-; CHECK-NEXT:    global.set __stack_pointer, $pop14
-; CHECK-NEXT:    return
   %a = add i128 %x, %y
   %b = mul i128 %a, %y
   %c = urem i128 %b, %y
   ret i128 %c
 }
 
 define i32 @f64libcalls(double %x, double %y, i32 %z) {
-; CHECK-LABEL: f64libcalls:
-; CHECK:         .functype f64libcalls (f64, f64, i32) -> (i32)
-; CHECK-NEXT:  # %bb.0:
-; CHECK-NEXT:    local.get $push8=, 0
-; CHECK-NEXT:    call $push0=, cos, $pop8
-; CHECK-NEXT:    call $push1=, log10, $pop0
-; CHECK-NEXT:    local.get $push9=, 1
-; CHECK-NEXT:    call $push2=, pow, $pop1, $pop9
-; CHECK-NEXT:    local.get $push10=, 2
-; CHECK-NEXT:    call $push3=, __powidf2, $pop2, $pop10
-; CHECK-NEXT:    call $push4=, log, $pop3
-; CHECK-NEXT:    call $push5=, exp, $pop4
-; CHECK-NEXT:    call $push6=, cbrt, $pop5
-; CHECK-NEXT:    call $push7=, lround, $pop6
-; CHECK-NEXT:    return $pop7
  %a = call double @llvm.cos.f64(double %x)
  %b = call double @llvm.log10.f64(double %a)
  %c = call double @llvm.pow.f64(double %b, double %y)
  %d = call double @llvm.powi.f64.i32(double %c, i32 %z)
  %e = call double @llvm.log.f64(double %d)
  %f = call double @llvm.exp.f64(double %e)
  %g = call fast double @llvm.pow.f64(double %f, double 0x3FD5555555555555)
  %h = call i32 @llvm.lround(double %g)
  ret i32 %h
 }
 
 ; fcmp ord and unord (RTLIB::O_F32 / RTLIB::UO_F32 etc) are a special case (see
 ; comment in WebAssemblyRunimeLibcallSignatures.cpp) so check them separately.
 ; no libcalls are needed for f32 and f64
 
 define i1 @unordd(double %x, double %y) {
-; CHECK-LABEL: unordd:
-; CHECK:         .functype unordd (f64, f64) -> (i32)
-; CHECK-NEXT:  # %bb.0:
-; CHECK-NEXT:    local.get $push8=, 0
-; CHECK-NEXT:    local.get $push7=, 0
-; CHECK-NEXT:    f64.ne $push4=, $pop8, $pop7
-; CHECK-NEXT:    local.get $push10=, 1
-; CHECK-NEXT:    local.get $push9=, 1
-; CHECK-NEXT:    f64.ne $push3=, $pop10, $pop9
-; CHECK-NEXT:    i32.or $push5=, $pop4, $pop3
-; CHECK-NEXT:    local.get $push12=, 0
-; CHECK-NEXT:    local.get $push11=, 0
-; CHECK-NEXT:    f64.eq $push1=, $pop12, $pop11
-; CHECK-NEXT:    local.get $push14=, 1
-; CHECK-NEXT:    local.get $push13=, 1
-; CHECK-NEXT:    f64.eq $push0=, $pop14, $pop13
-; CHECK-NEXT:    i32.and $push2=, $pop1, $pop0
-; CHECK-NEXT:    i32.xor $push6=, $pop5, $pop2
-; CHECK-NEXT:    return $pop6
  %a = fcmp uno double %x, %y
  %b = fcmp ord double %x, %y
  %c = xor i1 %a, %b
  ret i1 %c
 }
 
 define i1 @unordf(float %x, float %y) {
-; CHECK-LABEL: unordf:
-; CHECK:         .functype unordf (f32, f32) -> (i32)
-; CHECK-NEXT:  # %bb.0:
-; CHECK-NEXT:    local.get $push8=, 0
-; CHECK-NEXT:    local.get $push7=, 0
-; CHECK-NEXT:    f32.ne $push4=, $pop8, $pop7
-; CHECK-NEXT:    local.get $push10=, 1
-; CHECK-NEXT:    local.get $push9=, 1
-; CHECK-NEXT:    f32.ne $push3=, $pop10, $pop9
-; CHECK-NEXT:    i32.or $push5=, $pop4, $pop3
-; CHECK-NEXT:    local.get $push12=, 0
-; CHECK-NEXT:    local.get $push11=, 0
-; CHECK-NEXT:    f32.eq $push1=, $pop12, $pop11
-; CHECK-NEXT:    local.get $push14=, 1
-; CHECK-NEXT:    local.get $push13=, 1
-; CHECK-NEXT:    f32.eq $push0=, $pop14, $pop13
-; CHECK-NEXT:    i32.and $push2=, $pop1, $pop0
-; CHECK-NEXT:    i32.xor $push6=, $pop5, $pop2
-; CHECK-NEXT:    return $pop6
  %a = fcmp uno float %x, %y
  %b = fcmp ord float %x, %y
  %c = xor i1 %a, %b
  ret i1 %c
 }
 
 define i1 @unordt(fp128 %x, fp128 %y) {
-; CHECK-LABEL: unordt:
-; CHECK:         .functype unordt (i64, i64, i64, i64) -> (i32)
-; CHECK-NEXT:  # %bb.0:
-; CHECK-NEXT:    local.get $push6=, 0
-; CHECK-NEXT:    local.get $push5=, 1
-; CHECK-NEXT:    local.get $push4=, 2
-; CHECK-NEXT:    local.get $push3=, 3
-; CHECK-NEXT:    call $push1=, __unordtf2, $pop6, $pop5, $pop4, $pop3
-; CHECK-NEXT:    i32.const $push0=, 0
-; CHECK-NEXT:    i32.ne $push2=, $pop1, $pop0
-; CHECK-NEXT:    return $pop2
  %a = fcmp uno fp128 %x, %y
  ret i1 %a
 }
 
 define i1 @ordt(fp128 %x, fp128 %y) {
-; CHECK-LABEL: ordt:
-; CHECK:         .functype ordt (i64, i64, i64, i64) -> (i32)
-; CHECK-NEXT:  # %bb.0:
-; CHECK-NEXT:    local.get $push5=, 0
-; CHECK-NEXT:    local.get $push4=, 1
-; CHECK-NEXT:    local.get $push3=, 2
-; CHECK-NEXT:    local.get $push2=, 3
-; CHECK-NEXT:    call $push0=, __unordtf2, $pop5, $pop4, $pop3, $pop2
-; CHECK-NEXT:    i32.eqz $push1=, $pop0
-; CHECK-NEXT:    return $pop1
  %a = fcmp ord fp128 %x, %y
  ret i1 %a
 }
+;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:
+; CHECK: {{.*}}

llvm/test/CodeGen/WebAssembly/load-store-pic.ll

; CHECK-NEXT:    end_function
   ret i32 %1
 }
 
 define i32 @load_hidden_global_offset() {
 ; CHECK-LABEL: load_hidden_global_offset:
 ; PIC:         global.get $push[[L0:[0-9]+]]=, __memory_base{{$}}
 ; PIC-NEXT:    [[PTR]].const $push[[L1:[0-9]+]]=, hidden_global_array@MBREL{{$}}
 ; PIC-NEXT:    [[PTR]].add $push[[L2:[0-9]+]]=, $pop[[L0]], $pop[[L1:[0-9]+]]{{$}}
-; PIC-NEXT:    [[PTR]].const $push[[L3:[0-9]+]]=, 20{{$}}
-; PIC-NEXT:    [[PTR]].add $push[[L4:[0-9]+]]=, $pop[[L2]], $pop[[L3]]{{$}}
-; PIC-NEXT:    i32.load $push{{[0-9]+}}=, 0($pop[[L4]]){{$}}
+; PIC-NEXT:    i32.load $push{{[0-9]+}}=, 20($pop[[L2]]){{$}}
 ; CHECK-NEXT:  end_function
 
   %1 = getelementptr [10 x i32], ptr @hidden_global_array, i32 0, i32 5
   %2 = load i32, ptr %1
   ret i32 %2
 }
 
 ; Store to a hidden global
 
 define void @store_hidden_global(i32 %n) {
 ; CHECK-LABEL: store_hidden_global:
 ; PIC:         global.get $push[[L0:[0-9]+]]=, __memory_base{{$}}
-; PIC-NEXT:    [[PTR]].const $push[[L1:[0-9]+]]=, hidden_global@MBREL{{$}}
+; PIC-NEXT:    [[PTR]].const $push[[L1:[0-9]+]]=, hidden_global@MBREL
 ; PIC-NEXT:    [[PTR]].add $push[[L2:[0-9]+]]=, $pop[[L0]], $pop[[L1]]{{$}}
 ; PIC-NEXT:    i32.store 0($pop[[L2]]), $0{{$}}
-; CHECK-NEXT:    end_function
+; CHECK-NEXT:   end_function
 
   store i32 %n, ptr @hidden_global
   ret void
 }
 
 define void @store_hidden_global_offset(i32 %n) {
 ; CHECK-LABEL: store_hidden_global_offset:
 ; PIC:         global.get $push[[L0:[0-9]+]]=, __memory_base{{$}}
 ; PIC-NEXT:    [[PTR]].const $push[[L1:[0-9]+]]=, hidden_global_array@MBREL{{$}}
 ; PIC-NEXT:    [[PTR]].add $push[[L2:[0-9]+]]=, $pop[[L0]], $pop[[L1]]{{$}}
-; PIC-NEXT:    [[PTR]].const $push[[L3:[0-9]+]]=, 20{{$}}
-; PIC-NEXT:    [[PTR]].add $push[[L4:[0-9]+]]=, $pop[[L2]], $pop[[L3]]{{$}}
-; PIC-NEXT:    i32.store 0($pop[[L4]]), $0{{$}}
+; PIC-NEXT:    i32.store 20($pop[[L2]]), $0{{$}}
 
 ; CHECK-NEXT:   end_function
 
   %1 = getelementptr [10 x i32], ptr @hidden_global_array, i32 0, i32 5
   store i32 %n, ptr %1
   ret void
 }
 
; CHECK-NEXT:   end_function
 
   %1 = load i32, ptr @external_global
   ret i32 %1
 }
 
 define i32 @load_external_global_offset() {
 ; CHECK-LABEL:  load_external_global_offset:
 ; PIC:          global.get $push[[L0:[0-9]+]]=, external_global_array@GOT{{$}}
-; PIC-NEXT:     [[PTR]].const $push[[L1:[0-9]+]]=, 20{{$}}
-; PIC-NEXT:     [[PTR]].add $push[[L2:[0-9]+]]=, $pop[[L0]], $pop[[L1]]{{$}}
-; PIC-NEXT:     i32.load $push{{[0-9]+}}=, 0($pop[[L2]]){{$}}
+; PIC-NEXT:     i32.load $push{{[0-9]+}}=, 20($pop[[L0]]){{$}}
 
 ; CHECK-NEXT:   end_function
 
   %1 = getelementptr [10 x i32], ptr @external_global_array, i32 0, i32 5
   %2 = load i32, ptr %1
   ret i32 %2
 }
 
 ; Store to a non-hidden global via the wasm global.
 
 define void @store_external_global(i32 %n) {
 ; CHECK-LABEL:  store_external_global:
 ; PIC:          global.get $push[[L0:[0-9]+]]=, external_global@GOT{{$}}
 ; PIC-NEXT:     i32.store 0($pop[[L0]]), $0{{$}}
 
 ; CHECK-NEXT:   end_function
 
   store i32 %n, ptr @external_global
   ret void
 }
 
 define void @store_external_global_offset(i32 %n) {
 ; CHECK-LABEL:  store_external_global_offset:
 ; PIC:          global.get $push[[L0:[0-9]+]]=, external_global_array@GOT{{$}}
-; PIC-NEXT:     [[PTR]].const $push[[L1:[0-9]+]]=, 20{{$}}
-; PIC-NEXT:     [[PTR]].add $push[[L2:[0-9]+]]=, $pop[[L0]], $pop[[L1]]{{$}}
-; PIC-NEXT:     i32.store 0($pop[[L2]]), $0{{$}}
+; PIC-NEXT:     i32.store 20($pop[[L0]]), $0{{$}}
 
 ; CHECK-NEXT:   end_function
 
   %1 = getelementptr [10 x i32], ptr @external_global_array, i32 0, i32 5
   store i32 %n, ptr %1
   ret void
 }

llvm/test/CodeGen/WebAssembly/mem-intrinsics.ll

; CHECK-NEXT:    return
   call void @llvm.memset.p0.i32(ptr %dst, i8 %src, i32 %len, i1 false)
   ret void
 }
 
 define void @frame_index() {
 ; CHECK-LABEL: frame_index:
 ; CHECK:         .functype frame_index () -> ()
 ; CHECK-NEXT:  # %bb.0: # %entry
-; CHECK-NEXT:    global.get $push3=, __stack_pointer
-; CHECK-NEXT:    i32.const $push4=, 4096
-; CHECK-NEXT:    i32.sub $push12=, $pop3, $pop4
-; CHECK-NEXT:    local.tee $push11=, $0=, $pop12
-; CHECK-NEXT:    global.set __stack_pointer, $pop11
-; CHECK-NEXT:    i32.const $push7=, 2048
-; CHECK-NEXT:    i32.add $push8=, $0, $pop7
-; CHECK-NEXT:    i32.const $push1=, 0
-; CHECK-NEXT:    i32.const $push0=, 1024
-; CHECK-NEXT:    call $drop=, memset, $pop8, $pop1, $pop0
-; CHECK-NEXT:    i32.const $push10=, 0
-; CHECK-NEXT:    i32.const $push9=, 1024
-; CHECK-NEXT:    call $push2=, memset, $0, $pop10, $pop9
-; CHECK-NEXT:    i32.const $push5=, 4096
-; CHECK-NEXT:    i32.add $push6=, $pop2, $pop5
-; CHECK-NEXT:    global.set __stack_pointer, $pop6
 ; CHECK-NEXT:    return
 entry:
   %a = alloca [2048 x i8], align 16
   %b = alloca [2048 x i8], align 16
   call void @llvm.memset.p0.i32(ptr align 16 %a, i8 256, i32 1024, i1 false)
   call void @llvm.memset.p0.i32(ptr align 16 %b, i8 256, i32 1024, i1 false)
   ret void
 }

llvm/test/CodeGen/WebAssembly/offset-fastisel.ll

; CHECK-NEXT:    # fallthrough-return
   store i8 0, ptr %s
   ret void
 }
 
 define hidden void @store_i8_with_array_alloca_gep(i32 %idx) {
 ; CHECK-LABEL: store_i8_with_array_alloca_gep:
 ; CHECK:         .functype store_i8_with_array_alloca_gep (i32) -> ()
 ; CHECK-NEXT:  # %bb.0:
-; CHECK-NEXT:    global.get $push3=, __stack_pointer
-; CHECK-NEXT:    i32.const $push4=, 32
-; CHECK-NEXT:    i32.sub $push5=, $pop3, $pop4
-; CHECK-NEXT:    local.copy $push1=, $pop5
-; CHECK-NEXT:    i32.add $push2=, $pop1, $0
-; CHECK-NEXT:    i32.const $push0=, 0
-; CHECK-NEXT:    i32.store8 0($pop2), $pop0
 ; CHECK-NEXT:    # fallthrough-return
   %A = alloca [30 x i8], align 16
   %s = getelementptr inbounds [30 x i8], ptr %A, i32 0, i32 %idx
   store i8 0, ptr %s, align 1
   ret void
 }
 
 define void @store_i32_with_unfolded_gep_offset(ptr %p) {

llvm/test/CodeGen/WebAssembly/pr47375.ll

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc < %s | FileCheck %s
 
 target triple = "wasm32-unknown-unknown"
 
 ; Regression test for pr47375, in which an assertion was triggering
 ; because WebAssemblyTargetLowering::isVectorLoadExtDesirable was
 ; improperly assuming the use of simple value types.
 
-define void @sext_vec() {
+define void @sext_vec(ptr %arg) {
 ; CHECK-LABEL: sext_vec:
-; CHECK:         .functype sext_vec () -> ()
+; CHECK:         .functype sext_vec (i32) -> ()
 ; CHECK-NEXT:    .local i32
 ; CHECK-NEXT:  # %bb.0:
 ; CHECK-NEXT:    local.get 0
-; CHECK-NEXT:    i32.load8_u 0
-; CHECK-NEXT:    local.set 0
-; CHECK-NEXT:    local.get 0
+; CHECK-NEXT:    i32.const 2
+; CHECK-NEXT:    i32.add
 ; CHECK-NEXT:    i32.const 0
 ; CHECK-NEXT:    i32.store8 0
 ; CHECK-NEXT:    local.get 0
 ; CHECK-NEXT:    local.get 0
-; CHECK-NEXT:    local.get 0
+; CHECK-NEXT:    i32.load8_u 0
+; CHECK-NEXT:    local.tee 1
+; CHECK-NEXT:    local.get 1
 ; CHECK-NEXT:    i32.const 7
 ; CHECK-NEXT:    i32.shl
 ; CHECK-NEXT:    i32.or
 ; CHECK-NEXT:    i32.const 7175
 ; CHECK-NEXT:    i32.and
 ; CHECK-NEXT:    i32.store16 0
 ; CHECK-NEXT:    # fallthrough-return
-  %L1 = load <2 x i3>, ptr undef, align 2
+  %L1 = load <2 x i3>, ptr %arg, align 2
   %zext = zext <2 x i3> %L1 to <2 x i10>
-  store <2 x i10> %zext, ptr undef, align 4
+  store <2 x i10> %zext, ptr %arg, align 4
   ret void
 }

llvm/test/CodeGen/WebAssembly/pr59626.ll

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc < %s -mtriple=wasm32-- -mattr=+simd128 | FileCheck --check-prefix=CHECK-32 %s
 ; RUN: llc < %s -mtriple=wasm64-- -mattr=+simd128 | FileCheck --check-prefix=CHECK-64 %s
 
 define i8 @f(ptr %0, ptr %1) {
 ; CHECK-32-LABEL: f:
 ; CHECK-32:         .functype f (i32, i32) -> (i32)
 ; CHECK-32-NEXT:  # %bb.0: # %BB
 ; CHECK-32-NEXT:    local.get 0
 ; CHECK-32-NEXT:    i32.const 0
 ; CHECK-32-NEXT:    i32.store8 2
 ; CHECK-32-NEXT:    local.get 0
 ; CHECK-32-NEXT:    i32.const 0
 ; CHECK-32-NEXT:    i32.store16 0
 ; CHECK-32-NEXT:    local.get 1
-; CHECK-32-NEXT:    local.get 0
-; CHECK-32-NEXT:    i8x16.splat
-; CHECK-32-NEXT:    v128.store16_lane 0, 0
-; CHECK-32-NEXT:    v128.const 0, 0
-; CHECK-32-NEXT:    i32x4.extract_lane 0
+; CHECK-32-NEXT:    i32.const 5
+; CHECK-32-NEXT:    i32.store8 2
+; CHECK-32-NEXT:    local.get 1
+; CHECK-32-NEXT:    i32.const 769
+; CHECK-32-NEXT:    i32.store16 0
+; CHECK-32-NEXT:    i32.const 1
 ; CHECK-32-NEXT:    # fallthrough-return
 ;
 ; CHECK-64-LABEL: f:
 ; CHECK-64:         .functype f (i64, i64) -> (i32)
-; CHECK-64-NEXT:    .local i32
 ; CHECK-64-NEXT:  # %bb.0: # %BB
 ; CHECK-64-NEXT:    local.get 0
 ; CHECK-64-NEXT:    i32.const 0
 ; CHECK-64-NEXT:    i32.store8 2
 ; CHECK-64-NEXT:    local.get 0
 ; CHECK-64-NEXT:    i32.const 0
 ; CHECK-64-NEXT:    i32.store16 0
 ; CHECK-64-NEXT:    local.get 1
-; CHECK-64-NEXT:    local.get 2
-; CHECK-64-NEXT:    i8x16.splat
-; CHECK-64-NEXT:    v128.store16_lane 0, 0
-; CHECK-64-NEXT:    drop
-; CHECK-64-NEXT:    v128.const 0, 0
-; CHECK-64-NEXT:    i32x4.extract_lane 0
+; CHECK-64-NEXT:    i32.const 5
+; CHECK-64-NEXT:    i32.store8 2
+; CHECK-64-NEXT:    local.get 1
+; CHECK-64-NEXT:    i32.const 769
+; CHECK-64-NEXT:    i32.store16 0
+; CHECK-64-NEXT:    i32.const 1
 ; CHECK-64-NEXT:    # fallthrough-return
 BB:
   store <3 x i8> zeroinitializer, ptr %0
   %S = shufflevector <3 x i128> zeroinitializer, <3 x i128> <i128 0, i128 1, i128 2>, <3 x i32> undef
   %C = icmp ule <3 x i128> %S, zeroinitializer
   %C1 = zext <3 x i1> %C to <3 x i8>
   %E = extractelement <3 x i8> %C1, i32 0
   %B = sdiv <3 x i8> <i8 1, i8 3, i8 5>, %C1
   store <3 x i8> %B, ptr %1
   ret i8 %E
 }

llvm/test/CodeGen/WebAssembly/reg-stackify.ll

-; RUN: llc < %s -asm-verbose=false -disable-wasm-fallthrough-return-opt -wasm-disable-explicit-locals -wasm-keep-registers -verify-machineinstrs | FileCheck %s
-; RUN: llc < %s -asm-verbose=false -disable-wasm-fallthrough-return-opt -verify-machineinstrs | FileCheck %s --check-prefix=NOREGS
+; RUN: llc < %s -asm-verbose=false -start-after=licm -disable-wasm-fallthrough-return-opt -wasm-disable-explicit-locals -wasm-keep-registers -verify-machineinstrs | FileCheck %s
+; RUN: llc < %s -asm-verbose=false -start-after=licm -disable-wasm-fallthrough-return-opt -verify-machineinstrs | FileCheck %s --check-prefix=NOREGS
 
 ; Test the register stackifier pass.
 
 ; We have two sets of tests, one with registers and implicit locals, and
 ; a stack / explicit locals based version (NOREGS).
 
 target triple = "wasm32-unknown-unknown"
 

llvm/test/CodeGen/WebAssembly/return-int32.ll

; FAST-NEXT:    # fallthrough-return
   ret i32 %p
 }
 
 define i32 @return_i32_twice(i32 %a) {
 ; CHECK-LABEL: return_i32_twice:
 ; CHECK:         .functype return_i32_twice (i32) -> (i32)
 ; CHECK-NEXT:  # %bb.0:
 ; CHECK-NEXT:    block
-; CHECK-NEXT:    local.get $push6=, 0
-; CHECK-NEXT:    i32.eqz $push7=, $pop6
-; CHECK-NEXT:    br_if 0, $pop7 # 0: down to label0
+; CHECK-NEXT:    local.get $push2=, 0
+; CHECK-NEXT:    i32.eqz $push3=, $pop2
+; CHECK-NEXT:    br_if 0, $pop3 # 0: down to label0
 ; CHECK-NEXT:  # %bb.1: # %true
-; CHECK-NEXT:    i32.const $push3=, 0
-; CHECK-NEXT:    i32.const $push5=, 0
-; CHECK-NEXT:    i32.store 0($pop3), $pop5
-; CHECK-NEXT:    i32.const $push4=, 1
-; CHECK-NEXT:    return $pop4
+; CHECK-NEXT:    i32.const $push0=, 1
+; CHECK-NEXT:    return $pop0
 ; CHECK-NEXT:  .LBB1_2: # %false
 ; CHECK-NEXT:    end_block # label0:
-; CHECK-NEXT:    i32.const $push1=, 0
-; CHECK-NEXT:    i32.const $push0=, 2
-; CHECK-NEXT:    i32.store 0($pop1), $pop0
-; CHECK-NEXT:    i32.const $push2=, 3
+; CHECK-NEXT:    i32.const $push1=, 3
 ; CHECK-NEXT:    # fallthrough-return
 ;
 ; FAST-LABEL: return_i32_twice:
 ; FAST:         .functype return_i32_twice (i32) -> (i32)
 ; FAST-NEXT:  # %bb.0:
 ; FAST-NEXT:    block
-; FAST-NEXT:    local.get $push6=, 0
-; FAST-NEXT:    i32.eqz $push7=, $pop6
+; FAST-NEXT:    local.get $push4=, 0
+; FAST-NEXT:    i32.eqz $push7=, $pop4
 ; FAST-NEXT:    br_if 0, $pop7 # 0: down to label0
 ; FAST-NEXT:  # %bb.1: # %true
-; FAST-NEXT:    i32.const $push4=, 0
-; FAST-NEXT:    i32.const $push5=, 0
-; FAST-NEXT:    i32.store 0($pop4), $pop5
-; FAST-NEXT:    i32.const $push3=, 1
-; FAST-NEXT:    return $pop3
+; FAST-NEXT:    i32.const $push1=, 0
+; FAST-NEXT:    local.get $push5=, 0
+; FAST-NEXT:    i32.store 0($pop1), $pop5
+; FAST-NEXT:    i32.const $push0=, 1
+; FAST-NEXT:    return $pop0
 ; FAST-NEXT:  .LBB1_2: # %false
 ; FAST-NEXT:    end_block # label0:
-; FAST-NEXT:    i32.const $push1=, 0
-; FAST-NEXT:    i32.const $push2=, 2
-; FAST-NEXT:    i32.store 0($pop1), $pop2
-; FAST-NEXT:    i32.const $push0=, 3
+; FAST-NEXT:    i32.const $push3=, 0
+; FAST-NEXT:    local.get $push6=, 0
+; FAST-NEXT:    i32.store 0($pop3), $pop6
+; FAST-NEXT:    i32.const $push2=, 3
 ; FAST-NEXT:    # fallthrough-return
   %b = icmp ne i32 %a, 0
   br i1 %b, label %true, label %false
 
 true:
   store i32 0, ptr null
   ret i32 1
 
 false:
   store i32 2, ptr null
   ret i32 3
 }

llvm/test/CodeGen/WebAssembly/return-void.ll

 ; FAST-NEXT:    # fallthrough-return
   ret void
 }
 
 define void @return_void_twice(i32 %a) {
 ; CHECK-LABEL: return_void_twice:
 ; CHECK:         .functype return_void_twice (i32) -> ()
 ; CHECK-NEXT:  # %bb.0:
-; CHECK-NEXT:    block
-; CHECK-NEXT:    local.get 0
-; CHECK-NEXT:    i32.eqz
-; CHECK-NEXT:    br_if 0 # 0: down to label0
-; CHECK-NEXT:  # %bb.1: # %true
-; CHECK-NEXT:    i32.const 0
-; CHECK-NEXT:    i32.const 0
-; CHECK-NEXT:    i32.store 0
-; CHECK-NEXT:    return
-; CHECK-NEXT:  .LBB1_2: # %false
-; CHECK-NEXT:    end_block # label0:
-; CHECK-NEXT:    i32.const 0
-; CHECK-NEXT:    i32.const 1
-; CHECK-NEXT:    i32.store 0
 ; CHECK-NEXT:    # fallthrough-return
 ;
 ; FAST-LABEL: return_void_twice:
 ; FAST:         .functype return_void_twice (i32) -> ()
 ; FAST-NEXT:  # %bb.0:
 ; FAST-NEXT:    block
 ; FAST-NEXT:    local.get 0
 ; FAST-NEXT:    i32.eqz
 ; FAST-NEXT:    br_if 0 # 0: down to label0
 ; FAST-NEXT:  # %bb.1: # %true
 ; FAST-NEXT:    i32.const 0
-; FAST-NEXT:    i32.const 0
+; FAST-NEXT:    local.get 0
 ; FAST-NEXT:    i32.store 0
 ; FAST-NEXT:    return
 ; FAST-NEXT:  .LBB1_2: # %false
 ; FAST-NEXT:    end_block # label0:
 ; FAST-NEXT:    i32.const 0
-; FAST-NEXT:    i32.const 1
+; FAST-NEXT:    local.get 0
 ; FAST-NEXT:    i32.store 0
 ; FAST-NEXT:    # fallthrough-return
   %b = icmp ne i32 %a, 0
   br i1 %b, label %true, label %false
 
 true:
   store i32 0, ptr null
   ret void
 
 false:
   store i32 1, ptr null
   ret void
 }

llvm/test/CodeGen/WebAssembly/returned.ll

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc < %s -disable-wasm-fallthrough-return-opt -wasm-disable-explicit-locals -wasm-keep-registers | FileCheck %s
 
 ; Test that the "returned" attribute is optimized effectively.
 
 target triple = "wasm32-unknown-unknown"
 
 %class.Apple = type { i8 }
 declare noalias ptr @_Znwm(i32)
 declare ptr @_ZN5AppleC1Ev(ptr returned)
 define ptr @_Z3foov() {
 ; CHECK-LABEL: _Z3foov:
 ; CHECK:         .functype _Z3foov () -> (i32)
 ; CHECK-NEXT:  # %bb.0: # %entry
 ; CHECK-NEXT:    i32.const $push0=, 1
-; CHECK-NEXT:    call $push1=, _Znwm, $pop0
-; CHECK-NEXT:    call $push2=, _ZN5AppleC1Ev, $pop1
-; CHECK-NEXT:    return $pop2
+; CHECK-NEXT:    call $push2=, _Znwm, $pop0
+; CHECK-NEXT:    local.tee $push1=, $0=, $pop2
+; CHECK-NEXT:    call $drop=, _ZN5AppleC1Ev, $pop1
+; CHECK-NEXT:    return $0
 entry:
   %call = tail call noalias ptr @_Znwm(i32 1)
   %call1 = tail call ptr @_ZN5AppleC1Ev(ptr %call)
   ret ptr %call
 }
 
 declare ptr @memcpy(ptr returned, ptr, i32)
 define ptr @_Z3barPvS_l(ptr %p, ptr %s, i32 %n) {

 }
 
 ; Test that the optimization is performed on arguments other than the first.
 declare i32 @do_something_else(i32, i32 returned)
 define i32 @test_second_arg(i32 %a, i32 %b) {
 ; CHECK-LABEL: test_second_arg:
 ; CHECK:         .functype test_second_arg (i32, i32) -> (i32)
 ; CHECK-NEXT:  # %bb.0:
-; CHECK-NEXT:    call $push0=, do_something_else, $0, $1
-; CHECK-NEXT:    return $pop0
+; CHECK-NEXT:    call $drop=, do_something_else, $0, $1
+; CHECK-NEXT:    return $1
     %call = call i32 @do_something_else(i32 %a, i32 %b)
     ret i32 %b
 }

llvm/test/CodeGen/WebAssembly/select.ll

; FAST-NEXT:    return $pop0
   %c = select i1 %b, i32 %y, i32 %z
   ret i32 %c
 }
 
 define i64 @pr40805_i64(i64 %x, i64 %y, i64 %z) {
 ; CHECK-LABEL: pr40805_i64:
 ; CHECK:         .functype pr40805_i64 (i64, i64, i64) -> (i64)
 ; CHECK-NEXT:  # %bb.0:
-; CHECK-NEXT:    i32.wrap_i64 $push0=, $0
-; CHECK-NEXT:    i32.const $push1=, 1
-; CHECK-NEXT:    i32.and $push2=, $pop0, $pop1
-; CHECK-NEXT:    i64.select $push3=, $1, $2, $pop2
+; CHECK-NEXT:    i64.const $push0=, 1
+; CHECK-NEXT:    i64.and $push1=, $0, $pop0
+; CHECK-NEXT:    i64.eqz $push2=, $pop1
+; CHECK-NEXT:    i64.select $push3=, $2, $1, $pop2
 ; CHECK-NEXT:    return $pop3
 ;
 ; FAST-LABEL: pr40805_i64:
 ; FAST:         .functype pr40805_i64 (i64, i64, i64) -> (i64)
 ; FAST-NEXT:  # %bb.0:
 ; FAST-NEXT:    i64.const $push5=, 1
 ; FAST-NEXT:    i64.and $push6=, $0, $pop5
 ; FAST-NEXT:    i64.const $push3=, 0
-; FAST-NEXT:    i64.ne $push4=, $pop6, $pop3
+; FAST-NEXT:    i64.eq $push4=, $pop6, $pop3
 ; FAST-NEXT:    i32.const $push0=, 1
 ; FAST-NEXT:    i32.and $push1=, $pop4, $pop0
-; FAST-NEXT:    i64.select $push2=, $1, $2, $pop1
+; FAST-NEXT:    i64.select $push2=, $2, $1, $pop1
 ; FAST-NEXT:    return $pop2
   %a = and i64 %x, 1
   %b = icmp ne i64 %a, 0
   %c = select i1 %b, i64 %y, i64 %z
   ret i64 %c
 }
 
 define float @pr44012_i32(i32 %x, float %y, float %z) {
; FAST-NEXT:    return $pop0
   %c = select i1 %b, float %y, float %z
   ret float %c
 }
 
 define float @pr44012_i64(i64 %x, float %y, float %z) {
 ; CHECK-LABEL: pr44012_i64:
 ; CHECK:         .functype pr44012_i64 (i64, f32, f32) -> (f32)
 ; CHECK-NEXT:  # %bb.0:
-; CHECK-NEXT:    i32.wrap_i64 $push0=, $0
-; CHECK-NEXT:    i32.const $push1=, 1
-; CHECK-NEXT:    i32.and $push2=, $pop0, $pop1
-; CHECK-NEXT:    f32.select $push3=, $1, $2, $pop2
+; CHECK-NEXT:    i64.const $push0=, 1
+; CHECK-NEXT:    i64.and $push1=, $0, $pop0
+; CHECK-NEXT:    i64.eqz $push2=, $pop1
+; CHECK-NEXT:    f32.select $push3=, $2, $1, $pop2
 ; CHECK-NEXT:    return $pop3
 ;
 ; FAST-LABEL: pr44012_i64:
 ; FAST:         .functype pr44012_i64 (i64, f32, f32) -> (f32)
 ; FAST-NEXT:  # %bb.0:
 ; FAST-NEXT:    i64.const $push5=, 1
 ; FAST-NEXT:    i64.and $push6=, $0, $pop5
 ; FAST-NEXT:    i64.const $push3=, 0
-; FAST-NEXT:    i64.ne $push4=, $pop6, $pop3
+; FAST-NEXT:    i64.eq $push4=, $pop6, $pop3
 ; FAST-NEXT:    i32.const $push0=, 1
 ; FAST-NEXT:    i32.and $push1=, $pop4, $pop0
-; FAST-NEXT:    f32.select $push2=, $1, $2, $pop1
+; FAST-NEXT:    f32.select $push2=, $2, $1, $pop1
 ; FAST-NEXT:    return $pop2
   %a = and i64 %x, 1
   %b = icmp ne i64 %a, 0
   %c = select i1 %b, float %y, float %z
   ret float %c
 }

llvm/test/CodeGen/WebAssembly/simd-arith.ll

 ; SIMD128-FAST:         .functype min_s_v16i8 (v128, v128) -> (v128)
 ; SIMD128-FAST-NEXT:  # %bb.0:
 ; SIMD128-FAST-NEXT:    i8x16.min_s $push0=, $0, $1
 ; SIMD128-FAST-NEXT:    return $pop0
 ;
 ; NO-SIMD128-LABEL: min_s_v16i8:
 ; NO-SIMD128:         .functype min_s_v16i8 (i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32) -> ()
 ; NO-SIMD128-NEXT:  # %bb.0:
-; NO-SIMD128-NEXT:    i32.const $push4=, 15
-; NO-SIMD128-NEXT:    i32.add $push5=, $0, $pop4
-; NO-SIMD128-NEXT:    i32.extend8_s $push1=, $16
-; NO-SIMD128-NEXT:    i32.extend8_s $push0=, $32
-; NO-SIMD128-NEXT:    i32.lt_s $push2=, $pop1, $pop0
-; NO-SIMD128-NEXT:    i32.select $push3=, $16, $32, $pop2
-; NO-SIMD128-NEXT:    i32.store8 0($pop5), $pop3
-; NO-SIMD128-NEXT:    i32.const $push10=, 14
+; NO-SIMD128-NEXT:    i32.const $push2=, 15
+; NO-SIMD128-NEXT:    i32.add $push3=, $0, $pop2
+; NO-SIMD128-NEXT:    i32.extend8_s $push117=, $16
+; NO-SIMD128-NEXT:    local.tee $push116=, $16=, $pop117
+; NO-SIMD128-NEXT:    i32.extend8_s $push115=, $32
+; NO-SIMD128-NEXT:    local.tee $push114=, $32=, $pop115
+; NO-SIMD128-NEXT:    i32.lt_s $push0=, $16, $32
+; NO-SIMD128-NEXT:    i32.select $push1=, $pop116, $pop114, $pop0
+; NO-SIMD128-NEXT:    i32.store8 0($pop3), $pop1
+; NO-SIMD128-NEXT:    i32.const $push6=, 14
+; NO-SIMD128-NEXT:    i32.add $push7=, $0, $pop6
+; NO-SIMD128-NEXT:    i32.extend8_s $push113=, $15
+; NO-SIMD128-NEXT:    local.tee $push112=, $32=, $pop113
+; NO-SIMD128-NEXT:    i32.extend8_s $push111=, $31
+; NO-SIMD128-NEXT:    local.tee $push110=, $16=, $pop111
+; NO-SIMD128-NEXT:    i32.lt_s $push4=, $32, $16
+; NO-SIMD128-NEXT:    i32.select $push5=, $pop112, $pop110, $pop4
+; NO-SIMD128-NEXT:    i32.store8 0($pop7), $pop5
+; NO-SIMD128-NEXT:    i32.const $push10=, 13
 ; NO-SIMD128-NEXT:    i32.add $push11=, $0, $pop10
-; NO-SIMD128-NEXT:    i32.extend8_s $push7=, $15
-; NO-SIMD128-NEXT:    i32.extend8_s $push6=, $31
-; NO-SIMD128-NEXT:    i32.lt_s $push8=, $pop7, $pop6
-; NO-SIMD128-NEXT:    i32.select $push9=, $15, $31, $pop8
+; NO-SIMD128-NEXT:    i32.extend8_s $push109=, $14
+; NO-SIMD128-NEXT:    local.tee $push108=, $32=, $pop109
+; NO-SIMD128-NEXT:    i32.extend8_s $push107=, $30
+; NO-SIMD128-NEXT:    local.tee $push106=, $16=, $pop107
+; NO-SIMD128-NEXT:    i32.lt_s $push8=, $32, $16
+; NO-SIMD128-NEXT:    i32.select $push9=, $pop108, $pop106, $pop8
 ; NO-SIMD128-NEXT:    i32.store8 0($pop11), $pop9
-; NO-SIMD128-NEXT:    i32.const $push16=, 13
-; NO-SIMD128-NEXT:    i32.add $push17=, $0, $pop16
-; NO-SIMD128-NEXT:    i32.extend8_s $push13=, $14
-; NO-SIMD128-NEXT:    i32.extend8_s $push12=, $30
-; NO-SIMD128-NEXT:    i32.lt_s $push14=, $pop13, $pop12
-; NO-SIMD128-NEXT:    i32.select $push15=, $14, $30, $pop14
-; NO-SIMD128-NEXT:    i32.store8 0($pop17), $pop15
-; NO-SIMD128-NEXT:    i32.const $push22=, 12
+; NO-SIMD128-NEXT:    i32.const $push14=, 12
+; NO-SIMD128-NEXT:    i32.add $push15=, $0, $pop14
+; NO-SIMD128-NEXT:    i32.extend8_s $push105=, $13
+; NO-SIMD128-NEXT:    local.tee $push104=, $32=, $pop105
+; NO-SIMD128-NEXT:    i32.extend8_s $push103=, $29
+; NO-SIMD128-NEXT:    local.tee $push102=, $16=, $pop103
+; NO-SIMD128-NEXT:    i32.lt_s $push12=, $32, $16
+; NO-SIMD128-NEXT:    i32.select $push13=, $pop104, $pop102, $pop12
+; NO-SIMD128-NEXT:    i32.store8 0($pop15), $pop13
+; NO-SIMD128-NEXT:    i32.const $push18=, 11
+; NO-SIMD128-NEXT:    i32.add $push19=, $0, $pop18
+; NO-SIMD128-NEXT:    i32.extend8_s $push101=, $12
+; NO-SIMD128-NEXT:    local.tee $push100=, $32=, $pop101
+; NO-SIMD128-NEXT:    i32.extend8_s $push99=, $28
+; NO-SIMD128-NEXT:    local.tee $push98=, $16=, $pop99
+; NO-SIMD128-NEXT:    i32.lt_s $push16=, $32, $16
+; NO-SIMD128-NEXT:    i32.select $push17=, $pop100, $pop98, $pop16
+; NO-SIMD128-NEXT:    i32.store8 0($pop19), $pop17
+; NO-SIMD128-NEXT:    i32.const $push22=, 10
 ; NO-SIMD128-NEXT:    i32.add $push23=, $0, $pop22
-; NO-SIMD128-NEXT:    i32.extend8_s $push19=, $13
-; NO-SIMD128-NEXT:    i32.extend8_s $push18=, $29
-; NO-SIMD128-NEXT:    i32.lt_s $push20=, $pop19, $pop18
-; NO-SIMD128-NEXT:    i32.select $push21=, $13, $29, $pop20
+; NO-SIMD128-NEXT:    i32.extend8_s $push97=, $11
+; NO-SIMD128-NEXT:    local.tee $push96=, $32=, $pop97
+; NO-SIMD128-NEXT:    i32.extend8_s $push95=, $27
+; NO-SIMD128-NEXT:    local.tee $push94=, $16=, $pop95
+; NO-SIMD128-NEXT:    i32.lt_s $push20=, $32, $16
+; NO-SIMD128-NEXT:    i32.select $push21=, $pop96, $pop94, $pop20
 ; NO-SIMD128-NEXT:    i32.store8 0($pop23), $pop21
-; NO-SIMD128-NEXT:    i32.const $push28=, 11
-; NO-SIMD128-NEXT:    i32.add $push29=, $0, $pop28
-; NO-SIMD128-NEXT:    i32.extend8_s $push25=, $12
-; NO-SIMD128-NEXT:    i32.extend8_s $push24=, $28
-; NO-SIMD128-NEXT:    i32.lt_s $push26=, $pop25, $pop24
-; NO-SIMD128-NEXT:    i32.select $push27=, $12, $28, $pop26
-; NO-SIMD128-NEXT:    i32.store8 0($pop29), $pop27
-; NO-SIMD128-NEXT:    i32.const $push34=, 10
-; NO-SIMD128-NEXT:    i32.add $push35=, $0, $pop34
-; NO-SIMD128-NEXT:    i32.extend8_s $push31=, $11
-; NO-SIMD128-NEXT:    i32.extend8_s $push30=, $27
-; NO-SIMD128-NEXT:    i32.lt_s $push32=, $pop31, $pop30
-; NO-SIMD128-NEXT:    i32.select $push33=, $11, $27, $pop32
-; NO-SIMD128-NEXT:    i32.store8 0($pop35), $pop33
-; NO-SIMD128-NEXT:    i32.const $push40=, 9
+; NO-SIMD128-NEXT:    i32.const $push26=, 9
+; NO-SIMD128-NEXT:    i32.add $push27=, $0, $pop26
+; NO-SIMD128-NEXT:    i32.extend8_s $push93=, $10
+; NO-SIMD128-NEXT:    local.tee $push92=, $32=, $pop93
+; NO-SIMD128-NEXT:    i32.extend8_s $push91=, $26
+; NO-SIMD128-NEXT:    local.tee $push90=, $16=, $pop91
+; NO-SIMD128-NEXT:    i32.lt_s $push24=, $32, $16
+; NO-SIMD128-NEXT:    i32.select $push25=, $pop92, $pop90, $pop24
+; NO-SIMD128-NEXT:    i32.store8 0($pop27), $pop25
+; NO-SIMD128-NEXT:    i32.extend8_s $push89=, $9
+; NO-SIMD128-NEXT:    local.tee $push88=, $32=, $pop89
+; NO-SIMD128-NEXT:    i32.extend8_s $push87=, $25
+; NO-SIMD128-NEXT:    local.tee $push86=, $16=, $pop87
+; NO-SIMD128-NEXT:    i32.lt_s $push28=, $32, $16
+; NO-SIMD128-NEXT:    i32.select $push29=, $pop88, $pop86, $pop28
+; NO-SIMD128-NEXT:    i32.store8 8($0), $pop29
+; NO-SIMD128-NEXT:    i32.const $push32=, 7
+; NO-SIMD128-NEXT:    i32.add $push33=, $0, $pop32
+; NO-SIMD128-NEXT:    i32.extend8_s $push85=, $8
+; NO-SIMD128-NEXT:    local.tee $push84=, $32=, $pop85
+; NO-SIMD128-NEXT:    i32.extend8_s $push83=, $24
+; NO-SIMD128-NEXT:    local.tee $push82=, $16=, $pop83
+; NO-SIMD128-NEXT:    i32.lt_s $push30=, $32, $16
+; NO-SIMD128-NEXT:    i32.select $push31=, $pop84, $pop82, $pop30
+; NO-SIMD128-NEXT:    i32.store8 0($pop33), $pop31
+; NO-SIMD128-NEXT:    i32.const $push36=, 6
+; NO-SIMD128-NEXT:    i32.add $push37=, $0, $pop36
+; NO-SIMD128-NEXT:    i32.extend8_s $push81=, $7
+; NO-SIMD128-NEXT:    local.tee $push80=, $32=, $pop81
+; NO-SIMD128-NEXT:    i32.extend8_s $push79=, $23
+; NO-SIMD128-NEXT:    local.tee $push78=, $16=, $pop79
+; NO-SIMD128-NEXT:    i32.lt_s $push34=, $32, $16
+; NO-SIMD128-NEXT:    i32.select $push35=, $pop80, $pop78, $pop34
+; NO-SIMD128-NEXT:    i32.store8 0($pop37), $pop35
+; NO-SIMD128-NEXT:    i32.const $push40=, 5
 ; NO-SIMD128-NEXT:    i32.add $push41=, $0, $pop40
-; NO-SIMD128-NEXT:    i32.extend8_s $push37=, $10
-; NO-SIMD128-NEXT:    i32.extend8_s $push36=, $26
-; NO-SIMD128-NEXT:    i32.lt_s $push38=, $pop37, $pop36
-; NO-SIMD128-NEXT:    i32.select $push39=, $10, $26, $pop38
+; NO-SIMD128-NEXT:    i32.extend8_s $push77=, $6
+; NO-SIMD128-NEXT:    local.tee $push76=, $32=, $pop77
+; NO-SIMD128-NEXT:    i32.extend8_s $push75=, $22
+; NO-SIMD128-NEXT:    local.tee $push74=, $16=, $pop75
+; NO-SIMD128-NEXT:    i32.lt_s $push38=, $32, $16
+; NO-SIMD128-NEXT:    i32.select $push39=, $pop76, $pop74, $pop38
 ; NO-SIMD128-NEXT:    i32.store8 0($pop41), $pop39
-; NO-SIMD128-NEXT:    i32.extend8_s $push43=, $9
-; NO-SIMD128-NEXT:    i32.extend8_s $push42=, $25
-; NO-SIMD128-NEXT:    i32.lt_s $push44=, $pop43, $pop42
-; NO-SIMD128-NEXT:    i32.select $push45=, $9, $25, $pop44
-; NO-SIMD128-NEXT:    i32.store8 8($0), $pop45
-; NO-SIMD128-NEXT:    i32.const $push50=, 7
-; NO-SIMD128-NEXT:    i32.add $push51=, $0, $pop50
-; NO-SIMD128-NEXT:    i32.extend8_s $push47=, $8
-; NO-SIMD128-NEXT:    i32.extend8_s $push46=, $24
-; NO-SIMD128-NEXT:    i32.lt_s $push48=, $pop47, $pop46
-; NO-SIMD128-NEXT:    i32.select $push49=, $8, $24, $pop48
-; NO-SIMD128-NEXT:    i32.store8 0($pop51), $pop49
-; NO-SIMD128-NEXT:    i32.const $push56=, 6
-; NO-SIMD128-NEXT:    i32.add $push57=, $0, $pop56
-; NO-SIMD128-NEXT:    i32.extend8_s $push53=, $7
-; NO-SIMD128-NEXT:    i32.extend8_s $push52=, $23
-; NO-SIMD128-NEXT:    i32.lt_s $push54=, $pop53, $pop52
-; NO-SIMD128-NEXT:    i32.select $push55=, $7, $23, $pop54
-; NO-SIMD128-NEXT:    i32.store8 0($pop57), $pop55
-; NO-SIMD128-NEXT:    i32.const $push62=, 5
-; NO-SIMD128-NEXT:    i32.add $push63=, $0, $pop62
-; NO-SIMD128-NEXT:    i32.extend8_s $push59=, $6
-; NO-SIMD128-NEXT:    i32.extend8_s $push58=, $22
-; NO-SIMD128-NEXT:    i32.lt_s $push60=, $pop59, $pop58
-; NO-SIMD128-NEXT:    i32.select $push61=, $6, $22, $pop60
-; NO-SIMD128-NEXT:    i32.store8 0($pop63), $pop61
-; NO-SIMD128-NEXT:    i32.extend8_s $push65=, $5
-; NO-SIMD128-NEXT:    i32.extend8_s $push64=, $21
-; NO-SIMD128-NEXT:    i32.lt_s $push66=, $pop65, $pop64
-; NO-SIMD128-NEXT:    i32.select $push67=, $5, $21, $pop66
-; NO-SIMD128-NEXT:    i32.store8 4($0), $pop67
-; NO-SIMD128-NEXT:    i32.const $push72=, 3
-; NO-SIMD128-NEXT:    i32.add $push73=, $0, $pop72
+; NO-SIMD128-NEXT:    i32.extend8_s $push73=, $5
+; NO-SIMD128-NEXT:    local.tee $push72=, $32=, $pop73
+; NO-SIMD128-NEXT:    i32.extend8_s $push71=, $21
+; NO-SIMD128-NEXT:    local.tee $push70=, $16=, $pop71
+; NO-SIMD128-NEXT:    i32.lt_s $push42=, $32, $16
+; NO-SIMD128-NEXT:    i32.select $push43=, $pop72, $pop70, $pop42
+; NO-SIMD128-NEXT:    i32.store8 4($0), $pop43
+; NO-SIMD128-NEXT:    i32.const $push46=, 3
+; NO-SIMD128-NEXT:    i32.add $push47=, $0, $pop46
 ; NO-SIMD128-NEXT:    i32.extend8_s $push69=, $4
-; NO-SIMD128-NEXT:    i32.extend8_s $push68=, $20
-; NO-SIMD128-NEXT:    i32.lt_s $push70=, $pop69, $pop68
-; NO-SIMD128-NEXT:    i32.select $push71=, $4, $20, $pop70
-; NO-SIMD128-NEXT:    i32.store8 0($pop73), $pop71
-; NO-SIMD128-NEXT:    i32.extend8_s $push75=, $3
-; NO-SIMD128-NEXT:    i32.extend8_s $push74=, $19
-; NO-SIMD128-NEXT:    i32.lt_s $push76=, $pop75, $pop74
-; NO-SIMD128-NEXT:    i32.select $push77=, $3, $19, $pop76
-; NO-SIMD128-NEXT:    i32.store8 2($0), $pop77
-; NO-SIMD128-NEXT:    i32.extend8_s $push79=, $2
-; NO-SIMD128-NEXT:    i32.extend8_s $push78=, $18
-; NO-SIMD128-NEXT:    i32.lt_s $push80=, $pop79, $pop78
-; NO-SIMD128-NEXT:    i32.select $push81=, $2, $18, $pop80
-; NO-SIMD128-NEXT:    i32.store8 1($0), $pop81
-; NO-SIMD128-NEXT:    i32.extend8_s $push83=, $1
-; NO-SIMD128-NEXT:    i32.extend8_s $push82=, $17
-; NO-SIMD128-NEXT:    i32.lt_s $push84=, $pop83, $pop82
-; NO-SIMD128-NEXT:    i32.select $push85=, $1, $17, $pop84
-; NO-SIMD128-NEXT:    i32.store8 0($0), $pop85
+; NO-SIMD128-NEXT:    local.tee $push68=, $32=, $pop69
+; NO-SIMD128-NEXT:    i32.extend8_s $push67=, $20
+; NO-SIMD128-NEXT:    local.tee $push66=, $16=, $pop67
+; NO-SIMD128-NEXT:    i32.lt_s $push44=, $32, $16
+; NO-SIMD128-NEXT:    i32.select $push45=, $pop68, $pop66, $pop44
+; NO-SIMD128-NEXT:    i32.store8 0($pop47), $pop45
+; NO-SIMD128-NEXT:    i32.extend8_s $push65=, $3
+; NO-SIMD128-NEXT:    local.tee $push64=, $32=, $pop65
+; NO-SIMD128-NEXT:    i32.extend8_s $push63=, $19
+; NO-SIMD128-NEXT:    local.tee $push62=, $16=, $pop63
+; NO-SIMD128-NEXT:    i32.lt_s $push48=, $32, $16
+; NO-SIMD128-NEXT:    i32.select $push49=, $pop64, $pop62, $pop48
+; NO-SIMD128-NEXT:    i32.store8 2($0), $pop49
+; NO-SIMD128-NEXT:    i32.extend8_s $push61=, $2
+; NO-SIMD128-NEXT:    local.tee $push60=, $32=, $pop61
+; NO-SIMD128-NEXT:    i32.extend8_s $push59=, $18
+; NO-SIMD128-NEXT:    local.tee $push58=, $16=, $pop59
+; NO-SIMD128-NEXT:    i32.lt_s $push50=, $32, $16
+; NO-SIMD128-NEXT:    i32.select $push51=, $pop60, $pop58, $pop50
+; NO-SIMD128-NEXT:    i32.store8 1($0), $pop51
+; NO-SIMD128-NEXT:    i32.extend8_s $push57=, $1
+; NO-SIMD128-NEXT:    local.tee $push56=, $32=, $pop57
+; NO-SIMD128-NEXT:    i32.extend8_s $push55=, $17
+; NO-SIMD128-NEXT:    local.tee $push54=, $16=, $pop55
+; NO-SIMD128-NEXT:    i32.lt_s $push52=, $32, $16
+; NO-SIMD128-NEXT:    i32.select $push53=, $pop56, $pop54, $pop52
+; NO-SIMD128-NEXT:    i32.store8 0($0), $pop53
 ; NO-SIMD128-NEXT:    return
 ;
 ; NO-SIMD128-FAST-LABEL: min_s_v16i8:
 ; NO-SIMD128-FAST:         .functype min_s_v16i8 (i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32) -> ()
 ; NO-SIMD128-FAST-NEXT:  # %bb.0:
-; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push1=, $1
-; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push0=, $17
-; NO-SIMD128-FAST-NEXT:    i32.lt_s $push2=, $pop1, $pop0
-; NO-SIMD128-FAST-NEXT:    i32.select $push3=, $1, $17, $pop2
-; NO-SIMD128-FAST-NEXT:    i32.store8 0($0), $pop3
-; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push5=, $2
-; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push4=, $18
-; NO-SIMD128-FAST-NEXT:    i32.lt_s $push6=, $pop5, $pop4
-; NO-SIMD128-FAST-NEXT:    i32.select $push7=, $2, $18, $pop6
-; NO-SIMD128-FAST-NEXT:    i32.store8 1($0), $pop7
-; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push9=, $3
-; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push8=, $19
-; NO-SIMD128-FAST-NEXT:    i32.lt_s $push10=, $pop9, $pop8
-; NO-SIMD128-FAST-NEXT:    i32.select $push11=, $3, $19, $pop10
-; NO-SIMD128-FAST-NEXT:    i32.store8 2($0), $pop11
-; NO-SIMD128-FAST-NEXT:    i32.const $push16=, 3
-; NO-SIMD128-FAST-NEXT:    i32.add $push17=, $0, $pop16
-; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push13=, $4
-; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push12=, $20
-; NO-SIMD128-FAST-NEXT:    i32.lt_s $push14=, $pop13, $pop12
-; NO-SIMD128-FAST-NEXT:    i32.select $push15=, $4, $20, $pop14
-; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop17), $pop15
-; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push19=, $5
-; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push18=, $21
-; NO-SIMD128-FAST-NEXT:    i32.lt_s $push20=, $pop19, $pop18
-; NO-SIMD128-FAST-NEXT:    i32.select $push21=, $5, $21, $pop20
-; NO-SIMD128-FAST-NEXT:    i32.store8 4($0), $pop21
-; NO-SIMD128-FAST-NEXT:    i32.const $push26=, 5
-; NO-SIMD128-FAST-NEXT:    i32.add $push27=, $0, $pop26
-; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push23=, $6
-; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push22=, $22
-; NO-SIMD128-FAST-NEXT:    i32.lt_s $push24=, $pop23, $pop22
-; NO-SIMD128-FAST-NEXT:    i32.select $push25=, $6, $22, $pop24
-; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop27), $pop25
-; NO-SIMD128-FAST-NEXT:    i32.const $push32=, 6
+; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push117=, $1
+; NO-SIMD128-FAST-NEXT:    local.tee $push116=, $1=, $pop117
+; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push115=, $17
+; NO-SIMD128-FAST-NEXT:    local.tee $push114=, $17=, $pop115
+; NO-SIMD128-FAST-NEXT:    i32.lt_s $push0=, $1, $17
+; NO-SIMD128-FAST-NEXT:    i32.select $push1=, $pop116, $pop114, $pop0
+; NO-SIMD128-FAST-NEXT:    i32.store8 0($0), $pop1
+; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push113=, $2
+; NO-SIMD128-FAST-NEXT:    local.tee $push112=, $17=, $pop113
+; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push111=, $18
+; NO-SIMD128-FAST-NEXT:    local.tee $push110=, $18=, $pop111
+; NO-SIMD128-FAST-NEXT:    i32.lt_s $push2=, $17, $18
+; NO-SIMD128-FAST-NEXT:    i32.select $push3=, $pop112, $pop110, $pop2
+; NO-SIMD128-FAST-NEXT:    i32.store8 1($0), $pop3
+; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push109=, $3
+; NO-SIMD128-FAST-NEXT:    local.tee $push108=, $18=, $pop109
+; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push107=, $19
+; NO-SIMD128-FAST-NEXT:    local.tee $push106=, $19=, $pop107
+; NO-SIMD128-FAST-NEXT:    i32.lt_s $push4=, $18, $19
+; NO-SIMD128-FAST-NEXT:    i32.select $push5=, $pop108, $pop106, $pop4
+; NO-SIMD128-FAST-NEXT:    i32.store8 2($0), $pop5
+; NO-SIMD128-FAST-NEXT:    i32.const $push8=, 3
+; NO-SIMD128-FAST-NEXT:    i32.add $push9=, $0, $pop8
+; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push105=, $4
+; NO-SIMD128-FAST-NEXT:    local.tee $push104=, $19=, $pop105
+; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push103=, $20
+; NO-SIMD128-FAST-NEXT:    local.tee $push102=, $20=, $pop103
+; NO-SIMD128-FAST-NEXT:    i32.lt_s $push6=, $19, $20
+; NO-SIMD128-FAST-NEXT:    i32.select $push7=, $pop104, $pop102, $pop6
+; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop9), $pop7
+; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push101=, $5
+; NO-SIMD128-FAST-NEXT:    local.tee $push100=, $20=, $pop101
+; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push99=, $21
+; NO-SIMD128-FAST-NEXT:    local.tee $push98=, $21=, $pop99
+; NO-SIMD128-FAST-NEXT:    i32.lt_s $push10=, $20, $21
+; NO-SIMD128-FAST-NEXT:    i32.select $push11=, $pop100, $pop98, $pop10
+; NO-SIMD128-FAST-NEXT:    i32.store8 4($0), $pop11
+; NO-SIMD128-FAST-NEXT:    i32.const $push14=, 5
+; NO-SIMD128-FAST-NEXT:    i32.add $push15=, $0, $pop14
+; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push97=, $6
+; NO-SIMD128-FAST-NEXT:    local.tee $push96=, $21=, $pop97
+; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push95=, $22
+; NO-SIMD128-FAST-NEXT:    local.tee $push94=, $22=, $pop95
+; NO-SIMD128-FAST-NEXT:    i32.lt_s $push12=, $21, $22
+; NO-SIMD128-FAST-NEXT:    i32.select $push13=, $pop96, $pop94, $pop12
+; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop15), $pop13
+; NO-SIMD128-FAST-NEXT:    i32.const $push18=, 6
+; NO-SIMD128-FAST-NEXT:    i32.add $push19=, $0, $pop18
+; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push93=, $7
+; NO-SIMD128-FAST-NEXT:    local.tee $push92=, $22=, $pop93
+; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push91=, $23
+; NO-SIMD128-FAST-NEXT:    local.tee $push90=, $23=, $pop91
+; NO-SIMD128-FAST-NEXT:    i32.lt_s $push16=, $22, $23
+; NO-SIMD128-FAST-NEXT:    i32.select $push17=, $pop92, $pop90, $pop16
+; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop19), $pop17
+; NO-SIMD128-FAST-NEXT:    i32.const $push22=, 7
+; NO-SIMD128-FAST-NEXT:    i32.add $push23=, $0, $pop22
+; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push89=, $8
+; NO-SIMD128-FAST-NEXT:    local.tee $push88=, $23=, $pop89
+; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push87=, $24
+; NO-SIMD128-FAST-NEXT:    local.tee $push86=, $24=, $pop87
+; NO-SIMD128-FAST-NEXT:    i32.lt_s $push20=, $23, $24
+; NO-SIMD128-FAST-NEXT:    i32.select $push21=, $pop88, $pop86, $pop20
+; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop23), $pop21
+; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push85=, $9
+; NO-SIMD128-FAST-NEXT:    local.tee $push84=, $24=, $pop85
+; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push83=, $25
+; NO-SIMD128-FAST-NEXT:    local.tee $push82=, $25=, $pop83
+; NO-SIMD128-FAST-NEXT:    i32.lt_s $push24=, $24, $25
+; NO-SIMD128-FAST-NEXT:    i32.select $push25=, $pop84, $pop82, $pop24
+; NO-SIMD128-FAST-NEXT:    i32.store8 8($0), $pop25
+; NO-SIMD128-FAST-NEXT:    i32.const $push28=, 9
+; NO-SIMD128-FAST-NEXT:    i32.add $push29=, $0, $pop28
+; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push81=, $10
+; NO-SIMD128-FAST-NEXT:    local.tee $push80=, $25=, $pop81
+; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push79=, $26
+; NO-SIMD128-FAST-NEXT:    local.tee $push78=, $26=, $pop79
+; NO-SIMD128-FAST-NEXT:    i32.lt_s $push26=, $25, $26
+; NO-SIMD128-FAST-NEXT:    i32.select $push27=, $pop80, $pop78, $pop26
+; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop29), $pop27
+; NO-SIMD128-FAST-NEXT:    i32.const $push32=, 10
 ; NO-SIMD128-FAST-NEXT:    i32.add $push33=, $0, $pop32
-; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push29=, $7
-; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push28=, $23
-; NO-SIMD128-FAST-NEXT:    i32.lt_s $push30=, $pop29, $pop28
-; NO-SIMD128-FAST-NEXT:    i32.select $push31=, $7, $23, $pop30
+; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push77=, $11
+; NO-SIMD128-FAST-NEXT:    local.tee $push76=, $26=, $pop77
+; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push75=, $27
+; NO-SIMD128-FAST-NEXT:    local.tee $push74=, $27=, $pop75
+; NO-SIMD128-FAST-NEXT:    i32.lt_s $push30=, $26, $27
+; NO-SIMD128-FAST-NEXT:    i32.select $push31=, $pop76, $pop74, $pop30
 ; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop33), $pop31
-; NO-SIMD128-FAST-NEXT:    i32.const $push38=, 7
-; NO-SIMD128-FAST-NEXT:    i32.add $push39=, $0, $pop38
-; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push35=, $8
-; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push34=, $24
-; NO-SIMD128-FAST-NEXT:    i32.lt_s $push36=, $pop35, $pop34
-; NO-SIMD128-FAST-NEXT:    i32.select $push37=, $8, $24, $pop36
-; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop39), $pop37
-; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push41=, $9
-; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push40=, $25
-; NO-SIMD128-FAST-NEXT:    i32.lt_s $push42=, $pop41, $pop40
-; NO-SIMD128-FAST-NEXT:    i32.select $push43=, $9, $25, $pop42
-; NO-SIMD128-FAST-NEXT:    i32.store8 8($0), $pop43
-; NO-SIMD128-FAST-NEXT:    i32.const $push48=, 9
+; NO-SIMD128-FAST-NEXT:    i32.const $push36=, 11
+; NO-SIMD128-FAST-NEXT:    i32.add $push37=, $0, $pop36
+; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push73=, $12
+; NO-SIMD128-FAST-NEXT:    local.tee $push72=, $27=, $pop73
+; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push71=, $28
+; NO-SIMD128-FAST-NEXT:    local.tee $push70=, $28=, $pop71
+; NO-SIMD128-FAST-NEXT:    i32.lt_s $push34=, $27, $28
+; NO-SIMD128-FAST-NEXT:    i32.select $push35=, $pop72, $pop70, $pop34
+; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop37), $pop35
+; NO-SIMD128-FAST-NEXT:    i32.const $push40=, 12
+; NO-SIMD128-FAST-NEXT:    i32.add $push41=, $0, $pop40
+; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push69=, $13
+; NO-SIMD128-FAST-NEXT:    local.tee $push68=, $28=, $pop69
+; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push67=, $29
+; NO-SIMD128-FAST-NEXT:    local.tee $push66=, $29=, $pop67
+; NO-SIMD128-FAST-NEXT:    i32.lt_s $push38=, $28, $29
+; NO-SIMD128-FAST-NEXT:    i32.select $push39=, $pop68, $pop66, $pop38
+; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop41), $pop39
+; NO-SIMD128-FAST-NEXT:    i32.const $push44=, 13
+; NO-SIMD128-FAST-NEXT:    i32.add $push45=, $0, $pop44
+; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push65=, $14
+; NO-SIMD128-FAST-NEXT:    local.tee $push64=, $29=, $pop65
+; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push63=, $30
+; NO-SIMD128-FAST-NEXT:    local.tee $push62=, $30=, $pop63
+; NO-SIMD128-FAST-NEXT:    i32.lt_s $push42=, $29, $30
+; NO-SIMD128-FAST-NEXT:    i32.select $push43=, $pop64, $pop62, $pop42
+; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop45), $pop43
+; NO-SIMD128-FAST-NEXT:    i32.const $push48=, 14
 ; NO-SIMD128-FAST-NEXT:    i32.add $push49=, $0, $pop48
-; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push45=, $10
-; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push44=, $26
-; NO-SIMD128-FAST-NEXT:    i32.lt_s $push46=, $pop45, $pop44
-; NO-SIMD128-FAST-NEXT:    i32.select $push47=, $10, $26, $pop46
+; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push61=, $15
+; NO-SIMD128-FAST-NEXT:    local.tee $push60=, $30=, $pop61
+; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push59=, $31
+; NO-SIMD128-FAST-NEXT:    local.tee $push58=, $31=, $pop59
+; NO-SIMD128-FAST-NEXT:    i32.lt_s $push46=, $30, $31
+; NO-SIMD128-FAST-NEXT:    i32.select $push47=, $pop60, $pop58, $pop46
 ; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop49), $pop47
-; NO-SIMD128-FAST-NEXT:    i32.const $push54=, 10
-; NO-SIMD128-FAST-NEXT:    i32.add $push55=, $0, $pop54
-; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push51=, $11
-; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push50=, $27
-; NO-SIMD128-FAST-NEXT:    i32.lt_s $push52=, $pop51, $pop50
-; NO-SIMD128-FAST-NEXT:    i32.select $push53=, $11, $27, $pop52
-; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop55), $pop53
-; NO-SIMD128-FAST-NEXT:    i32.const $push60=, 11
-; NO-SIMD128-FAST-NEXT:    i32.add $push61=, $0, $pop60
-; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push57=, $12
-; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push56=, $28
-; NO-SIMD128-FAST-NEXT:    i32.lt_s $push58=, $pop57, $pop56
-; NO-SIMD128-FAST-NEXT:    i32.select $push59=, $12, $28, $pop58
-; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop61), $pop59
-; NO-SIMD128-FAST-NEXT:    i32.const $push66=, 12
-; NO-SIMD128-FAST-NEXT:    i32.add $push67=, $0, $pop66
-; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push63=, $13
-; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push62=, $29
-; NO-SIMD128-FAST-NEXT:    i32.lt_s $push64=, $pop63, $pop62
-; NO-SIMD128-FAST-NEXT:    i32.select $push65=, $13, $29, $pop64
-; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop67), $pop65
-; NO-SIMD128-FAST-NEXT:    i32.const $push72=, 13
-; NO-SIMD128-FAST-NEXT:    i32.add $push73=, $0, $pop72
-; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push69=, $14
-; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push68=, $30
-; NO-SIMD128-FAST-NEXT:    i32.lt_s $push70=, $pop69, $pop68
-; NO-SIMD128-FAST-NEXT:    i32.select $push71=, $14, $30, $pop70
-; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop73), $pop71
-; NO-SIMD128-FAST-NEXT:    i32.const $push78=, 14
-; NO-SIMD128-FAST-NEXT:    i32.add $push79=, $0, $pop78
-; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push75=, $15
-; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push74=, $31
-; NO-SIMD128-FAST-NEXT:    i32.lt_s $push76=, $pop75, $pop74
-; NO-SIMD128-FAST-NEXT:    i32.select $push77=, $15, $31, $pop76
-; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop79), $pop77
-; NO-SIMD128-FAST-NEXT:    i32.const $push84=, 15
-; NO-SIMD128-FAST-NEXT:    i32.add $push85=, $0, $pop84
-; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push81=, $16
-; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push80=, $32
-; NO-SIMD128-FAST-NEXT:    i32.lt_s $push82=, $pop81, $pop80
-; NO-SIMD128-FAST-NEXT:    i32.select $push83=, $16, $32, $pop82
-; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop85), $pop83
+; NO-SIMD128-FAST-NEXT:    i32.const $push52=, 15
+; NO-SIMD128-FAST-NEXT:    i32.add $push53=, $0, $pop52
+; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push57=, $16
+; NO-SIMD128-FAST-NEXT:    local.tee $push56=, $0=, $pop57
+; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push55=, $32
+; NO-SIMD128-FAST-NEXT:    local.tee $push54=, $32=, $pop55
+; NO-SIMD128-FAST-NEXT:    i32.lt_s $push50=, $0, $32
+; NO-SIMD128-FAST-NEXT:    i32.select $push51=, $pop56, $pop54, $pop50
+; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop53), $pop51
 ; NO-SIMD128-FAST-NEXT:    return
   %c = icmp slt <16 x i8> %x, %y
   %a = select <16 x i1> %c, <16 x i8> %x, <16 x i8> %y
   ret <16 x i8> %a
 }
 
 define <16 x i8> @min_u_v16i8(<16 x i8> %x, <16 x i8> %y) {
 ; SIMD128-LABEL: min_u_v16i8:
 ; SIMD128:         .functype min_u_v16i8 (v128, v128) -> (v128)
 ; SIMD128-NEXT:  # %bb.0:
 ; SIMD128-NEXT:    i8x16.min_u $push0=, $0, $1
 ; SIMD128-NEXT:    return $pop0
 ;
 ; SIMD128-FAST-LABEL: min_u_v16i8:
 ; SIMD128-FAST:         .functype min_u_v16i8 (v128, v128) -> (v128)
 ; SIMD128-FAST-NEXT:  # %bb.0:
 ; SIMD128-FAST-NEXT:    i8x16.min_u $push0=, $0, $1
 ; SIMD128-FAST-NEXT:    return $pop0
 ;
 ; NO-SIMD128-LABEL: min_u_v16i8:
 ; NO-SIMD128:         .functype min_u_v16i8 (i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32) -> ()
 ; NO-SIMD128-NEXT:  # %bb.0:
-; NO-SIMD128-NEXT:    i32.const $push5=, 15
-; NO-SIMD128-NEXT:    i32.add $push6=, $0, $pop5
+; NO-SIMD128-NEXT:    i32.const $push3=, 15
+; NO-SIMD128-NEXT:    i32.add $push4=, $0, $pop3
 ; NO-SIMD128-NEXT:    i32.const $push0=, 255
-; NO-SIMD128-NEXT:    i32.and $push2=, $16, $pop0
-; NO-SIMD128-NEXT:    i32.const $push117=, 255
-; NO-SIMD128-NEXT:    i32.and $push1=, $32, $pop117
-; NO-SIMD128-NEXT:    i32.lt_u $push3=, $pop2, $pop1
-; NO-SIMD128-NEXT:    i32.select $push4=, $16, $32, $pop3
-; NO-SIMD128-NEXT:    i32.store8 0($pop6), $pop4
-; NO-SIMD128-NEXT:    i32.const $push11=, 14
+; NO-SIMD128-NEXT:    i32.and $push149=, $16, $pop0
+; NO-SIMD128-NEXT:    local.tee $push148=, $16=, $pop149
+; NO-SIMD128-NEXT:    i32.const $push147=, 255
+; NO-SIMD128-NEXT:    i32.and $push146=, $32, $pop147
+; NO-SIMD128-NEXT:    local.tee $push145=, $32=, $pop146
+; NO-SIMD128-NEXT:    i32.lt_u $push1=, $16, $32
+; NO-SIMD128-NEXT:    i32.select $push2=, $pop148, $pop145, $pop1
+; NO-SIMD128-NEXT:    i32.store8 0($pop4), $pop2
+; NO-SIMD128-NEXT:    i32.const $push7=, 14
+; NO-SIMD128-NEXT:    i32.add $push8=, $0, $pop7
+; NO-SIMD128-NEXT:    i32.const $push144=, 255
+; NO-SIMD128-NEXT:    i32.and $push143=, $15, $pop144
+; NO-SIMD128-NEXT:    local.tee $push142=, $32=, $pop143
+; NO-SIMD128-NEXT:    i32.const $push141=, 255
+; NO-SIMD128-NEXT:    i32.and $push140=, $31, $pop141
+; NO-SIMD128-NEXT:    local.tee $push139=, $16=, $pop140
+; NO-SIMD128-NEXT:    i32.lt_u $push5=, $32, $16
+; NO-SIMD128-NEXT:    i32.select $push6=, $pop142, $pop139, $pop5
+; NO-SIMD128-NEXT:    i32.store8 0($pop8), $pop6
+; NO-SIMD128-NEXT:    i32.const $push11=, 13
 ; NO-SIMD128-NEXT:    i32.add $push12=, $0, $pop11
-; NO-SIMD128-NEXT:    i32.const $push116=, 255
-; NO-SIMD128-NEXT:    i32.and $push8=, $15, $pop116
-; NO-SIMD128-NEXT:    i32.const $push115=, 255
-; NO-SIMD128-NEXT:    i32.and $push7=, $31, $pop115
-; NO-SIMD128-NEXT:    i32.lt_u $push9=, $pop8, $pop7
-; NO-SIMD128-NEXT:    i32.select $push10=, $15, $31, $pop9
+; NO-SIMD128-NEXT:    i32.const $push138=, 255
+; NO-SIMD128-NEXT:    i32.and $push137=, $14, $pop138
+; NO-SIMD128-NEXT:    local.tee $push136=, $32=, $pop137
+; NO-SIMD128-NEXT:    i32.const $push135=, 255
+; NO-SIMD128-NEXT:    i32.and $push134=, $30, $pop135
+; NO-SIMD128-NEXT:    local.tee $push133=, $16=, $pop134
+; NO-SIMD128-NEXT:    i32.lt_u $push9=, $32, $16
+; NO-SIMD128-NEXT:    i32.select $push10=, $pop136, $pop133, $pop9
 ; NO-SIMD128-NEXT:    i32.store8 0($pop12), $pop10
-; NO-SIMD128-NEXT:    i32.const $push17=, 13
-; NO-SIMD128-NEXT:    i32.add $push18=, $0, $pop17
-; NO-SIMD128-NEXT:    i32.const $push114=, 255
-; NO-SIMD128-NEXT:    i32.and $push14=, $14, $pop114
-; NO-SIMD128-NEXT:    i32.const $push113=, 255
-; NO-SIMD128-NEXT:    i32.and $push13=, $30, $pop113
-; NO-SIMD128-NEXT:    i32.lt_u $push15=, $pop14, $pop13
-; NO-SIMD128-NEXT:    i32.select $push16=, $14, $30, $pop15
-; NO-SIMD128-NEXT:    i32.store8 0($pop18), $pop16
-; NO-SIMD128-NEXT:    i32.const $push23=, 12
+; NO-SIMD128-NEXT:    i32.const $push15=, 12
+; NO-SIMD128-NEXT:    i32.add $push16=, $0, $pop15
+; NO-SIMD128-NEXT:    i32.const $push132=, 255
+; NO-SIMD128-NEXT:    i32.and $push131=, $13, $pop132
+; NO-SIMD128-NEXT:    local.tee $push130=, $32=, $pop131
+; NO-SIMD128-NEXT:    i32.const $push129=, 255
+; NO-SIMD128-NEXT:    i32.and $push128=, $29, $pop129
+; NO-SIMD128-NEXT:    local.tee $push127=, $16=, $pop128
+; NO-SIMD128-NEXT:    i32.lt_u $push13=, $32, $16
+; NO-SIMD128-NEXT:    i32.select $push14=, $pop130, $pop127, $pop13
+; NO-SIMD128-NEXT:    i32.store8 0($pop16), $pop14
+; NO-SIMD128-NEXT:    i32.const $push19=, 11
+; NO-SIMD128-NEXT:    i32.add $push20=, $0, $pop19
+; NO-SIMD128-NEXT:    i32.const $push126=, 255
+; NO-SIMD128-NEXT:    i32.and $push125=, $12, $pop126
+; NO-SIMD128-NEXT:    local.tee $push124=, $32=, $pop125
+; NO-SIMD128-NEXT:    i32.const $push123=, 255
+; NO-SIMD128-NEXT:    i32.and $push122=, $28, $pop123
+; NO-SIMD128-NEXT:    local.tee $push121=, $16=, $pop122
+; NO-SIMD128-NEXT:    i32.lt_u $push17=, $32, $16
+; NO-SIMD128-NEXT:    i32.select $push18=, $pop124, $pop121, $pop17
+; NO-SIMD128-NEXT:    i32.store8 0($pop20), $pop18
+; NO-SIMD128-NEXT:    i32.const $push23=, 10
 ; NO-SIMD128-NEXT:    i32.add $push24=, $0, $pop23
-; NO-SIMD128-NEXT:    i32.const $push112=, 255
-; NO-SIMD128-NEXT:    i32.and $push20=, $13, $pop112
-; NO-SIMD128-NEXT:    i32.const $push111=, 255
-; NO-SIMD128-NEXT:    i32.and $push19=, $29, $pop111
-; NO-SIMD128-NEXT:    i32.lt_u $push21=, $pop20, $pop19
-; NO-SIMD128-NEXT:    i32.select $push22=, $13, $29, $pop21
+; NO-SIMD128-NEXT:    i32.const $push120=, 255
+; NO-SIMD128-NEXT:    i32.and $push119=, $11, $pop120
+; NO-SIMD128-NEXT:    local.tee $push118=, $32=, $pop119
+; NO-SIMD128-NEXT:    i32.const $push117=, 255
+; NO-SIMD128-NEXT:    i32.and $push116=, $27, $pop117
+; NO-SIMD128-NEXT:    local.tee $push115=, $16=, $pop116
+; NO-SIMD128-NEXT:    i32.lt_u $push21=, $32, $16
+; NO-SIMD128-NEXT:    i32.select $push22=, $pop118, $pop115, $pop21
 ; NO-SIMD128-NEXT:    i32.store8 0($pop24), $pop22
-; NO-SIMD128-NEXT:    i32.const $push29=, 11
-; NO-SIMD128-NEXT:    i32.add $push30=, $0, $pop29
-; NO-SIMD128-NEXT:    i32.const $push110=, 255
-; NO-SIMD128-NEXT:    i32.and $push26=, $12, $pop110
-; NO-SIMD128-NEXT:    i32.const $push109=, 255
-; NO-SIMD128-NEXT:    i32.and $push25=, $28, $pop109
-; NO-SIMD128-NEXT:    i32.lt_u $push27=, $pop26, $pop25
-; NO-SIMD128-NEXT:    i32.select $push28=, $12, $28, $pop27
-; NO-SIMD128-NEXT:    i32.store8 0($pop30), $pop28
-; NO-SIMD128-NEXT:    i32.const $push35=, 10
-; NO-SIMD128-NEXT:    i32.add $push36=, $0, $pop35
+; NO-SIMD128-NEXT:    i32.const $push27=, 9
+; NO-SIMD128-NEXT:    i32.add $push28=, $0, $pop27
+; NO-SIMD128-NEXT:    i32.const $push114=, 255
+; NO-SIMD128-NEXT:    i32.and $push113=, $10, $pop114
+; NO-SIMD128-NEXT:    local.tee $push112=, $32=, $pop113
+; NO-SIMD128-NEXT:    i32.const $push111=, 255
+; NO-SIMD128-NEXT:    i32.and $push110=, $26, $pop111
+; NO-SIMD128-NEXT:    local.tee $push109=, $16=, $pop110
+; NO-SIMD128-NEXT:    i32.lt_u $push25=, $32, $16
+; NO-SIMD128-NEXT:    i32.select $push26=, $pop112, $pop109, $pop25
+; NO-SIMD128-NEXT:    i32.store8 0($pop28), $pop26
 ; NO-SIMD128-NEXT:    i32.const $push108=, 255
-; NO-SIMD128-NEXT:    i32.and $push32=, $11, $pop108
-; NO-SIMD128-NEXT:    i32.const $push107=, 255
-; NO-SIMD128-NEXT:    i32.and $push31=, $27, $pop107
-; NO-SIMD128-NEXT:    i32.lt_u $push33=, $pop32, $pop31
-; NO-SIMD128-NEXT:    i32.select $push34=, $11, $27, $pop33
-; NO-SIMD128-NEXT:    i32.store8 0($pop36), $pop34
-; NO-SIMD128-NEXT:    i32.const $push41=, 9
-; NO-SIMD128-NEXT:    i32.add $push42=, $0, $pop41
-; NO-SIMD128-NEXT:    i32.const $push106=, 255
-; NO-SIMD128-NEXT:    i32.and $push38=, $10, $pop106
+; NO-SIMD128-NEXT:    i32.and $push107=, $9, $pop108
+; NO-SIMD128-NEXT:    local.tee $push106=, $32=, $pop107
 ; NO-SIMD128-NEXT:    i32.const $push105=, 255
-; NO-SIMD128-NEXT:    i32.and $push37=, $26, $pop105
-; NO-SIMD128-NEXT:    i32.lt_u $push39=, $pop38, $pop37
-; NO-SIMD128-NEXT:    i32.select $push40=, $10, $26, $pop39
-; NO-SIMD128-NEXT:    i32.store8 0($pop42), $pop40
-; NO-SIMD128-NEXT:    i32.const $push104=, 255
-; NO-SIMD128-NEXT:    i32.and $push44=, $9, $pop104
-; NO-SIMD128-NEXT:    i32.const $push103=, 255
-; NO-SIMD128-NEXT:    i32.and $push43=, $25, $pop103
-; NO-SIMD128-NEXT:    i32.lt_u $push45=, $pop44, $pop43
-; NO-SIMD128-NEXT:    i32.select $push46=, $9, $25, $pop45
-; NO-SIMD128-NEXT:    i32.store8 8($0), $pop46
-; NO-SIMD128-NEXT:    i32.const $push51=, 7
-; NO-SIMD128-NEXT:    i32.add $push52=, $0, $pop51
+; NO-SIMD128-NEXT:    i32.and $push104=, $25, $pop105
+; NO-SIMD128-NEXT:    local.tee $push103=, $16=, $pop104
+; NO-SIMD128-NEXT:    i32.lt_u $push29=, $32, $16
+; NO-SIMD128-NEXT:    i32.select $push30=, $pop106, $pop103, $pop29
+; NO-SIMD128-NEXT:    i32.store8 8($0), $pop30
+; NO-SIMD128-NEXT:    i32.const $push33=, 7
+; NO-SIMD128-NEXT:    i32.add $push34=, $0, $pop33
 ; NO-SIMD128-NEXT:    i32.const $push102=, 255
-; NO-SIMD128-NEXT:    i32.and $push48=, $8, $pop102
-; NO-SIMD128-NEXT:    i32.const $push101=, 255
-; NO-SIMD128-NEXT:    i32.and $push47=, $24, $pop101
-; NO-SIMD128-NEXT:    i32.lt_u $push49=, $pop48, $pop47
-; NO-SIMD128-NEXT:    i32.select $push50=, $8, $24, $pop49
-; NO-SIMD128-NEXT:    i32.store8 0($pop52), $pop50
-; NO-SIMD128-NEXT:    i32.const $push57=, 6
-; NO-SIMD128-NEXT:    i32.add $push58=, $0, $pop57
-; NO-SIMD128-NEXT:    i32.const $push100=, 255
-; NO-SIMD128-NEXT:    i32.and $push54=, $7, $pop100
+; NO-SIMD128-NEXT:    i32.and $push101=, $8, $pop102
+; NO-SIMD128-NEXT:    local.tee $push100=, $32=, $pop101
 ; NO-SIMD128-NEXT:    i32.const $push99=, 255
-; NO-SIMD128-NEXT:    i32.and $push53=, $23, $pop99
-; NO-SIMD128-NEXT:    i32.lt_u $push55=, $pop54, $pop53
-; NO-SIMD128-NEXT:    i32.select $push56=, $7, $23, $pop55
-; NO-SIMD128-NEXT:    i32.store8 0($pop58), $pop56
-; NO-SIMD128-NEXT:    i32.const $push63=, 5
-; NO-SIMD128-NEXT:    i32.add $push64=, $0, $pop63
-; NO-SIMD128-NEXT:    i32.const $push98=, 255
-; NO-SIMD128-NEXT:    i32.and $push60=, $6, $pop98
-; NO-SIMD128-NEXT:    i32.const $push97=, 255
-; NO-SIMD128-NEXT:    i32.and $push59=, $22, $pop97
-; NO-SIMD128-NEXT:    i32.lt_u $push61=, $pop60, $pop59
-; NO-SIMD128-NEXT:    i32.select $push62=, $6, $22, $pop61
-; NO-SIMD128-NEXT:    i32.store8 0($pop64), $pop62
+; NO-SIMD128-NEXT:    i32.and $push98=, $24, $pop99
+; NO-SIMD128-NEXT:    local.tee $push97=, $16=, $pop98
+; NO-SIMD128-NEXT:    i32.lt_u $push31=, $32, $16
+; NO-SIMD128-NEXT:    i32.select $push32=, $pop100, $pop97, $pop31
+; NO-SIMD128-NEXT:    i32.store8 0($pop34), $pop32
+; NO-SIMD128-NEXT:    i32.const $push37=, 6
+; NO-SIMD128-NEXT:    i32.add $push38=, $0, $pop37
 ; NO-SIMD128-NEXT:    i32.const $push96=, 255
-; NO-SIMD128-NEXT:    i32.and $push66=, $5, $pop96
-; NO-SIMD128-NEXT:    i32.const $push95=, 255
-; NO-SIMD128-NEXT:    i32.and $push65=, $21, $pop95
-; NO-SIMD128-NEXT:    i32.lt_u $push67=, $pop66, $pop65
-; NO-SIMD128-NEXT:    i32.select $push68=, $5, $21, $pop67
-; NO-SIMD128-NEXT:    i32.store8 4($0), $pop68
-; NO-SIMD128-NEXT:    i32.const $push73=, 3
-; NO-SIMD128-NEXT:    i32.add $push74=, $0, $pop73
-; NO-SIMD128-NEXT:    i32.const $push94=, 255
-; NO-SIMD128-NEXT:    i32.and $push70=, $4, $pop94
+; NO-SIMD128-NEXT:    i32.and $push95=, $7, $pop96
+; NO-SIMD128-NEXT:    local.tee $push94=, $32=, $pop95
 ; NO-SIMD128-NEXT:    i32.const $push93=, 255
-; NO-SIMD128-NEXT:    i32.and $push69=, $20, $pop93
-; NO-SIMD128-NEXT:    i32.lt_u $push71=, $pop70, $pop69
-; NO-SIMD128-NEXT:    i32.select $push72=, $4, $20, $pop71
-; NO-SIMD128-NEXT:    i32.store8 0($pop74), $pop72
-; NO-SIMD128-NEXT:    i32.const $push92=, 255
-; NO-SIMD128-NEXT:    i32.and $push76=, $3, $pop92
-; NO-SIMD128-NEXT:    i32.const $push91=, 255
-; NO-SIMD128-NEXT:    i32.and $push75=, $19, $pop91
-; NO-SIMD128-NEXT:    i32.lt_u $push77=, $pop76, $pop75
-; NO-SIMD128-NEXT:    i32.select $push78=, $3, $19, $pop77
-; NO-SIMD128-NEXT:    i32.store8 2($0), $pop78
+; NO-SIMD128-NEXT:    i32.and $push92=, $23, $pop93
+; NO-SIMD128-NEXT:    local.tee $push91=, $16=, $pop92
+; NO-SIMD128-NEXT:    i32.lt_u $push35=, $32, $16
+; NO-SIMD128-NEXT:    i32.select $push36=, $pop94, $pop91, $pop35
+; NO-SIMD128-NEXT:    i32.store8 0($pop38), $pop36
+; NO-SIMD128-NEXT:    i32.const $push41=, 5
+; NO-SIMD128-NEXT:    i32.add $push42=, $0, $pop41
 ; NO-SIMD128-NEXT:    i32.const $push90=, 255
-; NO-SIMD128-NEXT:    i32.and $push80=, $2, $pop90
-; NO-SIMD128-NEXT:    i32.const $push89=, 255
-; NO-SIMD128-NEXT:    i32.and $push79=, $18, $pop89
-; NO-SIMD128-NEXT:    i32.lt_u $push81=, $pop80, $pop79
-; NO-SIMD128-NEXT:    i32.select $push82=, $2, $18, $pop81
-; NO-SIMD128-NEXT:    i32.store8 1($0), $pop82
-; NO-SIMD128-NEXT:    i32.const $push88=, 255
-; NO-SIMD128-NEXT:    i32.and $push84=, $1, $pop88
+; NO-SIMD128-NEXT:    i32.and $push89=, $6, $pop90
+; NO-SIMD128-NEXT:    local.tee $push88=, $32=, $pop89
 ; NO-SIMD128-NEXT:    i32.const $push87=, 255
-; NO-SIMD128-NEXT:    i32.and $push83=, $17, $pop87
-; NO-SIMD128-NEXT:    i32.lt_u $push85=, $pop84, $pop83
-; NO-SIMD128-NEXT:    i32.select $push86=, $1, $17, $pop85
-; NO-SIMD128-NEXT:    i32.store8 0($0), $pop86
+; NO-SIMD128-NEXT:    i32.and $push86=, $22, $pop87
+; NO-SIMD128-NEXT:    local.tee $push85=, $16=, $pop86
+; NO-SIMD128-NEXT:    i32.lt_u $push39=, $32, $16
+; NO-SIMD128-NEXT:    i32.select $push40=, $pop88, $pop85, $pop39
+; NO-SIMD128-NEXT:    i32.store8 0($pop42), $pop40
+; NO-SIMD128-NEXT:    i32.const $push84=, 255
+; NO-SIMD128-NEXT:    i32.and $push83=, $5, $pop84
+; NO-SIMD128-NEXT:    local.tee $push82=, $32=, $pop83
+; NO-SIMD128-NEXT:    i32.const $push81=, 255
+; NO-SIMD128-NEXT:    i32.and $push80=, $21, $pop81
+; NO-SIMD128-NEXT:    local.tee $push79=, $16=, $pop80
+; NO-SIMD128-NEXT:    i32.lt_u $push43=, $32, $16
+; NO-SIMD128-NEXT:    i32.select $push44=, $pop82, $pop79, $pop43
+; NO-SIMD128-NEXT:    i32.store8 4($0), $pop44
+; NO-SIMD128-NEXT:    i32.const $push47=, 3
+; NO-SIMD128-NEXT:    i32.add $push48=, $0, $pop47
+; NO-SIMD128-NEXT:    i32.const $push78=, 255
+; NO-SIMD128-NEXT:    i32.and $push77=, $4, $pop78
+; NO-SIMD128-NEXT:    local.tee $push76=, $32=, $pop77
+; NO-SIMD128-NEXT:    i32.const $push75=, 255
+; NO-SIMD128-NEXT:    i32.and $push74=, $20, $pop75
+; NO-SIMD128-NEXT:    local.tee $push73=, $16=, $pop74
+; NO-SIMD128-NEXT:    i32.lt_u $push45=, $32, $16
+; NO-SIMD128-NEXT:    i32.select $push46=, $pop76, $pop73, $pop45
+; NO-SIMD128-NEXT:    i32.store8 0($pop48), $pop46
+; NO-SIMD128-NEXT:    i32.const $push72=, 255
+; NO-SIMD128-NEXT:    i32.and $push71=, $3, $pop72
+; NO-SIMD128-NEXT:    local.tee $push70=, $32=, $pop71
+; NO-SIMD128-NEXT:    i32.const $push69=, 255
+; NO-SIMD128-NEXT:    i32.and $push68=, $19, $pop69
+; NO-SIMD128-NEXT:    local.tee $push67=, $16=, $pop68
+; NO-SIMD128-NEXT:    i32.lt_u $push49=, $32, $16
+; NO-SIMD128-NEXT:    i32.select $push50=, $pop70, $pop67, $pop49
+; NO-SIMD128-NEXT:    i32.store8 2($0), $pop50
+; NO-SIMD128-NEXT:    i32.const $push66=, 255
+; NO-SIMD128-NEXT:    i32.and $push65=, $2, $pop66
+; NO-SIMD128-NEXT:    local.tee $push64=, $32=, $pop65
+; NO-SIMD128-NEXT:    i32.const $push63=, 255
+; NO-SIMD128-NEXT:    i32.and $push62=, $18, $pop63
+; NO-SIMD128-NEXT:    local.tee $push61=, $16=, $pop62
+; NO-SIMD128-NEXT:    i32.lt_u $push51=, $32, $16
+; NO-SIMD128-NEXT:    i32.select $push52=, $pop64, $pop61, $pop51
+; NO-SIMD128-NEXT:    i32.store8 1($0), $pop52
+; NO-SIMD128-NEXT:    i32.const $push60=, 255
+; NO-SIMD128-NEXT:    i32.and $push59=, $1, $pop60
+; NO-SIMD128-NEXT:    local.tee $push58=, $32=, $pop59
+; NO-SIMD128-NEXT:    i32.const $push57=, 255
+; NO-SIMD128-NEXT:    i32.and $push56=, $17, $pop57
+; NO-SIMD128-NEXT:    local.tee $push55=, $16=, $pop56
+; NO-SIMD128-NEXT:    i32.lt_u $push53=, $32, $16
+; NO-SIMD128-NEXT:    i32.select $push54=, $pop58, $pop55, $pop53
+; NO-SIMD128-NEXT:    i32.store8 0($0), $pop54
 ; NO-SIMD128-NEXT:    return
 ;
 ; NO-SIMD128-FAST-LABEL: min_u_v16i8:
 ; NO-SIMD128-FAST:         .functype min_u_v16i8 (i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32) -> ()
 ; NO-SIMD128-FAST-NEXT:  # %bb.0:
 ; NO-SIMD128-FAST-NEXT:    i32.const $push0=, 255
-; NO-SIMD128-FAST-NEXT:    i32.and $push2=, $1, $pop0
+; NO-SIMD128-FAST-NEXT:    i32.and $push149=, $1, $pop0
+; NO-SIMD128-FAST-NEXT:    local.tee $push148=, $1=, $pop149
+; NO-SIMD128-FAST-NEXT:    i32.const $push147=, 255
+; NO-SIMD128-FAST-NEXT:    i32.and $push146=, $17, $pop147
+; NO-SIMD128-FAST-NEXT:    local.tee $push145=, $17=, $pop146
+; NO-SIMD128-FAST-NEXT:    i32.lt_u $push1=, $1, $17
+; NO-SIMD128-FAST-NEXT:    i32.select $push2=, $pop148, $pop145, $pop1
+; NO-SIMD128-FAST-NEXT:    i32.store8 0($0), $pop2
+; NO-SIMD128-FAST-NEXT:    i32.const $push144=, 255
+; NO-SIMD128-FAST-NEXT:    i32.and $push143=, $2, $pop144
+; NO-SIMD128-FAST-NEXT:    local.tee $push142=, $17=, $pop143
+; NO-SIMD128-FAST-NEXT:    i32.const $push141=, 255
+; NO-SIMD128-FAST-NEXT:    i32.and $push140=, $18, $pop141
+; NO-SIMD128-FAST-NEXT:    local.tee $push139=, $18=, $pop140
+; NO-SIMD128-FAST-NEXT:    i32.lt_u $push3=, $17, $18
+; NO-SIMD128-FAST-NEXT:    i32.select $push4=, $pop142, $pop139, $pop3
+; NO-SIMD128-FAST-NEXT:    i32.store8 1($0), $pop4
+; NO-SIMD128-FAST-NEXT:    i32.const $push138=, 255
+; NO-SIMD128-FAST-NEXT:    i32.and $push137=, $3, $pop138
+; NO-SIMD128-FAST-NEXT:    local.tee $push136=, $18=, $pop137
+; NO-SIMD128-FAST-NEXT:    i32.const $push135=, 255
+; NO-SIMD128-FAST-NEXT:    i32.and $push134=, $19, $pop135
+; NO-SIMD128-FAST-NEXT:    local.tee $push133=, $19=, $pop134
+; NO-SIMD128-FAST-NEXT:    i32.lt_u $push5=, $18, $19
+; NO-SIMD128-FAST-NEXT:    i32.select $push6=, $pop136, $pop133, $pop5
+; NO-SIMD128-FAST-NEXT:    i32.store8 2($0), $pop6
+; NO-SIMD128-FAST-NEXT:    i32.const $push9=, 3
+; NO-SIMD128-FAST-NEXT:    i32.add $push10=, $0, $pop9
+; NO-SIMD128-FAST-NEXT:    i32.const $push132=, 255
+; NO-SIMD128-FAST-NEXT:    i32.and $push131=, $4, $pop132
+; NO-SIMD128-FAST-NEXT:    local.tee $push130=, $19=, $pop131
+; NO-SIMD128-FAST-NEXT:    i32.const $push129=, 255
+; NO-SIMD128-FAST-NEXT:    i32.and $push128=, $20, $pop129
+; NO-SIMD128-FAST-NEXT:    local.tee $push127=, $20=, $pop128
+; NO-SIMD128-FAST-NEXT:    i32.lt_u $push7=, $19, $20
+; NO-SIMD128-FAST-NEXT:    i32.select $push8=, $pop130, $pop127, $pop7
+; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop10), $pop8
+; NO-SIMD128-FAST-NEXT:    i32.const $push126=, 255
+; NO-SIMD128-FAST-NEXT:    i32.and $push125=, $5, $pop126
+; NO-SIMD128-FAST-NEXT:    local.tee $push124=, $20=, $pop125
+; NO-SIMD128-FAST-NEXT:    i32.const $push123=, 255
+; NO-SIMD128-FAST-NEXT:    i32.and $push122=, $21, $pop123
+; NO-SIMD128-FAST-NEXT:    local.tee $push121=, $21=, $pop122
+; NO-SIMD128-FAST-NEXT:    i32.lt_u $push11=, $20, $21
+; NO-SIMD128-FAST-NEXT:    i32.select $push12=, $pop124, $pop121, $pop11
+; NO-SIMD128-FAST-NEXT:    i32.store8 4($0), $pop12
+; NO-SIMD128-FAST-NEXT:    i32.const $push15=, 5
+; NO-SIMD128-FAST-NEXT:    i32.add $push16=, $0, $pop15
+; NO-SIMD128-FAST-NEXT:    i32.const $push120=, 255
+; NO-SIMD128-FAST-NEXT:    i32.and $push119=, $6, $pop120
+; NO-SIMD128-FAST-NEXT:    local.tee $push118=, $21=, $pop119
 ; NO-SIMD128-FAST-NEXT:    i32.const $push117=, 255
-; NO-SIMD128-FAST-NEXT:    i32.and $push1=, $17, $pop117
-; NO-SIMD128-FAST-NEXT:    i32.lt_u $push3=, $pop2, $pop1
-; NO-SIMD128-FAST-NEXT:    i32.select $push4=, $1, $17, $pop3
-; NO-SIMD128-FAST-NEXT:    i32.store8 0($0), $pop4
-; NO-SIMD128-FAST-NEXT:    i32.const $push116=, 255
-; NO-SIMD128-FAST-NEXT:    i32.and $push6=, $2, $pop116
-; NO-SIMD128-FAST-NEXT:    i32.const $push115=, 255
-; NO-SIMD128-FAST-NEXT:    i32.and $push5=, $18, $pop115
-; NO-SIMD128-FAST-NEXT:    i32.lt_u $push7=, $pop6, $pop5
-; NO-SIMD128-FAST-NEXT:    i32.select $push8=, $2, $18, $pop7
-; NO-SIMD128-FAST-NEXT:    i32.store8 1($0), $pop8
+; NO-SIMD128-FAST-NEXT:    i32.and $push116=, $22, $pop117
+; NO-SIMD128-FAST-NEXT:    local.tee $push115=, $22=, $pop116
+; NO-SIMD128-FAST-NEXT:    i32.lt_u $push13=, $21, $22
+; NO-SIMD128-FAST-NEXT:    i32.select $push14=, $pop118, $pop115, $pop13
+; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop16), $pop14
+; NO-SIMD128-FAST-NEXT:    i32.const $push19=, 6
+; NO-SIMD128-FAST-NEXT:    i32.add $push20=, $0, $pop19
 ; NO-SIMD128-FAST-NEXT:    i32.const $push114=, 255
-; NO-SIMD128-FAST-NEXT:    i32.and $push10=, $3, $pop114
-; NO-SIMD128-FAST-NEXT:    i32.const $push113=, 255
-; NO-SIMD128-FAST-NEXT:    i32.and $push9=, $19, $pop113
-; NO-SIMD128-FAST-NEXT:    i32.lt_u $push11=, $pop10, $pop9
-; NO-SIMD128-FAST-NEXT:    i32.select $push12=, $3, $19, $pop11
-; NO-SIMD128-FAST-NEXT:    i32.store8 2($0), $pop12
-; NO-SIMD128-FAST-NEXT:    i32.const $push17=, 3
-; NO-SIMD128-FAST-NEXT:    i32.add $push18=, $0, $pop17
-; NO-SIMD128-FAST-NEXT:    i32.const $push112=, 255
-; NO-SIMD128-FAST-NEXT:    i32.and $push14=, $4, $pop112
+; NO-SIMD128-FAST-NEXT:    i32.and $push113=, $7, $pop114
+; NO-SIMD128-FAST-NEXT:    local.tee $push112=, $22=, $pop113
 ; NO-SIMD128-FAST-NEXT:    i32.const $push111=, 255
-; NO-SIMD128-FAST-NEXT:    i32.and $push13=, $20, $pop111
-; NO-SIMD128-FAST-NEXT:    i32.lt_u $push15=, $pop14, $pop13
-; NO-SIMD128-FAST-NEXT:    i32.select $push16=, $4, $20, $pop15
-; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop18), $pop16
-; NO-SIMD128-FAST-NEXT:    i32.const $push110=, 255
-; NO-SIMD128-FAST-NEXT:    i32.and $push20=, $5, $pop110
-; NO-SIMD128-FAST-NEXT:    i32.const $push109=, 255
-; NO-SIMD128-FAST-NEXT:    i32.and $push19=, $21, $pop109
-; NO-SIMD128-FAST-NEXT:    i32.lt_u $push21=, $pop20, $pop19
-; NO-SIMD128-FAST-NEXT:    i32.select $push22=, $5, $21, $pop21
-; NO-SIMD128-FAST-NEXT:    i32.store8 4($0), $pop22
-; NO-SIMD128-FAST-NEXT:    i32.const $push27=, 5
-; NO-SIMD128-FAST-NEXT:    i32.add $push28=, $0, $pop27
+; NO-SIMD128-FAST-NEXT:    i32.and $push110=, $23, $pop111
+; NO-SIMD128-FAST-NEXT:    local.tee $push109=, $23=, $pop110
+; NO-SIMD128-FAST-NEXT:    i32.lt_u $push17=, $22, $23
+; NO-SIMD128-FAST-NEXT:    i32.select $push18=, $pop112, $pop109, $pop17
+; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop20), $pop18
+; NO-SIMD128-FAST-NEXT:    i32.const $push23=, 7
+; NO-SIMD128-FAST-NEXT:    i32.add $push24=, $0, $pop23
 ; NO-SIMD128-FAST-NEXT:    i32.const $push108=, 255
-; NO-SIMD128-FAST-NEXT:    i32.and $push24=, $6, $pop108
-; NO-SIMD128-FAST-NEXT:    i32.const $push107=, 255
-; NO-SIMD128-FAST-NEXT:    i32.and $push23=, $22, $pop107
-; NO-SIMD128-FAST-NEXT:    i32.lt_u $push25=, $pop24, $pop23
-; NO-SIMD128-FAST-NEXT:    i32.select $push26=, $6, $22, $pop25
-; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop28), $pop26
-; NO-SIMD128-FAST-NEXT:    i32.const $push33=, 6
-; NO-SIMD128-FAST-NEXT:    i32.add $push34=, $0, $pop33
-; NO-SIMD128-FAST-NEXT:    i32.const $push106=, 255
-; NO-SIMD128-FAST-NEXT:    i32.and $push30=, $7, $pop106
+; NO-SIMD128-FAST-NEXT:    i32.and $push107=, $8, $pop108
+; NO-SIMD128-FAST-NEXT:    local.tee $push106=, $23=, $pop107
 ; NO-SIMD128-FAST-NEXT:    i32.const $push105=, 255
-; NO-SIMD128-FAST-NEXT:    i32.and $push29=, $23, $pop105
-; NO-SIMD128-FAST-NEXT:    i32.lt_u $push31=, $pop30, $pop29
-; NO-SIMD128-FAST-NEXT:    i32.select $push32=, $7, $23, $pop31
-; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop34), $pop32
-; NO-SIMD128-FAST-NEXT:    i32.const $push39=, 7
-; NO-SIMD128-FAST-NEXT:    i32.add $push40=, $0, $pop39
-; NO-SIMD128-FAST-NEXT:    i32.const $push104=, 255
-; NO-SIMD128-FAST-NEXT:    i32.and $push36=, $8, $pop104
-; NO-SIMD128-FAST-NEXT:    i32.const $push103=, 255
-; NO-SIMD128-FAST-NEXT:    i32.and $push35=, $24, $pop103
-; NO-SIMD128-FAST-NEXT:    i32.lt_u $push37=, $pop36, $pop35
-; NO-SIMD128-FAST-NEXT:    i32.select $push38=, $8, $24, $pop37
-; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop40), $pop38
+; NO-SIMD128-FAST-NEXT:    i32.and $push104=, $24, $pop105
+; NO-SIMD128-FAST-NEXT:    local.tee $push103=, $24=, $pop104
+; NO-SIMD128-FAST-NEXT:    i32.lt_u $push21=, $23, $24
+; NO-SIMD128-FAST-NEXT:    i32.select $push22=, $pop106, $pop103, $pop21
+; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop24), $pop22
 ; NO-SIMD128-FAST-NEXT:    i32.const $push102=, 255
-; NO-SIMD128-FAST-NEXT:    i32.and $push42=, $9, $pop102
-; NO-SIMD128-FAST-NEXT:    i32.const $push101=, 255
-; NO-SIMD128-FAST-NEXT:    i32.and $push41=, $25, $pop101
-; NO-SIMD128-FAST-NEXT:    i32.lt_u $push43=, $pop42, $pop41
-; NO-SIMD128-FAST-NEXT:    i32.select $push44=, $9, $25, $pop43
-; NO-SIMD128-FAST-NEXT:    i32.store8 8($0), $pop44
-; NO-SIMD128-FAST-NEXT:    i32.const $push49=, 9
-; NO-SIMD128-FAST-NEXT:    i32.add $push50=, $0, $pop49
-; NO-SIMD128-FAST-NEXT:    i32.const $push100=, 255
-; NO-SIMD128-FAST-NEXT:    i32.and $push46=, $10, $pop100
+; NO-SIMD128-FAST-NEXT:    i32.and $push101=, $9, $pop102
+; NO-SIMD128-FAST-NEXT:    local.tee $push100=, $24=, $pop101
 ; NO-SIMD128-FAST-NEXT:    i32.const $push99=, 255
-; NO-SIMD128-FAST-NEXT:    i32.and $push45=, $26, $pop99
-; NO-SIMD128-FAST-NEXT:    i32.lt_u $push47=, $pop46, $pop45
-; NO-SIMD128-FAST-NEXT:    i32.select $push48=, $10, $26, $pop47
-; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop50), $pop48
-; NO-SIMD128-FAST-NEXT:    i32.const $push55=, 10
-; NO-SIMD128-FAST-NEXT:    i32.add $push56=, $0, $pop55
-; NO-SIMD128-FAST-NEXT:    i32.const $push98=, 255
-; NO-SIMD128-FAST-NEXT:    i32.and $push52=, $11, $pop98
-; NO-SIMD128-FAST-NEXT:    i32.const $push97=, 255
-; NO-SIMD128-FAST-NEXT:    i32.and $push51=, $27, $pop97
-; NO-SIMD128-FAST-NEXT:    i32.lt_u $push53=, $pop52, $pop51
-; NO-SIMD128-FAST-NEXT:    i32.select $push54=, $11, $27, $pop53
-; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop56), $pop54
-; NO-SIMD128-FAST-NEXT:    i32.const $push61=, 11
-; NO-SIMD128-FAST-NEXT:    i32.add $push62=, $0, $pop61
+; NO-SIMD128-FAST-NEXT:    i32.and $push98=, $25, $pop99
+; NO-SIMD128-FAST-NEXT:    local.tee $push97=, $25=, $pop98
+; NO-SIMD128-FAST-NEXT:    i32.lt_u $push25=, $24, $25
+; NO-SIMD128-FAST-NEXT:    i32.select $push26=, $pop100, $pop97, $pop25
+; NO-SIMD128-FAST-NEXT:    i32.store8 8($0), $pop26
+; NO-SIMD128-FAST-NEXT:    i32.const $push29=, 9
+; NO-SIMD128-FAST-NEXT:    i32.add $push30=, $0, $pop29
 ; NO-SIMD128-FAST-NEXT:    i32.const $push96=, 255
-; NO-SIMD128-FAST-NEXT:    i32.and $push58=, $12, $pop96
-; NO-SIMD128-FAST-NEXT:    i32.const $push95=, 255
-; NO-SIMD128-FAST-NEXT:    i32.and $push57=, $28, $pop95
-; NO-SIMD128-FAST-NEXT:    i32.lt_u $push59=, $pop58, $pop57
-; NO-SIMD128-FAST-NEXT:    i32.select $push60=, $12, $28, $pop59
-; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop62), $pop60
-; NO-SIMD128-FAST-NEXT:    i32.const $push67=, 12
-; NO-SIMD128-FAST-NEXT:    i32.add $push68=, $0, $pop67
-; NO-SIMD128-FAST-NEXT:    i32.const $push94=, 255
-; NO-SIMD128-FAST-NEXT:    i32.and $push64=, $13, $pop94
+; NO-SIMD128-FAST-NEXT:    i32.and $push95=, $10, $pop96
+; NO-SIMD128-FAST-NEXT:    local.tee $push94=, $25=, $pop95
 ; NO-SIMD128-FAST-NEXT:    i32.const $push93=, 255
-; NO-SIMD128-FAST-NEXT:    i32.and $push63=, $29, $pop93
-; NO-SIMD128-FAST-NEXT:    i32.lt_u $push65=, $pop64, $pop63
-; NO-SIMD128-FAST-NEXT:    i32.select $push66=, $13, $29, $pop65
-; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop68), $pop66
-; NO-SIMD128-FAST-NEXT:    i32.const $push73=, 13
-; NO-SIMD128-FAST-NEXT:    i32.add $push74=, $0, $pop73
-; NO-SIMD128-FAST-NEXT:    i32.const $push92=, 255
-; NO-SIMD128-FAST-NEXT:    i32.and $push70=, $14, $pop92
-; NO-SIMD128-FAST-NEXT:    i32.const $push91=, 255
-; NO-SIMD128-FAST-NEXT:    i32.and $push69=, $30, $pop91
-; NO-SIMD128-FAST-NEXT:    i32.lt_u $push71=, $pop70, $pop69
-; NO-SIMD128-FAST-NEXT:    i32.select $push72=, $14, $30, $pop71
-; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop74), $pop72
-; NO-SIMD128-FAST-NEXT:    i32.const $push79=, 14
-; NO-SIMD128-FAST-NEXT:    i32.add $push80=, $0, $pop79
+; NO-SIMD128-FAST-NEXT:    i32.and $push92=, $26, $pop93
+; NO-SIMD128-FAST-NEXT:    local.tee $push91=, $26=, $pop92
+; NO-SIMD128-FAST-NEXT:    i32.lt_u $push27=, $25, $26
+; NO-SIMD128-FAST-NEXT:    i32.select $push28=, $pop94, $pop91, $pop27
+; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop30), $pop28
+; NO-SIMD128-FAST-NEXT:    i32.const $push33=, 10
+; NO-SIMD128-FAST-NEXT:    i32.add $push34=, $0, $pop33
 ; NO-SIMD128-FAST-NEXT:    i32.const $push90=, 255
-; NO-SIMD128-FAST-NEXT:    i32.and $push76=, $15, $pop90
-; NO-SIMD128-FAST-NEXT:    i32.const $push89=, 255
-; NO-SIMD128-FAST-NEXT:    i32.and $push75=, $31, $pop89
-; NO-SIMD128-FAST-NEXT:    i32.lt_u $push77=, $pop76, $pop75
-; NO-SIMD128-FAST-NEXT:    i32.select $push78=, $15, $31, $pop77
-; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop80), $pop78
-; NO-SIMD128-FAST-NEXT:    i32.const $push85=, 15
-; NO-SIMD128-FAST-NEXT:    i32.add $push86=, $0, $pop85
-; NO-SIMD128-FAST-NEXT:    i32.const $push88=, 255
-; NO-SIMD128-FAST-NEXT:    i32.and $push82=, $16, $pop88
+; NO-SIMD128-FAST-NEXT:    i32.and $push89=, $11, $pop90
+; NO-SIMD128-FAST-NEXT:    local.tee $push88=, $26=, $pop89
 ; NO-SIMD128-FAST-NEXT:    i32.const $push87=, 255
-; NO-SIMD128-FAST-NEXT:    i32.and $push81=, $32, $pop87
-; NO-SIMD128-FAST-NEXT:    i32.lt_u $push83=, $pop82, $pop81
-; NO-SIMD128-FAST-NEXT:    i32.select $push84=, $16, $32, $pop83
-; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop86), $pop84
+; NO-SIMD128-FAST-NEXT:    i32.and $push86=, $27, $pop87
+; NO-SIMD128-FAST-NEXT:    local.tee $push85=, $27=, $pop86
+; NO-SIMD128-FAST-NEXT:    i32.lt_u $push31=, $26, $27
+; NO-SIMD128-FAST-NEXT:    i32.select $push32=, $pop88, $pop85, $pop31
+; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop34), $pop32
+; NO-SIMD128-FAST-NEXT:    i32.const $push37=, 11
+; NO-SIMD128-FAST-NEXT:    i32.add $push38=, $0, $pop37
+; NO-SIMD128-FAST-NEXT:    i32.const $push84=, 255
+; NO-SIMD128-FAST-NEXT:    i32.and $push83=, $12, $pop84
+; NO-SIMD128-FAST-NEXT:    local.tee $push82=, $27=, $pop83
+; NO-SIMD128-FAST-NEXT:    i32.const $push81=, 255
+; NO-SIMD128-FAST-NEXT:    i32.and $push80=, $28, $pop81
+; NO-SIMD128-FAST-NEXT:    local.tee $push79=, $28=, $pop80
+; NO-SIMD128-FAST-NEXT:    i32.lt_u $push35=, $27, $28
+; NO-SIMD128-FAST-NEXT:    i32.select $push36=, $pop82, $pop79, $pop35
+; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop38), $pop36
+; NO-SIMD128-FAST-NEXT:    i32.const $push41=, 12
+; NO-SIMD128-FAST-NEXT:    i32.add $push42=, $0, $pop41
+; NO-SIMD128-FAST-NEXT:    i32.const $push78=, 255
+; NO-SIMD128-FAST-NEXT:    i32.and $push77=, $13, $pop78
+; NO-SIMD128-FAST-NEXT:    local.tee $push76=, $28=, $pop77
+; NO-SIMD128-FAST-NEXT:    i32.const $push75=, 255
+; NO-SIMD128-FAST-NEXT:    i32.and $push74=, $29, $pop75
+; NO-SIMD128-FAST-NEXT:    local.tee $push73=, $29=, $pop74
+; NO-SIMD128-FAST-NEXT:    i32.lt_u $push39=, $28, $29
+; NO-SIMD128-FAST-NEXT:    i32.select $push40=, $pop76, $pop73, $pop39
+; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop42), $pop40
+; NO-SIMD128-FAST-NEXT:    i32.const $push45=, 13
+; NO-SIMD128-FAST-NEXT:    i32.add $push46=, $0, $pop45
+; NO-SIMD128-FAST-NEXT:    i32.const $push72=, 255
+; NO-SIMD128-FAST-NEXT:    i32.and $push71=, $14, $pop72
+; NO-SIMD128-FAST-NEXT:    local.tee $push70=, $29=, $pop71
+; NO-SIMD128-FAST-NEXT:    i32.const $push69=, 255
+; NO-SIMD128-FAST-NEXT:    i32.and $push68=, $30, $pop69
+; NO-SIMD128-FAST-NEXT:    local.tee $push67=, $30=, $pop68
+; NO-SIMD128-FAST-NEXT:    i32.lt_u $push43=, $29, $30
+; NO-SIMD128-FAST-NEXT:    i32.select $push44=, $pop70, $pop67, $pop43
+; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop46), $pop44
+; NO-SIMD128-FAST-NEXT:    i32.const $push49=, 14
+; NO-SIMD128-FAST-NEXT:    i32.add $push50=, $0, $pop49
+; NO-SIMD128-FAST-NEXT:    i32.const $push66=, 255
+; NO-SIMD128-FAST-NEXT:    i32.and $push65=, $15, $pop66
+; NO-SIMD128-FAST-NEXT:    local.tee $push64=, $30=, $pop65
+; NO-SIMD128-FAST-NEXT:    i32.const $push63=, 255
+; NO-SIMD128-FAST-NEXT:    i32.and $push62=, $31, $pop63
+; NO-SIMD128-FAST-NEXT:    local.tee $push61=, $31=, $pop62
+; NO-SIMD128-FAST-NEXT:    i32.lt_u $push47=, $30, $31
+; NO-SIMD128-FAST-NEXT:    i32.select $push48=, $pop64, $pop61, $pop47
+; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop50), $pop48
+; NO-SIMD128-FAST-NEXT:    i32.const $push53=, 15
+; NO-SIMD128-FAST-NEXT:    i32.add $push54=, $0, $pop53
+; NO-SIMD128-FAST-NEXT:    i32.const $push60=, 255
+; NO-SIMD128-FAST-NEXT:    i32.and $push59=, $16, $pop60
+; NO-SIMD128-FAST-NEXT:    local.tee $push58=, $0=, $pop59
+; NO-SIMD128-FAST-NEXT:    i32.const $push57=, 255
+; NO-SIMD128-FAST-NEXT:    i32.and $push56=, $32, $pop57
+; NO-SIMD128-FAST-NEXT:    local.tee $push55=, $32=, $pop56
+; NO-SIMD128-FAST-NEXT:    i32.lt_u $push51=, $0, $32
+; NO-SIMD128-FAST-NEXT:    i32.select $push52=, $pop58, $pop55, $pop51
+; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop54), $pop52
 ; NO-SIMD128-FAST-NEXT:    return
   %c = icmp ult <16 x i8> %x, %y
   %a = select <16 x i1> %c, <16 x i8> %x, <16 x i8> %y
   ret <16 x i8> %a
 }
 
 define <16 x i8> @max_s_v16i8(<16 x i8> %x, <16 x i8> %y) {
 ; SIMD128-LABEL: max_s_v16i8:
 ; SIMD128:         .functype max_s_v16i8 (v128, v128) -> (v128)
 ; SIMD128-NEXT:  # %bb.0:
 ; SIMD128-NEXT:    i8x16.max_s $push0=, $0, $1
 ; SIMD128-NEXT:    return $pop0
 ;
 ; SIMD128-FAST-LABEL: max_s_v16i8:
 ; SIMD128-FAST:         .functype max_s_v16i8 (v128, v128) -> (v128)
 ; SIMD128-FAST-NEXT:  # %bb.0:
 ; SIMD128-FAST-NEXT:    i8x16.max_s $push0=, $0, $1
 ; SIMD128-FAST-NEXT:    return $pop0
 ;
 ; NO-SIMD128-LABEL: max_s_v16i8:
 ; NO-SIMD128:         .functype max_s_v16i8 (i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32) -> ()
 ; NO-SIMD128-NEXT:  # %bb.0:
-; NO-SIMD128-NEXT:    i32.const $push4=, 15
-; NO-SIMD128-NEXT:    i32.add $push5=, $0, $pop4
-; NO-SIMD128-NEXT:    i32.extend8_s $push1=, $16
-; NO-SIMD128-NEXT:    i32.extend8_s $push0=, $32
-; NO-SIMD128-NEXT:    i32.gt_s $push2=, $pop1, $pop0
-; NO-SIMD128-NEXT:    i32.select $push3=, $16, $32, $pop2
-; NO-SIMD128-NEXT:    i32.store8 0($pop5), $pop3
-; NO-SIMD128-NEXT:    i32.const $push10=, 14
+; NO-SIMD128-NEXT:    i32.const $push2=, 15
+; NO-SIMD128-NEXT:    i32.add $push3=, $0, $pop2
+; NO-SIMD128-NEXT:    i32.extend8_s $push117=, $16
+; NO-SIMD128-NEXT:    local.tee $push116=, $16=, $pop117
+; NO-SIMD128-NEXT:    i32.extend8_s $push115=, $32
+; NO-SIMD128-NEXT:    local.tee $push114=, $32=, $pop115
+; NO-SIMD128-NEXT:    i32.gt_s $push0=, $16, $32
+; NO-SIMD128-NEXT:    i32.select $push1=, $pop116, $pop114, $pop0
+; NO-SIMD128-NEXT:    i32.store8 0($pop3), $pop1
+; NO-SIMD128-NEXT:    i32.const $push6=, 14
+; NO-SIMD128-NEXT:    i32.add $push7=, $0, $pop6
+; NO-SIMD128-NEXT:    i32.extend8_s $push113=, $15
+; NO-SIMD128-NEXT:    local.tee $push112=, $32=, $pop113
+; NO-SIMD128-NEXT:    i32.extend8_s $push111=, $31
+; NO-SIMD128-NEXT:    local.tee $push110=, $16=, $pop111
+; NO-SIMD128-NEXT:    i32.gt_s $push4=, $32, $16
+; NO-SIMD128-NEXT:    i32.select $push5=, $pop112, $pop110, $pop4
+; NO-SIMD128-NEXT:    i32.store8 0($pop7), $pop5
+; NO-SIMD128-NEXT:    i32.const $push10=, 13
 ; NO-SIMD128-NEXT:    i32.add $push11=, $0, $pop10
-; NO-SIMD128-NEXT:    i32.extend8_s $push7=, $15
-; NO-SIMD128-NEXT:    i32.extend8_s $push6=, $31
-; NO-SIMD128-NEXT:    i32.gt_s $push8=, $pop7, $pop6
-; NO-SIMD128-NEXT:    i32.select $push9=, $15, $31, $pop8
+; NO-SIMD128-NEXT:    i32.extend8_s $push109=, $14
+; NO-SIMD128-NEXT:    local.tee $push108=, $32=, $pop109
+; NO-SIMD128-NEXT:    i32.extend8_s $push107=, $30
+; NO-SIMD128-NEXT:    local.tee $push106=, $16=, $pop107
+; NO-SIMD128-NEXT:    i32.gt_s $push8=, $32, $16
+; NO-SIMD128-NEXT:    i32.select $push9=, $pop108, $pop106, $pop8
 ; NO-SIMD128-NEXT:    i32.store8 0($pop11), $pop9
-; NO-SIMD128-NEXT:    i32.const $push16=, 13
-; NO-SIMD128-NEXT:    i32.add $push17=, $0, $pop16
-; NO-SIMD128-NEXT:    i32.extend8_s $push13=, $14
-; NO-SIMD128-NEXT:    i32.extend8_s $push12=, $30
-; NO-SIMD128-NEXT:    i32.gt_s $push14=, $pop13, $pop12
-; NO-SIMD128-NEXT:    i32.select $push15=, $14, $30, $pop14
-; NO-SIMD128-NEXT:    i32.store8 0($pop17), $pop15
-; NO-SIMD128-NEXT:    i32.const $push22=, 12
+; NO-SIMD128-NEXT:    i32.const $push14=, 12
+; NO-SIMD128-NEXT:    i32.add $push15=, $0, $pop14
+; NO-SIMD128-NEXT:    i32.extend8_s $push105=, $13
+; NO-SIMD128-NEXT:    local.tee $push104=, $32=, $pop105
+; NO-SIMD128-NEXT:    i32.extend8_s $push103=, $29
+; NO-SIMD128-NEXT:    local.tee $push102=, $16=, $pop103
+; NO-SIMD128-NEXT:    i32.gt_s $push12=, $32, $16
+; NO-SIMD128-NEXT:    i32.select $push13=, $pop104, $pop102, $pop12
+; NO-SIMD128-NEXT:    i32.store8 0($pop15), $pop13
+; NO-SIMD128-NEXT:    i32.const $push18=, 11
+; NO-SIMD128-NEXT:    i32.add $push19=, $0, $pop18
+; NO-SIMD128-NEXT:    i32.extend8_s $push101=, $12
+; NO-SIMD128-NEXT:    local.tee $push100=, $32=, $pop101
+; NO-SIMD128-NEXT:    i32.extend8_s $push99=, $28
+; NO-SIMD128-NEXT:    local.tee $push98=, $16=, $pop99
+; NO-SIMD128-NEXT:    i32.gt_s $push16=, $32, $16
+; NO-SIMD128-NEXT:    i32.select $push17=, $pop100, $pop98, $pop16
+; NO-SIMD128-NEXT:    i32.store8 0($pop19), $pop17
+; NO-SIMD128-NEXT:    i32.const $push22=, 10
 ; NO-SIMD128-NEXT:    i32.add $push23=, $0, $pop22
-; NO-SIMD128-NEXT:    i32.extend8_s $push19=, $13
-; NO-SIMD128-NEXT:    i32.extend8_s $push18=, $29
-; NO-SIMD128-NEXT:    i32.gt_s $push20=, $pop19, $pop18
-; NO-SIMD128-NEXT:    i32.select $push21=, $13, $29, $pop20
+; NO-SIMD128-NEXT:    i32.extend8_s $push97=, $11
+; NO-SIMD128-NEXT:    local.tee $push96=, $32=, $pop97
+; NO-SIMD128-NEXT:    i32.extend8_s $push95=, $27
+; NO-SIMD128-NEXT:    local.tee $push94=, $16=, $pop95
+; NO-SIMD128-NEXT:    i32.gt_s $push20=, $32, $16
+; NO-SIMD128-NEXT:    i32.select $push21=, $pop96, $pop94, $pop20
 ; NO-SIMD128-NEXT:    i32.store8 0($pop23), $pop21
-; NO-SIMD128-NEXT:    i32.const $push28=, 11
-; NO-SIMD128-NEXT:    i32.add $push29=, $0, $pop28
-; NO-SIMD128-NEXT:    i32.extend8_s $push25=, $12
-; NO-SIMD128-NEXT:    i32.extend8_s $push24=, $28
-; NO-SIMD128-NEXT:    i32.gt_s $push26=, $pop25, $pop24
-; NO-SIMD128-NEXT:    i32.select $push27=, $12, $28, $pop26
-; NO-SIMD128-NEXT:    i32.store8 0($pop29), $pop27
-; NO-SIMD128-NEXT:    i32.const $push34=, 10
-; NO-SIMD128-NEXT:    i32.add $push35=, $0, $pop34
-; NO-SIMD128-NEXT:    i32.extend8_s $push31=, $11
-; NO-SIMD128-NEXT:    i32.extend8_s $push30=, $27
-; NO-SIMD128-NEXT:    i32.gt_s $push32=, $pop31, $pop30
-; NO-SIMD128-NEXT:    i32.select $push33=, $11, $27, $pop32
-; NO-SIMD128-NEXT:    i32.store8 0($pop35), $pop33
-; NO-SIMD128-NEXT:    i32.const $push40=, 9
+; NO-SIMD128-NEXT:    i32.const $push26=, 9
+; NO-SIMD128-NEXT:    i32.add $push27=, $0, $pop26
+; NO-SIMD128-NEXT:    i32.extend8_s $push93=, $10
+; NO-SIMD128-NEXT:    local.tee $push92=, $32=, $pop93
+; NO-SIMD128-NEXT:    i32.extend8_s $push91=, $26
+; NO-SIMD128-NEXT:    local.tee $push90=, $16=, $pop91
+; NO-SIMD128-NEXT:    i32.gt_s $push24=, $32, $16
+; NO-SIMD128-NEXT:    i32.select $push25=, $pop92, $pop90, $pop24
+; NO-SIMD128-NEXT:    i32.store8 0($pop27), $pop25
+; NO-SIMD128-NEXT:    i32.extend8_s $push89=, $9
+; NO-SIMD128-NEXT:    local.tee $push88=, $32=, $pop89
+; NO-SIMD128-NEXT:    i32.extend8_s $push87=, $25
+; NO-SIMD128-NEXT:    local.tee $push86=, $16=, $pop87
+; NO-SIMD128-NEXT:    i32.gt_s $push28=, $32, $16
+; NO-SIMD128-NEXT:    i32.select $push29=, $pop88, $pop86, $pop28
+; NO-SIMD128-NEXT:    i32.store8 8($0), $pop29
+; NO-SIMD128-NEXT:    i32.const $push32=, 7
+; NO-SIMD128-NEXT:    i32.add $push33=, $0, $pop32
+; NO-SIMD128-NEXT:    i32.extend8_s $push85=, $8
+; NO-SIMD128-NEXT:    local.tee $push84=, $32=, $pop85
+; NO-SIMD128-NEXT:    i32.extend8_s $push83=, $24
+; NO-SIMD128-NEXT:    local.tee $push82=, $16=, $pop83
+; NO-SIMD128-NEXT:    i32.gt_s $push30=, $32, $16
+; NO-SIMD128-NEXT:    i32.select $push31=, $pop84, $pop82, $pop30
+; NO-SIMD128-NEXT:    i32.store8 0($pop33), $pop31
+; NO-SIMD128-NEXT:    i32.const $push36=, 6
+; NO-SIMD128-NEXT:    i32.add $push37=, $0, $pop36
+; NO-SIMD128-NEXT:    i32.extend8_s $push81=, $7
+; NO-SIMD128-NEXT:    local.tee $push80=, $32=, $pop81
+; NO-SIMD128-NEXT:    i32.extend8_s $push79=, $23
+; NO-SIMD128-NEXT:    local.tee $push78=, $16=, $pop79
+; NO-SIMD128-NEXT:    i32.gt_s $push34=, $32, $16
+; NO-SIMD128-NEXT:    i32.select $push35=, $pop80, $pop78, $pop34
+; NO-SIMD128-NEXT:    i32.store8 0($pop37), $pop35
+; NO-SIMD128-NEXT:    i32.const $push40=, 5
 ; NO-SIMD128-NEXT:    i32.add $push41=, $0, $pop40
-; NO-SIMD128-NEXT:    i32.extend8_s $push37=, $10
-; NO-SIMD128-NEXT:    i32.extend8_s $push36=, $26
-; NO-SIMD128-NEXT:    i32.gt_s $push38=, $pop37, $pop36
-; NO-SIMD128-NEXT:    i32.select $push39=, $10, $26, $pop38
+; NO-SIMD128-NEXT:    i32.extend8_s $push77=, $6
+; NO-SIMD128-NEXT:    local.tee $push76=, $32=, $pop77
+; NO-SIMD128-NEXT:    i32.extend8_s $push75=, $22
+; NO-SIMD128-NEXT:    local.tee $push74=, $16=, $pop75
+; NO-SIMD128-NEXT:    i32.gt_s $push38=, $32, $16
+; NO-SIMD128-NEXT:    i32.select $push39=, $pop76, $pop74, $pop38
 ; NO-SIMD128-NEXT:    i32.store8 0($pop41), $pop39
-; NO-SIMD128-NEXT:    i32.extend8_s $push43=, $9
-; NO-SIMD128-NEXT:    i32.extend8_s $push42=, $25
-; NO-SIMD128-NEXT:    i32.gt_s $push44=, $pop43, $pop42
-; NO-SIMD128-NEXT:    i32.select $push45=, $9, $25, $pop44
-; NO-SIMD128-NEXT:    i32.store8 8($0), $pop45
-; NO-SIMD128-NEXT:    i32.const $push50=, 7
-; NO-SIMD128-NEXT:    i32.add $push51=, $0, $pop50
-; NO-SIMD128-NEXT:    i32.extend8_s $push47=, $8
-; NO-SIMD128-NEXT:    i32.extend8_s $push46=, $24
-; NO-SIMD128-NEXT:    i32.gt_s $push48=, $pop47, $pop46
-; NO-SIMD128-NEXT:    i32.select $push49=, $8, $24, $pop48
-; NO-SIMD128-NEXT:    i32.store8 0($pop51), $pop49
-; NO-SIMD128-NEXT:    i32.const $push56=, 6
-; NO-SIMD128-NEXT:    i32.add $push57=, $0, $pop56
-; NO-SIMD128-NEXT:    i32.extend8_s $push53=, $7
-; NO-SIMD128-NEXT:    i32.extend8_s $push52=, $23
-; NO-SIMD128-NEXT:    i32.gt_s $push54=, $pop53, $pop52
-; NO-SIMD128-NEXT:    i32.select $push55=, $7, $23, $pop54
-; NO-SIMD128-NEXT:    i32.store8 0($pop57), $pop55
-; NO-SIMD128-NEXT:    i32.const $push62=, 5
-; NO-SIMD128-NEXT:    i32.add $push63=, $0, $pop62
-; NO-SIMD128-NEXT:    i32.extend8_s $push59=, $6
-; NO-SIMD128-NEXT:    i32.extend8_s $push58=, $22
-; NO-SIMD128-NEXT:    i32.gt_s $push60=, $pop59, $pop58
-; NO-SIMD128-NEXT:    i32.select $push61=, $6, $22, $pop60
-; NO-SIMD128-NEXT:    i32.store8 0($pop63), $pop61
-; NO-SIMD128-NEXT:    i32.extend8_s $push65=, $5
-; NO-SIMD128-NEXT:    i32.extend8_s $push64=, $21
-; NO-SIMD128-NEXT:    i32.gt_s $push66=, $pop65, $pop64
-; NO-SIMD128-NEXT:    i32.select $push67=, $5, $21, $pop66
-; NO-SIMD128-NEXT:    i32.store8 4($0), $pop67
-; NO-SIMD128-NEXT:    i32.const $push72=, 3
-; NO-SIMD128-NEXT:    i32.add $push73=, $0, $pop72
+; NO-SIMD128-NEXT:    i32.extend8_s $push73=, $5
+; NO-SIMD128-NEXT:    local.tee $push72=, $32=, $pop73
+; NO-SIMD128-NEXT:    i32.extend8_s $push71=, $21
+; NO-SIMD128-NEXT:    local.tee $push70=, $16=, $pop71
+; NO-SIMD128-NEXT:    i32.gt_s $push42=, $32, $16
+; NO-SIMD128-NEXT:    i32.select $push43=, $pop72, $pop70, $pop42
+; NO-SIMD128-NEXT:    i32.store8 4($0), $pop43
+; NO-SIMD128-NEXT:    i32.const $push46=, 3
+; NO-SIMD128-NEXT:    i32.add $push47=, $0, $pop46
 ; NO-SIMD128-NEXT:    i32.extend8_s $push69=, $4
-; NO-SIMD128-NEXT:    i32.extend8_s $push68=, $20
-; NO-SIMD128-NEXT:    i32.gt_s $push70=, $pop69, $pop68
-; NO-SIMD128-NEXT:    i32.select $push71=, $4, $20, $pop70
-; NO-SIMD128-NEXT:    i32.store8 0($pop73), $pop71
-; NO-SIMD128-NEXT:    i32.extend8_s $push75=, $3
-; NO-SIMD128-NEXT:    i32.extend8_s $push74=, $19
-; NO-SIMD128-NEXT:    i32.gt_s $push76=, $pop75, $pop74
-; NO-SIMD128-NEXT:    i32.select $push77=, $3, $19, $pop76
-; NO-SIMD128-NEXT:    i32.store8 2($0), $pop77
-; NO-SIMD128-NEXT:    i32.extend8_s $push79=, $2
-; NO-SIMD128-NEXT:    i32.extend8_s $push78=, $18
-; NO-SIMD128-NEXT:    i32.gt_s $push80=, $pop79, $pop78
-; NO-SIMD128-NEXT:    i32.select $push81=, $2, $18, $pop80
-; NO-SIMD128-NEXT:    i32.store8 1($0), $pop81
-; NO-SIMD128-NEXT:    i32.extend8_s $push83=, $1
-; NO-SIMD128-NEXT:    i32.extend8_s $push82=, $17
-; NO-SIMD128-NEXT:    i32.gt_s $push84=, $pop83, $pop82
-; NO-SIMD128-NEXT:    i32.select $push85=, $1, $17, $pop84
-; NO-SIMD128-NEXT:    i32.store8 0($0), $pop85
+; NO-SIMD128-NEXT:    local.tee $push68=, $32=, $pop69
+; NO-SIMD128-NEXT:    i32.extend8_s $push67=, $20
+; NO-SIMD128-NEXT:    local.tee $push66=, $16=, $pop67
+; NO-SIMD128-NEXT:    i32.gt_s $push44=, $32, $16
+; NO-SIMD128-NEXT:    i32.select $push45=, $pop68, $pop66, $pop44
+; NO-SIMD128-NEXT:    i32.store8 0($pop47), $pop45
+; NO-SIMD128-NEXT:    i32.extend8_s $push65=, $3
+; NO-SIMD128-NEXT:    local.tee $push64=, $32=, $pop65
+; NO-SIMD128-NEXT:    i32.extend8_s $push63=, $19
+; NO-SIMD128-NEXT:    local.tee $push62=, $16=, $pop63
+; NO-SIMD128-NEXT:    i32.gt_s $push48=, $32, $16
+; NO-SIMD128-NEXT:    i32.select $push49=, $pop64, $pop62, $pop48
+; NO-SIMD128-NEXT:    i32.store8 2($0), $pop49
+; NO-SIMD128-NEXT:    i32.extend8_s $push61=, $2
+; NO-SIMD128-NEXT:    local.tee $push60=, $32=, $pop61
+; NO-SIMD128-NEXT:    i32.extend8_s $push59=, $18
+; NO-SIMD128-NEXT:    local.tee $push58=, $16=, $pop59
+; NO-SIMD128-NEXT:    i32.gt_s $push50=, $32, $16
+; NO-SIMD128-NEXT:    i32.select $push51=, $pop60, $pop58, $pop50
+; NO-SIMD128-NEXT:    i32.store8 1($0), $pop51
+; NO-SIMD128-NEXT:    i32.extend8_s $push57=, $1
+; NO-SIMD128-NEXT:    local.tee $push56=, $32=, $pop57
+; NO-SIMD128-NEXT:    i32.extend8_s $push55=, $17
+; NO-SIMD128-NEXT:    local.tee $push54=, $16=, $pop55
+; NO-SIMD128-NEXT:    i32.gt_s $push52=, $32, $16
+; NO-SIMD128-NEXT:    i32.select $push53=, $pop56, $pop54, $pop52
+; NO-SIMD128-NEXT:    i32.store8 0($0), $pop53
 ; NO-SIMD128-NEXT:    return
 ;
 ; NO-SIMD128-FAST-LABEL: max_s_v16i8:
 ; NO-SIMD128-FAST:         .functype max_s_v16i8 (i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32) -> ()
 ; NO-SIMD128-FAST-NEXT:  # %bb.0:
-; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push1=, $1
-; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push0=, $17
-; NO-SIMD128-FAST-NEXT:    i32.gt_s $push2=, $pop1, $pop0
-; NO-SIMD128-FAST-NEXT:    i32.select $push3=, $1, $17, $pop2
-; NO-SIMD128-FAST-NEXT:    i32.store8 0($0), $pop3
-; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push5=, $2
-; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push4=, $18
-; NO-SIMD128-FAST-NEXT:    i32.gt_s $push6=, $pop5, $pop4
-; NO-SIMD128-FAST-NEXT:    i32.select $push7=, $2, $18, $pop6
-; NO-SIMD128-FAST-NEXT:    i32.store8 1($0), $pop7
-; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push9=, $3
-; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push8=, $19
-; NO-SIMD128-FAST-NEXT:    i32.gt_s $push10=, $pop9, $pop8
-; NO-SIMD128-FAST-NEXT:    i32.select $push11=, $3, $19, $pop10
-; NO-SIMD128-FAST-NEXT:    i32.store8 2($0), $pop11
-; NO-SIMD128-FAST-NEXT:    i32.const $push16=, 3
-; NO-SIMD128-FAST-NEXT:    i32.add $push17=, $0, $pop16
-; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push13=, $4
-; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push12=, $20
-; NO-SIMD128-FAST-NEXT:    i32.gt_s $push14=, $pop13, $pop12
-; NO-SIMD128-FAST-NEXT:    i32.select $push15=, $4, $20, $pop14
-; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop17), $pop15
-; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push19=, $5
-; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push18=, $21
-; NO-SIMD128-FAST-NEXT:    i32.gt_s $push20=, $pop19, $pop18
-; NO-SIMD128-FAST-NEXT:    i32.select $push21=, $5, $21, $pop20
-; NO-SIMD128-FAST-NEXT:    i32.store8 4($0), $pop21
-; NO-SIMD128-FAST-NEXT:    i32.const $push26=, 5
-; NO-SIMD128-FAST-NEXT:    i32.add $push27=, $0, $pop26
-; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push23=, $6
-; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push22=, $22
-; NO-SIMD128-FAST-NEXT:    i32.gt_s $push24=, $pop23, $pop22
-; NO-SIMD128-FAST-NEXT:    i32.select $push25=, $6, $22, $pop24
-; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop27), $pop25
-; NO-SIMD128-FAST-NEXT:    i32.const $push32=, 6
+; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push117=, $1
+; NO-SIMD128-FAST-NEXT:    local.tee $push116=, $1=, $pop117
+; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push115=, $17
+; NO-SIMD128-FAST-NEXT:    local.tee $push114=, $17=, $pop115
+; NO-SIMD128-FAST-NEXT:    i32.gt_s $push0=, $1, $17
+; NO-SIMD128-FAST-NEXT:    i32.select $push1=, $pop116, $pop114, $pop0
+; NO-SIMD128-FAST-NEXT:    i32.store8 0($0), $pop1
+; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push113=, $2
+; NO-SIMD128-FAST-NEXT:    local.tee $push112=, $17=, $pop113
+; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push111=, $18
+; NO-SIMD128-FAST-NEXT:    local.tee $push110=, $18=, $pop111
+; NO-SIMD128-FAST-NEXT:    i32.gt_s $push2=, $17, $18
+; NO-SIMD128-FAST-NEXT:    i32.select $push3=, $pop112, $pop110, $pop2
+; NO-SIMD128-FAST-NEXT:    i32.store8 1($0), $pop3
+; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push109=, $3
+; NO-SIMD128-FAST-NEXT:    local.tee $push108=, $18=, $pop109
+; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push107=, $19
+; NO-SIMD128-FAST-NEXT:    local.tee $push106=, $19=, $pop107
+; NO-SIMD128-FAST-NEXT:    i32.gt_s $push4=, $18, $19
+; NO-SIMD128-FAST-NEXT:    i32.select $push5=, $pop108, $pop106, $pop4
+; NO-SIMD128-FAST-NEXT:    i32.store8 2($0), $pop5
+; NO-SIMD128-FAST-NEXT:    i32.const $push8=, 3
+; NO-SIMD128-FAST-NEXT:    i32.add $push9=, $0, $pop8
+; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push105=, $4
+; NO-SIMD128-FAST-NEXT:    local.tee $push104=, $19=, $pop105
+; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push103=, $20
+; NO-SIMD128-FAST-NEXT:    local.tee $push102=, $20=, $pop103
+; NO-SIMD128-FAST-NEXT:    i32.gt_s $push6=, $19, $20
+; NO-SIMD128-FAST-NEXT:    i32.select $push7=, $pop104, $pop102, $pop6
+; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop9), $pop7
+; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push101=, $5
+; NO-SIMD128-FAST-NEXT:    local.tee $push100=, $20=, $pop101
+; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push99=, $21
+; NO-SIMD128-FAST-NEXT:    local.tee $push98=, $21=, $pop99
+; NO-SIMD128-FAST-NEXT:    i32.gt_s $push10=, $20, $21
+; NO-SIMD128-FAST-NEXT:    i32.select $push11=, $pop100, $pop98, $pop10
+; NO-SIMD128-FAST-NEXT:    i32.store8 4($0), $pop11
+; NO-SIMD128-FAST-NEXT:    i32.const $push14=, 5
+; NO-SIMD128-FAST-NEXT:    i32.add $push15=, $0, $pop14
+; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push97=, $6
+; NO-SIMD128-FAST-NEXT:    local.tee $push96=, $21=, $pop97
+; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push95=, $22
+; NO-SIMD128-FAST-NEXT:    local.tee $push94=, $22=, $pop95
+; NO-SIMD128-FAST-NEXT:    i32.gt_s $push12=, $21, $22
+; NO-SIMD128-FAST-NEXT:    i32.select $push13=, $pop96, $pop94, $pop12
+; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop15), $pop13
+; NO-SIMD128-FAST-NEXT:    i32.const $push18=, 6
+; NO-SIMD128-FAST-NEXT:    i32.add $push19=, $0, $pop18
+; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push93=, $7
+; NO-SIMD128-FAST-NEXT:    local.tee $push92=, $22=, $pop93
+; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push91=, $23
+; NO-SIMD128-FAST-NEXT:    local.tee $push90=, $23=, $pop91
+; NO-SIMD128-FAST-NEXT:    i32.gt_s $push16=, $22, $23
+; NO-SIMD128-FAST-NEXT:    i32.select $push17=, $pop92, $pop90, $pop16
+; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop19), $pop17
+; NO-SIMD128-FAST-NEXT:    i32.const $push22=, 7
+; NO-SIMD128-FAST-NEXT:    i32.add $push23=, $0, $pop22
+; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push89=, $8
+; NO-SIMD128-FAST-NEXT:    local.tee $push88=, $23=, $pop89
+; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push87=, $24
+; NO-SIMD128-FAST-NEXT:    local.tee $push86=, $24=, $pop87
+; NO-SIMD128-FAST-NEXT:    i32.gt_s $push20=, $23, $24
+; NO-SIMD128-FAST-NEXT:    i32.select $push21=, $pop88, $pop86, $pop20
+; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop23), $pop21
+; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push85=, $9
+; NO-SIMD128-FAST-NEXT:    local.tee $push84=, $24=, $pop85
+; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push83=, $25
+; NO-SIMD128-FAST-NEXT:    local.tee $push82=, $25=, $pop83
+; NO-SIMD128-FAST-NEXT:    i32.gt_s $push24=, $24, $25
+; NO-SIMD128-FAST-NEXT:    i32.select $push25=, $pop84, $pop82, $pop24
+; NO-SIMD128-FAST-NEXT:    i32.store8 8($0), $pop25
+; NO-SIMD128-FAST-NEXT:    i32.const $push28=, 9
+; NO-SIMD128-FAST-NEXT:    i32.add $push29=, $0, $pop28
+; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push81=, $10
+; NO-SIMD128-FAST-NEXT:    local.tee $push80=, $25=, $pop81
+; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push79=, $26
+; NO-SIMD128-FAST-NEXT:    local.tee $push78=, $26=, $pop79
+; NO-SIMD128-FAST-NEXT:    i32.gt_s $push26=, $25, $26
+; NO-SIMD128-FAST-NEXT:    i32.select $push27=, $pop80, $pop78, $pop26
+; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop29), $pop27
+; NO-SIMD128-FAST-NEXT:    i32.const $push32=, 10
 ; NO-SIMD128-FAST-NEXT:    i32.add $push33=, $0, $pop32
-; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push29=, $7
-; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push28=, $23
-; NO-SIMD128-FAST-NEXT:    i32.gt_s $push30=, $pop29, $pop28
-; NO-SIMD128-FAST-NEXT:    i32.select $push31=, $7, $23, $pop30
+; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push77=, $11
+; NO-SIMD128-FAST-NEXT:    local.tee $push76=, $26=, $pop77
+; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push75=, $27
+; NO-SIMD128-FAST-NEXT:    local.tee $push74=, $27=, $pop75
+; NO-SIMD128-FAST-NEXT:    i32.gt_s $push30=, $26, $27
+; NO-SIMD128-FAST-NEXT:    i32.select $push31=, $pop76, $pop74, $pop30
 ; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop33), $pop31
-; NO-SIMD128-FAST-NEXT:    i32.const $push38=, 7
-; NO-SIMD128-FAST-NEXT:    i32.add $push39=, $0, $pop38
-; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push35=, $8
-; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push34=, $24
-; NO-SIMD128-FAST-NEXT:    i32.gt_s $push36=, $pop35, $pop34
-; NO-SIMD128-FAST-NEXT:    i32.select $push37=, $8, $24, $pop36
-; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop39), $pop37
-; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push41=, $9
-; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push40=, $25
-; NO-SIMD128-FAST-NEXT:    i32.gt_s $push42=, $pop41, $pop40
-; NO-SIMD128-FAST-NEXT:    i32.select $push43=, $9, $25, $pop42
-; NO-SIMD128-FAST-NEXT:    i32.store8 8($0), $pop43
-; NO-SIMD128-FAST-NEXT:    i32.const $push48=, 9
+; NO-SIMD128-FAST-NEXT:    i32.const $push36=, 11
+; NO-SIMD128-FAST-NEXT:    i32.add $push37=, $0, $pop36
+; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push73=, $12
+; NO-SIMD128-FAST-NEXT:    local.tee $push72=, $27=, $pop73
+; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push71=, $28
+; NO-SIMD128-FAST-NEXT:    local.tee $push70=, $28=, $pop71
+; NO-SIMD128-FAST-NEXT:    i32.gt_s $push34=, $27, $28
+; NO-SIMD128-FAST-NEXT:    i32.select $push35=, $pop72, $pop70, $pop34
+; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop37), $pop35
+; NO-SIMD128-FAST-NEXT:    i32.const $push40=, 12
+; NO-SIMD128-FAST-NEXT:    i32.add $push41=, $0, $pop40
+; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push69=, $13
+; NO-SIMD128-FAST-NEXT:    local.tee $push68=, $28=, $pop69
+; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push67=, $29
+; NO-SIMD128-FAST-NEXT:    local.tee $push66=, $29=, $pop67
+; NO-SIMD128-FAST-NEXT:    i32.gt_s $push38=, $28, $29
+; NO-SIMD128-FAST-NEXT:    i32.select $push39=, $pop68, $pop66, $pop38
+; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop41), $pop39
+; NO-SIMD128-FAST-NEXT:    i32.const $push44=, 13
+; NO-SIMD128-FAST-NEXT:    i32.add $push45=, $0, $pop44
+; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push65=, $14
+; NO-SIMD128-FAST-NEXT:    local.tee $push64=, $29=, $pop65
+; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push63=, $30
+; NO-SIMD128-FAST-NEXT:    local.tee $push62=, $30=, $pop63
+; NO-SIMD128-FAST-NEXT:    i32.gt_s $push42=, $29, $30
+; NO-SIMD128-FAST-NEXT:    i32.select $push43=, $pop64, $pop62, $pop42
+; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop45), $pop43
+; NO-SIMD128-FAST-NEXT:    i32.const $push48=, 14
 ; NO-SIMD128-FAST-NEXT:    i32.add $push49=, $0, $pop48
-; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push45=, $10
-; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push44=, $26
-; NO-SIMD128-FAST-NEXT:    i32.gt_s $push46=, $pop45, $pop44
-; NO-SIMD128-FAST-NEXT:    i32.select $push47=, $10, $26, $pop46
+; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push61=, $15
+; NO-SIMD128-FAST-NEXT:    local.tee $push60=, $30=, $pop61
+; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push59=, $31
+; NO-SIMD128-FAST-NEXT:    local.tee $push58=, $31=, $pop59
+; NO-SIMD128-FAST-NEXT:    i32.gt_s $push46=, $30, $31
+; NO-SIMD128-FAST-NEXT:    i32.select $push47=, $pop60, $pop58, $pop46
 ; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop49), $pop47
-; NO-SIMD128-FAST-NEXT:    i32.const $push54=, 10
-; NO-SIMD128-FAST-NEXT:    i32.add $push55=, $0, $pop54
-; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push51=, $11
-; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push50=, $27
-; NO-SIMD128-FAST-NEXT:    i32.gt_s $push52=, $pop51, $pop50
-; NO-SIMD128-FAST-NEXT:    i32.select $push53=, $11, $27, $pop52
-; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop55), $pop53
-; NO-SIMD128-FAST-NEXT:    i32.const $push60=, 11
-; NO-SIMD128-FAST-NEXT:    i32.add $push61=, $0, $pop60
-; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push57=, $12
-; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push56=, $28
-; NO-SIMD128-FAST-NEXT:    i32.gt_s $push58=, $pop57, $pop56
-; NO-SIMD128-FAST-NEXT:    i32.select $push59=, $12, $28, $pop58
-; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop61), $pop59
-; NO-SIMD128-FAST-NEXT:    i32.const $push66=, 12
-; NO-SIMD128-FAST-NEXT:    i32.add $push67=, $0, $pop66
-; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push63=, $13
-; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push62=, $29
-; NO-SIMD128-FAST-NEXT:    i32.gt_s $push64=, $pop63, $pop62
-; NO-SIMD128-FAST-NEXT:    i32.select $push65=, $13, $29, $pop64
-; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop67), $pop65
-; NO-SIMD128-FAST-NEXT:    i32.const $push72=, 13
-; NO-SIMD128-FAST-NEXT:    i32.add $push73=, $0, $pop72
-; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push69=, $14
-; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push68=, $30
-; NO-SIMD128-FAST-NEXT:    i32.gt_s $push70=, $pop69, $pop68
-; NO-SIMD128-FAST-NEXT:    i32.select $push71=, $14, $30, $pop70
-; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop73), $pop71
-; NO-SIMD128-FAST-NEXT:    i32.const $push78=, 14
-; NO-SIMD128-FAST-NEXT:    i32.add $push79=, $0, $pop78
-; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push75=, $15
-; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push74=, $31
-; NO-SIMD128-FAST-NEXT:    i32.gt_s $push76=, $pop75, $pop74
-; NO-SIMD128-FAST-NEXT:    i32.select $push77=, $15, $31, $pop76
-; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop79), $pop77
-; NO-SIMD128-FAST-NEXT:    i32.const $push84=, 15
-; NO-SIMD128-FAST-NEXT:    i32.add $push85=, $0, $pop84
-; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push81=, $16
-; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push80=, $32
-; NO-SIMD128-FAST-NEXT:    i32.gt_s $push82=, $pop81, $pop80
-; NO-SIMD128-FAST-NEXT:    i32.select $push83=, $16, $32, $pop82
-; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop85), $pop83
+; NO-SIMD128-FAST-NEXT:    i32.const $push52=, 15
+; NO-SIMD128-FAST-NEXT:    i32.add $push53=, $0, $pop52
+; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push57=, $16
+; NO-SIMD128-FAST-NEXT:    local.tee $push56=, $0=, $pop57
+; NO-SIMD128-FAST-NEXT:    i32.extend8_s $push55=, $32
+; NO-SIMD128-FAST-NEXT:    local.tee $push54=, $32=, $pop55
+; NO-SIMD128-FAST-NEXT:    i32.gt_s $push50=, $0, $32
+; NO-SIMD128-FAST-NEXT:    i32.select $push51=, $pop56, $pop54, $pop50
+; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop53), $pop51
 ; NO-SIMD128-FAST-NEXT:    return
   %c = icmp sgt <16 x i8> %x, %y
   %a = select <16 x i1> %c, <16 x i8> %x, <16 x i8> %y
   ret <16 x i8> %a
 }
 
 define <16 x i8> @max_u_v16i8(<16 x i8> %x, <16 x i8> %y) {
 ; SIMD128-LABEL: max_u_v16i8:
 ; SIMD128:         .functype max_u_v16i8 (v128, v128) -> (v128)
 ; SIMD128-NEXT:  # %bb.0:
 ; SIMD128-NEXT:    i8x16.max_u $push0=, $0, $1
 ; SIMD128-NEXT:    return $pop0
 ;
 ; SIMD128-FAST-LABEL: max_u_v16i8:
 ; SIMD128-FAST:         .functype max_u_v16i8 (v128, v128) -> (v128)
 ; SIMD128-FAST-NEXT:  # %bb.0:
 ; SIMD128-FAST-NEXT:    i8x16.max_u $push0=, $0, $1
 ; SIMD128-FAST-NEXT:    return $pop0
 ;
 ; NO-SIMD128-LABEL: max_u_v16i8:
 ; NO-SIMD128:         .functype max_u_v16i8 (i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32) -> ()
 ; NO-SIMD128-NEXT:  # %bb.0:
-; NO-SIMD128-NEXT:    i32.const $push5=, 15
-; NO-SIMD128-NEXT:    i32.add $push6=, $0, $pop5
+; NO-SIMD128-NEXT:    i32.const $push3=, 15
+; NO-SIMD128-NEXT:    i32.add $push4=, $0, $pop3
 ; NO-SIMD128-NEXT:    i32.const $push0=, 255
-; NO-SIMD128-NEXT:    i32.and $push2=, $16, $pop0
-; NO-SIMD128-NEXT:    i32.const $push117=, 255
-; NO-SIMD128-NEXT:    i32.and $push1=, $32, $pop117
-; NO-SIMD128-NEXT:    i32.gt_u $push3=, $pop2, $pop1
-; NO-SIMD128-NEXT:    i32.select $push4=, $16, $32, $pop3
-; NO-SIMD128-NEXT:    i32.store8 0($pop6), $pop4
-; NO-SIMD128-NEXT:    i32.const $push11=, 14
+; NO-SIMD128-NEXT:    i32.and $push149=, $16, $pop0
+; NO-SIMD128-NEXT:    local.tee $push148=, $16=, $pop149
+; NO-SIMD128-NEXT:    i32.const $push147=, 255
+; NO-SIMD128-NEXT:    i32.and $push146=, $32, $pop147
+; NO-SIMD128-NEXT:    local.tee $push145=, $32=, $pop146
+; NO-SIMD128-NEXT:    i32.gt_u $push1=, $16, $32
+; NO-SIMD128-NEXT:    i32.select $push2=, $pop148, $pop145, $pop1
+; NO-SIMD128-NEXT:    i32.store8 0($pop4), $pop2
+; NO-SIMD128-NEXT:    i32.const $push7=, 14
+; NO-SIMD128-NEXT:    i32.add $push8=, $0, $pop7
+; NO-SIMD128-NEXT:    i32.const $push144=, 255
+; NO-SIMD128-NEXT:    i32.and $push143=, $15, $pop144
+; NO-SIMD128-NEXT:    local.tee $push142=, $32=, $pop143
+; NO-SIMD128-NEXT:    i32.const $push141=, 255
+; NO-SIMD128-NEXT:    i32.and $push140=, $31, $pop141
+; NO-SIMD128-NEXT:    local.tee $push139=, $16=, $pop140
+; NO-SIMD128-NEXT:    i32.gt_u $push5=, $32, $16
+; NO-SIMD128-NEXT:    i32.select $push6=, $pop142, $pop139, $pop5
+; NO-SIMD128-NEXT:    i32.store8 0($pop8), $pop6
+; NO-SIMD128-NEXT:    i32.const $push11=, 13
 ; NO-SIMD128-NEXT:    i32.add $push12=, $0, $pop11
-; NO-SIMD128-NEXT:    i32.const $push116=, 255
-; NO-SIMD128-NEXT:    i32.and $push8=, $15, $pop116
-; NO-SIMD128-NEXT:    i32.const $push115=, 255
-; NO-SIMD128-NEXT:    i32.and $push7=, $31, $pop115
-; NO-SIMD128-NEXT:    i32.gt_u $push9=, $pop8, $pop7
-; NO-SIMD128-NEXT:    i32.select $push10=, $15, $31, $pop9
+; NO-SIMD128-NEXT:    i32.const $push138=, 255
+; NO-SIMD128-NEXT:    i32.and $push137=, $14, $pop138
+; NO-SIMD128-NEXT:    local.tee $push136=, $32=, $pop137
+; NO-SIMD128-NEXT:    i32.const $push135=, 255
+; NO-SIMD128-NEXT:    i32.and $push134=, $30, $pop135
+; NO-SIMD128-NEXT:    local.tee $push133=, $16=, $pop134
+; NO-SIMD128-NEXT:    i32.gt_u $push9=, $32, $16
+; NO-SIMD128-NEXT:    i32.select $push10=, $pop136, $pop133, $pop9
 ; NO-SIMD128-NEXT:    i32.store8 0($pop12), $pop10
-; NO-SIMD128-NEXT:    i32.const $push17=, 13
-; NO-SIMD128-NEXT:    i32.add $push18=, $0, $pop17
-; NO-SIMD128-NEXT:    i32.const $push114=, 255
-; NO-SIMD128-NEXT:    i32.and $push14=, $14, $pop114
-; NO-SIMD128-NEXT:    i32.const $push113=, 255
-; NO-SIMD128-NEXT:    i32.and $push13=, $30, $pop113
-; NO-SIMD128-NEXT:    i32.gt_u $push15=, $pop14, $pop13
-; NO-SIMD128-NEXT:    i32.select $push16=, $14, $30, $pop15
-; NO-SIMD128-NEXT:    i32.store8 0($pop18), $pop16
-; NO-SIMD128-NEXT:    i32.const $push23=, 12
+; NO-SIMD128-NEXT:    i32.const $push15=, 12
+; NO-SIMD128-NEXT:    i32.add $push16=, $0, $pop15
+; NO-SIMD128-NEXT:    i32.const $push132=, 255
+; NO-SIMD128-NEXT:    i32.and $push131=, $13, $pop132
+; NO-SIMD128-NEXT:    local.tee $push130=, $32=, $pop131
+; NO-SIMD128-NEXT:    i32.const $push129=, 255
+; NO-SIMD128-NEXT:    i32.and $push128=, $29, $pop129
+; NO-SIMD128-NEXT:    local.tee $push127=, $16=, $pop128
+; NO-SIMD128-NEXT:    i32.gt_u $push13=, $32, $16
+; NO-SIMD128-NEXT:    i32.select $push14=, $pop130, $pop127, $pop13
+; NO-SIMD128-NEXT:    i32.store8 0($pop16), $pop14
+; NO-SIMD128-NEXT:    i32.const $push19=, 11
+; NO-SIMD128-NEXT:    i32.add $push20=, $0, $pop19
+; NO-SIMD128-NEXT:    i32.const $push126=, 255
+; NO-SIMD128-NEXT:    i32.and $push125=, $12, $pop126
+; NO-SIMD128-NEXT:    local.tee $push124=, $32=, $pop125
+; NO-SIMD128-NEXT:    i32.const $push123=, 255
+; NO-SIMD128-NEXT:    i32.and $push122=, $28, $pop123
+; NO-SIMD128-NEXT:    local.tee $push121=, $16=, $pop122
+; NO-SIMD128-NEXT:    i32.gt_u $push17=, $32, $16
+; NO-SIMD128-NEXT:    i32.select $push18=, $pop124, $pop121, $pop17
+; NO-SIMD128-NEXT:    i32.store8 0($pop20), $pop18
+; NO-SIMD128-NEXT:    i32.const $push23=, 10
 ; NO-SIMD128-NEXT:    i32.add $push24=, $0, $pop23
-; NO-SIMD128-NEXT:    i32.const $push112=, 255
-; NO-SIMD128-NEXT:    i32.and $push20=, $13, $pop112
-; NO-SIMD128-NEXT:    i32.const $push111=, 255
-; NO-SIMD128-NEXT:    i32.and $push19=, $29, $pop111
-; NO-SIMD128-NEXT:    i32.gt_u $push21=, $pop20, $pop19
-; NO-SIMD128-NEXT:    i32.select $push22=, $13, $29, $pop21
+; NO-SIMD128-NEXT:    i32.const $push120=, 255
+; NO-SIMD128-NEXT:    i32.and $push119=, $11, $pop120
+; NO-SIMD128-NEXT:    local.tee $push118=, $32=, $pop119
+; NO-SIMD128-NEXT:    i32.const $push117=, 255
+; NO-SIMD128-NEXT:    i32.and $push116=, $27, $pop117
+; NO-SIMD128-NEXT:    local.tee $push115=, $16=, $pop116
+; NO-SIMD128-NEXT:    i32.gt_u $push21=, $32, $16
+; NO-SIMD128-NEXT:    i32.select $push22=, $pop118, $pop115, $pop21
 ; NO-SIMD128-NEXT:    i32.store8 0($pop24), $pop22
-; NO-SIMD128-NEXT:    i32.const $push29=, 11
-; NO-SIMD128-NEXT:    i32.add $push30=, $0, $pop29
-; NO-SIMD128-NEXT:    i32.const $push110=, 255
-; NO-SIMD128-NEXT:    i32.and $push26=, $12, $pop110
-; NO-SIMD128-NEXT:    i32.const $push109=, 255
-; NO-SIMD128-NEXT:    i32.and $push25=, $28, $pop109
-; NO-SIMD128-NEXT:    i32.gt_u $push27=, $pop26, $pop25
-; NO-SIMD128-NEXT:    i32.select $push28=, $12, $28, $pop27
-; NO-SIMD128-NEXT:    i32.store8 0($pop30), $pop28
-; NO-SIMD128-NEXT:    i32.const $push35=, 10
-; NO-SIMD128-NEXT:    i32.add $push36=, $0, $pop35
+; NO-SIMD128-NEXT:    i32.const $push27=, 9
+; NO-SIMD128-NEXT:    i32.add $push28=, $0, $pop27
+; NO-SIMD128-NEXT:    i32.const $push114=, 255
+; NO-SIMD128-NEXT:    i32.and $push113=, $10, $pop114
+; NO-SIMD128-NEXT:    local.tee $push112=, $32=, $pop113
+; NO-SIMD128-NEXT:    i32.const $push111=, 255
+; NO-SIMD128-NEXT:    i32.and $push110=, $26, $pop111
+; NO-SIMD128-NEXT:    local.tee $push109=, $16=, $pop110
+; NO-SIMD128-NEXT:    i32.gt_u $push25=, $32, $16
+; NO-SIMD128-NEXT:    i32.select $push26=, $pop112, $pop109, $pop25
+; NO-SIMD128-NEXT:    i32.store8 0($pop28), $pop26
 ; NO-SIMD128-NEXT:    i32.const $push108=, 255
-; NO-SIMD128-NEXT:    i32.and $push32=, $11, $pop108
-; NO-SIMD128-NEXT:    i32.const $push107=, 255
-; NO-SIMD128-NEXT:    i32.and $push31=, $27, $pop107
-; NO-SIMD128-NEXT:    i32.gt_u $push33=, $pop32, $pop31
-; NO-SIMD128-NEXT:    i32.select $push34=, $11, $27, $pop33
-; NO-SIMD128-NEXT:    i32.store8 0($pop36), $pop34
-; NO-SIMD128-NEXT:    i32.const $push41=, 9
-; NO-SIMD128-NEXT:    i32.add $push42=, $0, $pop41
-; NO-SIMD128-NEXT:    i32.const $push106=, 255
-; NO-SIMD128-NEXT:    i32.and $push38=, $10, $pop106
+; NO-SIMD128-NEXT:    i32.and $push107=, $9, $pop108
+; NO-SIMD128-NEXT:    local.tee $push106=, $32=, $pop107
 ; NO-SIMD128-NEXT:    i32.const $push105=, 255
-; NO-SIMD128-NEXT:    i32.and $push37=, $26, $pop105
-; NO-SIMD128-NEXT:    i32.gt_u $push39=, $pop38, $pop37
-; NO-SIMD128-NEXT:    i32.select $push40=, $10, $26, $pop39
-; NO-SIMD128-NEXT:    i32.store8 0($pop42), $pop40
-; NO-SIMD128-NEXT:    i32.const $push104=, 255
-; NO-SIMD128-NEXT:    i32.and $push44=, $9, $pop104
-; NO-SIMD128-NEXT:    i32.const $push103=, 255
-; NO-SIMD128-NEXT:    i32.and $push43=, $25, $pop103
-; NO-SIMD128-NEXT:    i32.gt_u $push45=, $pop44, $pop43
-; NO-SIMD128-NEXT:    i32.select $push46=, $9, $25, $pop45
-; NO-SIMD128-NEXT:    i32.store8 8($0), $pop46
-; NO-SIMD128-NEXT:    i32.const $push51=, 7
-; NO-SIMD128-NEXT:    i32.add $push52=, $0, $pop51
+; NO-SIMD128-NEXT:    i32.and $push104=, $25, $pop105
+; NO-SIMD128-NEXT:    local.tee $push103=, $16=, $pop104
+; NO-SIMD128-NEXT:    i32.gt_u $push29=, $32, $16
+; NO-SIMD128-NEXT:    i32.select $push30=, $pop106, $pop103, $pop29
+; NO-SIMD128-NEXT:    i32.store8 8($0), $pop30
+; NO-SIMD128-NEXT:    i32.const $push33=, 7
+; NO-SIMD128-NEXT:    i32.add $push34=, $0, $pop33
 ; NO-SIMD128-NEXT:    i32.const $push102=, 255
-; NO-SIMD128-NEXT:    i32.and $push48=, $8, $pop102
-; NO-SIMD128-NEXT:    i32.const $push101=, 255
-; NO-SIMD128-NEXT:    i32.and $push47=, $24, $pop101
-; NO-SIMD128-NEXT:    i32.gt_u $push49=, $pop48, $pop47
-; NO-SIMD128-NEXT:    i32.select $push50=, $8, $24, $pop49
-; NO-SIMD128-NEXT:    i32.store8 0($pop52), $pop50
-; NO-SIMD128-NEXT:    i32.const $push57=, 6
-; NO-SIMD128-NEXT:    i32.add $push58=, $0, $pop57
-; NO-SIMD128-NEXT:    i32.const $push100=, 255
-; NO-SIMD128-NEXT:    i32.and $push54=, $7, $pop100
+; NO-SIMD128-NEXT:    i32.and $push101=, $8, $pop102
+; NO-SIMD128-NEXT:    local.tee $push100=, $32=, $pop101
 ; NO-SIMD128-NEXT:    i32.const $push99=, 255
-; NO-SIMD128-NEXT:    i32.and $push53=, $23, $pop99
-; NO-SIMD128-NEXT:    i32.gt_u $push55=, $pop54, $pop53
-; NO-SIMD128-NEXT:    i32.select $push56=, $7, $23, $pop55
-; NO-SIMD128-NEXT:    i32.store8 0($pop58), $pop56
-; NO-SIMD128-NEXT:    i32.const $push63=, 5
-; NO-SIMD128-NEXT:    i32.add $push64=, $0, $pop63
-; NO-SIMD128-NEXT:    i32.const $push98=, 255
-; NO-SIMD128-NEXT:    i32.and $push60=, $6, $pop98
-; NO-SIMD128-NEXT:    i32.const $push97=, 255
-; NO-SIMD128-NEXT:    i32.and $push59=, $22, $pop97
-; NO-SIMD128-NEXT:    i32.gt_u $push61=, $pop60, $pop59
-; NO-SIMD128-NEXT:    i32.select $push62=, $6, $22, $pop61
-; NO-SIMD128-NEXT:    i32.store8 0($pop64), $pop62
+; NO-SIMD128-NEXT:    i32.and $push98=, $24, $pop99
+; NO-SIMD128-NEXT:    local.tee $push97=, $16=, $pop98
+; NO-SIMD128-NEXT:    i32.gt_u $push31=, $32, $16
+; NO-SIMD128-NEXT:    i32.select $push32=, $pop100, $pop97, $pop31
+; NO-SIMD128-NEXT:    i32.store8 0($pop34), $pop32
+; NO-SIMD128-NEXT:    i32.const $push37=, 6
+; NO-SIMD128-NEXT:    i32.add $push38=, $0, $pop37
 ; NO-SIMD128-NEXT:    i32.const $push96=, 255
-; NO-SIMD128-NEXT:    i32.and $push66=, $5, $pop96
-; NO-SIMD128-NEXT:    i32.const $push95=, 255
-; NO-SIMD128-NEXT:    i32.and $push65=, $21, $pop95
-; NO-SIMD128-NEXT:    i32.gt_u $push67=, $pop66, $pop65
-; NO-SIMD128-NEXT:    i32.select $push68=, $5, $21, $pop67
-; NO-SIMD128-NEXT:    i32.store8 4($0), $pop68
-; NO-SIMD128-NEXT:    i32.const $push73=, 3
-; NO-SIMD128-NEXT:    i32.add $push74=, $0, $pop73
-; NO-SIMD128-NEXT:    i32.const $push94=, 255
-; NO-SIMD128-NEXT:    i32.and $push70=, $4, $pop94
-; NO-SIMD128-NEXT:    i32.const $push93=, 255
-; NO-SIMD128-NEXT:    i32.and $push69=, $20, $pop93
-; NO-SIMD128-NEXT:    i32.gt_u $push71=, $pop70, $pop69
-; NO-SIMD128-NEXT:    i32.select $push72=, $4, $20, $pop71
-; NO-SIMD128-NEXT:    i32.store8 0($pop74), $pop72
-; NO-SIMD128-NEXT:    i32.const $push92=, 255
-; NO-SIMD128-NEXT:    i32.and $push76=, $3, $pop92
-; NO-SIMD128-NEXT:    i32.const $push91=, 255
-; NO-SIMD128-NEXT:    i32.and $push75=, $19, $pop91
-; NO-SIMD128-NEXT:    i32.gt_u $push77=, $pop76, $pop75
-; NO-SIMD128-NEXT:    i32.select $push78=, $3, $19, $pop77
-; NO-SIMD128-NEXT:    i32.store8 2($0), $pop78
+; NO-SIMD128-NEXT:    i32.and $push95=, $7, $pop96
+; NO-SIMD128-NEXT:    local.tee $push94=, $32=, $pop95
+; NO-SIMD128-NEXT:    i32.const $push93=, 255
+; NO-SIMD128-NEXT:    i32.and $push92=, $23, $pop93
+; NO-SIMD128-NEXT:    local.tee $push91=, $16=, $pop92
+; NO-SIMD128-NEXT:    i32.gt_u $push35=, $32, $16
+; NO-SIMD128-NEXT:    i32.select $push36=, $pop94, $pop91, $pop35
+; NO-SIMD128-NEXT:    i32.store8 0($pop38), $pop36
+; NO-SIMD128-NEXT:    i32.const $push41=, 5
+; NO-SIMD128-NEXT:    i32.add $push42=, $0, $pop41
 ; NO-SIMD128-NEXT:    i32.const $push90=, 255
-; NO-SIMD128-NEXT:    i32.and $push80=, $2, $pop90
-; NO-SIMD128-NEXT:    i32.const $push89=, 255
-; NO-SIMD128-NEXT:    i32.and $push79=, $18, $pop89
-; NO-SIMD128-NEXT:    i32.gt_u $push81=, $pop80, $pop79
-; NO-SIMD128-NEXT:    i32.select $push82=, $2, $18, $pop81
-; NO-SIMD128-NEXT:    i32.store8 1($0), $pop82
-; NO-SIMD128-NEXT:    i32.const $push88=, 255
-; NO-SIMD128-NEXT:    i32.and $push84=, $1, $pop88
+; NO-SIMD128-NEXT:    i32.and $push89=, $6, $pop90
+; NO-SIMD128-NEXT:    local.tee $push88=, $32=, $pop89
 ; NO-SIMD128-NEXT:    i32.const $push87=, 255
-; NO-SIMD128-NEXT:    i32.and $push83=, $17, $pop87
-; NO-SIMD128-NEXT:    i32.gt_u $push85=, $pop84, $pop83
-; NO-SIMD128-NEXT:    i32.select $push86=, $1, $17, $pop85
-; NO-SIMD128-NEXT:    i32.store8 0($0), $pop86
+; NO-SIMD128-NEXT:    i32.and $push86=, $22, $pop87
+; NO-SIMD128-NEXT:    local.tee $push85=, $16=, $pop86
+; NO-SIMD128-NEXT:    i32.gt_u $push39=, $32, $16
+; NO-SIMD128-NEXT:    i32.select $push40=, $pop88, $pop85, $pop39
+; NO-SIMD128-NEXT:    i32.store8 0($pop42), $pop40
+; NO-SIMD128-NEXT:    i32.const $push84=, 255
+; NO-SIMD128-NEXT:    i32.and $push83=, $5, $pop84
+; NO-SIMD128-NEXT:    local.tee $push82=, $32=, $pop83
+; NO-SIMD128-NEXT:    i32.const $push81=, 255
+; NO-SIMD128-NEXT:    i32.and $push80=, $21, $pop81
+; NO-SIMD128-NEXT:    local.tee $push79=, $16=, $pop80
+; NO-SIMD128-NEXT:    i32.gt_u $push43=, $32, $16
+; NO-SIMD128-NEXT:    i32.select $push44=, $pop82, $pop79, $pop43
+; NO-SIMD128-NEXT:    i32.store8 4($0), $pop44
+; NO-SIMD128-NEXT:    i32.const $push47=, 3
+; NO-SIMD128-NEXT:    i32.add $push48=, $0, $pop47
+; NO-SIMD128-NEXT:    i32.const $push78=, 255
+; NO-SIMD128-NEXT:    i32.and $push77=, $4, $pop78
+; NO-SIMD128-NEXT:    local.tee $push76=, $32=, $pop77
+; NO-SIMD128-NEXT:    i32.const $push75=, 255
+; NO-SIMD128-NEXT:    i32.and $push74=, $20, $pop75
+; NO-SIMD128-NEXT:    local.tee $push73=, $16=, $pop74
+; NO-SIMD128-NEXT:    i32.gt_u $push45=, $32, $16
+; NO-SIMD128-NEXT:    i32.select $push46=, $pop76, $pop73, $pop45
+; NO-SIMD128-NEXT:    i32.store8 0($pop48), $pop46
+; NO-SIMD128-NEXT:    i32.const $push72=, 255
+; NO-SIMD128-NEXT:    i32.and $push71=, $3, $pop72
+; NO-SIMD128-NEXT:    local.tee $push70=, $32=, $pop71
+; NO-SIMD128-NEXT:    i32.const $push69=, 255
+; NO-SIMD128-NEXT:    i32.and $push68=, $19, $pop69
+; NO-SIMD128-NEXT:    local.tee $push67=, $16=, $pop68
+; NO-SIMD128-NEXT:    i32.gt_u $push49=, $32, $16
+; NO-SIMD128-NEXT:    i32.select $push50=, $pop70, $pop67, $pop49
+; NO-SIMD128-NEXT:    i32.store8 2($0), $pop50
+; NO-SIMD128-NEXT:    i32.const $push66=, 255
+; NO-SIMD128-NEXT:    i32.and $push65=, $2, $pop66
+; NO-SIMD128-NEXT:    local.tee $push64=, $32=, $pop65
+; NO-SIMD128-NEXT:    i32.const $push63=, 255
+; NO-SIMD128-NEXT:    i32.and $push62=, $18, $pop63
+; NO-SIMD128-NEXT:    local.tee $push61=, $16=, $pop62
+; NO-SIMD128-NEXT:    i32.gt_u $push51=, $32, $16
+; NO-SIMD128-NEXT:    i32.select $push52=, $pop64, $pop61, $pop51
+; NO-SIMD128-NEXT:    i32.store8 1($0), $pop52
+; NO-SIMD128-NEXT:    i32.const $push60=, 255
+; NO-SIMD128-NEXT:    i32.and $push59=, $1, $pop60
+; NO-SIMD128-NEXT:    local.tee $push58=, $32=, $pop59
+; NO-SIMD128-NEXT:    i32.const $push57=, 255
+; NO-SIMD128-NEXT:    i32.and $push56=, $17, $pop57
+; NO-SIMD128-NEXT:    local.tee $push55=, $16=, $pop56
+; NO-SIMD128-NEXT:    i32.gt_u $push53=, $32, $16
+; NO-SIMD128-NEXT:    i32.select $push54=, $pop58, $pop55, $pop53
+; NO-SIMD128-NEXT:    i32.store8 0($0), $pop54
 ; NO-SIMD128-NEXT:    return
 ;
 ; NO-SIMD128-FAST-LABEL: max_u_v16i8:
 ; NO-SIMD128-FAST:         .functype max_u_v16i8 (i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32) -> ()
 ; NO-SIMD128-FAST-NEXT:  # %bb.0:
 ; NO-SIMD128-FAST-NEXT:    i32.const $push0=, 255
-; NO-SIMD128-FAST-NEXT:    i32.and $push2=, $1, $pop0
+; NO-SIMD128-FAST-NEXT:    i32.and $push149=, $1, $pop0
+; NO-SIMD128-FAST-NEXT:    local.tee $push148=, $1=, $pop149
+; NO-SIMD128-FAST-NEXT:    i32.const $push147=, 255
+; NO-SIMD128-FAST-NEXT:    i32.and $push146=, $17, $pop147
+; NO-SIMD128-FAST-NEXT:    local.tee $push145=, $17=, $pop146
+; NO-SIMD128-FAST-NEXT:    i32.gt_u $push1=, $1, $17
+; NO-SIMD128-FAST-NEXT:    i32.select $push2=, $pop148, $pop145, $pop1
+; NO-SIMD128-FAST-NEXT:    i32.store8 0($0), $pop2
+; NO-SIMD128-FAST-NEXT:    i32.const $push144=, 255
+; NO-SIMD128-FAST-NEXT:    i32.and $push143=, $2, $pop144
+; NO-SIMD128-FAST-NEXT:    local.tee $push142=, $17=, $pop143
+; NO-SIMD128-FAST-NEXT:    i32.const $push141=, 255
+; NO-SIMD128-FAST-NEXT:    i32.and $push140=, $18, $pop141
+; NO-SIMD128-FAST-NEXT:    local.tee $push139=, $18=, $pop140
+; NO-SIMD128-FAST-NEXT:    i32.gt_u $push3=, $17, $18
+; NO-SIMD128-FAST-NEXT:    i32.select $push4=, $pop142, $pop139, $pop3
+; NO-SIMD128-FAST-NEXT:    i32.store8 1($0), $pop4
+; NO-SIMD128-FAST-NEXT:    i32.const $push138=, 255
+; NO-SIMD128-FAST-NEXT:    i32.and $push137=, $3, $pop138
+; NO-SIMD128-FAST-NEXT:    local.tee $push136=, $18=, $pop137
+; NO-SIMD128-FAST-NEXT:    i32.const $push135=, 255
+; NO-SIMD128-FAST-NEXT:    i32.and $push134=, $19, $pop135
+; NO-SIMD128-FAST-NEXT:    local.tee $push133=, $19=, $pop134
+; NO-SIMD128-FAST-NEXT:    i32.gt_u $push5=, $18, $19
+; NO-SIMD128-FAST-NEXT:    i32.select $push6=, $pop136, $pop133, $pop5
+; NO-SIMD128-FAST-NEXT:    i32.store8 2($0), $pop6
+; NO-SIMD128-FAST-NEXT:    i32.const $push9=, 3
+; NO-SIMD128-FAST-NEXT:    i32.add $push10=, $0, $pop9
+; NO-SIMD128-FAST-NEXT:    i32.const $push132=, 255
+; NO-SIMD128-FAST-NEXT:    i32.and $push131=, $4, $pop132
+; NO-SIMD128-FAST-NEXT:    local.tee $push130=, $19=, $pop131
+; NO-SIMD128-FAST-NEXT:    i32.const $push129=, 255
+; NO-SIMD128-FAST-NEXT:    i32.and $push128=, $20, $pop129
+; NO-SIMD128-FAST-NEXT:    local.tee $push127=, $20=, $pop128
+; NO-SIMD128-FAST-NEXT:    i32.gt_u $push7=, $19, $20
+; NO-SIMD128-FAST-NEXT:    i32.select $push8=, $pop130, $pop127, $pop7
+; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop10), $pop8
+; NO-SIMD128-FAST-NEXT:    i32.const $push126=, 255
+; NO-SIMD128-FAST-NEXT:    i32.and $push125=, $5, $pop126
+; NO-SIMD128-FAST-NEXT:    local.tee $push124=, $20=, $pop125
+; NO-SIMD128-FAST-NEXT:    i32.const $push123=, 255
+; NO-SIMD128-FAST-NEXT:    i32.and $push122=, $21, $pop123
+; NO-SIMD128-FAST-NEXT:    local.tee $push121=, $21=, $pop122
+; NO-SIMD128-FAST-NEXT:    i32.gt_u $push11=, $20, $21
+; NO-SIMD128-FAST-NEXT:    i32.select $push12=, $pop124, $pop121, $pop11
+; NO-SIMD128-FAST-NEXT:    i32.store8 4($0), $pop12
+; NO-SIMD128-FAST-NEXT:    i32.const $push15=, 5
+; NO-SIMD128-FAST-NEXT:    i32.add $push16=, $0, $pop15
+; NO-SIMD128-FAST-NEXT:    i32.const $push120=, 255
+; NO-SIMD128-FAST-NEXT:    i32.and $push119=, $6, $pop120
+; NO-SIMD128-FAST-NEXT:    local.tee $push118=, $21=, $pop119
 ; NO-SIMD128-FAST-NEXT:    i32.const $push117=, 255
-; NO-SIMD128-FAST-NEXT:    i32.and $push1=, $17, $pop117
-; NO-SIMD128-FAST-NEXT:    i32.gt_u $push3=, $pop2, $pop1
-; NO-SIMD128-FAST-NEXT:    i32.select $push4=, $1, $17, $pop3
-; NO-SIMD128-FAST-NEXT:    i32.store8 0($0), $pop4
-; NO-SIMD128-FAST-NEXT:    i32.const $push116=, 255
-; NO-SIMD128-FAST-NEXT:    i32.and $push6=, $2, $pop116
-; NO-SIMD128-FAST-NEXT:    i32.const $push115=, 255
-; NO-SIMD128-FAST-NEXT:    i32.and $push5=, $18, $pop115
-; NO-SIMD128-FAST-NEXT:    i32.gt_u $push7=, $pop6, $pop5
-; NO-SIMD128-FAST-NEXT:    i32.select $push8=, $2, $18, $pop7
-; NO-SIMD128-FAST-NEXT:    i32.store8 1($0), $pop8
+; NO-SIMD128-FAST-NEXT:    i32.and $push116=, $22, $pop117
+; NO-SIMD128-FAST-NEXT:    local.tee $push115=, $22=, $pop116
+; NO-SIMD128-FAST-NEXT:    i32.gt_u $push13=, $21, $22
+; NO-SIMD128-FAST-NEXT:    i32.select $push14=, $pop118, $pop115, $pop13
+; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop16), $pop14
+; NO-SIMD128-FAST-NEXT:    i32.const $push19=, 6
+; NO-SIMD128-FAST-NEXT:    i32.add $push20=, $0, $pop19
 ; NO-SIMD128-FAST-NEXT:    i32.const $push114=, 255
-; NO-SIMD128-FAST-NEXT:    i32.and $push10=, $3, $pop114
-; NO-SIMD128-FAST-NEXT:    i32.const $push113=, 255
-; NO-SIMD128-FAST-NEXT:    i32.and $push9=, $19, $pop113
-; NO-SIMD128-FAST-NEXT:    i32.gt_u $push11=, $pop10, $pop9
-; NO-SIMD128-FAST-NEXT:    i32.select $push12=, $3, $19, $pop11
-; NO-SIMD128-FAST-NEXT:    i32.store8 2($0), $pop12
-; NO-SIMD128-FAST-NEXT:    i32.const $push17=, 3
-; NO-SIMD128-FAST-NEXT:    i32.add $push18=, $0, $pop17
-; NO-SIMD128-FAST-NEXT:    i32.const $push112=, 255
-; NO-SIMD128-FAST-NEXT:    i32.and $push14=, $4, $pop112
+; NO-SIMD128-FAST-NEXT:    i32.and $push113=, $7, $pop114
+; NO-SIMD128-FAST-NEXT:    local.tee $push112=, $22=, $pop113
 ; NO-SIMD128-FAST-NEXT:    i32.const $push111=, 255
-; NO-SIMD128-FAST-NEXT:    i32.and $push13=, $20, $pop111
-; NO-SIMD128-FAST-NEXT:    i32.gt_u $push15=, $pop14, $pop13
-; NO-SIMD128-FAST-NEXT:    i32.select $push16=, $4, $20, $pop15
-; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop18), $pop16
-; NO-SIMD128-FAST-NEXT:    i32.const $push110=, 255
-; NO-SIMD128-FAST-NEXT:    i32.and $push20=, $5, $pop110
-; NO-SIMD128-FAST-NEXT:    i32.const $push109=, 255
-; NO-SIMD128-FAST-NEXT:    i32.and $push19=, $21, $pop109
-; NO-SIMD128-FAST-NEXT:    i32.gt_u $push21=, $pop20, $pop19
-; NO-SIMD128-FAST-NEXT:    i32.select $push22=, $5, $21, $pop21
-; NO-SIMD128-FAST-NEXT:    i32.store8 4($0), $pop22
-; NO-SIMD128-FAST-NEXT:    i32.const $push27=, 5
-; NO-SIMD128-FAST-NEXT:    i32.add $push28=, $0, $pop27
+; NO-SIMD128-FAST-NEXT:    i32.and $push110=, $23, $pop111
+; NO-SIMD128-FAST-NEXT:    local.tee $push109=, $23=, $pop110
+; NO-SIMD128-FAST-NEXT:    i32.gt_u $push17=, $22, $23
+; NO-SIMD128-FAST-NEXT:    i32.select $push18=, $pop112, $pop109, $pop17
+; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop20), $pop18
+; NO-SIMD128-FAST-NEXT:    i32.const $push23=, 7
+; NO-SIMD128-FAST-NEXT:    i32.add $push24=, $0, $pop23
 ; NO-SIMD128-FAST-NEXT:    i32.const $push108=, 255
-; NO-SIMD128-FAST-NEXT:    i32.and $push24=, $6, $pop108
-; NO-SIMD128-FAST-NEXT:    i32.const $push107=, 255
-; NO-SIMD128-FAST-NEXT:    i32.and $push23=, $22, $pop107
-; NO-SIMD128-FAST-NEXT:    i32.gt_u $push25=, $pop24, $pop23
-; NO-SIMD128-FAST-NEXT:    i32.select $push26=, $6, $22, $pop25
-; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop28), $pop26
-; NO-SIMD128-FAST-NEXT:    i32.const $push33=, 6
-; NO-SIMD128-FAST-NEXT:    i32.add $push34=, $0, $pop33
-; NO-SIMD128-FAST-NEXT:    i32.const $push106=, 255
-; NO-SIMD128-FAST-NEXT:    i32.and $push30=, $7, $pop106
+; NO-SIMD128-FAST-NEXT:    i32.and $push107=, $8, $pop108
+; NO-SIMD128-FAST-NEXT:    local.tee $push106=, $23=, $pop107
 ; NO-SIMD128-FAST-NEXT:    i32.const $push105=, 255
-; NO-SIMD128-FAST-NEXT:    i32.and $push29=, $23, $pop105
-; NO-SIMD128-FAST-NEXT:    i32.gt_u $push31=, $pop30, $pop29
-; NO-SIMD128-FAST-NEXT:    i32.select $push32=, $7, $23, $pop31
-; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop34), $pop32
-; NO-SIMD128-FAST-NEXT:    i32.const $push39=, 7
-; NO-SIMD128-FAST-NEXT:    i32.add $push40=, $0, $pop39
-; NO-SIMD128-FAST-NEXT:    i32.const $push104=, 255
-; NO-SIMD128-FAST-NEXT:    i32.and $push36=, $8, $pop104
-; NO-SIMD128-FAST-NEXT:    i32.const $push103=, 255
-; NO-SIMD128-FAST-NEXT:    i32.and $push35=, $24, $pop103
-; NO-SIMD128-FAST-NEXT:    i32.gt_u $push37=, $pop36, $pop35
-; NO-SIMD128-FAST-NEXT:    i32.select $push38=, $8, $24, $pop37
-; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop40), $pop38
+; NO-SIMD128-FAST-NEXT:    i32.and $push104=, $24, $pop105
+; NO-SIMD128-FAST-NEXT:    local.tee $push103=, $24=, $pop104
+; NO-SIMD128-FAST-NEXT:    i32.gt_u $push21=, $23, $24
+; NO-SIMD128-FAST-NEXT:    i32.select $push22=, $pop106, $pop103, $pop21
+; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop24), $pop22
 ; NO-SIMD128-FAST-NEXT:    i32.const $push102=, 255
-; NO-SIMD128-FAST-NEXT:    i32.and $push42=, $9, $pop102
-; NO-SIMD128-FAST-NEXT:    i32.const $push101=, 255
-; NO-SIMD128-FAST-NEXT:    i32.and $push41=, $25, $pop101
-; NO-SIMD128-FAST-NEXT:    i32.gt_u $push43=, $pop42, $pop41
-; NO-SIMD128-FAST-NEXT:    i32.select $push44=, $9, $25, $pop43
-; NO-SIMD128-FAST-NEXT:    i32.store8 8($0), $pop44
-; NO-SIMD128-FAST-NEXT:    i32.const $push49=, 9
-; NO-SIMD128-FAST-NEXT:    i32.add $push50=, $0, $pop49
-; NO-SIMD128-FAST-NEXT:    i32.const $push100=, 255
-; NO-SIMD128-FAST-NEXT:    i32.and $push46=, $10, $pop100
+; NO-SIMD128-FAST-NEXT:    i32.and $push101=, $9, $pop102
+; NO-SIMD128-FAST-NEXT:    local.tee $push100=, $24=, $pop101
 ; NO-SIMD128-FAST-NEXT:    i32.const $push99=, 255
-; NO-SIMD128-FAST-NEXT:    i32.and $push45=, $26, $pop99
-; NO-SIMD128-FAST-NEXT:    i32.gt_u $push47=, $pop46, $pop45
-; NO-SIMD128-FAST-NEXT:    i32.select $push48=, $10, $26, $pop47
-; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop50), $pop48
-; NO-SIMD128-FAST-NEXT:    i32.const $push55=, 10
-; NO-SIMD128-FAST-NEXT:    i32.add $push56=, $0, $pop55
-; NO-SIMD128-FAST-NEXT:    i32.const $push98=, 255
-; NO-SIMD128-FAST-NEXT:    i32.and $push52=, $11, $pop98
-; NO-SIMD128-FAST-NEXT:    i32.const $push97=, 255
-; NO-SIMD128-FAST-NEXT:    i32.and $push51=, $27, $pop97
-; NO-SIMD128-FAST-NEXT:    i32.gt_u $push53=, $pop52, $pop51
-; NO-SIMD128-FAST-NEXT:    i32.select $push54=, $11, $27, $pop53
-; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop56), $pop54
-; NO-SIMD128-FAST-NEXT:    i32.const $push61=, 11
-; NO-SIMD128-FAST-NEXT:    i32.add $push62=, $0, $pop61
+; NO-SIMD128-FAST-NEXT:    i32.and $push98=, $25, $pop99
+; NO-SIMD128-FAST-NEXT:    local.tee $push97=, $25=, $pop98
+; NO-SIMD128-FAST-NEXT:    i32.gt_u $push25=, $24, $25
+; NO-SIMD128-FAST-NEXT:    i32.select $push26=, $pop100, $pop97, $pop25
+; NO-SIMD128-FAST-NEXT:    i32.store8 8($0), $pop26
+; NO-SIMD128-FAST-NEXT:    i32.const $push29=, 9
+; NO-SIMD128-FAST-NEXT:    i32.add $push30=, $0, $pop29
 ; NO-SIMD128-FAST-NEXT:    i32.const $push96=, 255
-; NO-SIMD128-FAST-NEXT:    i32.and $push58=, $12, $pop96
-; NO-SIMD128-FAST-NEXT:    i32.const $push95=, 255
-; NO-SIMD128-FAST-NEXT:    i32.and $push57=, $28, $pop95
-; NO-SIMD128-FAST-NEXT:    i32.gt_u $push59=, $pop58, $pop57
-; NO-SIMD128-FAST-NEXT:    i32.select $push60=, $12, $28, $pop59
-; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop62), $pop60
-; NO-SIMD128-FAST-NEXT:    i32.const $push67=, 12
-; NO-SIMD128-FAST-NEXT:    i32.add $push68=, $0, $pop67
-; NO-SIMD128-FAST-NEXT:    i32.const $push94=, 255
-; NO-SIMD128-FAST-NEXT:    i32.and $push64=, $13, $pop94
+; NO-SIMD128-FAST-NEXT:    i32.and $push95=, $10, $pop96
+; NO-SIMD128-FAST-NEXT:    local.tee $push94=, $25=, $pop95
 ; NO-SIMD128-FAST-NEXT:    i32.const $push93=, 255
-; NO-SIMD128-FAST-NEXT:    i32.and $push63=, $29, $pop93
-; NO-SIMD128-FAST-NEXT:    i32.gt_u $push65=, $pop64, $pop63
-; NO-SIMD128-FAST-NEXT:    i32.select $push66=, $13, $29, $pop65
-; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop68), $pop66
-; NO-SIMD128-FAST-NEXT:    i32.const $push73=, 13
-; NO-SIMD128-FAST-NEXT:    i32.add $push74=, $0, $pop73
-; NO-SIMD128-FAST-NEXT:    i32.const $push92=, 255
-; NO-SIMD128-FAST-NEXT:    i32.and $push70=, $14, $pop92
-; NO-SIMD128-FAST-NEXT:    i32.const $push91=, 255
-; NO-SIMD128-FAST-NEXT:    i32.and $push69=, $30, $pop91
-; NO-SIMD128-FAST-NEXT:    i32.gt_u $push71=, $pop70, $pop69
-; NO-SIMD128-FAST-NEXT:    i32.select $push72=, $14, $30, $pop71
-; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop74), $pop72
-; NO-SIMD128-FAST-NEXT:    i32.const $push79=, 14
-; NO-SIMD128-FAST-NEXT:    i32.add $push80=, $0, $pop79
+; NO-SIMD128-FAST-NEXT:    i32.and $push92=, $26, $pop93
+; NO-SIMD128-FAST-NEXT:    local.tee $push91=, $26=, $pop92
+; NO-SIMD128-FAST-NEXT:    i32.gt_u $push27=, $25, $26
+; NO-SIMD128-FAST-NEXT:    i32.select $push28=, $pop94, $pop91, $pop27
+; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop30), $pop28
+; NO-SIMD128-FAST-NEXT:    i32.const $push33=, 10
+; NO-SIMD128-FAST-NEXT:    i32.add $push34=, $0, $pop33
 ; NO-SIMD128-FAST-NEXT:    i32.const $push90=, 255
-; NO-SIMD128-FAST-NEXT:    i32.and $push76=, $15, $pop90
-; NO-SIMD128-FAST-NEXT:    i32.const $push89=, 255
-; NO-SIMD128-FAST-NEXT:    i32.and $push75=, $31, $pop89
-; NO-SIMD128-FAST-NEXT:    i32.gt_u $push77=, $pop76, $pop75
-; NO-SIMD128-FAST-NEXT:    i32.select $push78=, $15, $31, $pop77
-; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop80), $pop78
-; NO-SIMD128-FAST-NEXT:    i32.const $push85=, 15
-; NO-SIMD128-FAST-NEXT:    i32.add $push86=, $0, $pop85
-; NO-SIMD128-FAST-NEXT:    i32.const $push88=, 255
-; NO-SIMD128-FAST-NEXT:    i32.and $push82=, $16, $pop88
+; NO-SIMD128-FAST-NEXT:    i32.and $push89=, $11, $pop90
+; NO-SIMD128-FAST-NEXT:    local.tee $push88=, $26=, $pop89
 ; NO-SIMD128-FAST-NEXT:    i32.const $push87=, 255
-; NO-SIMD128-FAST-NEXT:    i32.and $push81=, $32, $pop87
-; NO-SIMD128-FAST-NEXT:    i32.gt_u $push83=, $pop82, $pop81
-; NO-SIMD128-FAST-NEXT:    i32.select $push84=, $16, $32, $pop83
-; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop86), $pop84
+; NO-SIMD128-FAST-NEXT:    i32.and $push86=, $27, $pop87
+; NO-SIMD128-FAST-NEXT:    local.tee $push85=, $27=, $pop86
+; NO-SIMD128-FAST-NEXT:    i32.gt_u $push31=, $26, $27
+; NO-SIMD128-FAST-NEXT:    i32.select $push32=, $pop88, $pop85, $pop31
+; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop34), $pop32
+; NO-SIMD128-FAST-NEXT:    i32.const $push37=, 11
+; NO-SIMD128-FAST-NEXT:    i32.add $push38=, $0, $pop37
+; NO-SIMD128-FAST-NEXT:    i32.const $push84=, 255
+; NO-SIMD128-FAST-NEXT:    i32.and $push83=, $12, $pop84
+; NO-SIMD128-FAST-NEXT:    local.tee $push82=, $27=, $pop83
+; NO-SIMD128-FAST-NEXT:    i32.const $push81=, 255
+; NO-SIMD128-FAST-NEXT:    i32.and $push80=, $28, $pop81
+; NO-SIMD128-FAST-NEXT:    local.tee $push79=, $28=, $pop80
+; NO-SIMD128-FAST-NEXT:    i32.gt_u $push35=, $27, $28
+; NO-SIMD128-FAST-NEXT:    i32.select $push36=, $pop82, $pop79, $pop35
+; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop38), $pop36
+; NO-SIMD128-FAST-NEXT:    i32.const $push41=, 12
+; NO-SIMD128-FAST-NEXT:    i32.add $push42=, $0, $pop41
+; NO-SIMD128-FAST-NEXT:    i32.const $push78=, 255
+; NO-SIMD128-FAST-NEXT:    i32.and $push77=, $13, $pop78
+; NO-SIMD128-FAST-NEXT:    local.tee $push76=, $28=, $pop77
+; NO-SIMD128-FAST-NEXT:    i32.const $push75=, 255
+; NO-SIMD128-FAST-NEXT:    i32.and $push74=, $29, $pop75
+; NO-SIMD128-FAST-NEXT:    local.tee $push73=, $29=, $pop74
+; NO-SIMD128-FAST-NEXT:    i32.gt_u $push39=, $28, $29
+; NO-SIMD128-FAST-NEXT:    i32.select $push40=, $pop76, $pop73, $pop39
+; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop42), $pop40
+; NO-SIMD128-FAST-NEXT:    i32.const $push45=, 13
+; NO-SIMD128-FAST-NEXT:    i32.add $push46=, $0, $pop45
+; NO-SIMD128-FAST-NEXT:    i32.const $push72=, 255
+; NO-SIMD128-FAST-NEXT:    i32.and $push71=, $14, $pop72
+; NO-SIMD128-FAST-NEXT:    local.tee $push70=, $29=, $pop71
+; NO-SIMD128-FAST-NEXT:    i32.const $push69=, 255
+; NO-SIMD128-FAST-NEXT:    i32.and $push68=, $30, $pop69
+; NO-SIMD128-FAST-NEXT:    local.tee $push67=, $30=, $pop68
+; NO-SIMD128-FAST-NEXT:    i32.gt_u $push43=, $29, $30
+; NO-SIMD128-FAST-NEXT:    i32.select $push44=, $pop70, $pop67, $pop43
+; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop46), $pop44
+; NO-SIMD128-FAST-NEXT:    i32.const $push49=, 14
+; NO-SIMD128-FAST-NEXT:    i32.add $push50=, $0, $pop49
+; NO-SIMD128-FAST-NEXT:    i32.const $push66=, 255
+; NO-SIMD128-FAST-NEXT:    i32.and $push65=, $15, $pop66
+; NO-SIMD128-FAST-NEXT:    local.tee $push64=, $30=, $pop65
+; NO-SIMD128-FAST-NEXT:    i32.const $push63=, 255
+; NO-SIMD128-FAST-NEXT:    i32.and $push62=, $31, $pop63
+; NO-SIMD128-FAST-NEXT:    local.tee $push61=, $31=, $pop62
+; NO-SIMD128-FAST-NEXT:    i32.gt_u $push47=, $30, $31
+; NO-SIMD128-FAST-NEXT:    i32.select $push48=, $pop64, $pop61, $pop47
+; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop50), $pop48
+; NO-SIMD128-FAST-NEXT:    i32.const $push53=, 15
+; NO-SIMD128-FAST-NEXT:    i32.add $push54=, $0, $pop53
+; NO-SIMD128-FAST-NEXT:    i32.const $push60=, 255
+; NO-SIMD128-FAST-NEXT:    i32.and $push59=, $16, $pop60
+; NO-SIMD128-FAST-NEXT:    local.tee $push58=, $0=, $pop59
+; NO-SIMD128-FAST-NEXT:    i32.const $push57=, 255
+; NO-SIMD128-FAST-NEXT:    i32.and $push56=, $32, $pop57
+; NO-SIMD128-FAST-NEXT:    local.tee $push55=, $32=, $pop56
+; NO-SIMD128-FAST-NEXT:    i32.gt_u $push51=, $0, $32
+; NO-SIMD128-FAST-NEXT:    i32.select $push52=, $pop58, $pop55, $pop51
+; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop54), $pop52
 ; NO-SIMD128-FAST-NEXT:    return
   %c = icmp ugt <16 x i8> %x, %y
   %a = select <16 x i1> %c, <16 x i8> %x, <16 x i8> %y
   ret <16 x i8> %a
 }
 
 define <16 x i8> @avgr_u_v16i8(<16 x i8> %x, <16 x i8> %y) {
 ; SIMD128-LABEL: avgr_u_v16i8:

 ; SIMD128-NEXT:  # %bb.0:
 ; SIMD128-NEXT:    v128.andnot $push0=, $0, $1
 ; SIMD128-NEXT:    return $pop0
 ;
 ; SIMD128-FAST-LABEL: andnot_v16i8:
 ; SIMD128-FAST:         .functype andnot_v16i8 (v128, v128) -> (v128)
 ; SIMD128-FAST-NEXT:  # %bb.0:
 ; SIMD128-FAST-NEXT:    v128.not $push0=, $1
-; SIMD128-FAST-NEXT:    v128.and $push1=, $0, $pop0
+; SIMD128-FAST-NEXT:    v128.and $push1=, $pop0, $0
 ; SIMD128-FAST-NEXT:    return $pop1
 ;
 ; NO-SIMD128-LABEL: andnot_v16i8:
 ; NO-SIMD128:         .functype andnot_v16i8 (i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32) -> ()
 ; NO-SIMD128-NEXT:  # %bb.0:
 ; NO-SIMD128-NEXT:    i32.const $push0=, -1
 ; NO-SIMD128-NEXT:    i32.xor $push1=, $25, $pop0
-; NO-SIMD128-NEXT:    i32.and $push2=, $9, $pop1
+; NO-SIMD128-NEXT:    i32.and $push2=, $pop1, $9
 ; NO-SIMD128-NEXT:    i32.store8 8($0), $pop2
 ; NO-SIMD128-NEXT:    i32.const $push69=, -1
 ; NO-SIMD128-NEXT:    i32.xor $push3=, $21, $pop69
-; NO-SIMD128-NEXT:    i32.and $push4=, $5, $pop3
+; NO-SIMD128-NEXT:    i32.and $push4=, $pop3, $5
 ; NO-SIMD128-NEXT:    i32.store8 4($0), $pop4
 ; NO-SIMD128-NEXT:    i32.const $push68=, -1
 ; NO-SIMD128-NEXT:    i32.xor $push5=, $19, $pop68
-; NO-SIMD128-NEXT:    i32.and $push6=, $3, $pop5
+; NO-SIMD128-NEXT:    i32.and $push6=, $pop5, $3
 ; NO-SIMD128-NEXT:    i32.store8 2($0), $pop6
 ; NO-SIMD128-NEXT:    i32.const $push67=, -1
 ; NO-SIMD128-NEXT:    i32.xor $push7=, $18, $pop67
-; NO-SIMD128-NEXT:    i32.and $push8=, $2, $pop7
+; NO-SIMD128-NEXT:    i32.and $push8=, $pop7, $2
 ; NO-SIMD128-NEXT:    i32.store8 1($0), $pop8
 ; NO-SIMD128-NEXT:    i32.const $push66=, -1
 ; NO-SIMD128-NEXT:    i32.xor $push9=, $17, $pop66
-; NO-SIMD128-NEXT:    i32.and $push10=, $1, $pop9
+; NO-SIMD128-NEXT:    i32.and $push10=, $pop9, $1
 ; NO-SIMD128-NEXT:    i32.store8 0($0), $pop10
 ; NO-SIMD128-NEXT:    i32.const $push13=, 15
 ; NO-SIMD128-NEXT:    i32.add $push14=, $0, $pop13
 ; NO-SIMD128-NEXT:    i32.const $push65=, -1
 ; NO-SIMD128-NEXT:    i32.xor $push11=, $32, $pop65
-; NO-SIMD128-NEXT:    i32.and $push12=, $16, $pop11
+; NO-SIMD128-NEXT:    i32.and $push12=, $pop11, $16
 ; NO-SIMD128-NEXT:    i32.store8 0($pop14), $pop12
 ; NO-SIMD128-NEXT:    i32.const $push17=, 14
 ; NO-SIMD128-NEXT:    i32.add $push18=, $0, $pop17
 ; NO-SIMD128-NEXT:    i32.const $push64=, -1
 ; NO-SIMD128-NEXT:    i32.xor $push15=, $31, $pop64
-; NO-SIMD128-NEXT:    i32.and $push16=, $15, $pop15
+; NO-SIMD128-NEXT:    i32.and $push16=, $pop15, $15
 ; NO-SIMD128-NEXT:    i32.store8 0($pop18), $pop16
 ; NO-SIMD128-NEXT:    i32.const $push21=, 13
 ; NO-SIMD128-NEXT:    i32.add $push22=, $0, $pop21
 ; NO-SIMD128-NEXT:    i32.const $push63=, -1
 ; NO-SIMD128-NEXT:    i32.xor $push19=, $30, $pop63
-; NO-SIMD128-NEXT:    i32.and $push20=, $14, $pop19
+; NO-SIMD128-NEXT:    i32.and $push20=, $pop19, $14
 ; NO-SIMD128-NEXT:    i32.store8 0($pop22), $pop20
 ; NO-SIMD128-NEXT:    i32.const $push25=, 12
 ; NO-SIMD128-NEXT:    i32.add $push26=, $0, $pop25
 ; NO-SIMD128-NEXT:    i32.const $push62=, -1
 ; NO-SIMD128-NEXT:    i32.xor $push23=, $29, $pop62
-; NO-SIMD128-NEXT:    i32.and $push24=, $13, $pop23
+; NO-SIMD128-NEXT:    i32.and $push24=, $pop23, $13
 ; NO-SIMD128-NEXT:    i32.store8 0($pop26), $pop24
 ; NO-SIMD128-NEXT:    i32.const $push29=, 11
 ; NO-SIMD128-NEXT:    i32.add $push30=, $0, $pop29
 ; NO-SIMD128-NEXT:    i32.const $push61=, -1
 ; NO-SIMD128-NEXT:    i32.xor $push27=, $28, $pop61
-; NO-SIMD128-NEXT:    i32.and $push28=, $12, $pop27
+; NO-SIMD128-NEXT:    i32.and $push28=, $pop27, $12
 ; NO-SIMD128-NEXT:    i32.store8 0($pop30), $pop28
 ; NO-SIMD128-NEXT:    i32.const $push33=, 10
 ; NO-SIMD128-NEXT:    i32.add $push34=, $0, $pop33
 ; NO-SIMD128-NEXT:    i32.const $push60=, -1
 ; NO-SIMD128-NEXT:    i32.xor $push31=, $27, $pop60
-; NO-SIMD128-NEXT:    i32.and $push32=, $11, $pop31
+; NO-SIMD128-NEXT:    i32.and $push32=, $pop31, $11
 ; NO-SIMD128-NEXT:    i32.store8 0($pop34), $pop32
 ; NO-SIMD128-NEXT:    i32.const $push37=, 9
 ; NO-SIMD128-NEXT:    i32.add $push38=, $0, $pop37
 ; NO-SIMD128-NEXT:    i32.const $push59=, -1
 ; NO-SIMD128-NEXT:    i32.xor $push35=, $26, $pop59
-; NO-SIMD128-NEXT:    i32.and $push36=, $10, $pop35
+; NO-SIMD128-NEXT:    i32.and $push36=, $pop35, $10
 ; NO-SIMD128-NEXT:    i32.store8 0($pop38), $pop36
 ; NO-SIMD128-NEXT:    i32.const $push41=, 7
 ; NO-SIMD128-NEXT:    i32.add $push42=, $0, $pop41
 ; NO-SIMD128-NEXT:    i32.const $push58=, -1
 ; NO-SIMD128-NEXT:    i32.xor $push39=, $24, $pop58
-; NO-SIMD128-NEXT:    i32.and $push40=, $8, $pop39
+; NO-SIMD128-NEXT:    i32.and $push40=, $pop39, $8
 ; NO-SIMD128-NEXT:    i32.store8 0($pop42), $pop40
 ; NO-SIMD128-NEXT:    i32.const $push45=, 6
 ; NO-SIMD128-NEXT:    i32.add $push46=, $0, $pop45
 ; NO-SIMD128-NEXT:    i32.const $push57=, -1
 ; NO-SIMD128-NEXT:    i32.xor $push43=, $23, $pop57
-; NO-SIMD128-NEXT:    i32.and $push44=, $7, $pop43
+; NO-SIMD128-NEXT:    i32.and $push44=, $pop43, $7
 ; NO-SIMD128-NEXT:    i32.store8 0($pop46), $pop44
 ; NO-SIMD128-NEXT:    i32.const $push49=, 5
 ; NO-SIMD128-NEXT:    i32.add $push50=, $0, $pop49
 ; NO-SIMD128-NEXT:    i32.const $push56=, -1
 ; NO-SIMD128-NEXT:    i32.xor $push47=, $22, $pop56
-; NO-SIMD128-NEXT:    i32.and $push48=, $6, $pop47
+; NO-SIMD128-NEXT:    i32.and $push48=, $pop47, $6
 ; NO-SIMD128-NEXT:    i32.store8 0($pop50), $pop48
 ; NO-SIMD128-NEXT:    i32.const $push53=, 3
 ; NO-SIMD128-NEXT:    i32.add $push54=, $0, $pop53
 ; NO-SIMD128-NEXT:    i32.const $push55=, -1
 ; NO-SIMD128-NEXT:    i32.xor $push51=, $20, $pop55
-; NO-SIMD128-NEXT:    i32.and $push52=, $4, $pop51
+; NO-SIMD128-NEXT:    i32.and $push52=, $pop51, $4
 ; NO-SIMD128-NEXT:    i32.store8 0($pop54), $pop52
 ; NO-SIMD128-NEXT:    return
 ;
 ; NO-SIMD128-FAST-LABEL: andnot_v16i8:
 ; NO-SIMD128-FAST:         .functype andnot_v16i8 (i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32) -> ()
 ; NO-SIMD128-FAST-NEXT:  # %bb.0:
 ; NO-SIMD128-FAST-NEXT:    i32.const $push0=, -1
 ; NO-SIMD128-FAST-NEXT:    i32.xor $push1=, $17, $pop0
-; NO-SIMD128-FAST-NEXT:    i32.and $push2=, $1, $pop1
+; NO-SIMD128-FAST-NEXT:    i32.and $push2=, $pop1, $1
 ; NO-SIMD128-FAST-NEXT:    i32.store8 0($0), $pop2
 ; NO-SIMD128-FAST-NEXT:    i32.const $push69=, -1
 ; NO-SIMD128-FAST-NEXT:    i32.xor $push3=, $18, $pop69
-; NO-SIMD128-FAST-NEXT:    i32.and $push4=, $2, $pop3
+; NO-SIMD128-FAST-NEXT:    i32.and $push4=, $pop3, $2
 ; NO-SIMD128-FAST-NEXT:    i32.store8 1($0), $pop4
 ; NO-SIMD128-FAST-NEXT:    i32.const $push68=, -1
 ; NO-SIMD128-FAST-NEXT:    i32.xor $push5=, $19, $pop68
-; NO-SIMD128-FAST-NEXT:    i32.and $push6=, $3, $pop5
+; NO-SIMD128-FAST-NEXT:    i32.and $push6=, $pop5, $3
 ; NO-SIMD128-FAST-NEXT:    i32.store8 2($0), $pop6
 ; NO-SIMD128-FAST-NEXT:    i32.const $push7=, 3
 ; NO-SIMD128-FAST-NEXT:    i32.add $push8=, $0, $pop7
 ; NO-SIMD128-FAST-NEXT:    i32.const $push67=, -1
 ; NO-SIMD128-FAST-NEXT:    i32.xor $push9=, $20, $pop67
-; NO-SIMD128-FAST-NEXT:    i32.and $push10=, $4, $pop9
+; NO-SIMD128-FAST-NEXT:    i32.and $push10=, $pop9, $4
 ; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop8), $pop10
 ; NO-SIMD128-FAST-NEXT:    i32.const $push66=, -1
 ; NO-SIMD128-FAST-NEXT:    i32.xor $push11=, $21, $pop66
-; NO-SIMD128-FAST-NEXT:    i32.and $push12=, $5, $pop11
+; NO-SIMD128-FAST-NEXT:    i32.and $push12=, $pop11, $5
 ; NO-SIMD128-FAST-NEXT:    i32.store8 4($0), $pop12
 ; NO-SIMD128-FAST-NEXT:    i32.const $push13=, 5
 ; NO-SIMD128-FAST-NEXT:    i32.add $push14=, $0, $pop13
 ; NO-SIMD128-FAST-NEXT:    i32.const $push65=, -1
 ; NO-SIMD128-FAST-NEXT:    i32.xor $push15=, $22, $pop65
-; NO-SIMD128-FAST-NEXT:    i32.and $push16=, $6, $pop15
+; NO-SIMD128-FAST-NEXT:    i32.and $push16=, $pop15, $6
 ; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop14), $pop16
 ; NO-SIMD128-FAST-NEXT:    i32.const $push17=, 6
 ; NO-SIMD128-FAST-NEXT:    i32.add $push18=, $0, $pop17
 ; NO-SIMD128-FAST-NEXT:    i32.const $push64=, -1
 ; NO-SIMD128-FAST-NEXT:    i32.xor $push19=, $23, $pop64
-; NO-SIMD128-FAST-NEXT:    i32.and $push20=, $7, $pop19
+; NO-SIMD128-FAST-NEXT:    i32.and $push20=, $pop19, $7
 ; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop18), $pop20
 ; NO-SIMD128-FAST-NEXT:    i32.const $push21=, 7
 ; NO-SIMD128-FAST-NEXT:    i32.add $push22=, $0, $pop21
 ; NO-SIMD128-FAST-NEXT:    i32.const $push63=, -1
 ; NO-SIMD128-FAST-NEXT:    i32.xor $push23=, $24, $pop63
-; NO-SIMD128-FAST-NEXT:    i32.and $push24=, $8, $pop23
+; NO-SIMD128-FAST-NEXT:    i32.and $push24=, $pop23, $8
 ; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop22), $pop24
 ; NO-SIMD128-FAST-NEXT:    i32.const $push62=, -1
 ; NO-SIMD128-FAST-NEXT:    i32.xor $push25=, $25, $pop62
-; NO-SIMD128-FAST-NEXT:    i32.and $push26=, $9, $pop25
+; NO-SIMD128-FAST-NEXT:    i32.and $push26=, $pop25, $9
 ; NO-SIMD128-FAST-NEXT:    i32.store8 8($0), $pop26
 ; NO-SIMD128-FAST-NEXT:    i32.const $push27=, 9
 ; NO-SIMD128-FAST-NEXT:    i32.add $push28=, $0, $pop27
 ; NO-SIMD128-FAST-NEXT:    i32.const $push61=, -1
 ; NO-SIMD128-FAST-NEXT:    i32.xor $push29=, $26, $pop61
-; NO-SIMD128-FAST-NEXT:    i32.and $push30=, $10, $pop29
+; NO-SIMD128-FAST-NEXT:    i32.and $push30=, $pop29, $10
 ; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop28), $pop30
 ; NO-SIMD128-FAST-NEXT:    i32.const $push31=, 10
 ; NO-SIMD128-FAST-NEXT:    i32.add $push32=, $0, $pop31
 ; NO-SIMD128-FAST-NEXT:    i32.const $push60=, -1
 ; NO-SIMD128-FAST-NEXT:    i32.xor $push33=, $27, $pop60
-; NO-SIMD128-FAST-NEXT:    i32.and $push34=, $11, $pop33
+; NO-SIMD128-FAST-NEXT:    i32.and $push34=, $pop33, $11
 ; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop32), $pop34
 ; NO-SIMD128-FAST-NEXT:    i32.const $push35=, 11
 ; NO-SIMD128-FAST-NEXT:    i32.add $push36=, $0, $pop35
 ; NO-SIMD128-FAST-NEXT:    i32.const $push59=, -1
 ; NO-SIMD128-FAST-NEXT:    i32.xor $push37=, $28, $pop59
-; NO-SIMD128-FAST-NEXT:    i32.and $push38=, $12, $pop37
+; NO-SIMD128-FAST-NEXT:    i32.and $push38=, $pop37, $12
 ; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop36), $pop38
 ; NO-SIMD128-FAST-NEXT:    i32.const $push39=, 12
 ; NO-SIMD128-FAST-NEXT:    i32.add $push40=, $0, $pop39
 ; NO-SIMD128-FAST-NEXT:    i32.const $push58=, -1
 ; NO-SIMD128-FAST-NEXT:    i32.xor $push41=, $29, $pop58
-; NO-SIMD128-FAST-NEXT:    i32.and $push42=, $13, $pop41
+; NO-SIMD128-FAST-NEXT:    i32.and $push42=, $pop41, $13
 ; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop40), $pop42
 ; NO-SIMD128-FAST-NEXT:    i32.const $push43=, 13
 ; NO-SIMD128-FAST-NEXT:    i32.add $push44=, $0, $pop43
 ; NO-SIMD128-FAST-NEXT:    i32.const $push57=, -1
 ; NO-SIMD128-FAST-NEXT:    i32.xor $push45=, $30, $pop57
-; NO-SIMD128-FAST-NEXT:    i32.and $push46=, $14, $pop45
+; NO-SIMD128-FAST-NEXT:    i32.and $push46=, $pop45, $14
 ; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop44), $pop46
 ; NO-SIMD128-FAST-NEXT:    i32.const $push47=, 14
 ; NO-SIMD128-FAST-NEXT:    i32.add $push48=, $0, $pop47
 ; NO-SIMD128-FAST-NEXT:    i32.const $push56=, -1
 ; NO-SIMD128-FAST-NEXT:    i32.xor $push49=, $31, $pop56
-; NO-SIMD128-FAST-NEXT:    i32.and $push50=, $15, $pop49
+; NO-SIMD128-FAST-NEXT:    i32.and $push50=, $pop49, $15
 ; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop48), $pop50
 ; NO-SIMD128-FAST-NEXT:    i32.const $push51=, 15
 ; NO-SIMD128-FAST-NEXT:    i32.add $push52=, $0, $pop51
 ; NO-SIMD128-FAST-NEXT:    i32.const $push55=, -1
 ; NO-SIMD128-FAST-NEXT:    i32.xor $push53=, $32, $pop55
-; NO-SIMD128-FAST-NEXT:    i32.and $push54=, $16, $pop53
+; NO-SIMD128-FAST-NEXT:    i32.and $push54=, $pop53, $16
 ; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop52), $pop54
 ; NO-SIMD128-FAST-NEXT:    return
  %inv_y = xor <16 x i8> %y,
    <i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1,
     i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1>
  %a = and <16 x i8> %x, %inv_y
  ret <16 x i8> %a
 }

 ; SIMD128:         .functype bitselect_xor_reversed_v16i8 (v128, v128, v128) -> (v128)
 ; SIMD128-NEXT:  # %bb.0:
 ; SIMD128-NEXT:    v128.bitselect $push0=, $2, $1, $0
 ; SIMD128-NEXT:    return $pop0
 ;
 ; SIMD128-FAST-LABEL: bitselect_xor_reversed_v16i8:
 ; SIMD128-FAST:         .functype bitselect_xor_reversed_v16i8 (v128, v128, v128) -> (v128)
 ; SIMD128-FAST-NEXT:  # %bb.0:
-; SIMD128-FAST-NEXT:    v128.xor $push1=, $1, $2
-; SIMD128-FAST-NEXT:    v128.not $push2=, $0
-; SIMD128-FAST-NEXT:    v128.and $push3=, $pop1, $pop2
-; SIMD128-FAST-NEXT:    v128.xor $push0=, $pop3, $2
+; SIMD128-FAST-NEXT:    v128.xor $push2=, $1, $2
+; SIMD128-FAST-NEXT:    v128.and $push1=, $pop2, $0
+; SIMD128-FAST-NEXT:    v128.xor $push0=, $pop1, $1
 ; SIMD128-FAST-NEXT:    return $pop0
 ;
 ; NO-SIMD128-LABEL: bitselect_xor_reversed_v16i8:
 ; NO-SIMD128:         .functype bitselect_xor_reversed_v16i8 (i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32) -> ()
 ; NO-SIMD128-NEXT:  # %bb.0:
-; NO-SIMD128-NEXT:    i32.const $push5=, 15
-; NO-SIMD128-NEXT:    i32.add $push6=, $0, $pop5
-; NO-SIMD128-NEXT:    i32.xor $push2=, $32, $48
-; NO-SIMD128-NEXT:    i32.const $push0=, -1
-; NO-SIMD128-NEXT:    i32.xor $push1=, $16, $pop0
-; NO-SIMD128-NEXT:    i32.and $push3=, $pop2, $pop1
-; NO-SIMD128-NEXT:    i32.xor $push4=, $pop3, $48
-; NO-SIMD128-NEXT:    i32.store8 0($pop6), $pop4
-; NO-SIMD128-NEXT:    i32.const $push11=, 14
-; NO-SIMD128-NEXT:    i32.add $push12=, $0, $pop11
-; NO-SIMD128-NEXT:    i32.xor $push8=, $31, $47
-; NO-SIMD128-NEXT:    i32.const $push101=, -1
-; NO-SIMD128-NEXT:    i32.xor $push7=, $15, $pop101
-; NO-SIMD128-NEXT:    i32.and $push9=, $pop8, $pop7
-; NO-SIMD128-NEXT:    i32.xor $push10=, $pop9, $47
-; NO-SIMD128-NEXT:    i32.store8 0($pop12), $pop10
-; NO-SIMD128-NEXT:    i32.const $push17=, 13
-; NO-SIMD128-NEXT:    i32.add $push18=, $0, $pop17
-; NO-SIMD128-NEXT:    i32.xor $push14=, $30, $46
-; NO-SIMD128-NEXT:    i32.const $push100=, -1
-; NO-SIMD128-NEXT:    i32.xor $push13=, $14, $pop100
-; NO-SIMD128-NEXT:    i32.and $push15=, $pop14, $pop13
-; NO-SIMD128-NEXT:    i32.xor $push16=, $pop15, $46
-; NO-SIMD128-NEXT:    i32.store8 0($pop18), $pop16
-; NO-SIMD128-NEXT:    i32.const $push23=, 12
+; NO-SIMD128-NEXT:    i32.const $push3=, 15
+; NO-SIMD128-NEXT:    i32.add $push4=, $0, $pop3
+; NO-SIMD128-NEXT:    i32.xor $push0=, $32, $48
+; NO-SIMD128-NEXT:    i32.and $push1=, $pop0, $16
+; NO-SIMD128-NEXT:    i32.xor $push2=, $pop1, $32
+; NO-SIMD128-NEXT:    i32.store8 0($pop4), $pop2
+; NO-SIMD128-NEXT:    i32.const $push8=, 14
+; NO-SIMD128-NEXT:    i32.add $push9=, $0, $pop8
+; NO-SIMD128-NEXT:    i32.xor $push5=, $31, $47
+; NO-SIMD128-NEXT:    i32.and $push6=, $pop5, $15
+; NO-SIMD128-NEXT:    i32.xor $push7=, $pop6, $31
+; NO-SIMD128-NEXT:    i32.store8 0($pop9), $pop7
+; NO-SIMD128-NEXT:    i32.const $push13=, 13
+; NO-SIMD128-NEXT:    i32.add $push14=, $0, $pop13
+; NO-SIMD128-NEXT:    i32.xor $push10=, $30, $46
+; NO-SIMD128-NEXT:    i32.and $push11=, $pop10, $14
+; NO-SIMD128-NEXT:    i32.xor $push12=, $pop11, $30
+; NO-SIMD128-NEXT:    i32.store8 0($pop14), $pop12
+; NO-SIMD128-NEXT:    i32.const $push18=, 12
+; NO-SIMD128-NEXT:    i32.add $push19=, $0, $pop18
+; NO-SIMD128-NEXT:    i32.xor $push15=, $29, $45
+; NO-SIMD128-NEXT:    i32.and $push16=, $pop15, $13
+; NO-SIMD128-NEXT:    i32.xor $push17=, $pop16, $29
+; NO-SIMD128-NEXT:    i32.store8 0($pop19), $pop17
+; NO-SIMD128-NEXT:    i32.const $push23=, 11
 ; NO-SIMD128-NEXT:    i32.add $push24=, $0, $pop23
-; NO-SIMD128-NEXT:    i32.xor $push20=, $29, $45
-; NO-SIMD128-NEXT:    i32.const $push99=, -1
-; NO-SIMD128-NEXT:    i32.xor $push19=, $13, $pop99
-; NO-SIMD128-NEXT:    i32.and $push21=, $pop20, $pop19
-; NO-SIMD128-NEXT:    i32.xor $push22=, $pop21, $45
+; NO-SIMD128-NEXT:    i32.xor $push20=, $28, $44
+; NO-SIMD128-NEXT:    i32.and $push21=, $pop20, $12
+; NO-SIMD128-NEXT:    i32.xor $push22=, $pop21, $28
 ; NO-SIMD128-NEXT:    i32.store8 0($pop24), $pop22
-; NO-SIMD128-NEXT:    i32.const $push29=, 11
-; NO-SIMD128-NEXT:    i32.add $push30=, $0, $pop29
-; NO-SIMD128-NEXT:    i32.xor $push26=, $28, $44
-; NO-SIMD128-NEXT:    i32.const $push98=, -1
-; NO-SIMD128-NEXT:    i32.xor $push25=, $12, $pop98
-; NO-SIMD128-NEXT:    i32.and $push27=, $pop26, $pop25
-; NO-SIMD128-NEXT:    i32.xor $push28=, $pop27, $44
-; NO-SIMD128-NEXT:    i32.store8 0($pop30), $pop28
-; NO-SIMD128-NEXT:    i32.const $push35=, 10
-; NO-SIMD128-NEXT:    i32.add $push36=, $0, $pop35
-; NO-SIMD128-NEXT:    i32.xor $push32=, $27, $43
-; NO-SIMD128-NEXT:    i32.const $push97=, -1
-; NO-SIMD128-NEXT:    i32.xor $push31=, $11, $pop97
-; NO-SIMD128-NEXT:    i32.and $push33=, $pop32, $pop31
-; NO-SIMD128-NEXT:    i32.xor $push34=, $pop33, $43
-; NO-SIMD128-NEXT:    i32.store8 0($pop36), $pop34
-; NO-SIMD128-NEXT:    i32.const $push41=, 9
+; NO-SIMD128-NEXT:    i32.const $push28=, 10
+; NO-SIMD128-NEXT:    i32.add $push29=, $0, $pop28
+; NO-SIMD128-NEXT:    i32.xor $push25=, $27, $43
+; NO-SIMD128-NEXT:    i32.and $push26=, $pop25, $11
+; NO-SIMD128-NEXT:    i32.xor $push27=, $pop26, $27
+; NO-SIMD128-NEXT:    i32.store8 0($pop29), $pop27
+; NO-SIMD128-NEXT:    i32.const $push33=, 9
+; NO-SIMD128-NEXT:    i32.add $push34=, $0, $pop33
+; NO-SIMD128-NEXT:    i32.xor $push30=, $26, $42
+; NO-SIMD128-NEXT:    i32.and $push31=, $pop30, $10
+; NO-SIMD128-NEXT:    i32.xor $push32=, $pop31, $26
+; NO-SIMD128-NEXT:    i32.store8 0($pop34), $pop32
+; NO-SIMD128-NEXT:    i32.xor $push35=, $25, $41
+; NO-SIMD128-NEXT:    i32.and $push36=, $pop35, $9
+; NO-SIMD128-NEXT:    i32.xor $push37=, $pop36, $25
+; NO-SIMD128-NEXT:    i32.store8 8($0), $pop37
+; NO-SIMD128-NEXT:    i32.const $push41=, 7
 ; NO-SIMD128-NEXT:    i32.add $push42=, $0, $pop41
-; NO-SIMD128-NEXT:    i32.xor $push38=, $26, $42
-; NO-SIMD128-NEXT:    i32.const $push96=, -1
-; NO-SIMD128-NEXT:    i32.xor $push37=, $10, $pop96
-; NO-SIMD128-NEXT:    i32.and $push39=, $pop38, $pop37
-; NO-SIMD128-NEXT:    i32.xor $push40=, $pop39, $42
+; NO-SIMD128-NEXT:    i32.xor $push38=, $24, $40
+; NO-SIMD128-NEXT:    i32.and $push39=, $pop38, $8
+; NO-SIMD128-NEXT:    i32.xor $push40=, $pop39, $24
 ; NO-SIMD128-NEXT:    i32.store8 0($pop42), $pop40
-; NO-SIMD128-NEXT:    i32.xor $push44=, $25, $41
-; NO-SIMD128-NEXT:    i32.const $push95=, -1
-; NO-SIMD128-NEXT:    i32.xor $push43=, $9, $pop95
-; NO-SIMD128-NEXT:    i32.and $push45=, $pop44, $pop43
-; NO-SIMD128-NEXT:    i32.xor $push46=, $pop45, $41
-; NO-SIMD128-NEXT:    i32.store8 8($0), $pop46
-; NO-SIMD128-NEXT:    i32.const $push51=, 7
+; NO-SIMD128-NEXT:    i32.const $push46=, 6
+; NO-SIMD128-NEXT:    i32.add $push47=, $0, $pop46
+; NO-SIMD128-NEXT:    i32.xor $push43=, $23, $39
+; NO-SIMD128-NEXT:    i32.and $push44=, $pop43, $7
+; NO-SIMD128-NEXT:    i32.xor $push45=, $pop44, $23
+; NO-SIMD128-NEXT:    i32.store8 0($pop47), $pop45
+; NO-SIMD128-NEXT:    i32.const $push51=, 5
 ; NO-SIMD128-NEXT:    i32.add $push52=, $0, $pop51
-; NO-SIMD128-NEXT:    i32.xor $push48=, $24, $40
-; NO-SIMD128-NEXT:    i32.const $push94=, -1
-; NO-SIMD128-NEXT:    i32.xor $push47=, $8, $pop94
-; NO-SIMD128-NEXT:    i32.and $push49=, $pop48, $pop47
-; NO-SIMD128-NEXT:    i32.xor $push50=, $pop49, $40
+; NO-SIMD128-NEXT:    i32.xor $push48=, $22, $38
+; NO-SIMD128-NEXT:    i32.and $push49=, $pop48, $6
+; NO-SIMD128-NEXT:    i32.xor $push50=, $pop49, $22
 ; NO-SIMD128-NEXT:    i32.store8 0($pop52), $pop50
-; NO-SIMD128-NEXT:    i32.const $push57=, 6
-; NO-SIMD128-NEXT:    i32.add $push58=, $0, $pop57
-; NO-SIMD128-NEXT:    i32.xor $push54=, $23, $39
-; NO-SIMD128-NEXT:    i32.const $push93=, -1
-; NO-SIMD128-NEXT:    i32.xor $push53=, $7, $pop93
-; NO-SIMD128-NEXT:    i32.and $push55=, $pop54, $pop53
-; NO-SIMD128-NEXT:    i32.xor $push56=, $pop55, $39
-; NO-SIMD128-NEXT:    i32.store8 0($pop58), $pop56
-; NO-SIMD128-NEXT:    i32.const $push63=, 5
-; NO-SIMD128-NEXT:    i32.add $push64=, $0, $pop63
-; NO-SIMD128-NEXT:    i32.xor $push60=, $22, $38
-; NO-SIMD128-NEXT:    i32.const $push92=, -1
-; NO-SIMD128-NEXT:    i32.xor $push59=, $6, $pop92
-; NO-SIMD128-NEXT:    i32.and $push61=, $pop60, $pop59
-; NO-SIMD128-NEXT:    i32.xor $push62=, $pop61, $38
-; NO-SIMD128-NEXT:    i32.store8 0($pop64), $pop62
-; NO-SIMD128-NEXT:    i32.xor $push66=, $21, $37
-; NO-SIMD128-NEXT:    i32.const $push91=, -1
-; NO-SIMD128-NEXT:    i32.xor $push65=, $5, $pop91
-; NO-SIMD128-NEXT:    i32.and $push67=, $pop66, $pop65
-; NO-SIMD128-NEXT:    i32.xor $push68=, $pop67, $37
-; NO-SIMD128-NEXT:    i32.store8 4($0), $pop68
-; NO-SIMD128-NEXT:    i32.const $push73=, 3
-; NO-SIMD128-NEXT:    i32.add $push74=, $0, $pop73
-; NO-SIMD128-NEXT:    i32.xor $push70=, $20, $36
-; NO-SIMD128-NEXT:    i32.const $push90=, -1
-; NO-SIMD128-NEXT:    i32.xor $push69=, $4, $pop90
-; NO-SIMD128-NEXT:    i32.and $push71=, $pop70, $pop69
-; NO-SIMD128-NEXT:    i32.xor $push72=, $pop71, $36
-; NO-SIMD128-NEXT:    i32.store8 0($pop74), $pop72
-; NO-SIMD128-NEXT:    i32.xor $push76=, $19, $35
-; NO-SIMD128-NEXT:    i32.const $push89=, -1
-; NO-SIMD128-NEXT:    i32.xor $push75=, $3, $pop89
-; NO-SIMD128-NEXT:    i32.and $push77=, $pop76, $pop75
-; NO-SIMD128-NEXT:    i32.xor $push78=, $pop77, $35
-; NO-SIMD128-NEXT:    i32.store8 2($0), $pop78
-; NO-SIMD128-NEXT:    i32.xor $push80=, $18, $34
-; NO-SIMD128-NEXT:    i32.const $push88=, -1
-; NO-SIMD128-NEXT:    i32.xor $push79=, $2, $pop88
-; NO-SIMD128-NEXT:    i32.and $push81=, $pop80, $pop79
-; NO-SIMD128-NEXT:    i32.xor $push82=, $pop81, $34
-; NO-SIMD128-NEXT:    i32.store8 1($0), $pop82
-; NO-SIMD128-NEXT:    i32.xor $push84=, $17, $33
-; NO-SIMD128-NEXT:    i32.const $push87=, -1
-; NO-SIMD128-NEXT:    i32.xor $push83=, $1, $pop87
-; NO-SIMD128-NEXT:    i32.and $push85=, $pop84, $pop83
-; NO-SIMD128-NEXT:    i32.xor $push86=, $pop85, $33
-; NO-SIMD128-NEXT:    i32.store8 0($0), $pop86
+; NO-SIMD128-NEXT:    i32.xor $push53=, $21, $37
+; NO-SIMD128-NEXT:    i32.and $push54=, $pop53, $5
+; NO-SIMD128-NEXT:    i32.xor $push55=, $pop54, $21
+; NO-SIMD128-NEXT:    i32.store8 4($0), $pop55
+; NO-SIMD128-NEXT:    i32.const $push59=, 3
+; NO-SIMD128-NEXT:    i32.add $push60=, $0, $pop59
+; NO-SIMD128-NEXT:    i32.xor $push56=, $20, $36
+; NO-SIMD128-NEXT:    i32.and $push57=, $pop56, $4
+; NO-SIMD128-NEXT:    i32.xor $push58=, $pop57, $20
+; NO-SIMD128-NEXT:    i32.store8 0($pop60), $pop58
+; NO-SIMD128-NEXT:    i32.xor $push61=, $19, $35
+; NO-SIMD128-NEXT:    i32.and $push62=, $pop61, $3
+; NO-SIMD128-NEXT:    i32.xor $push63=, $pop62, $19
+; NO-SIMD128-NEXT:    i32.store8 2($0), $pop63
+; NO-SIMD128-NEXT:    i32.xor $push64=, $18, $34
+; NO-SIMD128-NEXT:    i32.and $push65=, $pop64, $2
+; NO-SIMD128-NEXT:    i32.xor $push66=, $pop65, $18
+; NO-SIMD128-NEXT:    i32.store8 1($0), $pop66
+; NO-SIMD128-NEXT:    i32.xor $push67=, $17, $33
+; NO-SIMD128-NEXT:    i32.and $push68=, $pop67, $1
+; NO-SIMD128-NEXT:    i32.xor $push69=, $pop68, $17
+; NO-SIMD128-NEXT:    i32.store8 0($0), $pop69
 ; NO-SIMD128-NEXT:    return
 ;
 ; NO-SIMD128-FAST-LABEL: bitselect_xor_reversed_v16i8:
 ; NO-SIMD128-FAST:         .functype bitselect_xor_reversed_v16i8 (i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32) -> ()
 ; NO-SIMD128-FAST-NEXT:  # %bb.0:
-; NO-SIMD128-FAST-NEXT:    i32.xor $push2=, $17, $33
-; NO-SIMD128-FAST-NEXT:    i32.const $push0=, -1
-; NO-SIMD128-FAST-NEXT:    i32.xor $push1=, $1, $pop0
-; NO-SIMD128-FAST-NEXT:    i32.and $push3=, $pop2, $pop1
-; NO-SIMD128-FAST-NEXT:    i32.xor $push4=, $pop3, $33
-; NO-SIMD128-FAST-NEXT:    i32.store8 0($0), $pop4
-; NO-SIMD128-FAST-NEXT:    i32.xor $push6=, $18, $34
-; NO-SIMD128-FAST-NEXT:    i32.const $push101=, -1
-; NO-SIMD128-FAST-NEXT:    i32.xor $push5=, $2, $pop101
-; NO-SIMD128-FAST-NEXT:    i32.and $push7=, $pop6, $pop5
-; NO-SIMD128-FAST-NEXT:    i32.xor $push8=, $pop7, $34
-; NO-SIMD128-FAST-NEXT:    i32.store8 1($0), $pop8
-; NO-SIMD128-FAST-NEXT:    i32.xor $push10=, $19, $35
-; NO-SIMD128-FAST-NEXT:    i32.const $push100=, -1
-; NO-SIMD128-FAST-NEXT:    i32.xor $push9=, $3, $pop100
-; NO-SIMD128-FAST-NEXT:    i32.and $push11=, $pop10, $pop9
-; NO-SIMD128-FAST-NEXT:    i32.xor $push12=, $pop11, $35
-; NO-SIMD128-FAST-NEXT:    i32.store8 2($0), $pop12
-; NO-SIMD128-FAST-NEXT:    i32.const $push17=, 3
+; NO-SIMD128-FAST-NEXT:    i32.xor $push0=, $17, $33
+; NO-SIMD128-FAST-NEXT:    i32.and $push1=, $pop0, $1
+; NO-SIMD128-FAST-NEXT:    i32.xor $push2=, $pop1, $17
+; NO-SIMD128-FAST-NEXT:    i32.store8 0($0), $pop2
+; NO-SIMD128-FAST-NEXT:    i32.xor $push3=, $18, $34
+; NO-SIMD128-FAST-NEXT:    i32.and $push4=, $pop3, $2
+; NO-SIMD128-FAST-NEXT:    i32.xor $push5=, $pop4, $18
+; NO-SIMD128-FAST-NEXT:    i32.store8 1($0), $pop5
+; NO-SIMD128-FAST-NEXT:    i32.xor $push6=, $19, $35
+; NO-SIMD128-FAST-NEXT:    i32.and $push7=, $pop6, $3
+; NO-SIMD128-FAST-NEXT:    i32.xor $push8=, $pop7, $19
+; NO-SIMD128-FAST-NEXT:    i32.store8 2($0), $pop8
+; NO-SIMD128-FAST-NEXT:    i32.const $push9=, 3
+; NO-SIMD128-FAST-NEXT:    i32.add $push10=, $0, $pop9
+; NO-SIMD128-FAST-NEXT:    i32.xor $push11=, $20, $36
+; NO-SIMD128-FAST-NEXT:    i32.and $push12=, $pop11, $4
+; NO-SIMD128-FAST-NEXT:    i32.xor $push13=, $pop12, $20
+; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop10), $pop13
+; NO-SIMD128-FAST-NEXT:    i32.xor $push14=, $21, $37
+; NO-SIMD128-FAST-NEXT:    i32.and $push15=, $pop14, $5
+; NO-SIMD128-FAST-NEXT:    i32.xor $push16=, $pop15, $21
+; NO-SIMD128-FAST-NEXT:    i32.store8 4($0), $pop16
+; NO-SIMD128-FAST-NEXT:    i32.const $push17=, 5
 ; NO-SIMD128-FAST-NEXT:    i32.add $push18=, $0, $pop17
-; NO-SIMD128-FAST-NEXT:    i32.xor $push14=, $20, $36
-; NO-SIMD128-FAST-NEXT:    i32.const $push99=, -1
-; NO-SIMD128-FAST-NEXT:    i32.xor $push13=, $4, $pop99
-; NO-SIMD128-FAST-NEXT:    i32.and $push15=, $pop14, $pop13
-; NO-SIMD128-FAST-NEXT:    i32.xor $push16=, $pop15, $36
-; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop18), $pop16
-; NO-SIMD128-FAST-NEXT:    i32.xor $push20=, $21, $37
-; NO-SIMD128-FAST-NEXT:    i32.const $push98=, -1
-; NO-SIMD128-FAST-NEXT:    i32.xor $push19=, $5, $pop98
-; NO-SIMD128-FAST-NEXT:    i32.and $push21=, $pop20, $pop19
-; NO-SIMD128-FAST-NEXT:    i32.xor $push22=, $pop21, $37
-; NO-SIMD128-FAST-NEXT:    i32.store8 4($0), $pop22
-; NO-SIMD128-FAST-NEXT:    i32.const $push27=, 5
+; NO-SIMD128-FAST-NEXT:    i32.xor $push19=, $22, $38
+; NO-SIMD128-FAST-NEXT:    i32.and $push20=, $pop19, $6
+; NO-SIMD128-FAST-NEXT:    i32.xor $push21=, $pop20, $22
+; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop18), $pop21
+; NO-SIMD128-FAST-NEXT:    i32.const $push22=, 6
+; NO-SIMD128-FAST-NEXT:    i32.add $push23=, $0, $pop22
+; NO-SIMD128-FAST-NEXT:    i32.xor $push24=, $23, $39
+; NO-SIMD128-FAST-NEXT:    i32.and $push25=, $pop24, $7
+; NO-SIMD128-FAST-NEXT:    i32.xor $push26=, $pop25, $23
+; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop23), $pop26
+; NO-SIMD128-FAST-NEXT:    i32.const $push27=, 7
 ; NO-SIMD128-FAST-NEXT:    i32.add $push28=, $0, $pop27
-; NO-SIMD128-FAST-NEXT:    i32.xor $push24=, $22, $38
-; NO-SIMD128-FAST-NEXT:    i32.const $push97=, -1
-; NO-SIMD128-FAST-NEXT:    i32.xor $push23=, $6, $pop97
-; NO-SIMD128-FAST-NEXT:    i32.and $push25=, $pop24, $pop23
-; NO-SIMD128-FAST-NEXT:    i32.xor $push26=, $pop25, $38
-; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop28), $pop26
-; NO-SIMD128-FAST-NEXT:    i32.const $push33=, 6
-; NO-SIMD128-FAST-NEXT:    i32.add $push34=, $0, $pop33
-; NO-SIMD128-FAST-NEXT:    i32.xor $push30=, $23, $39
-; NO-SIMD128-FAST-NEXT:    i32.const $push96=, -1
-; NO-SIMD128-FAST-NEXT:    i32.xor $push29=, $7, $pop96
-; NO-SIMD128-FAST-NEXT:    i32.and $push31=, $pop30, $pop29
-; NO-SIMD128-FAST-NEXT:    i32.xor $push32=, $pop31, $39
-; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop34), $pop32
-; NO-SIMD128-FAST-NEXT:    i32.const $push39=, 7
-; NO-SIMD128-FAST-NEXT:    i32.add $push40=, $0, $pop39
-; NO-SIMD128-FAST-NEXT:    i32.xor $push36=, $24, $40
-; NO-SIMD128-FAST-NEXT:    i32.const $push95=, -1
-; NO-SIMD128-FAST-NEXT:    i32.xor $push35=, $8, $pop95
-; NO-SIMD128-FAST-NEXT:    i32.and $push37=, $pop36, $pop35
-; NO-SIMD128-FAST-NEXT:    i32.xor $push38=, $pop37, $40
-; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop40), $pop38
-; NO-SIMD128-FAST-NEXT:    i32.xor $push42=, $25, $41
-; NO-SIMD128-FAST-NEXT:    i32.const $push94=, -1
-; NO-SIMD128-FAST-NEXT:    i32.xor $push41=, $9, $pop94
-; NO-SIMD128-FAST-NEXT:    i32.and $push43=, $pop42, $pop41
-; NO-SIMD128-FAST-NEXT:    i32.xor $push44=, $pop43, $41
-; NO-SIMD128-FAST-NEXT:    i32.store8 8($0), $pop44
-; NO-SIMD128-FAST-NEXT:    i32.const $push49=, 9
-; NO-SIMD128-FAST-NEXT:    i32.add $push50=, $0, $pop49
-; NO-SIMD128-FAST-NEXT:    i32.xor $push46=, $26, $42
-; NO-SIMD128-FAST-NEXT:    i32.const $push93=, -1
-; NO-SIMD128-FAST-NEXT:    i32.xor $push45=, $10, $pop93
-; NO-SIMD128-FAST-NEXT:    i32.and $push47=, $pop46, $pop45
-; NO-SIMD128-FAST-NEXT:    i32.xor $push48=, $pop47, $42
-; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop50), $pop48
-; NO-SIMD128-FAST-NEXT:    i32.const $push55=, 10
+; NO-SIMD128-FAST-NEXT:    i32.xor $push29=, $24, $40
+; NO-SIMD128-FAST-NEXT:    i32.and $push30=, $pop29, $8
+; NO-SIMD128-FAST-NEXT:    i32.xor $push31=, $pop30, $24
+; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop28), $pop31
+; NO-SIMD128-FAST-NEXT:    i32.xor $push32=, $25, $41
+; NO-SIMD128-FAST-NEXT:    i32.and $push33=, $pop32, $9
+; NO-SIMD128-FAST-NEXT:    i32.xor $push34=, $pop33, $25
+; NO-SIMD128-FAST-NEXT:    i32.store8 8($0), $pop34
+; NO-SIMD128-FAST-NEXT:    i32.const $push35=, 9
+; NO-SIMD128-FAST-NEXT:    i32.add $push36=, $0, $pop35
+; NO-SIMD128-FAST-NEXT:    i32.xor $push37=, $26, $42
+; NO-SIMD128-FAST-NEXT:    i32.and $push38=, $pop37, $10
+; NO-SIMD128-FAST-NEXT:    i32.xor $push39=, $pop38, $26
+; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop36), $pop39
+; NO-SIMD128-FAST-NEXT:    i32.const $push40=, 10
+; NO-SIMD128-FAST-NEXT:    i32.add $push41=, $0, $pop40
+; NO-SIMD128-FAST-NEXT:    i32.xor $push42=, $27, $43
+; NO-SIMD128-FAST-NEXT:    i32.and $push43=, $pop42, $11
+; NO-SIMD128-FAST-NEXT:    i32.xor $push44=, $pop43, $27
+; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop41), $pop44
+; NO-SIMD128-FAST-NEXT:    i32.const $push45=, 11
+; NO-SIMD128-FAST-NEXT:    i32.add $push46=, $0, $pop45
+; NO-SIMD128-FAST-NEXT:    i32.xor $push47=, $28, $44
+; NO-SIMD128-FAST-NEXT:    i32.and $push48=, $pop47, $12
+; NO-SIMD128-FAST-NEXT:    i32.xor $push49=, $pop48, $28
+; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop46), $pop49
+; NO-SIMD128-FAST-NEXT:    i32.const $push50=, 12
+; NO-SIMD128-FAST-NEXT:    i32.add $push51=, $0, $pop50
+; NO-SIMD128-FAST-NEXT:    i32.xor $push52=, $29, $45
+; NO-SIMD128-FAST-NEXT:    i32.and $push53=, $pop52, $13
+; NO-SIMD128-FAST-NEXT:    i32.xor $push54=, $pop53, $29
+; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop51), $pop54
+; NO-SIMD128-FAST-NEXT:    i32.const $push55=, 13
 ; NO-SIMD128-FAST-NEXT:    i32.add $push56=, $0, $pop55
-; NO-SIMD128-FAST-NEXT:    i32.xor $push52=, $27, $43
-; NO-SIMD128-FAST-NEXT:    i32.const $push92=, -1
-; NO-SIMD128-FAST-NEXT:    i32.xor $push51=, $11, $pop92
-; NO-SIMD128-FAST-NEXT:    i32.and $push53=, $pop52, $pop51
-; NO-SIMD128-FAST-NEXT:    i32.xor $push54=, $pop53, $43
-; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop56), $pop54
-; NO-SIMD128-FAST-NEXT:    i32.const $push61=, 11
-; NO-SIMD128-FAST-NEXT:    i32.add $push62=, $0, $pop61
-; NO-SIMD128-FAST-NEXT:    i32.xor $push58=, $28, $44
-; NO-SIMD128-FAST-NEXT:    i32.const $push91=, -1
-; NO-SIMD128-FAST-NEXT:    i32.xor $push57=, $12, $pop91
-; NO-SIMD128-FAST-NEXT:    i32.and $push59=, $pop58, $pop57
-; NO-SIMD128-FAST-NEXT:    i32.xor $push60=, $pop59, $44
-; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop62), $pop60
-; NO-SIMD128-FAST-NEXT:    i32.const $push67=, 12
-; NO-SIMD128-FAST-NEXT:    i32.add $push68=, $0, $pop67
-; NO-SIMD128-FAST-NEXT:    i32.xor $push64=, $29, $45
-; NO-SIMD128-FAST-NEXT:    i32.const $push90=, -1
-; NO-SIMD128-FAST-NEXT:    i32.xor $push63=, $13, $pop90
-; NO-SIMD128-FAST-NEXT:    i32.and $push65=, $pop64, $pop63
-; NO-SIMD128-FAST-NEXT:    i32.xor $push66=, $pop65, $45
-; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop68), $pop66
-; NO-SIMD128-FAST-NEXT:    i32.const $push73=, 13
-; NO-SIMD128-FAST-NEXT:    i32.add $push74=, $0, $pop73
-; NO-SIMD128-FAST-NEXT:    i32.xor $push70=, $30, $46
-; NO-SIMD128-FAST-NEXT:    i32.const $push89=, -1
-; NO-SIMD128-FAST-NEXT:    i32.xor $push69=, $14, $pop89
-; NO-SIMD128-FAST-NEXT:    i32.and $push71=, $pop70, $pop69
-; NO-SIMD128-FAST-NEXT:    i32.xor $push72=, $pop71, $46
-; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop74), $pop72
-; NO-SIMD128-FAST-NEXT:    i32.const $push79=, 14
-; NO-SIMD128-FAST-NEXT:    i32.add $push80=, $0, $pop79
-; NO-SIMD128-FAST-NEXT:    i32.xor $push76=, $31, $47
-; NO-SIMD128-FAST-NEXT:    i32.const $push88=, -1
-; NO-SIMD128-FAST-NEXT:    i32.xor $push75=, $15, $pop88
-; NO-SIMD128-FAST-NEXT:    i32.and $push77=, $pop76, $pop75
-; NO-SIMD128-FAST-NEXT:    i32.xor $push78=, $pop77, $47
-; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop80), $pop78
-; NO-SIMD128-FAST-NEXT:    i32.const $push85=, 15
-; NO-SIMD128-FAST-NEXT:    i32.add $push86=, $0, $pop85
-; NO-SIMD128-FAST-NEXT:    i32.xor $push82=, $32, $48
-; NO-SIMD128-FAST-NEXT:    i32.const $push87=, -1
-; NO-SIMD128-FAST-NEXT:    i32.xor $push81=, $16, $pop87
-; NO-SIMD128-FAST-NEXT:    i32.and $push83=, $pop82, $pop81
-; NO-SIMD128-FAST-NEXT:    i32.xor $push84=, $pop83, $48
-; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop86), $pop84
+; NO-SIMD128-FAST-NEXT:    i32.xor $push57=, $30, $46
+; NO-SIMD128-FAST-NEXT:    i32.and $push58=, $pop57, $14
+; NO-SIMD128-FAST-NEXT:    i32.xor $push59=, $pop58, $30
+; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop56), $pop59
+; NO-SIMD128-FAST-NEXT:    i32.const $push60=, 14
+; NO-SIMD128-FAST-NEXT:    i32.add $push61=, $0, $pop60
+; NO-SIMD128-FAST-NEXT:    i32.xor $push62=, $31, $47
+; NO-SIMD128-FAST-NEXT:    i32.and $push63=, $pop62, $15
+; NO-SIMD128-FAST-NEXT:    i32.xor $push64=, $pop63, $31
+; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop61), $pop64
+; NO-SIMD128-FAST-NEXT:    i32.const $push65=, 15
+; NO-SIMD128-FAST-NEXT:    i32.add $push66=, $0, $pop65
+; NO-SIMD128-FAST-NEXT:    i32.xor $push67=, $32, $48
+; NO-SIMD128-FAST-NEXT:    i32.and $push68=, $pop67, $16
+; NO-SIMD128-FAST-NEXT:    i32.xor $push69=, $pop68, $32
+; NO-SIMD128-FAST-NEXT:    i32.store8 0($pop66), $pop69
 ; NO-SIMD128-FAST-NEXT:    return
  %xor1 = xor <16 x i8> %v1, %v2
  %notc = xor <16 x i8> %c, <i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1,
                             i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1>
  %and = and <16 x i8> %xor1, %notc
  %a = xor <16 x i8> %and, %v2
  ret <16 x i8> %a
 }

 ; SIMD128-FAST:         .functype min_s_v8i16 (v128, v128) -> (v128)
 ; SIMD128-FAST-NEXT:  # %bb.0:
 ; SIMD128-FAST-NEXT:    i16x8.min_s $push0=, $0, $1
 ; SIMD128-FAST-NEXT:    return $pop0
 ;
 ; NO-SIMD128-LABEL: min_s_v8i16:
 ; NO-SIMD128:         .functype min_s_v8i16 (i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32) -> ()
 ; NO-SIMD128-NEXT:  # %bb.0:
-; NO-SIMD128-NEXT:    i32.const $push4=, 14
-; NO-SIMD128-NEXT:    i32.add $push5=, $0, $pop4
-; NO-SIMD128-NEXT:    i32.extend16_s $push1=, $8
-; NO-SIMD128-NEXT:    i32.extend16_s $push0=, $16
-; NO-SIMD128-NEXT:    i32.lt_s $push2=, $pop1, $pop0
-; NO-SIMD128-NEXT:    i32.select $push3=, $8, $16, $pop2
-; NO-SIMD128-NEXT:    i32.store16 0($pop5), $pop3
-; NO-SIMD128-NEXT:    i32.const $push10=, 12
+; NO-SIMD128-NEXT:    i32.const $push2=, 14
+; NO-SIMD128-NEXT:    i32.add $push3=, $0, $pop2
+; NO-SIMD128-NEXT:    i32.extend16_s $push55=, $8
+; NO-SIMD128-NEXT:    local.tee $push54=, $8=, $pop55
+; NO-SIMD128-NEXT:    i32.extend16_s $push53=, $16
+; NO-SIMD128-NEXT:    local.tee $push52=, $16=, $pop53
+; NO-SIMD128-NEXT:    i32.lt_s $push0=, $8, $16
+; NO-SIMD128-NEXT:    i32.select $push1=, $pop54, $pop52, $pop0
+; NO-SIMD128-NEXT:    i32.store16 0($pop3), $pop1
+; NO-SIMD128-NEXT:    i32.const $push6=, 12
+; NO-SIMD128-NEXT:    i32.add $push7=, $0, $pop6
+; NO-SIMD128-NEXT:    i32.extend16_s $push51=, $7
+; NO-SIMD128-NEXT:    local.tee $push50=, $16=, $pop51
+; NO-SIMD128-NEXT:    i32.extend16_s $push49=, $15
+; NO-SIMD128-NEXT:    local.tee $push48=, $8=, $pop49
+; NO-SIMD128-NEXT:    i32.lt_s $push4=, $16, $8
+; NO-SIMD128-NEXT:    i32.select $push5=, $pop50, $pop48, $pop4
+; NO-SIMD128-NEXT:    i32.store16 0($pop7), $pop5
+; NO-SIMD128-NEXT:    i32.const $push10=, 10
 ; NO-SIMD128-NEXT:    i32.add $push11=, $0, $pop10
-; NO-SIMD128-NEXT:    i32.extend16_s $push7=, $7
-; NO-SIMD128-NEXT:    i32.extend16_s $push6=, $15
-; NO-SIMD128-NEXT:    i32.lt_s $push8=, $pop7, $pop6
-; NO-SIMD128-NEXT:    i32.select $push9=, $7, $15, $pop8
+; NO-SIMD128-NEXT:    i32.extend16_s $push47=, $6
+; NO-SIMD128-NEXT:    local.tee $push46=, $16=, $pop47
+; NO-SIMD128-NEXT:    i32.extend16_s $push45=, $14
+; NO-SIMD128-NEXT:    local.tee $push44=, $8=, $pop45
+; NO-SIMD128-NEXT:    i32.lt_s $push8=, $16, $8
+; NO-SIMD128-NEXT:    i32.select $push9=, $pop46, $pop44, $pop8
 ; NO-SIMD128-NEXT:    i32.store16 0($pop11), $pop9
-; NO-SIMD128-NEXT:    i32.const $push16=, 10
+; NO-SIMD128-NEXT:    i32.extend16_s $push43=, $5
+; NO-SIMD128-NEXT:    local.tee $push42=, $16=, $pop43
+; NO-SIMD128-NEXT:    i32.extend16_s $push41=, $13
+; NO-SIMD128-NEXT:    local.tee $push40=, $8=, $pop41
+; NO-SIMD128-NEXT:    i32.lt_s $push12=, $16, $8
+; NO-SIMD128-NEXT:    i32.select $push13=, $pop42, $pop40, $pop12
+; NO-SIMD128-NEXT:    i32.store16 8($0), $pop13
+; NO-SIMD128-NEXT:    i32.const $push16=, 6
 ; NO-SIMD128-NEXT:    i32.add $push17=, $0, $pop16
-; NO-SIMD128-NEXT:    i32.extend16_s $push13=, $6
-; NO-SIMD128-NEXT:    i32.extend16_s $push12=, $14
-; NO-SIMD128-NEXT:    i32.lt_s $push14=, $pop13, $pop12
-; NO-SIMD128-NEXT:    i32.select $push15=, $6, $14, $pop14
+; NO-SIMD128-NEXT:    i32.extend16_s $push39=, $4
+; NO-SIMD128-NEXT:    local.tee $push38=, $16=, $pop39
+; NO-SIMD128-NEXT:    i32.extend16_s $push37=, $12
+; NO-SIMD128-NEXT:    local.tee $push36=, $8=, $pop37
+; NO-SIMD128-NEXT:    i32.lt_s $push14=, $16, $8
+; NO-SIMD128-NEXT:    i32.select $push15=, $pop38, $pop36, $pop14
 ; NO-SIMD128-NEXT:    i32.store16 0($pop17), $pop15
-; NO-SIMD128-NEXT:    i32.extend16_s $push19=, $5
-; NO-SIMD128-NEXT:    i32.extend16_s $push18=, $13
-; NO-SIMD128-NEXT:    i32.lt_s $push20=, $pop19, $pop18
-; NO-SIMD128-NEXT:    i32.select $push21=, $5, $13, $pop20
-; NO-SIMD128-NEXT:    i32.store16 8($0), $pop21
-; NO-SIMD128-NEXT:    i32.const $push26=, 6
-; NO-SIMD128-NEXT:    i32.add $push27=, $0, $pop26
-; NO-SIMD128-NEXT:    i32.extend16_s $push23=, $4
-; NO-SIMD128-NEXT:    i32.extend16_s $push22=, $12
-; NO-SIMD128-NEXT:    i32.lt_s $push24=, $pop23, $pop22
-; NO-SIMD128-NEXT:    i32.select $push25=, $4, $12, $pop24
-; NO-SIMD128-NEXT:    i32.store16 0($pop27), $pop25
-; NO-SIMD128-NEXT:    i32.extend16_s $push29=, $3
-; NO-SIMD128-NEXT:    i32.extend16_s $push28=, $11
-; NO-SIMD128-NEXT:    i32.lt_s $push30=, $pop29, $pop28
-; NO-SIMD128-NEXT:    i32.select $push31=, $3, $11, $pop30
-; NO-SIMD128-NEXT:    i32.store16 4($0), $pop31
-; NO-SIMD128-NEXT:    i32.extend16_s $push33=, $2
-; NO-SIMD128-NEXT:    i32.extend16_s $push32=, $10
-; NO-SIMD128-NEXT:    i32.lt_s $push34=, $pop33, $pop32
-; NO-SIMD128-NEXT:    i32.select $push35=, $2, $10, $pop34
-; NO-SIMD128-NEXT:    i32.store16 2($0), $pop35
-; NO-SIMD128-NEXT:    i32.extend16_s $push37=, $1
-; NO-SIMD128-NEXT:    i32.extend16_s $push36=, $9
-; NO-SIMD128-NEXT:    i32.lt_s $push38=, $pop37, $pop36
-; NO-SIMD128-NEXT:    i32.select $push39=, $1, $9, $pop38
-; NO-SIMD128-NEXT:    i32.store16 0($0), $pop39
+; NO-SIMD128-NEXT:    i32.extend16_s $push35=, $3
+; NO-SIMD128-NEXT:    local.tee $push34=, $16=, $pop35
+; NO-SIMD128-NEXT:    i32.extend16_s $push33=, $11
+; NO-SIMD128-NEXT:    local.tee $push32=, $8=, $pop33
+; NO-SIMD128-NEXT:    i32.lt_s $push18=, $16, $8
+; NO-SIMD128-NEXT:    i32.select $push19=, $pop34, $pop32, $pop18
+; NO-SIMD128-NEXT:    i32.store16 4($0), $pop19
+; NO-SIMD128-NEXT:    i32.extend16_s $push31=, $2
+; NO-SIMD128-NEXT:    local.tee $push30=, $16=, $pop31
+; NO-SIMD128-NEXT:    i32.extend16_s $push29=, $10
+; NO-SIMD128-NEXT:    local.tee $push28=, $8=, $pop29
+; NO-SIMD128-NEXT:    i32.lt_s $push20=, $16, $8
+; NO-SIMD128-NEXT:    i32.select $push21=, $pop30, $pop28, $pop20
+; NO-SIMD128-NEXT:    i32.store16 2($0), $pop21
+; NO-SIMD128-NEXT:    i32.extend16_s $push27=, $1
+; NO-SIMD128-NEXT:    local.tee $push26=, $16=, $pop27
+; NO-SIMD128-NEXT:    i32.extend16_s $push25=, $9
+; NO-SIMD128-NEXT:    local.tee $push24=, $8=, $pop25
+; NO-SIMD128-NEXT:    i32.lt_s $push22=, $16, $8
+; NO-SIMD128-NEXT:    i32.select $push23=, $pop26, $pop24, $pop22
+; NO-SIMD128-NEXT:    i32.store16 0($0), $pop23
 ; NO-SIMD128-NEXT:    return
 ;
 ; NO-SIMD128-FAST-LABEL: min_s_v8i16:
 ; NO-SIMD128-FAST:         .functype min_s_v8i16 (i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32) -> ()
 ; NO-SIMD128-FAST-NEXT:  # %bb.0:
-; NO-SIMD128-FAST-NEXT:    i32.extend16_s $push1=, $1
-; NO-SIMD128-FAST-NEXT:    i32.extend16_s $push0=, $9
-; NO-SIMD128-FAST-NEXT:    i32.lt_s $push2=, $pop1, $pop0
-; NO-SIMD128-FAST-NEXT:    i32.select $push3=, $1, $9, $pop2
-; NO-SIMD128-FAST-NEXT:    i32.store16 0($0), $pop3
-; NO-SIMD128-FAST-NEXT:    i32.extend16_s $push5=, $2
-; NO-SIMD128-FAST-NEXT:    i32.extend16_s $push4=, $10
-; NO-SIMD128-FAST-NEXT:    i32.lt_s $push6=, $pop5, $pop4
-; NO-SIMD128-FAST-NEXT:    i32.select $push7=, $2, $10, $pop6
-; NO-SIMD128-FAST-NEXT:    i32.store16 2($0), $pop7
-; NO-SIMD128-FAST-NEXT:    i32.extend16_s $push9=, $3
-; NO-SIMD128-FAST-NEXT:    i32.extend16_s $push8=, $11
-; NO-SIMD128-FAST-NEXT:    i32.lt_s $push10=, $pop9, $pop8
-; NO-SIMD128-FAST-NEXT:    i32.select $push11=, $3, $11, $pop10
-; NO-SIMD128-FAST-NEXT:    i32.store16 4($0), $pop11
-; NO-SIMD128-FAST-NEXT:    i32.const $push16=, 6
-; NO-SIMD128-FAST-NEXT:    i32.add $push17=, $0, $pop16
-; NO-SIMD128-FAST-NEXT:    i32.extend16_s $push13=, $4
-; NO-SIMD128-FAST-NEXT:    i32.extend16_s $push12=, $12
-; NO-SIMD128-FAST-NEXT:    i32.lt_s $push14=, $pop13, $pop12
-; NO-SIMD128-FAST-NEXT:    i32.select $push15=, $4, $12, $pop14
-; NO-SIMD128-FAST-NEXT:    i32.store16 0($pop17), $pop15
-; NO-SIMD128-FAST-NEXT:    i32.extend16_s $push19=, $5
-; NO-SIMD128-FAST-NEXT:    i32.extend16_s $push18=, $13
-; NO-SIMD128-FAST-NEXT:    i32.lt_s $push20=, $pop19, $pop18
-; NO-SIMD128-FAST-NEXT:    i32.select $push21=, $5, $13, $pop20
-; NO-SIMD128-FAST-NEXT:    i32.store16 8($0), $pop21
-; NO-SIMD128-FAST-NEXT:    i32.const $push26=, 10
-; NO-SIMD128-FAST-NEXT:    i32.add $push27=, $0, $pop26
-; NO-SIMD128-FAST-NEXT:    i32.extend16_s $push23=, $6
-; NO-SIMD128-FAST-NEXT:    i32.extend16_s $push22=, $14
-; NO-SIMD128-FAST-NEXT:    i32.lt_s $push24=, $pop23, $pop22
-; NO-SIMD128-FAST-NEXT:    i32.select $push25=, $6, $14, $pop24
-; NO-SIMD128-FAST-NEXT:    i32.store16 0($pop27), $pop25
-; NO-SIMD128-FAST-NEXT:    i32.const $push32=, 12
-; NO-SIMD128-FAST-NEXT:    i32.add $push33=, $0, $pop32
-; NO-SIMD128-FAST-NEXT:    i32.extend16_s $push29=, $7
-; NO-SIMD128-FAST-NEXT:    i32.extend16_s $push28=, $15
-; NO-SIMD128-FAST-NEXT:    i32.lt_s $push30=, $pop29, $pop28
-; NO-SIMD128-FAST-NEXT:    i32.select $push31=, $7, $15, $pop30
-; NO-SIMD128-FAST-NEXT:    i32.store16 0($pop33), $pop31
-; NO-SIMD128-FAST-NEXT:    i32.const $push38=, 14
-; NO-SIMD128-FAST-NEXT:    i32.add $push39=, $0, $pop38
-; NO-SIMD128-FAST-NEXT:    i32.extend16_s $push35=, $8
-; NO-SIMD128-FAST-NEXT:    i32.extend16_s $push34=, $16
-; NO-SIMD128-FAST-NEXT:    i32.lt_s $push36=, $pop35, $pop34
-; NO-SIMD128-FAST-NEXT:    i32.select $push37=, $8, $16, $pop36
-; NO-SIMD128-FAST-NEXT:    i32.store16 0($pop39), $pop37
+; NO-SIMD128-FAST-NEXT:    i32.extend16_s $push55=, $1
+; NO-SIMD128-FAST-NEXT:    local.tee $push54=, $1=, $pop55
+; NO-SIMD128-FAST-NEXT:    i32.extend16_s $push53=, $9
+; NO-SIMD128-FAST-NEXT:    local.tee $push52=, $9=, $pop53
+; NO-SIMD128-FAST-NEXT:    i32.lt_s $push0=, $1, $9
+; NO-SIMD128-FAST-NEXT:    i32.select $push1=, $pop54, $pop52, $pop0
+; NO-SIMD128-FAST-NEXT:    i32.store16 0($0), $pop1
+; NO-SIMD128-FAST-NEXT:    i32.extend16_s $push51=, $2
+; NO-SIMD128-FAST-NEXT:    local.tee $push50=, $9=, $pop51
+; NO-SIMD128-FAST-NEXT:    i32.extend16_s $push49=, $10
+; NO-SIMD128-FAST-NEXT:    local.tee $push48=, $10=, $pop49
+; NO-SIMD128-FAST-NEXT:    i32.lt_s $push2=, $9, $10
+; NO-SIMD128-FAST-NEXT:    i32.select $push3=, $pop50, $pop48, $pop2
+; NO-SIMD128-FAST-NEXT:    i32.store16 2($0), $pop3
+; NO-SIMD128-FAST-NEXT:    i32.extend16_s $push47=, $3
+; NO-SIMD128-FAST-NEXT:    local.tee $push46=, $10=, $pop47
+; NO-SIMD128-FAST-NEXT:    i32.extend16_s $push45=, $11
+; NO-SIMD128-FAST-NEXT:    local.tee $push44=, $11=, $pop45
+; NO-SIMD128-FAST-NEXT:    i32.lt_s $push4=, $10, $11
+; NO-SIMD128-FAST-NEXT:    i32.select $push5=, $pop46, $pop44, $pop4
+; NO-SIMD128-FAST-NEXT:    i32.store16 4($0), $pop5
+; NO-SIMD128-FAST-NEXT:    i32.const $push8=, 6
+; NO-SIMD128-FAST-NEXT:    i32.add $push9=, $0, $pop8
+; NO-SIMD128-FAST-NEXT:    i32.extend16_s $push43=, $4
+; NO-SIMD128-FAST-NEXT:    local.tee $push42=, $11=, $pop43
+; NO-SIMD128-FAST-NEXT:    i32.extend16_s $push41=, $12
+; NO-SIMD128-FAST-NEXT:    local.tee $push40=, $12=, $pop41
+; NO-SIMD128-FAST-NEXT:    i32.lt_s $push6=, $11, $12
+; NO-SIMD128-FAST-NEXT:    i32.select $push7=, $pop42, $pop40, $pop6
+; NO-SIMD128-FAST-NEXT:    i32.store16 0($pop9), $pop7
+; NO-SIMD128-FAST-NEXT:    i32.extend16_s $push39=, $5
+; NO-SIMD128-FAST-NEXT:    local.tee $push38=, $12=, $pop39
+; NO-SIMD128-FAST-NEXT:    i32.extend16_s $push37=, $13
+; NO-SIMD128-FAST-NEXT:    local.tee $push36=, $13=, $pop37
+; NO-SIMD128-FAST-NEXT:    i32.lt_s $push10=, $12, $13
+; NO-SIMD128-FAST-NEXT:    i32.select $push11=, $pop38, $pop36, $pop10
+; NO-SIMD128-FAST-NEXT:    i32.store16 8($0), $pop11
+; NO-SIMD128-FAST-NEXT:    i32.const $push14=, 10
+; NO-SIMD128-FAST-NEXT:    i32.add $push15=, $0, $pop14
+; NO-SIMD128-FAST-NEXT:    i32.extend16_s $push35=, $6
+; NO-SIMD128-FAST-NEXT:    local.tee $push34=, $13=, $pop35
+; NO-SIMD128-FAST-NEXT:    i32.extend16_s $push33=, $14
+; NO-SIMD128-FAST-NEXT:    local.tee $push32=, $14=, $pop33
+; NO-SIMD128-FAST-NEXT:    i32.lt_s $push12=, $13, $14
+; NO-SIMD128-FAST-NEXT:    i32.select $push13=, $pop34, $pop32, $pop12
+; NO-SIMD128-FAST-NEXT:    i32.store16 0($pop15), $pop13
+; NO-SIMD128-FAST-NEXT:    i32.const $push18=, 12
+; NO-SIMD128-FAST-NEXT:    i32.add $push19=, $0, $pop18
+; NO-SIMD128-FAST-NEXT:    i32.extend16_s $push31=, $7
+; NO-SIMD128-FAST-NEXT:    local.tee $push30=, $14=, $pop31
+; NO-SIMD128-FAST-NEXT:    i32.extend16_s $push29=, $15
+; NO-SIMD128-FAST-NEXT:    local.tee $push28=, $15=, $pop29
+; NO-SIMD128-FAST-NEXT:    i32.lt_s $push16=, $14, $15
+; NO-SIMD128-FAST-NEXT:    i32.select $push17=, $pop30, $pop28, $pop16
+; NO-SIMD128-FAST-NEXT:    i32.store16 0($pop19), $pop17
+; NO-SIMD128-FAST-NEXT:    i32.const $push22=, 14
+; NO-SIMD128-FAST-NEXT:    i32.add $push23=, $0, $pop22
+; NO-SIMD128-FAST-NEXT:    i32.extend16_s $push27=, $8
+; NO-SIMD128-FAST-NEXT:    local.tee $push26=, $0=, $pop27
+; NO-SIMD128-FAST-NEXT:    i32.extend16_s $push25=, $16
+; NO-SIMD128-FAST-NEXT:    local.tee $push24=, $16=, $pop25
+; NO-SIMD128-FAST-NEXT:    i32.lt_s $push20=, $0, $16
+; NO-SIMD128-FAST-NEXT:    i32.select $push21=, $pop26, $pop24, $pop20
+; NO-SIMD128-FAST-NEXT:    i32.store16 0($pop23), $pop21
 ; NO-SIMD128-FAST-NEXT:    return
   %c = icmp slt <8 x i16> %x, %y
   %a = select <8 x i1> %c, <8 x i16> %x, <8 x i16> %y
   ret <8 x i16> %a
 }
 
 define <8 x i16> @min_u_v8i16(<8 x i16> %x, <8 x i16> %y) {
 ; SIMD128-LABEL: min_u_v8i16:
 ; SIMD128:         .functype min_u_v8i16 (v128, v128) -> (v128)
 ; SIMD128-NEXT:  # %bb.0:
 ; SIMD128-NEXT:    i16x8.min_u $push0=, $0, $1
 ; SIMD128-NEXT:    return $pop0
 ;
 ; SIMD128-FAST-LABEL: min_u_v8i16:
 ; SIMD128-FAST:         .functype min_u_v8i16 (v128, v128) -> (v128)
 ; SIMD128-FAST-NEXT:  # %bb.0:
 ; SIMD128-FAST-NEXT:    i16x8.min_u $push0=, $0, $1
 ; SIMD128-FAST-NEXT:    return $pop0
 ;
 ; NO-SIMD128-LABEL: min_u_v8i16:
 ; NO-SIMD128:         .functype min_u_v8i16 (i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32) -> ()
 ; NO-SIMD128-NEXT:  # %bb.0:
-; NO-SIMD128-NEXT:    i32.const $push5=, 14
-; NO-SIMD128-NEXT:    i32.add $push6=, $0, $pop5
+; NO-SIMD128-NEXT:    i32.const $push3=, 14
+; NO-SIMD128-NEXT:    i32.add $push4=, $0, $pop3
 ; NO-SIMD128-NEXT:    i32.const $push0=, 65535
-; NO-SIMD128-NEXT:    i32.and $push2=, $8, $pop0
-; NO-SIMD128-NEXT:    i32.const $push55=, 65535
-; NO-SIMD128-NEXT:    i32.and $push1=, $16, $pop55
-; NO-SIMD128-NEXT:    i32.lt_u $push3=, $pop2, $pop1
-; NO-SIMD128-NEXT:    i32.select $push4=, $8, $16, $pop3
-; NO-SIMD128-NEXT:    i32.store16 0($pop6), $pop4
-; NO-SIMD128-NEXT:    i32.const $push11=, 12
+; NO-SIMD128-NEXT:    i32.and $push71=, $8, $pop0
+; NO-SIMD128-NEXT:    local.tee $push70=, $8=, $pop71
+; NO-SIMD128-NEXT:    i32.const $push69=, 65535
+; NO-SIMD128-NEXT:    i32.and $push68=, $16, $pop69
+; NO-SIMD128-NEXT:    local.tee $push67=, $16=, $pop68
+; NO-SIMD128-NEXT:    i32.lt_u $push1=, $8, $16
+; NO-SIMD128-NEXT:    i32.select $push2=, $pop70, $pop67, $pop1
+; NO-SIMD128-NEXT:    i32.store16 0($pop4), $pop2
+; NO-SIMD128-NEXT:    i32.const $push7=, 12
+; NO-SIMD128-NEXT:    i32.add $push8=, $0, $pop7
+; NO-SIMD128-NEXT:    i32.const $push66=, 65535
+; NO-SIMD128-NEXT:    i32.and $push65=, $7, $pop66
+; NO-SIMD128-NEXT:    local.tee $push64=, $16=, $pop65
+; NO-SIMD128-NEXT:    i32.const $push63=, 65535
+; NO-SIMD128-NEXT:    i32.and $push62=, $15, $pop63
+; NO-SIMD128-NEXT:    local.tee $push61=, $8=, $pop62
+; NO-SIMD128-NEXT:    i32.lt_u $push5=, $16, $8
+; NO-SIMD128-NEXT:    i32.select $push6=, $pop64, $pop61, $pop5
+; NO-SIMD128-NEXT:    i32.store16 0($pop8), $pop6
+; NO-SIMD128-NEXT:    i32.const $push11=, 10
 ; NO-SIMD128-NEXT:    i32.add $push12=, $0, $pop11
-; NO-SIMD128-NEXT:    i32.const $push54=, 65535
-; NO-SIMD128-NEXT:    i32.and $push8=, $7, $pop54
-; NO-SIMD128-NEXT:    i32.const $push53=, 65535
-; NO-SIMD128-NEXT:    i32.and $push7=, $15, $pop53
-; NO-SIMD128-NEXT:    i32.lt_u $push9=, $pop8, $pop7
-; NO-SIMD128-NEXT:    i32.select $push10=, $7, $15, $pop9
+; NO-SIMD128-NEXT:    i32.const $push60=, 65535
+; NO-SIMD128-NEXT:    i32.and $push59=, $6, $pop60
+; NO-SIMD128-NEXT:    local.tee $push58=, $16=, $pop59
+; NO-SIMD128-NEXT:    i32.const $push57=, 65535
+; NO-SIMD128-NEXT:    i32.and $push56=, $14, $pop57
+; NO-SIMD128-NEXT:    local.tee $push55=, $8=, $pop56
+; NO-SIMD128-NEXT:    i32.lt_u $push9=, $16, $8
+; NO-SIMD128-NEXT:    i32.select $push10=, $pop58, $pop55, $pop9
 ; NO-SIMD128-NEXT:    i32.store16 0($pop12), $pop10
-; NO-SIMD128-NEXT:    i32.const $push17=, 10
-; NO-SIMD128-NEXT:    i32.add $push18=, $0, $pop17
-; NO-SIMD128-NEXT:    i32.const $push52=, 65535
-; NO-SIMD128-NEXT:    i32.and $push14=, $6, $pop52
+; NO-SIMD128-NEXT:    i32.const $push54=, 65535
+; NO-SIMD128-NEXT:    i32.and $push53=, $5, $pop54
+; NO-SIMD128-NEXT:    local.tee $push52=, $16=, $pop53
 ; NO-SIMD128-NEXT:    i32.const $push51=, 65535
-; NO-SIMD128-NEXT:    i32.and $push13=, $14, $pop51
-; NO-SIMD128-NEXT:    i32.lt_u $push15=, $pop14, $pop13
-; NO-SIMD128-NEXT:    i32.select $push16=, $6, $14, $pop15
-; NO-SIMD128-NEXT:    i32.store16 0($pop18), $pop16
-; NO-SIMD128-NEXT:    i32.const $push50=, 65535
-; NO-SIMD128-NEXT:    i32.and $push20=, $5, $pop50
-; NO-SIMD128-NEXT:    i32.const $push49=, 65535
-; NO-SIMD128-NEXT:    i32.and $push19=, $13, $pop49
-; NO-SIMD128-NEXT:    i32.lt_u $push21=, $pop20, $pop19
-; NO-SIMD128-NEXT:    i32.select $push22=, $5, $13, $pop21
-; NO-SIMD128-NEXT:    i32.store16 8($0), $pop22
-; NO-SIMD128-NEXT:    i32.const $push27=, 6
-; NO-SIMD128-NEXT:    i32.add $push28=, $0, $pop27
+; NO-SIMD128-NEXT:    i32.and $push50=, $13, $pop51
+; NO-SIMD128-NEXT:    local.tee $push49=, $8=, $pop50
+; NO-SIMD128-NEXT:    i32.lt_u $push13=, $16, $8
+; NO-SIMD128-NEXT:    i32.select $push14=, $pop52, $pop49, $pop13
+; NO-SIMD128-NEXT:    i32.store16 8($0), $pop14
+; NO-SIMD128-NEXT:    i32.const $push17=, 6
+; NO-SIMD128-NEXT:    i32.add $push18=, $0, $pop17
 ; NO-SIMD128-NEXT:    i32.const $push48=, 65535
-; NO-SIMD128-NEXT:    i32.and $push24=, $4, $pop48
-; NO-SIMD128-NEXT:    i32.const $push47=, 65535
-; NO-SIMD128-NEXT:    i32.and $push23=, $12, $pop47
-; NO-SIMD128-NEXT:    i32.lt_u $push25=, $pop24, $pop23
-; NO-SIMD128-NEXT:    i32.select $push26=, $4, $12, $pop25
-; NO-SIMD128-NEXT:    i32.store16 0($pop28), $pop26
-; NO-SIMD128-NEXT:    i32.const $push46=, 65535
-; NO-SIMD128-NEXT:    i32.and $push30=, $3, $pop46
+; NO-SIMD128-NEXT:    i32.and $push47=, $4, $pop48
+; NO-SIMD128-NEXT:    local.tee $push46=, $16=, $pop47
 ; NO-SIMD128-NEXT:    i32.const $push45=, 65535
-; NO-SIMD128-NEXT:    i32.and $push29=, $11, $pop45
-; NO-SIMD128-NEXT:    i32.lt_u $push31=, $pop30, $pop29
-; NO-SIMD128-NEXT:    i32.select $push32=, $3, $11, $pop31
-; NO-SIMD128-NEXT:    i32.store16 4($0), $pop32
-; NO-SIMD128-NEXT:    i32.const $push44=, 65535
-; NO-SIMD128-NEXT:    i32.and $push34=, $2, $pop44
-; NO-SIMD128-NEXT:    i32.const $push43=, 65535
-; NO-SIMD128-NEXT:    i32.and $push33=, $10, $pop43
-; NO-SIMD128-NEXT:    i32.lt_u $push35=, $pop34, $pop33
-; NO-SIMD128-NEXT:    i32.select $push36=, $2, $10, $pop35
-; NO-SIMD128-NEXT:    i32.store16 2($0), $pop36
+; NO-SIMD128-NEXT:    i32.and $push44=, $12, $pop45
+; NO-SIMD128-NEXT:    local.tee $push43=, $8=, $pop44
+; NO-SIMD128-NEXT:    i32.lt_u $push15=, $16, $8
+; NO-SIMD128-NEXT:    i32.select $push16=, $pop46, $pop43, $pop15
+; NO-SIMD128-NEXT:    i32.store16 0($pop18), $pop16
 ; NO-SIMD128-NEXT:    i32.const $push42=, 65535
-; NO-SIMD128-NEXT:    i32.and $push38=, $1, $pop42
-; NO-SIMD128-NEXT:    i32.const $push41=, 65535
-; NO-SIMD128-NEXT:    i32.and $push37=, $9, $pop41
-; NO-SIMD128-NEXT:    i32.lt_u $push39=, $pop38, $pop37
-; NO-SIMD128-NEXT:    i32.select $push40=, $1, $9, $pop39
-; NO-SIMD128-NEXT:    i32.store16 0($0), $pop40
+; NO-SIMD128-NEXT:    i32.and $push41=, $3, $pop42
+; NO-SIMD128-NEXT:    local.tee $push40=, $16=, $pop41
+; NO-SIMD128-NEXT:    i32.const $push39=, 65535
+; NO-SIMD128-NEXT:    i32.and $push38=, $11, $pop39
+; NO-SIMD128-NEXT:    local.tee $push37=, $8=, $pop38
+; NO-SIMD128-NEXT:    i32.lt_u $push19=, $16, $8
+; NO-SIMD128-NEXT:    i32.select $push20=, $pop40, $pop37, $pop19
+; NO-SIMD128-NEXT:    i32.store16 4($0), $pop20
+; NO-SIMD128-NEXT:    i32.const $push36=, 65535
+; NO-SIMD128-NEXT:    i32.and $push35=, $2, $pop36
+; NO-SIMD128-NEXT:    local.tee $push34=, $16=, $pop35
+; NO-SIMD128-NEXT:    i32.const $push33=, 65535
+; NO-SIMD128-NEXT:    i32.and $push32=, $10, $pop33
+; NO-SIMD128-NEXT:    local.tee $push31=, $8=, $pop32
+; NO-SIMD128-NEXT:    i32.lt_u $push21=, $16, $8
+; NO-SIMD128-NEXT:    i32.select $push22=, $pop34, $pop31, $pop21
+; NO-SIMD128-NEXT:    i32.store16 2($0), $pop22
+; NO-SIMD128-NEXT:    i32.const $push30=, 65535
+; NO-SIMD128-NEXT:    i32.and $push29=, $1, $pop30
+; NO-SIMD128-NEXT:    local.tee $push28=, $16=, $pop29
+; NO-SIMD128-NEXT:    i32.const $push27=, 65535
+; NO-SIMD128-NEXT:    i32.and $push26=, $9, $pop27
+; NO-SIMD128-NEXT:    local.tee $push25=, $8=, $pop26
+; NO-SIMD128-NEXT:    i32.lt_u $push23=, $16, $8
+; NO-SIMD128-NEXT:    i32.select $push24=, $pop28, $pop25, $pop23
+; NO-SIMD128-NEXT:    i32.store16 0($0), $pop24
 ; NO-SIMD128-NEXT:    return
 ;
 ; NO-SIMD128-FAST-LABEL: min_u_v8i16:
 ; NO-SIMD128-FAST:         .functype min_u_v8i16 (i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32) -> ()
 ; NO-SIMD128-FAST-NEXT:  # %bb.0:
 ; NO-SIMD128-FAST-NEXT:    i32.const $push0=, 65535
-; NO-SIMD128-FAST-NEXT:    i32.and $push2=, $1, $pop0
-; NO-SIMD128-FAST-NEXT:    i32.const $push55=, 65535
-; NO-SIMD128-FAST-NEXT:    i32.and $push1=, $9, $pop55
-; NO-SIMD128-FAST-NEXT:    i32.lt_u $push3=, $pop2, $pop1
-; NO-SIMD128-FAST-NEXT:    i32.select $push4=, $1, $9, $pop3
-; NO-SIMD128-FAST-NEXT:    i32.store16 0($0), $pop4
+; NO-SIMD128-FAST-NEXT:    i32.and $push71=, $1, $pop0
+; NO-SIMD128-FAST-NEXT:    local.tee $push70=, $1=, $pop71
+; NO-SIMD128-FAST-NEXT:    i32.const $push69=, 65535
+; NO-SIMD128-FAST-NEXT:    i32.and $push68=, $9, $pop69
+; NO-SIMD128-FAST-NEXT:    local.tee $push67=, $9=, $pop68
+; NO-SIMD128-FAST-NEXT:    i32.lt_u $push1=, $1, $9
+; NO-SIMD128-FAST-NEXT:    i32.select $push2=, $pop70, $pop67, $pop1
+; NO-SIMD128-FAST-NEXT:    i32.store16 0($0), $pop2
+; NO-SIMD128-FAST-NEXT:    i32.const $push66=, 65535
+; NO-SIMD128-FAST-NEXT:    i32.and $push65=, $2, $pop66
+; NO-SIMD128-FAST-NEXT:    local.tee $push64=, $9=, $pop65
+; NO-SIMD128-FAST-NEXT:    i32.const $push63=, 65535
+; NO-SIMD128-FAST-NEXT:    i32.and $push62=, $10, $pop63
+; NO-SIMD128-FAST-NEXT:    local.tee $push61=, $10=, $pop62
+; NO-SIMD128-FAST-NEXT:    i32.lt_u $push3=, $9, $10
+; NO-SIMD128-FAST-NEXT:    i32.select $push4=, $pop64, $pop61, $pop3
+; NO-SIMD128-FAST-NEXT:    i32.store16 2($0), $pop4
+; NO-SIMD128-FAST-NEXT:    i32.const $push60=, 65535
+; NO-SIMD128-FAST-NEXT:    i32.and $push59=, $3, $pop60
+; NO-SIMD128-FAST-NEXT:    local.tee $push58=, $10=, $pop59
+; NO-SIMD128-FAST-NEXT:    i32.const $push57=, 65535
+; NO-SIMD128-FAST-NEXT:    i32.and $push56=, $11, $pop57
+; NO-SIMD128-FAST-NEXT:    local.tee $push55=, $11=, $pop56
+; NO-SIMD128-FAST-NEXT:    i32.lt_u $push5=, $10, $11
+; NO-SIMD128-FAST-NEXT:    i32.select $push6=, $pop58, $pop55, $pop5
+; NO-SIMD128-FAST-NEXT:    i32.store16 4($0), $pop6
+; NO-SIMD128-FAST-NEXT:    i32.const $push9=, 6
+; NO-SIMD128-FAST-NEXT:    i32.add $push10=, $0, $pop9
 ; NO-SIMD128-FAST-NEXT:    i32.const $push54=, 65535
-; NO-SIMD128-FAST-NEXT:    i32.and $push6=, $2, $pop54
-; NO-SIMD128-FAST-NEXT:    i32.const $push53=, 65535
-; NO-SIMD128-FAST-NEXT:    i32.and $push5=, $10, $pop53
-; NO-SIMD128-FAST-NEXT:    i32.lt_u $push7=, $pop6, $pop5
-; NO-SIMD128-FAST-NEXT:    i32.select $push8=, $2, $10, $pop7
-; NO-SIMD128-FAST-NEXT:    i32.store16 2($0), $pop8
-; NO-SIMD128-FAST-NEXT:    i32.const $push52=, 65535
-; NO-SIMD128-FAST-NEXT:    i32.and $push10=, $3, $pop52
+; NO-SIMD128-FAST-NEXT:    i32.and $push53=, $4, $pop54
+; NO-SIMD128-FAST-NEXT:    local.tee $push52=, $11=, $pop53
 ; NO-SIMD128-FAST-NEXT:    i32.const $push51=, 65535
-; NO-SIMD128-FAST-NEXT:    i32.and $push9=, $11, $pop51
-; NO-SIMD128-FAST-NEXT:    i32.lt_u $push11=, $pop10, $pop9
-; NO-SIMD128-FAST-NEXT:    i32.select $push12=, $3, $11, $pop11
-; NO-SIMD128-FAST-NEXT:    i32.store16 4($0), $pop12
-; NO-SIMD128-FAST-NEXT:    i32.const $push17=, 6
-; NO-SIMD128-FAST-NEXT:    i32.add $push18=, $0, $pop17
-; NO-SIMD128-FAST-NEXT:    i32.const $push50=, 65535
-; NO-SIMD128-FAST-NEXT:    i32.and $push14=, $4, $pop50
-; NO-SIMD128-FAST-NEXT:    i32.const $push49=, 65535
-; NO-SIMD128-FAST-NEXT:    i32.and $push13=, $12, $pop49
-; NO-SIMD128-FAST-NEXT:    i32.lt_u $push15=, $pop14, $pop13
-; NO-SIMD128-FAST-NEXT:    i32.select $push16=, $4, $12, $pop15
-; NO-SIMD128-FAST-NEXT:    i32.store16 0($pop18), $pop16
+; NO-SIMD128-FAST-NEXT:    i32.and $push50=, $12, $pop51
+; NO-SIMD128-FAST-NEXT:    local.tee $push49=, $12=, $pop50
+; NO-SIMD128-FAST-NEXT:    i32.lt_u $push7=, $11, $12
+; NO-SIMD128-FAST-NEXT:    i32.select $push8=, $pop52, $pop49, $pop7
+; NO-SIMD128-FAST-NEXT:    i32.store16 0($pop10), $pop8
 ; NO-SIMD128-FAST-NEXT:    i32.const $push48=, 65535
-; NO-SIMD128-FAST-NEXT:    i32.and $push20=, $5, $pop48
-; NO-SIMD128-FAST-NEXT:    i32.const $push47=, 65535
-; NO-SIMD128-FAST-NEXT:    i32.and $push19=, $13, $pop47
-; NO-SIMD128-FAST-NEXT:    i32.lt_u $push21=, $pop20, $pop19
-; NO-SIMD128-FAST-NEXT:    i32.select $push22=, $5, $13, $pop21
-; NO-SIMD128-FAST-NEXT:    i32.store16 8($0), $pop22
-; NO-SIMD128-FAST-NEXT:    i32.const $push27=, 10
-; NO-SIMD128-FAST-NEXT:    i32.add $push28=, $0, $pop27
-; NO-SIMD128-FAST-NEXT:    i32.const $push46=, 65535
-; NO-SIMD128-FAST-NEXT:    i32.and $push24=, $6, $pop46
+; NO-SIMD128-FAST-NEXT:    i32.and $push47=, $5, $pop48
+; NO-SIMD128-FAST-NEXT:    local.tee $push46=, $12=, $pop47
 ; NO-SIMD128-FAST-NEXT:    i32.const $push45=, 65535
-; NO-SIMD128-FAST-NEXT:    i32.and $push23=, $14, $pop45
-; NO-SIMD128-FAST-NEXT:    i32.lt_u $push25=, $pop24, $pop23
-; NO-SIMD128-FAST-NEXT:    i32.select $push26=, $6, $14, $pop25
-; NO-SIMD128-FAST-NEXT:    i32.store16 0($pop28), $pop26
-; NO-SIMD128-FAST-NEXT:    i32.const $push33=, 12
-; NO-SIMD128-FAST-NEXT:    i32.add $push34=, $0, $pop33
-; NO-SIMD128-FAST-NEXT:    i32.const $push44=, 65535
-; NO-SIMD128-FAST-NEXT:    i32.and $push30=, $7, $pop44
-; NO-SIMD128-FAST-NEXT:    i32.const $push43=, 65535
-; NO-SIMD128-FAST-NEXT:    i32.and $push29=, $15, $pop43
-; NO-SIMD128-FAST-NEXT:    i32.lt_u $push31=, $pop30, $pop29
-; NO-SIMD128-FAST-NEXT:    i32.select $push32=, $7, $15, $pop31
-; NO-SIMD128-FAST-NEXT:    i32.store16 0($pop34), $pop32
-; NO-SIMD128-FAST-NEXT:    i32.const $push39=, 14
-; NO-SIMD128-FAST-NEXT:    i32.add $push40=, $0, $pop39
+; NO-SIMD128-FAST-NEXT:    i32.and $push44=, $13, $pop45
+; NO-SIMD128-FAST-NEXT:    local.tee $push43=, $13=, $pop44
+; NO-SIMD128-FAST-NEXT:    i32.lt_u $push11=, $12, $13
+; NO-SIMD128-FAST-NEXT:    i32.select $push12=, $pop46, $pop43, $pop11
+; NO-SIMD128-FAST-NEXT:    i32.store16 8($0), $pop12
+; NO-SIMD128-FAST-NEXT:    i32.const $push15=, 10
+; NO-SIMD128-FAST-NEXT:    i32.add $push16=, $0, $pop15
 ; NO-SIMD128-FAST-NEXT:    i32.const $push42=, 65535
-; NO-SIMD128-FAST-NEXT:    i32.and $push36=, $8, $pop42
-; NO-SIMD128-FAST-NEXT:    i32.const $push41=, 65535
-; NO-SIMD128-FAST-NEXT:    i32.and $push35=, $16, $pop41
-; NO-SIMD128-FAST-NEXT:    i32.lt_u $push37=, $pop36, $pop35
-; NO-SIMD128-FAST-NEXT:    i32.select $push38=, $8, $16, $pop37
-; NO-SIMD128-FAST-NEXT:    i32.store16 0($pop40), $pop38
+; NO-SIMD128-FAST-NEXT:    i32.and $push41=, $6, $pop42
+; NO-SIMD128-FAST-NEXT:    local.tee $push40=, $13=, $pop41
+; NO-SIMD128-FAST-NEXT:    i32.const $push39=, 65535
+; NO-SIMD128-FAST-NEXT:    i32.and $push38=, $14, $pop39
+; NO-SIMD128-FAST-NEXT:    local.tee $push37=, $14=, $pop38
+; NO-SIMD128-FAST-NEXT:    i32.lt_u $push13=, $13, $14
+; NO-SIMD128-FAST-NEXT:    i32.select $push14=, $pop40, $pop37, $pop13
+; NO-SIMD128-FAST-NEXT:    i32.store16 0($pop16), $pop14
+; NO-SIMD128-FAST-NEXT:    i32.const $push19=, 12
+; NO-SIMD128-FAST-NEXT:    i32.add $push20=, $0, $pop19
+; NO-SIMD128-FAST-NEXT:    i32.const $push36=, 65535
+; NO-SIMD128-FAST-NEXT:    i32.and $push35=, $7, $pop36
+; NO-SIMD128-FAST-NEXT:    local.tee $push34=, $14=, $pop35
+; NO-SIMD128-FAST-NEXT:    i32.const $push33=, 65535
+; NO-SIMD128-FAST-NEXT:    i32.and $push32=, $15, $pop33
+; NO-SIMD128-FAST-NEXT:    local.tee $push31=, $15=, $pop32
+; NO-SIMD128-FAST-NEXT:    i32.lt_u $push17=, $14, $15
+; NO-SIMD128-FAST-NEXT:    i32.select $push18=, $pop34, $pop31, $pop17
+; NO-SIMD128-FAST-NEXT:    i32.store16 0($pop20), $pop18
+; NO-SIMD128-FAST-NEXT:    i32.const $push23=, 14
+; NO-SIMD128-FAST-NEXT:    i32.add $push24=, $0, $pop23
+; NO-SIMD128-FAST-NEXT:    i32.const $push30=, 65535
+; NO-SIMD128-FAST-NEXT:    i32.and $push29=, $8, $pop30
+; NO-SIMD128-FAST-NEXT:    local.tee $push28=, $0=, $pop29
+; NO-SIMD128-FAST-NEXT:    i32.const $push27=, 65535
+; NO-SIMD128-FAST-NEXT:    i32.and $push26=, $16, $pop27
+; NO-SIMD128-FAST-NEXT:    local.tee $push25=, $16=, $pop26
+; NO-SIMD128-FAST-NEXT:    i32.lt_u $push21=, $0, $16
+; NO-SIMD128-FAST-NEXT:    i32.select $push22=, $pop28, $pop25, $pop21
+; NO-SIMD128-FAST-NEXT:    i32.store16 0($pop24), $pop22
 ; NO-SIMD128-FAST-NEXT:    return
   %c = icmp ult <8 x i16> %x, %y
   %a = select <8 x i1> %c, <8 x i16> %x, <8 x i16> %y
   ret <8 x i16> %a
 }
 
 define <8 x i16> @max_s_v8i16(<8 x i16> %x, <8 x i16> %y) {
 ; SIMD128-LABEL: max_s_v8i16:
 ; SIMD128:         .functype max_s_v8i16 (v128, v128) -> (v128)
 ; SIMD128-NEXT:  # %bb.0:
 ; SIMD128-NEXT:    i16x8.max_s $push0=, $0, $1
 ; SIMD128-NEXT:    return $pop0
 ;
 ; SIMD128-FAST-LABEL: max_s_v8i16:
 ; SIMD128-FAST:         .functype max_s_v8i16 (v128, v128) -> (v128)
 ; SIMD128-FAST-NEXT:  # %bb.0:
 ; SIMD128-FAST-NEXT:    i16x8.max_s $push0=, $0, $1
 ; SIMD128-FAST-NEXT:    return $pop0
 ;
 ; NO-SIMD128-LABEL: max_s_v8i16:
 ; NO-SIMD128:         .functype max_s_v8i16 (i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32) -> ()
 ; NO-SIMD128-NEXT:  # %bb.0:
-; NO-SIMD128-NEXT:    i32.const $push4=, 14
-; NO-SIMD128-NEXT:    i32.add $push5=, $0, $pop4
-; NO-SIMD128-NEXT:    i32.extend16_s $push1=, $8
-; NO-SIMD128-NEXT:    i32.extend16_s $push0=, $16
-; NO-SIMD128-NEXT:    i32.gt_s $push2=, $pop1, $pop0
-; NO-SIMD128-NEXT:    i32.select $push3=, $8, $16, $pop2
-; NO-SIMD128-NEXT:    i32.store16 0($pop5), $pop3
-; NO-SIMD128-NEXT:    i32.const $push10=, 12
+; NO-SIMD128-NEXT:    i32.const $push2=, 14
+; NO-SIMD128-NEXT:    i32.add $push3=, $0, $pop2
+; NO-SIMD128-NEXT:    i32.extend16_s $push55=, $8
+; NO-SIMD128-NEXT:    local.tee $push54=, $8=, $pop55
+; NO-SIMD128-NEXT:    i32.extend16_s $push53=, $16
+; NO-SIMD128-NEXT:    local.tee $push52=, $16=, $pop53
+; NO-SIMD128-NEXT:    i32.gt_s $push0=, $8, $16
+; NO-SIMD128-NEXT:    i32.select $push1=, $pop54, $pop52, $pop0
+; NO-SIMD128-NEXT:    i32.store16 0($pop3), $pop1
+; NO-SIMD128-NEXT:    i32.const $push6=, 12
+; NO-SIMD128-NEXT:    i32.add $push7=, $0, $pop6
+; NO-SIMD128-NEXT:    i32.extend16_s $push51=, $7
+; NO-SIMD128-NEXT:    local.tee $push50=, $16=, $pop51
+; NO-SIMD128-NEXT:    i32.extend16_s $push49=, $15
+; NO-SIMD128-NEXT:    local.tee $push48=, $8=, $pop49
+; NO-SIMD128-NEXT:    i32.gt_s $push4=, $16, $8
+; NO-SIMD128-NEXT:    i32.select $push5=, $pop50, $pop48, $pop4
+; NO-SIMD128-NEXT:    i32.store16 0($pop7), $pop5
+; NO-SIMD128-NEXT:    i32.const $push10=, 10
 ; NO-SIMD128-NEXT:    i32.add $push11=, $0, $pop10
-; NO-SIMD128-NEXT:    i32.extend16_s $push7=, $7
-; NO-SIMD128-NEXT:    i32.extend16_s $push6=, $15
-; NO-SIMD128-NEXT:    i32.gt_s $push8=, $pop7, $pop6
-; NO-SIMD128-NEXT:    i32.select $push9=, $7, $15, $pop8
+; NO-SIMD128-NEXT:    i32.extend16_s $push47=, $6
+; NO-SIMD128-NEXT:    local.tee $push46=, $16=, $pop47
+; NO-SIMD128-NEXT:    i32.extend16_s $push45=, $14
+; NO-SIMD128-NEXT:    local.tee $push44=, $8=, $pop45
+; NO-SIMD128-NEXT:    i32.gt_s $push8=, $16, $8
+; NO-SIMD128-NEXT:    i32.select $push9=, $pop46, $pop44, $pop8
 ; NO-SIMD128-NEXT:    i32.store16 0($pop11), $pop9
-; NO-SIMD128-NEXT:    i32.const $push16=, 10
+; NO-SIMD128-NEXT:    i32.extend16_s $push43=, $5
+; NO-SIMD128-NEXT:    local.tee $push42=, $16=, $pop43
+; NO-SIMD128-NEXT:    i32.extend16_s $push41=, $13
+; NO-SIMD128-NEXT:    local.tee $push40=, $8=, $pop41
+; NO-SIMD128-NEXT:    i32.gt_s $push12=, $16, $8
+; NO-SIMD128-NEXT:    i32.select $push13=, $pop42, $pop40, $pop12
+; NO-SIMD128-NEXT:    i32.store16 8($0), $pop13
+; NO-SIMD128-NEXT:    i32.const $push16=, 6
 ; NO-SIMD128-NEXT:    i32.add $push17=, $0, $pop16
-; NO-SIMD128-NEXT:    i32.extend16_s $push13=, $6
-; NO-SIMD128-NEXT:    i32.extend16_s $push12=, $14
-; NO-SIMD128-NEXT:    i32.gt_s $push14=, $pop13, $pop12
-; NO-SIMD128-NEXT:    i32.select $push15=, $6, $14, $pop14
+; NO-SIMD128-NEXT:    i32.extend16_s $push39=, $4
+; NO-SIMD128-NEXT:    local.tee $push38=, $16=, $pop39
+; NO-SIMD128-NEXT:    i32.extend16_s $push37=, $12
+; NO-SIMD128-NEXT:    local.tee $push36=, $8=, $pop37
+; NO-SIMD128-NEXT:    i32.gt_s $push14=, $16, $8
+; NO-SIMD128-NEXT:    i32.select $push15=, $pop38, $pop36, $pop14
 ; NO-SIMD128-NEXT:    i32.store16 0($pop17), $pop15
-; NO-SIMD128-NEXT:    i32.extend16_s $push19=, $5
-; NO-SIMD128-NEXT:    i32.extend16_s $push18=, $13
-; NO-SIMD128-NEXT:    i32.gt_s $push20=, $pop19, $pop18
-; NO-SIMD128-NEXT:    i32.select $push21=, $5, $13, $pop20
-; NO-SIMD128-NEXT:    i32.store16 8($0), $pop21
-; NO-SIMD128-NEXT:    i32.const $push26=, 6
-; NO-SIMD128-NEXT:    i32.add $push27=, $0, $pop26
-; NO-SIMD128-NEXT:    i32.extend16_s $push23=, $4
-; NO-SIMD128-NEXT:    i32.extend16_s $push22=, $12
-; NO-SIMD128-NEXT:    i32.gt_s $push24=, $pop23, $pop22
-; NO-SIMD128-NEXT:    i32.select $push25=, $4, $12, $pop24
-; NO-SIMD128-NEXT:    i32.store16 0($pop27), $pop25
-; NO-SIMD128-NEXT:    i32.extend16_s $push29=, $3
-; NO-SIMD128-NEXT:    i32.extend16_s $push28=, $11
-; NO-SIMD128-NEXT:    i32.gt_s $push30=, $pop29, $pop28
-; NO-SIMD128-NEXT:    i32.select $push31=, $3, $11, $pop30
-; NO-SIMD128-NEXT:    i32.store16 4($0), $pop31
-; NO-SIMD128-NEXT:    i32.extend16_s $push33=, $2
-; NO-SIMD128-NEXT:    i32.extend16_s $push32=, $10
-; NO-SIMD128-NEXT:    i32.gt_s $push34=, $pop33, $pop32
-; NO-SIMD128-NEXT:    i32.select $push35=, $2, $10, $pop34
-; NO-SIMD128-NEXT:    i32.store16 2($0), $pop35
-; NO-SIMD128-NEXT:    i32.extend16_s $push37=, $1
-; NO-SIMD128-NEXT:    i32.extend16_s $push36=, $9
-; NO-SIMD128-NEXT:    i32.gt_s $push38=, $pop37, $pop36
-; NO-SIMD128-NEXT:    i32.select $push39=, $1, $9, $pop38
-; NO-SIMD128-NEXT:    i32.store16 0($0), $pop39
+; NO-SIMD128-NEXT:    i32.extend16_s $push35=, $3
+; NO-SIMD128-NEXT:    local.tee $push34=, $16=, $pop35
+; NO-SIMD128-NEXT:    i32.extend16_s $push33=, $11
+; NO-SIMD128-NEXT:    local.tee $push32=, $8=, $pop33
+; NO-SIMD128-NEXT:    i32.gt_s $push18=, $16, $8
+; NO-SIMD128-NEXT:    i32.select $push19=, $pop34, $pop32, $pop18
+; NO-SIMD128-NEXT:    i32.store16 4($0), $pop19
+; NO-SIMD128-NEXT:    i32.extend16_s $push31=, $2
+; NO-SIMD128-NEXT:    local.tee $push30=, $16=, $pop31
+; NO-SIMD128-NEXT:    i32.extend16_s $push29=, $10
+; NO-SIMD128-NEXT:    local.tee $push28=, $8=, $pop29
+; NO-SIMD128-NEXT:    i32.gt_s $push20=, $16, $8
+; NO-SIMD128-NEXT:    i32.select $push21=, $pop30, $pop28, $pop20
+; NO-SIMD128-NEXT:    i32.store16 2($0), $pop21
+; NO-SIMD128-NEXT:    i32.extend16_s $push27=, $1
+; NO-SIMD128-NEXT:    local.tee $push26=, $16=, $pop27
+; NO-SIMD128-NEXT:    i32.extend16_s $push25=, $9
+; NO-SIMD128-NEXT:    local.tee $push24=, $8=, $pop25
+; NO-SIMD128-NEXT:    i32.gt_s $push22=, $16, $8
+; NO-SIMD128-NEXT:    i32.select $push23=, $pop26, $pop24, $pop22
+; NO-SIMD128-NEXT:    i32.store16 0($0), $pop23
 ; NO-SIMD128-NEXT:    return
 ;
 ; NO-SIMD128-FAST-LABEL: max_s_v8i16:
 ; NO-SIMD128-FAST:         .functype max_s_v8i16 (i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32) -> ()
 ; NO-SIMD128-FAST-NEXT:  # %bb.0:
-; NO-SIMD128-FAST-NEXT:    i32.extend16_s $push1=, $1
-; NO-SIMD128-FAST-NEXT:    i32.extend16_s $push0=, $9
-; NO-SIMD128-FAST-NEXT:    i32.gt_s $push2=, $pop1, $pop0
-; NO-SIMD128-FAST-NEXT:    i32.select $push3=, $1, $9, $pop2
-; NO-SIMD128-FAST-NEXT:    i32.store16 0($0), $pop3
-; NO-SIMD128-FAST-NEXT:    i32.extend16_s $push5=, $2
-; NO-SIMD128-FAST-NEXT:    i32.extend16_s $push4=, $10
-; NO-SIMD128-FAST-NEXT:    i32.gt_s $push6=, $pop5, $pop4
-; NO-SIMD128-FAST-NEXT:    i32.select $push7=, $2, $10, $pop6
-; NO-SIMD128-FAST-NEXT:    i32.store16 2($0), $pop7
-; NO-SIMD128-FAST-NEXT:    i32.extend16_s $push9=, $3
-; NO-SIMD128-FAST-NEXT:    i32.extend16_s $push8=, $11
-; NO-SIMD128-FAST-NEXT:    i32.gt_s $push10=, $pop9, $pop8
-; NO-SIMD128-FAST-NEXT:    i32.select $push11=, $3, $11, $pop10
-; NO-SIMD128-FAST-NEXT:    i32.store16 4($0), $pop11
-; NO-SIMD128-FAST-NEXT:    i32.const $push16=, 6
-; NO-SIMD128-FAST-NEXT:    i32.add $push17=, $0, $pop16
-; NO-SIMD128-FAST-NEXT:    i32.extend16_s $push13=, $4
-; NO-SIMD128-FAST-NEXT:    i32.extend16_s $push12=, $12
-; NO-SIMD128-FAST-NEXT:    i32.gt_s $push14=, $pop13, $pop12
-; NO-SIMD128-FAST-NEXT:    i32.select $push15=, $4, $12, $pop14
-; NO-SIMD128-FAST-NEXT:    i32.store16 0($pop17), $pop15
-; NO-SIMD128-FAST-NEXT:    i32.extend16_s $push19=, $5
-; NO-SIMD128-FAST-NEXT:    i32.extend16_s $push18=, $13
-; NO-SIMD128-FAST-NEXT:    i32.gt_s $push20=, $pop19, $pop18
-; NO-SIMD128-FAST-NEXT:    i32.select $push21=, $5, $13, $pop20
-; NO-SIMD128-FAST-NEXT:    i32.store16 8($0), $pop21
-; NO-SIMD128-FAST-NEXT:    i32.const $push26=, 10
-; NO-SIMD128-FAST-NEXT:    i32.add $push27=, $0, $pop26
-; NO-SIMD128-FAST-NEXT:    i32.extend16_s $push23=, $6
-; NO-SIMD128-FAST-NEXT:    i32.extend16_s $push22=, $14
-; NO-SIMD128-FAST-NEXT:    i32.gt_s $push24=, $pop23, $pop22
-; NO-SIMD128-FAST-NEXT:    i32.select $push25=, $6, $14, $pop24
-; NO-SIMD128-FAST-NEXT:    i32.store16 0($pop27), $pop25
-; NO-SIMD128-FAST-NEXT:    i32.const $push32=, 12
-; NO-SIMD128-FAST-NEXT:    i32.add $push33=, $0, $pop32
-; NO-SIMD128-FAST-NEXT:    i32.extend16_s $push29=, $7
-; NO-SIMD128-FAST-NEXT:    i32.extend16_s $push28=, $15
-; NO-SIMD128-FAST-NEXT:    i32.gt_s $push30=, $pop29, $pop28
-; NO-SIMD128-FAST-NEXT:    i32.select $push31=, $7, $15, $pop30
-; NO-SIMD128-FAST-NEXT:    i32.store16 0($pop33), $pop31
-; NO-SIMD128-FAST-NEXT:    i32.const $push38=, 14
-; NO-SIMD128-FAST-NEXT:    i32.add $push39=, $0, $pop38
-; NO-SIMD128-FAST-NEXT:    i32.extend16_s $push35=, $8
-; NO-SIMD128-FAST-NEXT:    i32.extend16_s $push34=, $16
-; NO-SIMD128-FAST-NEXT:    i32.gt_s $push36=, $pop35, $pop34
-; NO-SIMD128-FAST-NEXT:    i32.select $push37=, $8, $16, $pop36
-; NO-SIMD128-FAST-NEXT:    i32.store16 0($pop39), $pop37
+; NO-SIMD128-FAST-NEXT:    i32.extend16_s $push55=, $1
+; NO-SIMD128-FAST-NEXT:    local.tee $push54=, $1=, $pop55
+; NO-SIMD128-FAST-NEXT:    i32.extend16_s $push53=, $9
+; NO-SIMD128-FAST-NEXT:    local.tee $push52=, $9=, $pop53
+; NO-SIMD128-FAST-NEXT:    i32.gt_s $push0=, $1, $9
+; NO-SIMD128-FAST-NEXT:    i32.select $push1=, $pop54, $pop52, $pop0
+; NO-SIMD128-FAST-NEXT:    i32.store16 0($0), $pop1
+; NO-SIMD128-FAST-NEXT:    i32.extend16_s $push51=, $2
+; NO-SIMD128-FAST-NEXT:    local.tee $push50=, $9=, $pop51
+; NO-SIMD128-FAST-NEXT:    i32.extend16_s $push49=, $10
+; NO-SIMD128-FAST-NEXT:    local.tee $push48=, $10=, $pop49
+; NO-SIMD128-FAST-NEXT:    i32.gt_s $push2=, $9, $10
+; NO-SIMD128-FAST-NEXT:    i32.select $push3=, $pop50, $pop48, $pop2
+; NO-SIMD128-FAST-NEXT:    i32.store16 2($0), $pop3
+; NO-SIMD128-FAST-NEXT:    i32.extend16_s $push47=, $3
+; NO-SIMD128-FAST-NEXT:    local.tee $push46=, $10=, $pop47
+; NO-SIMD128-FAST-NEXT:    i32.extend16_s $push45=, $11
+; NO-SIMD128-FAST-NEXT:    local.tee $push44=, $11=, $pop45
+; NO-SIMD128-FAST-NEXT:    i32.gt_s $push4=, $10, $11
+; NO-SIMD128-FAST-NEXT:    i32.select $push5=, $pop46, $pop44, $pop4
+; NO-SIMD128-FAST-NEXT:    i32.store16 4($0), $pop5
+; NO-SIMD128-FAST-NEXT:    i32.const $push8=, 6
+; NO-SIMD128-FAST-NEXT:    i32.add $push9=, $0, $pop8
+; NO-SIMD128-FAST-NEXT:    i32.extend16_s $push43=, $4
+; NO-SIMD128-FAST-NEXT:    local.tee $push42=, $11=, $pop43
+; NO-SIMD128-FAST-NEXT:    i32.extend16_s $push41=, $12
+; NO-SIMD128-FAST-NEXT:    local.tee $push40=, $12=, $pop41
+; NO-SIMD128-FAST-NEXT:    i32.gt_s $push6=, $11, $12
+; NO-SIMD128-FAST-NEXT:    i32.select $push7=, $pop42, $pop40, $pop6
+; NO-SIMD128-FAST-NEXT:    i32.store16 0($pop9), $pop7
+; NO-SIMD128-FAST-NEXT:    i32.extend16_s $push39=, $5
+; NO-SIMD128-FAST-NEXT:    local.tee $push38=, $12=, $pop39
+; NO-SIMD128-FAST-NEXT:    i32.extend16_s $push37=, $13
+; NO-SIMD128-FAST-NEXT:    local.tee $push36=, $13=, $pop37
+; NO-SIMD128-FAST-NEXT:    i32.gt_s $push10=, $12, $13
+; NO-SIMD128-FAST-NEXT:    i32.select $push11=, $pop38, $pop36, $pop10
+; NO-SIMD128-FAST-NEXT:    i32.store16 8($0), $pop11
+; NO-SIMD128-FAST-NEXT:    i32.const $push14=, 10
+; NO-SIMD128-FAST-NEXT:    i32.add $push15=, $0, $pop14
+; NO-SIMD128-FAST-NEXT:    i32.extend16_s $push35=, $6
+; NO-SIMD128-FAST-NEXT:    local.tee $push34=, $13=, $pop35
+; NO-SIMD128-FAST-NEXT:    i32.extend16_s $push33=, $14
+; NO-SIMD128-FAST-NEXT:    local.tee $push32=, $14=, $pop33
+; NO-SIMD128-FAST-NEXT:    i32.gt_s $push12=, $13, $14
+; NO-SIMD128-FAST-NEXT:    i32.select $push13=, $pop34, $pop32, $pop12
+; NO-SIMD128-FAST-NEXT:    i32.store16 0($pop15), $pop13
+; NO-SIMD128-FAST-NEXT:    i32.const $push18=, 12
+; NO-SIMD128-FAST-NEXT:    i32.add $push19=, $0, $pop18
+; NO-SIMD128-FAST-NEXT:    i32.extend16_s $push31=, $7
+; NO-SIMD128-FAST-NEXT:    local.tee $push30=, $14=, $pop31
+; NO-SIMD128-FAST-NEXT:    i32.extend16_s $push29=, $15
+; NO-SIMD128-FAST-NEXT:    local.tee $push28=, $15=, $pop29
+; NO-SIMD128-FAST-NEXT:    i32.gt_s $push16=, $14, $15
+; NO-SIMD128-FAST-NEXT:    i32.select $push17=, $pop30, $pop28, $pop16
+; NO-SIMD128-FAST-NEXT:    i32.store16 0($pop19), $pop17
+; NO-SIMD128-FAST-NEXT:    i32.const $push22=, 14
+; NO-SIMD128-FAST-NEXT:    i32.add $push23=, $0, $pop22
+; NO-SIMD128-FAST-NEXT:    i32.extend16_s $push27=, $8
+; NO-SIMD128-FAST-NEXT:    local.tee $push26=, $0=, $pop27
+; NO-SIMD128-FAST-NEXT:    i32.extend16_s $push25=, $16
+; NO-SIMD128-FAST-NEXT:    local.tee $push24=, $16=, $pop25
+; NO-SIMD128-FAST-NEXT:    i32.gt_s $push20=, $0, $16
+; NO-SIMD128-FAST-NEXT:    i32.select $push21=, $pop26, $pop24, $pop20
+; NO-SIMD128-FAST-NEXT:    i32.store16 0($pop23), $pop21
 ; NO-SIMD128-FAST-NEXT:    return
   %c = icmp sgt <8 x i16> %x, %y
   %a = select <8 x i1> %c, <8 x i16> %x, <8 x i16> %y
   ret <8 x i16> %a
 }
 
 define <8 x i16> @max_u_v8i16(<8 x i16> %x, <8 x i16> %y) {
 ; SIMD128-LABEL: max_u_v8i16:
 ; SIMD128:         .functype max_u_v8i16 (v128, v128) -> (v128)
 ; SIMD128-NEXT:  # %bb.0:
 ; SIMD128-NEXT:    i16x8.max_u $push0=, $0, $1
 ; SIMD128-NEXT:    return $pop0
 ;
 ; SIMD128-FAST-LABEL: max_u_v8i16:
 ; SIMD128-FAST:         .functype max_u_v8i16 (v128, v128) -> (v128)
 ; SIMD128-FAST-NEXT:  # %bb.0:
 ; SIMD128-FAST-NEXT:    i16x8.max_u $push0=, $0, $1
 ; SIMD128-FAST-NEXT:    return $pop0
 ;
 ; NO-SIMD128-LABEL: max_u_v8i16:
 ; NO-SIMD128:         .functype max_u_v8i16 (i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32) -> ()
 ; NO-SIMD128-NEXT:  # %bb.0:
-; NO-SIMD128-NEXT:    i32.const $push5=, 14
-; NO-SIMD128-NEXT:    i32.add $push6=, $0, $pop5
+; NO-SIMD128-NEXT:    i32.const $push3=, 14
+; NO-SIMD128-NEXT:    i32.add $push4=, $0, $pop3
 ; NO-SIMD128-NEXT:    i32.const $push0=, 65535
-; NO-SIMD128-NEXT:    i32.and $push2=, $8, $pop0
-; NO-SIMD128-NEXT:    i32.const $push55=, 65535
-; NO-SIMD128-NEXT:    i32.and $push1=, $16, $pop55
-; NO-SIMD128-NEXT:    i32.gt_u $push3=, $pop2, $pop1
-; NO-SIMD128-NEXT:    i32.select $push4=, $8, $16, $pop3
-; NO-SIMD128-NEXT:    i32.store16 0($pop6), $pop4
-; NO-SIMD128-NEXT:    i32.const $push11=, 12
+; NO-SIMD128-NEXT:    i32.and $push71=, $8, $pop0
+; NO-SIMD128-NEXT:    local.tee $push70=, $8=, $pop71
+; NO-SIMD128-NEXT:    i32.const $push69=, 65535
+; NO-SIMD128-NEXT:    i32.and $push68=, $16, $pop69
+; NO-SIMD128-NEXT:    local.tee $push67=, $16=, $pop68
+; NO-SIMD128-NEXT:    i32.gt_u $push1=, $8, $16
+; NO-SIMD128-NEXT:    i32.select $push2=, $pop70, $pop67, $pop1
+; NO-SIMD128-NEXT:    i32.store16 0($pop4), $pop2
+; NO-SIMD128-NEXT:    i32.const $push7=, 12
+; NO-SIMD128-NEXT:    i32.add $push8=, $0, $pop7
+; NO-SIMD128-NEXT:    i32.const $push66=, 65535
+; NO-SIMD128-NEXT:    i32.and $push65=, $7, $pop66
+; NO-SIMD128-NEXT:    local.tee $push64=, $16=, $pop65
+; NO-SIMD128-NEXT:    i32.const $push63=, 65535
+; NO-SIMD128-NEXT:    i32.and $push62=, $15, $pop63
+; NO-SIMD128-NEXT:    local.tee $push61=, $8=, $pop62
+; NO-SIMD128-NEXT:    i32.gt_u $push5=, $16, $8
+; NO-SIMD128-NEXT:    i32.select $push6=, $pop64, $pop61, $pop5
+; NO-SIMD128-NEXT:    i32.store16 0($pop8), $pop6
+; NO-SIMD128-NEXT:    i32.const $push11=, 10
 ; NO-SIMD128-NEXT:    i32.add $push12=, $0, $pop11
-; NO-SIMD128-NEXT:    i32.const $push54=, 65535
-; NO-SIMD128-NEXT:    i32.and $push8=, $7, $pop54
-; NO-SIMD128-NEXT:    i32.const $push53=, 65535
-; NO-SIMD128-NEXT:    i32.and $push7=, $15, $pop53
-; NO-SIMD128-NEXT:    i32.gt_u $push9=, $pop8, $pop7
-; NO-SIMD128-NEXT:    i32.select $push10=, $7, $15, $pop9
+; NO-SIMD128-NEXT:    i32.const $push60=, 65535
+; NO-SIMD128-NEXT:    i32.and $push59=, $6, $pop60
+; NO-SIMD128-NEXT:    local.tee $push58=, $16=, $pop59
+; NO-SIMD128-NEXT:    i32.const $push57=, 65535
+; NO-SIMD128-NEXT:    i32.and $push56=, $14, $pop57
+; NO-SIMD128-NEXT:    local.tee $push55=, $8=, $pop56
+; NO-SIMD128-NEXT:    i32.gt_u $push9=, $16, $8
+; NO-SIMD128-NEXT:    i32.select $push10=, $pop58, $pop55, $pop9
 ; NO-SIMD128-NEXT:    i32.store16 0($pop12), $pop10
-; NO-SIMD128-NEXT:    i32.const $push17=, 10
-; NO-SIMD128-NEXT:    i32.add $push18=, $0, $pop17
-; NO-SIMD128-NEXT:    i32.const $push52=, 65535
-; NO-SIMD128-NEXT:    i32.and $push14=, $6, $pop52
+; NO-SIMD128-NEXT:    i32.const $push54=, 65535
+; NO-SIMD128-NEXT:    i32.and $push53=, $5, $pop54
+; NO-SIMD128-NEXT:    local.tee $push52=, $16=, $pop53
 ; NO-SIMD128-NEXT:    i32.const $push51=, 65535
-; NO-SIMD128-NEXT:    i32.and $push13=, $14, $pop51
-; NO-SIMD128-NEXT:    i32.gt_u $push15=, $pop14, $pop13
-; NO-SIMD128-NEXT:    i32.select $push16=, $6, $14, $pop15
-; NO-SIMD128-NEXT:    i32.store16 0($pop18), $pop16
-; NO-SIMD128-NEXT:    i32.const $push50=, 65535
-; NO-SIMD128-NEXT:    i32.and $push20=, $5, $pop50
-; NO-SIMD128-NEXT:    i32.const $push49=, 65535
-; NO-SIMD128-NEXT:    i32.and $push19=, $13, $pop49
-; NO-SIMD128-NEXT:    i32.gt_u $push21=, $pop20, $pop19
-; NO-SIMD128-NEXT:    i32.select $push22=, $5, $13, $pop21
-; NO-SIMD128-NEXT:    i32.store16 8($0), $pop22
-; NO-SIMD128-NEXT:    i32.const $push27=, 6
-; NO-SIMD128-NEXT:    i32.add $push28=, $0, $pop27
-; NO-SIMD128-NEXT:    i32.const $push48=, 65535
-; NO-SIMD128-NEXT:    i32.and $push24=, $4, $pop48
-; NO-SIMD128-NEXT:    i32.const $push47=, 65535
-; NO-SIMD128-NEXT:    i32.and $push23=, $12, $pop47
-; NO-SIMD128-NEXT:    i32.gt_u $push25=, $pop24, $pop23
-; NO-SIMD128-NEXT:    i32.select $push26=, $4, $12, $pop25
-; NO-SIMD128-NEXT:    i32.store16 0($pop28), $pop26
-; NO-SIMD128-NEXT:    i32.const $push46=, 65535
-; NO-SIMD128-NEXT:    i32.and $push30=, $3, $pop46
+; NO-SIMD128-NEXT:    i32.and $push50=, $13, $pop51
+; NO-SIMD128-NEXT:    local.tee $push49=, $8=, $pop50
+; NO-SIMD128-NEXT:    i32.gt_u $push13=, $16, $8
+; NO-SIMD128-NEXT:    i32.select $push14=, $pop52, $pop49, $pop13
+; NO-SIMD128-NEXT:    i32.store16 8($0), $pop14
+; NO-SIMD128-NEXT:    i32.const $push17=, 6
+; NO-SIMD128-NEXT:    i32.add $push18=, $0, $pop17
+; NO-SIMD128-NEXT:    i32.const $push48=, 65535
+; NO-SIMD128-NEXT:    i32.and $push47=, $4, $pop48
+; NO-SIMD128-NEXT:    local.tee $push46=, $16=, $pop47
 ; NO-SIMD128-NEXT:    i32.const $push45=, 65535
-; NO-SIMD128-NEXT:    i32.and $push29=, $11, $pop45
-; NO-SIMD128-NEXT:    i32.gt_u $push31=, $pop30, $pop29
-; NO-SIMD128-NEXT:    i32.select $push32=, $3, $11, $pop31
-; NO-SIMD128-NEXT:    i32.store16 4($0), $pop32
-; NO-SIMD128-NEXT:    i32.const $push44=, 65535
-; NO-SIMD128-NEXT:    i32.and $push34=, $2, $pop44
-; NO-SIMD128-NEXT:    i32.const $push43=, 65535
-; NO-SIMD128-NEXT:    i32.and $push33=, $10, $pop43
-; NO-SIMD128-NEXT:    i32.gt_u $push35=, $pop34, $pop33
-; NO-SIMD128-NEXT:    i32.select $push36=, $2, $10, $pop35
-; NO-SIMD128-NEXT:    i32.store16 2($0), $pop36
+; NO-SIMD128-NEXT:    i32.and $push44=, $12, $pop45
+; NO-SIMD128-NEXT:    local.tee $push43=, $8=, $pop44
+; NO-SIMD128-NEXT:    i32.gt_u $push15=, $16, $8
+; NO-SIMD128-NEXT:    i32.select $push16=, $pop46, $pop43, $pop15
+; NO-SIMD128-NEXT:    i32.store16 0($pop18), $pop16
 ; NO-SIMD128-NEXT:    i32.const $push42=, 65535
-; NO-SIMD128-NEXT:    i32.and $push38=, $1, $pop42
-; NO-SIMD128-NEXT:    i32.const $push41=, 65535
-; NO-SIMD128-NEXT:    i32.and $push37=, $9, $pop41
-; NO-SIMD128-NEXT:    i32.gt_u $push39=, $pop38, $pop37
-; NO-SIMD128-NEXT:    i32.select $push40=, $1, $9, $pop39
-; NO-SIMD128-NEXT:    i32.store16 0($0), $pop40
+; NO-SIMD128-NEXT:    i32.and $push41=, $3, $pop42
+; NO-SIMD128-NEXT:    local.tee $push40=, $16=, $pop41
+; NO-SIMD128-NEXT:    i32.const $push39=, 65535
+; NO-SIMD128-NEXT:    i32.and $push38=, $11, $pop39
+; NO-SIMD128-NEXT:    local.tee $push37=, $8=, $pop38
+; NO-SIMD128-NEXT:    i32.gt_u $push19=, $16, $8
+; NO-SIMD128-NEXT:    i32.select $push20=, $pop40, $pop37, $pop19
+; NO-SIMD128-NEXT:    i32.store16 4($0), $pop20
+; NO-SIMD128-NEXT:    i32.const $push36=, 65535
+; NO-SIMD128-NEXT:    i32.and $push35=, $2, $pop36
+; NO-SIMD128-NEXT:    local.tee $push34=, $16=, $pop35
+; NO-SIMD128-NEXT:    i32.const $push33=, 65535
+; NO-SIMD128-NEXT:    i32.and $push32=, $10, $pop33
+; NO-SIMD128-NEXT:    local.tee $push31=, $8=, $pop32
+; NO-SIMD128-NEXT:    i32.gt_u $push21=, $16, $8
+; NO-SIMD128-NEXT:    i32.select $push22=, $pop34, $pop31, $pop21
+; NO-SIMD128-NEXT:    i32.store16 2($0), $pop22
+; NO-SIMD128-NEXT:    i32.const $push30=, 65535
+; NO-SIMD128-NEXT:    i32.and $push29=, $1, $pop30
+; NO-SIMD128-NEXT:    local.tee $push28=, $16=, $pop29
+; NO-SIMD128-NEXT:    i32.const $push27=, 65535
+; NO-SIMD128-NEXT:    i32.and $push26=, $9, $pop27
+; NO-SIMD128-NEXT:    local.tee $push25=, $8=, $pop26
+; NO-SIMD128-NEXT:    i32.gt_u $push23=, $16, $8
+; NO-SIMD128-NEXT:    i32.select $push24=, $pop28, $pop25, $pop23
+; NO-SIMD128-NEXT:    i32.store16 0($0), $pop24
 ; NO-SIMD128-NEXT:    return
 ;
 ; NO-SIMD128-FAST-LABEL: max_u_v8i16:
 ; NO-SIMD128-FAST:         .functype max_u_v8i16 (i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32) -> ()
 ; NO-SIMD128-FAST-NEXT:  # %bb.0:
 ; NO-SIMD128-FAST-NEXT:    i32.const $push0=, 65535
-; NO-SIMD128-FAST-NEXT:    i32.and $push2=, $1, $pop0
-; NO-SIMD128-FAST-NEXT:    i32.const $push55=, 65535
-; NO-SIMD128-FAST-NEXT:    i32.and $push1=, $9, $pop55
-; NO-SIMD128-FAST-NEXT:    i32.gt_u $push3=, $pop2, $pop1
-; NO-SIMD128-FAST-NEXT:    i32.select $push4=, $1, $9, $pop3
-; NO-SIMD128-FAST-NEXT:    i32.store16 0($0), $pop4
+; NO-SIMD128-FAST-NEXT:    i32.and $push71=, $1, $pop0
+; NO-SIMD128-FAST-NEXT:    local.tee $push70=, $1=, $pop71
+; NO-SIMD128-FAST-NEXT:    i32.const $push69=, 65535
+; NO-SIMD128-FAST-NEXT:    i32.and $push68=, $9, $pop69
+; NO-SIMD128-FAST-NEXT:    local.tee $push67=, $9=, $pop68
+; NO-SIMD128-FAST-NEXT:    i32.gt_u $push1=, $1, $9
+; NO-SIMD128-FAST-NEXT:    i32.select $push2=, $pop70, $pop67, $pop1
+; NO-SIMD128-FAST-NEXT:    i32.store16 0($0), $pop2
+; NO-SIMD128-FAST-NEXT:    i32.const $push66=, 65535
+; NO-SIMD128-FAST-NEXT:    i32.and $push65=, $2, $pop66
+; NO-SIMD128-FAST-NEXT:    local.tee $push64=, $9=, $pop65
+; NO-SIMD128-FAST-NEXT:    i32.const $push63=, 65535
+; NO-SIMD128-FAST-NEXT:    i32.and $push62=, $10, $pop63
+; NO-SIMD128-FAST-NEXT:    local.tee $push61=, $10=, $pop62
+; NO-SIMD128-FAST-NEXT:    i32.gt_u $push3=, $9, $10
+; NO-SIMD128-FAST-NEXT:    i32.select $push4=, $pop64, $pop61, $pop3
+; NO-SIMD128-FAST-NEXT:    i32.store16 2($0), $pop4
+; NO-SIMD128-FAST-NEXT:    i32.const $push60=, 65535
+; NO-SIMD128-FAST-NEXT:    i32.and $push59=, $3, $pop60
+; NO-SIMD128-FAST-NEXT:    local.tee $push58=, $10=, $pop59
+; NO-SIMD128-FAST-NEXT:    i32.const $push57=, 65535
+; NO-SIMD128-FAST-NEXT:    i32.and $push56=, $11, $pop57
+; NO-SIMD128-FAST-NEXT:    local.tee $push55=, $11=, $pop56
+; NO-SIMD128-FAST-NEXT:    i32.gt_u $push5=, $10, $11
+; NO-SIMD128-FAST-NEXT:    i32.select $push6=, $pop58, $pop55, $pop5
+; NO-SIMD128-FAST-NEXT:    i32.store16 4($0), $pop6
+; NO-SIMD128-FAST-NEXT:    i32.const $push9=, 6
+; NO-SIMD128-FAST-NEXT:    i32.add $push10=, $0, $pop9
 ; NO-SIMD128-FAST-NEXT:    i32.const $push54=, 65535
-; NO-SIMD128-FAST-NEXT:    i32.and $push6=, $2, $pop54
-; NO-SIMD128-FAST-NEXT:    i32.const $push53=, 65535
-; NO-SIMD128-FAST-NEXT:    i32.and $push5=, $10, $pop53
-; NO-SIMD128-FAST-NEXT:    i32.gt_u $push7=, $pop6, $pop5
-; NO-SIMD128-FAST-NEXT:    i32.select $push8=, $2, $10, $pop7
-; NO-SIMD128-FAST-NEXT:    i32.store16 2($0), $pop8
-; NO-SIMD128-FAST-NEXT:    i32.const $push52=, 65535
-; NO-SIMD128-FAST-NEXT:    i32.and $push10=, $3, $pop52
+; NO-SIMD128-FAST-NEXT:    i32.and $push53=, $4, $pop54
+; NO-SIMD128-FAST-NEXT:    local.tee $push52=, $11=, $pop53
 ; NO-SIMD128-FAST-NEXT:    i32.const $push51=, 65535
-; NO-SIMD128-FAST-NEXT:    i32.and $push9=, $11, $pop51
-; NO-SIMD128-FAST-NEXT:    i32.gt_u $push11=, $pop10, $pop9
-; NO-SIMD128-FAST-NEXT:    i32.select $push12=, $3, $11, $pop11
-; NO-SIMD128-FAST-NEXT:    i32.store16 4($0), $pop12
-; NO-SIMD128-FAST-NEXT:    i32.const $push17=, 6
-; NO-SIMD128-FAST-NEXT:    i32.add $push18=, $0, $pop17
-; NO-SIMD128-FAST-NEXT:    i32.const $push50=, 65535
-; NO-SIMD128-FAST-NEXT:    i32.and $push14=, $4, $pop50
-; NO-SIMD128-FAST-NEXT:    i32.const $push49=, 65535
-; NO-SIMD128-FAST-NEXT:    i32.and $push13=, $12, $pop49
-; NO-SIMD128-FAST-NEXT:    i32.gt_u $push15=, $pop14, $pop13
-; NO-SIMD128-FAST-NEXT:    i32.select $push16=, $4, $12, $pop15
-; NO-SIMD128-FAST-NEXT:    i32.store16 0($pop18), $pop16
+; NO-SIMD128-FAST-NEXT:    i32.and $push50=, $12, $pop51
+; NO-SIMD128-FAST-NEXT:    local.tee $push49=, $12=, $pop50
+; NO-SIMD128-FAST-NEXT:    i32.gt_u $push7=, $11, $12
+; NO-SIMD128-FAST-NEXT:    i32.select $push8=, $pop52, $pop49, $pop7
+; NO-SIMD128-FAST-NEXT:    i32.store16 0($pop10), $pop8
 ; NO-SIMD128-FAST-NEXT:    i32.const $push48=, 65535
-; NO-SIMD128-FAST-NEXT:    i32.and $push20=, $5, $pop48
-; NO-SIMD128-FAST-NEXT:    i32.const $push47=, 65535
-; NO-SIMD128-FAST-NEXT:    i32.and $push19=, $13, $pop47
-; NO-SIMD128-FAST-NEXT:    i32.gt_u $push21=, $pop20, $pop19
-; NO-SIMD128-FAST-NEXT:    i32.select $push22=, $5, $13, $pop21
-; NO-SIMD128-FAST-NEXT:    i32.store16 8($0), $pop22
-; NO-SIMD128-FAST-NEXT:    i32.const $push27=, 10
-; NO-SIMD128-FAST-NEXT:    i32.add $push28=, $0, $pop27
-; NO-SIMD128-FAST-NEXT:    i32.const $push46=, 65535
-; NO-SIMD128-FAST-NEXT:    i32.and $push24=, $6, $pop46
+; NO-SIMD128-FAST-NEXT:    i32.and $push47=, $5, $pop48
+; NO-SIMD128-FAST-NEXT:    local.tee $push46=, $12=, $pop47
 ; NO-SIMD128-FAST-NEXT:    i32.const $push45=, 65535
-; NO-SIMD128-FAST-NEXT:    i32.and $push23=, $14, $pop45
-; NO-SIMD128-FAST-NEXT:    i32.gt_u $push25=, $pop24, $pop23
-; NO-SIMD128-FAST-NEXT:    i32.select $push26=, $6, $14, $pop25
-; NO-SIMD128-FAST-NEXT:    i32.store16 0($pop28), $pop26
-; NO-SIMD128-FAST-NEXT:    i32.const $push33=, 12
-; NO-SIMD128-FAST-NEXT:    i32.add $push34=, $0, $pop33
-; NO-SIMD128-FAST-NEXT:    i32.const $push44=, 65535
-; NO-SIMD128-FAST-NEXT:    i32.and $push30=, $7, $pop44
-; NO-SIMD128-FAST-NEXT:    i32.const $push43=, 65535
-; NO-SIMD128-FAST-NEXT:    i32.and $push29=, $15, $pop43
-; NO-SIMD128-FAST-NEXT:    i32.gt_u $push31=, $pop30, $pop29
-; NO-SIMD128-FAST-NEXT:    i32.select $push32=, $7, $15, $pop31
-; NO-SIMD128-FAST-NEXT:    i32.store16 0($pop34), $pop32
-; NO-SIMD128-FAST-NEXT:    i32.const $push39=, 14
-; NO-SIMD128-FAST-NEXT:    i32.add $push40=, $0, $pop39
+; NO-SIMD128-FAST-NEXT:    i32.and $push44=, $13, $pop45
+; NO-SIMD128-FAST-NEXT:    local.tee $push43=, $13=, $pop44
+; NO-SIMD128-FAST-NEXT:    i32.gt_u $push11=, $12, $13
+; NO-SIMD128-FAST-NEXT:    i32.select $push12=, $pop46, $pop43, $pop11
+; NO-SIMD128-FAST-NEXT:    i32.store16 8($0), $pop12
+; NO-SIMD128-FAST-NEXT:    i32.const $push15=, 10
+; NO-SIMD128-FAST-NEXT:    i32.add $push16=, $0, $pop15
 ; NO-SIMD128-FAST-NEXT:    i32.const $push42=, 65535
-; NO-SIMD128-FAST-NEXT:    i32.and $push36=, $8, $pop42
-; NO-SIMD128-FAST-NEXT:    i32.const $push41=, 65535
-; NO-SIMD128-FAST-NEXT:    i32.and $push35=, $16, $pop41
-; NO-SIMD128-FAST-NEXT:    i32.gt_u $push37=, $pop36, $pop35
-; NO-SIMD128-FAST-NEXT:    i32.select $push38=, $8, $16, $pop37
-; NO-SIMD128-FAST-NEXT:    i32.store16 0($pop40), $pop38
+; NO-SIMD128-FAST-NEXT:    i32.and $push41=, $6, $pop42
+; NO-SIMD128-FAST-NEXT:    local.tee $push40=, $13=, $pop41
+; NO-SIMD128-FAST-NEXT:    i32.const $push39=, 65535
+; NO-SIMD128-FAST-NEXT:    i32.and $push38=, $14, $pop39
+; NO-SIMD128-FAST-NEXT:    local.tee $push37=, $14=, $pop38
+; NO-SIMD128-FAST-NEXT:    i32.gt_u $push13=, $13, $14
+; NO-SIMD128-FAST-NEXT:    i32.select $push14=, $pop40, $pop37, $pop13
+; NO-SIMD128-FAST-NEXT:    i32.store16 0($pop16), $pop14
+; NO-SIMD128-FAST-NEXT:    i32.const $push19=, 12
+; NO-SIMD128-FAST-NEXT:    i32.add $push20=, $0, $pop19
+; NO-SIMD128-FAST-NEXT:    i32.const $push36=, 65535
+; NO-SIMD128-FAST-NEXT:    i32.and $push35=, $7, $pop36
+; NO-SIMD128-FAST-NEXT:    local.tee $push34=, $14=, $pop35
+; NO-SIMD128-FAST-NEXT:    i32.const $push33=, 65535
+; NO-SIMD128-FAST-NEXT:    i32.and $push32=, $15, $pop33
+; NO-SIMD128-FAST-NEXT:    local.tee $push31=, $15=, $pop32
+; NO-SIMD128-FAST-NEXT:    i32.gt_u $push17=, $14, $15
+; NO-SIMD128-FAST-NEXT:    i32.select $push18=, $pop34, $pop31, $pop17
+; NO-SIMD128-FAST-NEXT:    i32.store16 0($pop20), $pop18
+; NO-SIMD128-FAST-NEXT:    i32.const $push23=, 14
+; NO-SIMD128-FAST-NEXT:    i32.add $push24=, $0, $pop23
+; NO-SIMD128-FAST-NEXT:    i32.const $push30=, 65535
+; NO-SIMD128-FAST-NEXT:    i32.and $push29=, $8, $pop30
+; NO-SIMD128-FAST-NEXT:    local.tee $push28=, $0=, $pop29
+; NO-SIMD128-FAST-NEXT:    i32.const $push27=, 65535
+; NO-SIMD128-FAST-NEXT:    i32.and $push26=, $16, $pop27
+; NO-SIMD128-FAST-NEXT:    local.tee $push25=, $16=, $pop26
+; NO-SIMD128-FAST-NEXT:    i32.gt_u $push21=, $0, $16
+; NO-SIMD128-FAST-NEXT:    i32.select $push22=, $pop28, $pop25, $pop21
+; NO-SIMD128-FAST-NEXT:    i32.store16 0($pop24), $pop22
 ; NO-SIMD128-FAST-NEXT:    return
   %c = icmp ugt <8 x i16> %x, %y
   %a = select <8 x i1> %c, <8 x i16> %x, <8 x i16> %y
   ret <8 x i16> %a
 }
 
 define <8 x i16> @avgr_u_v8i16(<8 x i16> %x, <8 x i16> %y) {
 ; SIMD128-LABEL: avgr_u_v8i16:

 ; SIMD128-NEXT:  # %bb.0:
 ; SIMD128-NEXT:    v128.andnot $push0=, $0, $1
 ; SIMD128-NEXT:    return $pop0
 ;
 ; SIMD128-FAST-LABEL: andnot_v8i16:
 ; SIMD128-FAST:         .functype andnot_v8i16 (v128, v128) -> (v128)
 ; SIMD128-FAST-NEXT:  # %bb.0:
 ; SIMD128-FAST-NEXT:    v128.not $push0=, $1
-; SIMD128-FAST-NEXT:    v128.and $push1=, $0, $pop0
+; SIMD128-FAST-NEXT:    v128.and $push1=, $pop0, $0
 ; SIMD128-FAST-NEXT:    return $pop1
 ;
 ; NO-SIMD128-LABEL: andnot_v8i16:
 ; NO-SIMD128:         .functype andnot_v8i16 (i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32) -> ()
 ; NO-SIMD128-NEXT:  # %bb.0:
 ; NO-SIMD128-NEXT:    i32.const $push0=, -1
 ; NO-SIMD128-NEXT:    i32.xor $push1=, $13, $pop0
-; NO-SIMD128-NEXT:    i32.and $push2=, $5, $pop1
+; NO-SIMD128-NEXT:    i32.and $push2=, $pop1, $5
 ; NO-SIMD128-NEXT:    i32.store16 8($0), $pop2
 ; NO-SIMD128-NEXT:    i32.const $push31=, -1
 ; NO-SIMD128-NEXT:    i32.xor $push3=, $11, $pop31
-; NO-SIMD128-NEXT:    i32.and $push4=, $3, $pop3
+; NO-SIMD128-NEXT:    i32.and $push4=, $pop3, $3
 ; NO-SIMD128-NEXT:    i32.store16 4($0), $pop4
 ; NO-SIMD128-NEXT:    i32.const $push30=, -1
 ; NO-SIMD128-NEXT:    i32.xor $push5=, $10, $pop30
-; NO-SIMD128-NEXT:    i32.and $push6=, $2, $pop5
+; NO-SIMD128-NEXT:    i32.and $push6=, $pop5, $2
 ; NO-SIMD128-NEXT:    i32.store16 2($0), $pop6
 ; NO-SIMD128-NEXT:    i32.const $push29=, -1
 ; NO-SIMD128-NEXT:    i32.xor $push7=, $9, $pop29
-; NO-SIMD128-NEXT:    i32.and $push8=, $1, $pop7
+; NO-SIMD128-NEXT:    i32.and $push8=, $pop7, $1
 ; NO-SIMD128-NEXT:    i32.store16 0($0), $pop8
 ; NO-SIMD128-NEXT:    i32.const $push11=, 14
 ; NO-SIMD128-NEXT:    i32.add $push12=, $0, $pop11
 ; NO-SIMD128-NEXT:    i32.const $push28=, -1
 ; NO-SIMD128-NEXT:    i32.xor $push9=, $16, $pop28
-; NO-SIMD128-NEXT:    i32.and $push10=, $8, $pop9
+; NO-SIMD128-NEXT:    i32.and $push10=, $pop9, $8
 ; NO-SIMD128-NEXT:    i32.store16 0($pop12), $pop10
 ; NO-SIMD128-NEXT:    i32.const $push15=, 12
 ; NO-SIMD128-NEXT:    i32.add $push16=, $0, $pop15
 ; NO-SIMD128-NEXT:    i32.const $push27=, -1
 ; NO-SIMD128-NEXT:    i32.xor $push13=, $15, $pop27
-; NO-SIMD128-NEXT:    i32.and $push14=, $7, $pop13
+; NO-SIMD128-NEXT:    i32.and $push14=, $pop13, $7
 ; NO-SIMD128-NEXT:    i32.store16 0($pop16), $pop14
 ; NO-SIMD128-NEXT:    i32.const $push19=, 10
 ; NO-SIMD128-NEXT:    i32.add $push20=, $0, $pop19
 ; NO-SIMD128-NEXT:    i32.const $push26=, -1
 ; NO-SIMD128-NEXT:    i32.xor $push17=, $14, $pop26
-; NO-SIMD128-NEXT:    i32.and $push18=, $6, $pop17
+; NO-SIMD128-NEXT:    i32.and $push18=, $pop17, $6
 ; NO-SIMD128-NEXT:    i32.store16 0($pop20), $pop18
 ; NO-SIMD128-NEXT:    i32.const $push23=, 6
 ; NO-SIMD128-NEXT:    i32.add $push24=, $0, $pop23
 ; NO-SIMD128-NEXT:    i32.const $push25=, -1
 ; NO-SIMD128-NEXT:    i32.xor $push21=, $12, $pop25
-; NO-SIMD128-NEXT:    i32.and $push22=, $4, $pop21
+; NO-SIMD128-NEXT:    i32.and $push22=, $pop21, $4
 ; NO-SIMD128-NEXT:    i32.store16 0($pop24), $pop22
 ; NO-SIMD128-NEXT:    return
 ;
 ; NO-SIMD128-FAST-LABEL: andnot_v8i16:
 ; NO-SIMD128-FAST:         .functype andnot_v8i16 (i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32) -> ()
 ; NO-SIMD128-FAST-NEXT:  # %bb.0:
 ; NO-SIMD128-FAST-NEXT:    i32.const $push0=, -1
 ; NO-SIMD128-FAST-NEXT:    i32.xor $push1=, $9, $pop0
-; NO-SIMD128-FAST-NEXT:    i32.and $push2=, $1, $pop1
+; NO-SIMD128-FAST-NEXT:    i32.and $push2=, $pop1, $1
 ; NO-SIMD128-FAST-NEXT:    i32.store16 0($0), $pop2
 ; NO-SIMD128-FAST-NEXT:    i32.const $push31=, -1
 ; NO-SIMD128-FAST-NEXT:    i32.xor $push3=, $10, $pop31
-; NO-SIMD128-FAST-NEXT:    i32.and $push4=, $2, $pop3
+; NO-SIMD128-FAST-NEXT:    i32.and $push4=, $pop3, $2
 ; NO-SIMD128-FAST-NEXT:    i32.store16 2($0), $pop4
 ; NO-SIMD128-FAST-NEXT:    i32.const $push30=, -1
 ; NO-SIMD128-FAST-NEXT:    i32.xor $push5=, $11, $pop30
-; NO-SIMD128-FAST-NEXT:    i32.and $push6=, $3, $pop5
+; NO-SIMD128-FAST-NEXT:    i32.and $push6=, $pop5, $3
 ; NO-SIMD128-FAST-NEXT:    i32.store16 4($0), $pop6
 ; NO-SIMD128-FAST-NEXT:    i32.const $push7=, 6
 ; NO-SIMD128-FAST-NEXT:    i32.add $push8=, $0, $pop7
 ; NO-SIMD128-FAST-NEXT:    i32.const $push29=, -1
 ; NO-SIMD128-FAST-NEXT:    i32.xor $push9=, $12, $pop29
-; NO-SIMD128-FAST-NEXT:    i32.and $push10=, $4, $pop9
+; NO-SIMD128-FAST-NEXT:    i32.and $push10=, $pop9, $4
 ; NO-SIMD128-FAST-NEXT:    i32.store16 0($pop8), $pop10
 ; NO-SIMD128-FAST-NEXT:    i32.const $push28=, -1
 ; NO-SIMD128-FAST-NEXT:    i32.xor $push11=, $13, $pop28
-; NO-SIMD128-FAST-NEXT:    i32.and $push12=, $5, $pop11
+; NO-SIMD128-FAST-NEXT:    i32.and $push12=, $pop11, $5
 ; NO-SIMD128-FAST-NEXT:    i32.store16 8($0), $pop12
 ; NO-SIMD128-FAST-NEXT:    i32.const $push13=, 10
 ; NO-SIMD128-FAST-NEXT:    i32.add $push14=, $0, $pop13
 ; NO-SIMD128-FAST-NEXT:    i32.const $push27=, -1
 ; NO-SIMD128-FAST-NEXT:    i32.xor $push15=, $14, $pop27
-; NO-SIMD128-FAST-NEXT:    i32.and $push16=, $6, $pop15
+; NO-SIMD128-FAST-NEXT:    i32.and $push16=, $pop15, $6
 ; NO-SIMD128-FAST-NEXT:    i32.store16 0($pop14), $pop16
 ; NO-SIMD128-FAST-NEXT:    i32.const $push17=, 12
 ; NO-SIMD128-FAST-NEXT:    i32.add $push18=, $0, $pop17
 ; NO-SIMD128-FAST-NEXT:    i32.const $push26=, -1
 ; NO-SIMD128-FAST-NEXT:    i32.xor $push19=, $15, $pop26
-; NO-SIMD128-FAST-NEXT:    i32.and $push20=, $7, $pop19
+; NO-SIMD128-FAST-NEXT:    i32.and $push20=, $pop19, $7
 ; NO-SIMD128-FAST-NEXT:    i32.store16 0($pop18), $pop20
 ; NO-SIMD128-FAST-NEXT:    i32.const $push21=, 14
 ; NO-SIMD128-FAST-NEXT:    i32.add $push22=, $0, $pop21
 ; NO-SIMD128-FAST-NEXT:    i32.const $push25=, -1
 ; NO-SIMD128-FAST-NEXT:    i32.xor $push23=, $16, $pop25
-; NO-SIMD128-FAST-NEXT:    i32.and $push24=, $8, $pop23
+; NO-SIMD128-FAST-NEXT:    i32.and $push24=, $pop23, $8
 ; NO-SIMD128-FAST-NEXT:    i32.store16 0($pop22), $pop24
 ; NO-SIMD128-FAST-NEXT:    return
  %inv_y = xor <8 x i16> %y,
    <i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1>
  %a = and <8 x i16> %x, %inv_y
  ret <8 x i16> %a
 }
 
 define <8 x i16> @bitselect_v8i16(<8 x i16> %c, <8 x i16> %v1, <8 x i16> %v2) {
 ; SIMD128-LABEL: bitselect_v8i16:
 ; SIMD128:         .functype bitselect_v8i16 (v128, v128, v128) -> (v128)
 ; SIMD128-NEXT:  # %bb.0:
 ; SIMD128-NEXT:    v128.bitselect $push0=, $1, $2, $0
 ; SIMD128-NEXT:    return $pop0
 ;
 ; SIMD128-FAST-LABEL: bitselect_v8i16:
 ; SIMD128-FAST:         .functype bitselect_v8i16 (v128, v128, v128) -> (v128)
 ; SIMD128-FAST-NEXT:  # %bb.0:
 ; SIMD128-FAST-NEXT:    v128.and $push0=, $1, $0
 ; SIMD128-FAST-NEXT:    v128.not $push2=, $0
-; SIMD128-FAST-NEXT:    v128.and $push3=, $2, $pop2
+; SIMD128-FAST-NEXT:    v128.and $push3=, $pop2, $2
 ; SIMD128-FAST-NEXT:    v128.or $push1=, $pop0, $pop3
 ; SIMD128-FAST-NEXT:    return $pop1
 ;
 ; NO-SIMD128-LABEL: bitselect_v8i16:
 ; NO-SIMD128:         .functype bitselect_v8i16 (i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32) -> ()
 ; NO-SIMD128-NEXT:  # %bb.0:
 ; NO-SIMD128-NEXT:    i32.const $push5=, 14
 ; NO-SIMD128-NEXT:    i32.add $push6=, $0, $pop5
 ; NO-SIMD128-NEXT:    i32.and $push0=, $16, $8
 ; NO-SIMD128-NEXT:    i32.const $push1=, -1
 ; NO-SIMD128-NEXT:    i32.xor $push2=, $8, $pop1
-; NO-SIMD128-NEXT:    i32.and $push3=, $24, $pop2
+; NO-SIMD128-NEXT:    i32.and $push3=, $pop2, $24
 ; NO-SIMD128-NEXT:    i32.or $push4=, $pop0, $pop3
 ; NO-SIMD128-NEXT:    i32.store16 0($pop6), $pop4
 ; NO-SIMD128-NEXT:    i32.const $push11=, 12
 ; NO-SIMD128-NEXT:    i32.add $push12=, $0, $pop11
 ; NO-SIMD128-NEXT:    i32.and $push7=, $15, $7
 ; NO-SIMD128-NEXT:    i32.const $push47=, -1
 ; NO-SIMD128-NEXT:    i32.xor $push8=, $7, $pop47
-; NO-SIMD128-NEXT:    i32.and $push9=, $23, $pop8
+; NO-SIMD128-NEXT:    i32.and $push9=, $pop8, $23
 ; NO-SIMD128-NEXT:    i32.or $push10=, $pop7, $pop9
 ; NO-SIMD128-NEXT:    i32.store16 0($pop12), $pop10
 ; NO-SIMD128-NEXT:    i32.const $push17=, 10
 ; NO-SIMD128-NEXT:    i32.add $push18=, $0, $pop17
 ; NO-SIMD128-NEXT:    i32.and $push13=, $14, $6
 ; NO-SIMD128-NEXT:    i32.const $push46=, -1
 ; NO-SIMD128-NEXT:    i32.xor $push14=, $6, $pop46
-; NO-SIMD128-NEXT:    i32.and $push15=, $22, $pop14
+; NO-SIMD128-NEXT:    i32.and $push15=, $pop14, $22
 ; NO-SIMD128-NEXT:    i32.or $push16=, $pop13, $pop15
 ; NO-SIMD128-NEXT:    i32.store16 0($pop18), $pop16
 ; NO-SIMD128-NEXT:    i32.and $push19=, $13, $5
 ; NO-SIMD128-NEXT:    i32.const $push45=, -1
 ; NO-SIMD128-NEXT:    i32.xor $push20=, $5, $pop45
-; NO-SIMD128-NEXT:    i32.and $push21=, $21, $pop20
+; NO-SIMD128-NEXT:    i32.and $push21=, $pop20, $21
 ; NO-SIMD128-NEXT:    i32.or $push22=, $pop19, $pop21
 ; NO-SIMD128-NEXT:    i32.store16 8($0), $pop22
 ; NO-SIMD128-NEXT:    i32.const $push27=, 6
 ; NO-SIMD128-NEXT:    i32.add $push28=, $0, $pop27
 ; NO-SIMD128-NEXT:    i32.and $push23=, $12, $4
 ; NO-SIMD128-NEXT:    i32.const $push44=, -1
 ; NO-SIMD128-NEXT:    i32.xor $push24=, $4, $pop44
-; NO-SIMD128-NEXT:    i32.and $push25=, $20, $pop24
+; NO-SIMD128-NEXT:    i32.and $push25=, $pop24, $20
 ; NO-SIMD128-NEXT:    i32.or $push26=, $pop23, $pop25
 ; NO-SIMD128-NEXT:    i32.store16 0($pop28), $pop26
 ; NO-SIMD128-NEXT:    i32.and $push29=, $11, $3
 ; NO-SIMD128-NEXT:    i32.const $push43=, -1
 ; NO-SIMD128-NEXT:    i32.xor $push30=, $3, $pop43
-; NO-SIMD128-NEXT:    i32.and $push31=, $19, $pop30
+; NO-SIMD128-NEXT:    i32.and $push31=, $pop30, $19
 ; NO-SIMD128-NEXT:    i32.or $push32=, $pop29, $pop31
 ; NO-SIMD128-NEXT:    i32.store16 4($0), $pop32
 ; NO-SIMD128-NEXT:    i32.and $push33=, $10, $2
 ; NO-SIMD128-NEXT:    i32.const $push42=, -1
 ; NO-SIMD128-NEXT:    i32.xor $push34=, $2, $pop42
-; NO-SIMD128-NEXT:    i32.and $push35=, $18, $pop34
+; NO-SIMD128-NEXT:    i32.and $push35=, $pop34, $18
 ; NO-SIMD128-NEXT:    i32.or $push36=, $pop33, $pop35
 ; NO-SIMD128-NEXT:    i32.store16 2($0), $pop36
 ; NO-SIMD128-NEXT:    i32.and $push37=, $9, $1
 ; NO-SIMD128-NEXT:    i32.const $push41=, -1
 ; NO-SIMD128-NEXT:    i32.xor $push38=, $1, $pop41
-; NO-SIMD128-NEXT:    i32.and $push39=, $17, $pop38
+; NO-SIMD128-NEXT:    i32.and $push39=, $pop38, $17
 ; NO-SIMD128-NEXT:    i32.or $push40=, $pop37, $pop39
 ; NO-SIMD128-NEXT:    i32.store16 0($0), $pop40
 ; NO-SIMD128-NEXT:    return
 ;
 ; NO-SIMD128-FAST-LABEL: bitselect_v8i16:
 ; NO-SIMD128-FAST:         .functype bitselect_v8i16 (i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32) -> ()
 ; NO-SIMD128-FAST-NEXT:  # %bb.0:
 ; NO-SIMD128-FAST-NEXT:    i32.and $push0=, $9, $1
 ; NO-SIMD128-FAST-NEXT:    i32.const $push1=, -1
 ; NO-SIMD128-FAST-NEXT:    i32.xor $push2=, $1, $pop1
-; NO-SIMD128-FAST-NEXT:    i32.and $push3=, $17, $pop2
+; NO-SIMD128-FAST-NEXT:    i32.and $push3=, $pop2, $17
 ; NO-SIMD128-FAST-NEXT:    i32.or $push4=, $pop0, $pop3
 ; NO-SIMD128-FAST-NEXT:    i32.store16 0($0), $pop4
 ; NO-SIMD128-FAST-NEXT:    i32.and $push5=, $10, $2
 ; NO-SIMD128-FAST-NEXT:    i32.const $push47=, -1
 ; NO-SIMD128-FAST-NEXT:    i32.xor $push6=, $2, $pop47
-; NO-SIMD128-FAST-NEXT:    i32.and $push7=, $18, $pop6
+; NO-SIMD128-FAST-NEXT:    i32.and $push7=, $pop6, $18
 ; NO-SIMD128-FAST-NEXT:    i32.or $push8=, $pop5, $pop7
 ; NO-SIMD128-FAST-NEXT:    i32.store16 2($0), $pop8
 ; NO-SIMD128-FAST-NEXT:    i32.and $push9=, $11, $3
 ; NO-SIMD128-FAST-NEXT:    i32.const $push46=, -1
 ; NO-SIMD128-FAST-NEXT:    i32.xor $push10=, $3, $pop46
-; NO-SIMD128-FAST-NEXT:    i32.and $push11=, $19, $pop10
+; NO-SIMD128-FAST-NEXT:    i32.and $push11=, $pop10, $19
 ; NO-SIMD128-FAST-NEXT:    i32.or $push12=, $pop9, $pop11
 ; NO-SIMD128-FAST-NEXT:    i32.store16 4($0), $pop12
 ; NO-SIMD128-FAST-NEXT:    i32.const $push17=, 6
 ; NO-SIMD128-FAST-NEXT:    i32.add $push18=, $0, $pop17
 ; NO-SIMD128-FAST-NEXT:    i32.and $push13=, $12, $4
 ; NO-SIMD128-FAST-NEXT:    i32.const $push45=, -1
 ; NO-SIMD128-FAST-NEXT:    i32.xor $push14=, $4, $pop45
-; NO-SIMD128-FAST-NEXT:    i32.and $push15=, $20, $pop14
+; NO-SIMD128-FAST-NEXT:    i32.and $push15=, $pop14, $20
 ; NO-SIMD128-FAST-NEXT:    i32.or $push16=, $pop13, $pop15
 ; NO-SIMD128-FAST-NEXT:    i32.store16 0($pop18), $pop16
 ; NO-SIMD128-FAST-NEXT:    i32.and $push19=, $13, $5
 ; NO-SIMD128-FAST-NEXT:    i32.const $push44=, -1
 ; NO-SIMD128-FAST-NEXT:    i32.xor $push20=, $5, $pop44
-; NO-SIMD128-FAST-NEXT:    i32.and $push21=, $21, $pop20
+; NO-SIMD128-FAST-NEXT:    i32.and $push21=, $pop20, $21
 ; NO-SIMD128-FAST-NEXT:    i32.or $push22=, $pop19, $pop21
 ; NO-SIMD128-FAST-NEXT:    i32.store16 8($0), $pop22
 ; NO-SIMD128-FAST-NEXT:    i32.const $push27=, 10
 ; NO-SIMD128-FAST-NEXT:    i32.add $push28=, $0, $pop27
 ; NO-SIMD128-FAST-NEXT:    i32.and $push23=, $14, $6
 ; NO-SIMD128-FAST-NEXT:    i32.const $push43=, -1
 ; NO-SIMD128-FAST-NEXT:    i32.xor $push24=, $6, $pop43
-; NO-SIMD128-FAST-NEXT:    i32.and $push25=, $22, $pop24
+; NO-SIMD128-FAST-NEXT:    i32.and $push25=, $pop24, $22
 ; NO-SIMD128-FAST-NEXT:    i32.or $push26=, $pop23, $pop25
 ; NO-SIMD128-FAST-NEXT:    i32.store16 0($pop28), $pop26
 ; NO-SIMD128-FAST-NEXT:    i32.const $push33=, 12
 ; NO-SIMD128-FAST-NEXT:    i32.add $push34=, $0, $pop33
 ; NO-SIMD128-FAST-NEXT:    i32.and $push29=, $15, $7
 ; NO-SIMD128-FAST-NEXT:    i32.const $push42=, -1
 ; NO-SIMD128-FAST-NEXT:    i32.xor $push30=, $7, $pop42
-; NO-SIMD128-FAST-NEXT:    i32.and $push31=, $23, $pop30
+; NO-SIMD128-FAST-NEXT:    i32.and $push31=, $pop30, $23
 ; NO-SIMD128-FAST-NEXT:    i32.or $push32=, $pop29, $pop31
 ; NO-SIMD128-FAST-NEXT:    i32.store16 0($pop34), $pop32
 ; NO-SIMD128-FAST-NEXT:    i32.const $push39=, 14
 ; NO-SIMD128-FAST-NEXT:    i32.add $push40=, $0, $pop39
 ; NO-SIMD128-FAST-NEXT:    i32.and $push35=, $16, $8
 ; NO-SIMD128-FAST-NEXT:    i32.const $push41=, -1
 ; NO-SIMD128-FAST-NEXT:    i32.xor $push36=, $8, $pop41
-; NO-SIMD128-FAST-NEXT:    i32.and $push37=, $24, $pop36
+; NO-SIMD128-FAST-NEXT:    i32.and $push37=, $pop36, $24
 ; NO-SIMD128-FAST-NEXT:    i32.or $push38=, $pop35, $pop37
 ; NO-SIMD128-FAST-NEXT:    i32.store16 0($pop40), $pop38
 ; NO-SIMD128-FAST-NEXT:    return
   %masked_v1 = and <8 x i16> %v1, %c
   %inv_mask = xor <8 x i16>
     <i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1, i16 -1>,
     %c
   %masked_v2 = and <8 x i16> %v2, %inv_mask

 ; SIMD128:         .functype bitselect_xor_reversed_v8i16 (v128, v128, v128) -> (v128)
 ; SIMD128-NEXT:  # %bb.0:
 ; SIMD128-NEXT:    v128.bitselect $push0=, $2, $1, $0
 ; SIMD128-NEXT:    return $pop0
 ;
 ; SIMD128-FAST-LABEL: bitselect_xor_reversed_v8i16:
 ; SIMD128-FAST:         .functype bitselect_xor_reversed_v8i16 (v128, v128, v128) -> (v128)
 ; SIMD128-FAST-NEXT:  # %bb.0:
-; SIMD128-FAST-NEXT:    v128.xor $push1=, $1, $2
-; SIMD128-FAST-NEXT:    v128.not $push2=, $0
-; SIMD128-FAST-NEXT:    v128.and $push3=, $pop1, $pop2
-; SIMD128-FAST-NEXT:    v128.xor $push0=, $pop3, $2
+; SIMD128-FAST-NEXT:    v128.xor $push2=, $1, $2
+; SIMD128-FAST-NEXT:    v128.and $push1=, $pop2, $0
+; SIMD128-FAST-NEXT:    v128.xor $push0=, $pop1, $1
 ; SIMD128-FAST-NEXT:    return $pop0
 ;
 ; NO-SIMD128-LABEL: bitselect_xor_reversed_v8i16:
 ; NO-SIMD128:         .functype bitselect_xor_reversed_v8i16 (i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32) -> ()
 ; NO-SIMD128-NEXT:  # %bb.0:
-; NO-SIMD128-NEXT:    i32.const $push5=, 14
-; NO-SIMD128-NEXT:    i32.add $push6=, $0, $pop5
-; NO-SIMD128-NEXT:    i32.xor $push2=, $16, $24
-; NO-SIMD128-NEXT:    i32.const $push0=, -1
-; NO-SIMD128-NEXT:    i32.xor $push1=, $8, $pop0
-; NO-SIMD128-NEXT:    i32.and $push3=, $pop2, $pop1
-; NO-SIMD128-NEXT:    i32.xor $push4=, $pop3, $24
-; NO-SIMD128-NEXT:    i32.store16 0($pop6), $pop4
-; NO-SIMD128-NEXT:    i32.const $push11=, 12
-; NO-SIMD128-NEXT:    i32.add $push12=, $0, $pop11
-; NO-SIMD128-NEXT:    i32.xor $push8=, $15, $23
-; NO-SIMD128-NEXT:    i32.const $push47=, -1
-; NO-SIMD128-NEXT:    i32.xor $push7=, $7, $pop47
-; NO-SIMD128-NEXT:    i32.and $push9=, $pop8, $pop7
-; NO-SIMD128-NEXT:    i32.xor $push10=, $pop9, $23
-; NO-SIMD128-NEXT:    i32.store16 0($pop12), $pop10
-; NO-SIMD128-NEXT:    i32.const $push17=, 10
-; NO-SIMD128-NEXT:    i32.add $push18=, $0, $pop17
-; NO-SIMD128-NEXT:    i32.xor $push14=, $14, $22
-; NO-SIMD128-NEXT:    i32.const $push46=, -1
-; NO-SIMD128-NEXT:    i32.xor $push13=, $6, $pop46
-; NO-SIMD128-NEXT:    i32.and $push15=, $pop14, $pop13
-; NO-SIMD128-NEXT:    i32.xor $push16=, $pop15, $22
-; NO-SIMD128-NEXT:    i32.store16 0($pop18), $pop16
-; NO-SIMD128-NEXT:    i32.xor $push20=, $13, $21
-; NO-SIMD128-NEXT:    i32.const $push45=, -1
-; NO-SIMD128-NEXT:    i32.xor $push19=, $5, $pop45
-; NO-SIMD128-NEXT:    i32.and $push21=, $pop20, $pop19
-; NO-SIMD128-NEXT:    i32.xor $push22=, $pop21, $21
-; NO-SIMD128-NEXT:    i32.store16 8($0), $pop22
-; NO-SIMD128-NEXT:    i32.const $push27=, 6
-; NO-SIMD128-NEXT:    i32.add $push28=, $0, $pop27
-; NO-SIMD128-NEXT:    i32.xor $push24=, $12, $20
-; NO-SIMD128-NEXT:    i32.const $push44=, -1
-; NO-SIMD128-NEXT:    i32.xor $push23=, $4, $pop44
-; NO-SIMD128-NEXT:    i32.and $push25=, $pop24, $pop23
-; NO-SIMD128-NEXT:    i32.xor $push26=, $pop25, $20
-; NO-SIMD128-NEXT:    i32.store16 0($pop28), $pop26
-; NO-SIMD128-NEXT:    i32.xor $push30=, $11, $19
-; NO-SIMD128-NEXT:    i32.const $push43=, -1
-; NO-SIMD128-NEXT:    i32.xor $push29=, $3, $pop43
-; NO-SIMD128-NEXT:    i32.and $push31=, $pop30, $pop29
-; NO-SIMD128-NEXT:    i32.xor $push32=, $pop31, $19
-; NO-SIMD128-NEXT:    i32.store16 4($0), $pop32
-; NO-SIMD128-NEXT:    i32.xor $push34=, $10, $18
-; NO-SIMD128-NEXT:    i32.const $push42=, -1
-; NO-SIMD128-NEXT:    i32.xor $push33=, $2, $pop42
-; NO-SIMD128-NEXT:    i32.and $push35=, $pop34, $pop33
-; NO-SIMD128-NEXT:    i32.xor $push36=, $pop35, $18
-; NO-SIMD128-NEXT:    i32.store16 2($0), $pop36
-; NO-SIMD128-NEXT:    i32.xor $push38=, $9, $17
-; NO-SIMD128-NEXT:    i32.const $push41=, -1
-; NO-SIMD128-NEXT:    i32.xor $push37=, $1, $pop41
-; NO-SIMD128-NEXT:    i32.and $push39=, $pop38, $pop37
-; NO-SIMD128-NEXT:    i32.xor $push40=, $pop39, $17
-; NO-SIMD128-NEXT:    i32.store16 0($0), $pop40
+; NO-SIMD128-NEXT:    i32.const $push3=, 14
+; NO-SIMD128-NEXT:    i32.add $push4=, $0, $pop3
+; NO-SIMD128-NEXT:    i32.xor $push0=, $16, $24
+; NO-SIMD128-NEXT:    i32.and $push1=, $pop0, $8
+; NO-SIMD128-NEXT:    i32.xor $push2=, $pop1, $16
+; NO-SIMD128-NEXT:    i32.store16 0($pop4), $pop2
+; NO-SIMD128-NEXT:    i32.const $push8=, 12
+; NO-SIMD128-NEXT:    i32.add $push9=, $0, $pop8
+; NO-SIMD128-NEXT:    i32.xor $push5=, $15, $23
+; NO-SIMD128-NEXT:    i32.and $push6=, $pop5, $7
+; NO-SIMD128-NEXT:    i32.xor $push7=, $pop6, $15
+; NO-SIMD128-NEXT:    i32.store16 0($pop9), $pop7
+; NO-SIMD128-NEXT:    i32.const $push13=, 10
+; NO-SIMD128-NEXT:    i32.add $push14=, $0, $pop13
+; NO-SIMD128-NEXT:    i32.xor $push10=, $14, $22
+; NO-SIMD128-NEXT:    i32.and $push11=, $pop10, $6
+; NO-SIMD128-NEXT:    i32.xor $push12=, $pop11, $14
+; NO-SIMD128-NEXT:    i32.store16 0($pop14), $pop12
+; NO-SIMD128-NEXT:    i32.xor $push15=, $13, $21
+; NO-SIMD128-NEXT:    i32.and $push16=, $pop15, $5
+; NO-SIMD128-NEXT:    i32.xor $push17=, $pop16, $13
+; NO-SIMD128-NEXT:    i32.store16 8($0), $pop17
+; NO-SIMD128-NEXT:    i32.const $push21=, 6
+; NO-SIMD128-NEXT:    i32.add $push22=, $0, $pop21
+; NO-SIMD128-NEXT:    i32.xor $push18=, $12, $20
+; NO-SIMD128-NEXT:    i32.and $push19=, $pop18, $4
+; NO-SIMD128-NEXT:    i32.xor $push20=, $pop19, $12
+; NO-SIMD128-NEXT:    i32.store16 0($pop22), $pop20
+; NO-SIMD128-NEXT:    i32.xor $push23=, $11, $19
+; NO-SIMD128-NEXT:    i32.and $push24=, $pop23, $3
+; NO-SIMD128-NEXT:    i32.xor $push25=, $pop24, $11
+; NO-SIMD128-NEXT:    i32.store16 4($0), $pop25
+; NO-SIMD128-NEXT:    i32.xor $push26=, $10, $18
+; NO-SIMD128-NEXT:    i32.and $push27=, $pop26, $2
+; NO-SIMD128-NEXT:    i32.xor $push28=, $pop27, $10
+; NO-SIMD128-NEXT:    i32.store16 2($0), $pop28
+; NO-SIMD128-NEXT:    i32.xor $push29=, $9, $17
+; NO-SIMD128-NEXT:    i32.and $push30=, $pop29, $1
+; NO-SIMD128-NEXT:    i32.xor $push31=, $pop30, $9
+; NO-SIMD128-NEXT:    i32.store16 0($0), $pop31
 ; NO-SIMD128-NEXT:    return
 ;
 ; NO-SIMD128-FAST-LABEL: bitselect_xor_reversed_v8i16:
 ; NO-SIMD128-FAST:         .functype bitselect_xor_reversed_v8i16 (i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32) -> ()
 ; NO-SIMD128-FAST-NEXT:  # %bb.0:
-; NO-SIMD128-FAST-NEXT:    i32.xor $push2=, $9, $17
-; NO-SIMD128-FAST-NEXT:    i32.const $push0=, -1
-; NO-SIMD128-FAST-NEXT:    i32.xor $push1=, $1, $pop0
-; NO-SIMD128-FAST-NEXT:    i32.and $push3=, $pop2, $pop1
-; NO-SIMD128-FAST-NEXT:    i32.xor $push4=, $pop3, $17
-; NO-SIMD128-FAST-NEXT:    i32.store16 0($0), $pop4
-; NO-SIMD128-FAST-NEXT:    i32.xor $push6=, $10, $18
-; NO-SIMD128-FAST-NEXT:    i32.const $push47=, -1
-; NO-SIMD128-FAST-NEXT:    i32.xor $push5=, $2, $pop47
-; NO-SIMD128-FAST-NEXT:    i32.and $push7=, $pop6, $pop5
-; NO-SIMD128-FAST-NEXT:    i32.xor $push8=, $pop7, $18
-; NO-SIMD128-FAST-NEXT:    i32.store16 2($0), $pop8
-; NO-SIMD128-FAST-NEXT:    i32.xor $push10=, $11, $19
-; NO-SIMD128-FAST-NEXT:    i32.const $push46=, -1
-; NO-SIMD128-FAST-NEXT:    i32.xor $push9=, $3, $pop46
-; NO-SIMD128-FAST-NEXT:    i32.and $push11=, $pop10, $pop9
-; NO-SIMD128-FAST-NEXT:    i32.xor $push12=, $pop11, $19
-; NO-SIMD128-FAST-NEXT:    i32.store16 4($0), $pop12
-; NO-SIMD128-FAST-NEXT:    i32.const $push17=, 6
+; NO-SIMD128-FAST-NEXT:    i32.xor $push0=, $9, $17
+; NO-SIMD128-FAST-NEXT:    i32.and $push1=, $pop0, $1
+; NO-SIMD128-FAST-NEXT:    i32.xor $push2=, $pop1, $9
+; NO-SIMD128-FAST-NEXT:    i32.store16 0($0), $pop2
+; NO-SIMD128-FAST-NEXT:    i32.xor $push3=, $10, $18
+; NO-SIMD128-FAST-NEXT:    i32.and $push4=, $pop3, $2
+; NO-SIMD128-FAST-NEXT:    i32.xor $push5=, $pop4, $10
+; NO-SIMD128-FAST-NEXT:    i32.store16 2($0), $pop5
+; NO-SIMD128-FAST-NEXT:    i32.xor $push6=, $11, $19
+; NO-SIMD128-FAST-NEXT:    i32.and $push7=, $pop6, $3
+; NO-SIMD128-FAST-NEXT:    i32.xor $push8=, $pop7, $11
+; NO-SIMD128-FAST-NEXT:    i32.store16 4($0), $pop8
+; NO-SIMD128-FAST-NEXT:    i32.const $push9=, 6
+; NO-SIMD128-FAST-NEXT:    i32.add $push10=, $0, $pop9
+; NO-SIMD128-FAST-NEXT:    i32.xor $push11=, $12, $20
+; NO-SIMD128-FAST-NEXT:    i32.and $push12=, $pop11, $4
+; NO-SIMD128-FAST-NEXT:    i32.xor $push13=, $pop12, $12
+; NO-SIMD128-FAST-NEXT:    i32.store16 0($pop10), $pop13
+; NO-SIMD128-FAST-NEXT:    i32.xor $push14=, $13, $21
+; NO-SIMD128-FAST-NEXT:    i32.and $push15=, $pop14, $5
+; NO-SIMD128-FAST-NEXT:    i32.xor $push16=, $pop15, $13
+; NO-SIMD128-FAST-NEXT:    i32.store16 8($0), $pop16
+; NO-SIMD128-FAST-NEXT:    i32.const $push17=, 10
 ; NO-SIMD128-FAST-NEXT:    i32.add $push18=, $0, $pop17
-; NO-SIMD128-FAST-NEXT:    i32.xor $push14=, $12, $20
-; NO-SIMD128-FAST-NEXT:    i32.const $push45=, -1
-; NO-SIMD128-FAST-NEXT:    i32.xor $push13=, $4, $pop45
-; NO-SIMD128-FAST-NEXT:    i32.and $push15=, $pop14, $pop13
-; NO-SIMD128-FAST-NEXT:    i32.xor $push16=, $pop15, $20
-; NO-SIMD128-FAST-NEXT:    i32.store16 0($pop18), $pop16
-; NO-SIMD128-FAST-NEXT:    i32.xor $push20=, $13, $21
-; NO-SIMD128-FAST-NEXT:    i32.const $push44=, -1
-; NO-SIMD128-FAST-NEXT:    i32.xor $push19=, $5, $pop44
-; NO-SIMD128-FAST-NEXT:    i32.and $push21=, $pop20, $pop19
-; NO-SIMD128-FAST-NEXT:    i32.xor $push22=, $pop21, $21
-; NO-SIMD128-FAST-NEXT:    i32.store16 8($0), $pop22
-; NO-SIMD128-FAST-NEXT:    i32.const $push27=, 10
+; NO-SIMD128-FAST-NEXT:    i32.xor $push19=, $14, $22
+; NO-SIMD128-FAST-NEXT:    i32.and $push20=, $pop19, $6
+; NO-SIMD128-FAST-NEXT:    i32.xor $push21=, $pop20, $14
+; NO-SIMD128-FAST-NEXT:    i32.store16 0($pop18), $pop21
+; NO-SIMD128-FAST-NEXT:    i32.const $push22=, 12
+; NO-SIMD128-FAST-NEXT:    i32.add $push23=, $0, $pop22
+; NO-SIMD128-FAST-NEXT:    i32.xor $push24=, $15, $23
+; NO-SIMD128-FAST-NEXT:    i32.and $push25=, $pop24, $7
+; NO-SIMD128-FAST-NEXT:    i32.xor $push26=, $pop25, $15
+; NO-SIMD128-FAST-NEXT:    i32.store16 0($pop23), $pop26
+; NO-SIMD128-FAST-NEXT:    i32.const $push27=, 14
 ; NO-SIMD128-FAST-NEXT:    i32.add $push28=, $0, $pop27
-; NO-SIMD128-FAST-NEXT:    i32.xor $push24=, $14, $22
-; NO-SIMD128-FAST-NEXT:    i32.const $push43=, -1
-; NO-SIMD128-FAST-NEXT:    i32.xor $push23=, $6, $pop43
-; NO-SIMD128-FAST-NEXT:    i32.and $push25=, $pop24, $pop23
-; NO-SIMD128-FAST-NEXT:    i32.xor $push26=, $pop25, $22
-; NO-SIMD128-FAST-NEXT:    i32.store16 0($pop28), $pop26
-; NO-SIMD128-FAST-NEXT:    i32.const $push33=, 12
-; NO-SIMD128-FAST-NEXT:    i32.add $push34=, $0, $pop33
-; NO-SIMD128-FAST-NEXT:    i32.xor $push30=, $15, $23
-; NO-SIMD128-FAST-NEXT:    i32.const $push42=, -1
-; NO-SIMD128-FAST-NEXT:    i32.xor $push29=, $7, $pop42
-; NO-SIMD128-FAST-NEXT:    i32.and $push31=, $pop30, $pop29
-; NO-SIMD128-FAST-NEXT:    i32.xor $push32=, $pop31, $23
-; NO-SIMD128-FAST-NEXT:    i32.store16 0($pop34), $pop32
-; NO-SIMD128-FAST-NEXT:    i32.const $push39=, 14
-; NO-SIMD128-FAST-NEXT:    i32.add $push40=, $0, $pop39
-; NO-SIMD128-FAST-NEXT:    i32.xor $push36=, $16, $24
-; NO-SIMD128-FAST-NEXT:    i32.const $push41=, -1
-; NO-SIMD128-FAST-NEXT:    i32.xor $push35=, $8, $pop41
-; NO-SIMD128-FAST-NEXT:    i32.and $push37=, $pop36, $pop35
-; NO-SIMD128-FAST-NEXT:    i32.xor $push38=, $pop37, $24
-; NO-SIMD128-FAST-NEXT:    i32.store16 0($pop40), $pop38
+; NO-SIMD128-FAST-NEXT:    i32.xor $push29=, $16, $24
+; NO-SIMD128-FAST-NEXT:    i32.and $push30=, $pop29, $8
+; NO-SIMD128-FAST-NEXT:    i32.xor $push31=, $pop30, $16
+; NO-SIMD128-FAST-NEXT:    i32.store16 0($pop28), $pop31
 ; NO-SIMD128-FAST-NEXT:    return
  %xor1 = xor <8 x i16> %v1, %v2
  %notc = xor <8 x i16> %c, <i16 -1, i16 -1, i16 -1, i16 -1,
                             i16 -1, i16 -1, i16 -1, i16 -1>
  %and = and <8 x i16> %xor1, %notc
  %a = xor <8 x i16> %and, %v2
  ret <8 x i16> %a
 }

 ; SIMD128-NEXT:  # %bb.0:
 ; SIMD128-NEXT:    v128.andnot $push0=, $0, $1
 ; SIMD128-NEXT:    return $pop0
 ;
 ; SIMD128-FAST-LABEL: andnot_v4i32:
 ; SIMD128-FAST:         .functype andnot_v4i32 (v128, v128) -> (v128)
 ; SIMD128-FAST-NEXT:  # %bb.0:
 ; SIMD128-FAST-NEXT:    v128.not $push0=, $1
-; SIMD128-FAST-NEXT:    v128.and $push1=, $0, $pop0
+; SIMD128-FAST-NEXT:    v128.and $push1=, $pop0, $0
 ; SIMD128-FAST-NEXT:    return $pop1
 ;
 ; NO-SIMD128-LABEL: andnot_v4i32:
 ; NO-SIMD128:         .functype andnot_v4i32 (i32, i32, i32, i32, i32, i32, i32, i32, i32) -> ()
 ; NO-SIMD128-NEXT:  # %bb.0:
 ; NO-SIMD128-NEXT:    i32.const $push0=, -1
 ; NO-SIMD128-NEXT:    i32.xor $push1=, $7, $pop0
-; NO-SIMD128-NEXT:    i32.and $push2=, $3, $pop1
+; NO-SIMD128-NEXT:    i32.and $push2=, $pop1, $3
 ; NO-SIMD128-NEXT:    i32.store 8($0), $pop2
 ; NO-SIMD128-NEXT:    i32.const $push13=, -1
 ; NO-SIMD128-NEXT:    i32.xor $push3=, $6, $pop13
-; NO-SIMD128-NEXT:    i32.and $push4=, $2, $pop3
+; NO-SIMD128-NEXT:    i32.and $push4=, $pop3, $2
 ; NO-SIMD128-NEXT:    i32.store 4($0), $pop4
 ; NO-SIMD128-NEXT:    i32.const $push12=, -1
 ; NO-SIMD128-NEXT:    i32.xor $push5=, $5, $pop12
-; NO-SIMD128-NEXT:    i32.and $push6=, $1, $pop5
+; NO-SIMD128-NEXT:    i32.and $push6=, $pop5, $1
 ; NO-SIMD128-NEXT:    i32.store 0($0), $pop6
 ; NO-SIMD128-NEXT:    i32.const $push9=, 12
 ; NO-SIMD128-NEXT:    i32.add $push10=, $0, $pop9
 ; NO-SIMD128-NEXT:    i32.const $push11=, -1
 ; NO-SIMD128-NEXT:    i32.xor $push7=, $8, $pop11
-; NO-SIMD128-NEXT:    i32.and $push8=, $4, $pop7
+; NO-SIMD128-NEXT:    i32.and $push8=, $pop7, $4
 ; NO-SIMD128-NEXT:    i32.store 0($pop10), $pop8
 ; NO-SIMD128-NEXT:    return
 ;
 ; NO-SIMD128-FAST-LABEL: andnot_v4i32:
 ; NO-SIMD128-FAST:         .functype andnot_v4i32 (i32, i32, i32, i32, i32, i32, i32, i32, i32) -> ()
 ; NO-SIMD128-FAST-NEXT:  # %bb.0:
 ; NO-SIMD128-FAST-NEXT:    i32.const $push0=, -1
 ; NO-SIMD128-FAST-NEXT:    i32.xor $push1=, $5, $pop0
-; NO-SIMD128-FAST-NEXT:    i32.and $push2=, $1, $pop1
+; NO-SIMD128-FAST-NEXT:    i32.and $push2=, $pop1, $1
 ; NO-SIMD128-FAST-NEXT:    i32.store 0($0), $pop2
 ; NO-SIMD128-FAST-NEXT:    i32.const $push13=, -1
 ; NO-SIMD128-FAST-NEXT:    i32.xor $push3=, $6, $pop13
-; NO-SIMD128-FAST-NEXT:    i32.and $push4=, $2, $pop3
+; NO-SIMD128-FAST-NEXT:    i32.and $push4=, $pop3, $2
 ; NO-SIMD128-FAST-NEXT:    i32.store 4($0), $pop4
 ; NO-SIMD128-FAST-NEXT:    i32.const $push12=, -1
 ; NO-SIMD128-FAST-NEXT:    i32.xor $push5=, $7, $pop12
-; NO-SIMD128-FAST-NEXT:    i32.and $push6=, $3, $pop5
+; NO-SIMD128-FAST-NEXT:    i32.and $push6=, $pop5, $3
 ; NO-SIMD128-FAST-NEXT:    i32.store 8($0), $pop6
 ; NO-SIMD128-FAST-NEXT:    i32.const $push7=, 12
 ; NO-SIMD128-FAST-NEXT:    i32.add $push8=, $0, $pop7
 ; NO-SIMD128-FAST-NEXT:    i32.const $push11=, -1
 ; NO-SIMD128-FAST-NEXT:    i32.xor $push9=, $8, $pop11
-; NO-SIMD128-FAST-NEXT:    i32.and $push10=, $4, $pop9
+; NO-SIMD128-FAST-NEXT:    i32.and $push10=, $pop9, $4
 ; NO-SIMD128-FAST-NEXT:    i32.store 0($pop8), $pop10
 ; NO-SIMD128-FAST-NEXT:    return
  %inv_y = xor <4 x i32> %y, <i32 -1, i32 -1, i32 -1, i32 -1>
  %a = and <4 x i32> %x, %inv_y
  ret <4 x i32> %a
 }
 
 define <4 x i32> @bitselect_v4i32(<4 x i32> %c, <4 x i32> %v1, <4 x i32> %v2) {

 ; SIMD128:         .functype bitselect_xor_reversed_v4i32 (v128, v128, v128) -> (v128)
 ; SIMD128-NEXT:  # %bb.0:
 ; SIMD128-NEXT:    v128.bitselect $push0=, $2, $1, $0
 ; SIMD128-NEXT:    return $pop0
 ;
 ; SIMD128-FAST-LABEL: bitselect_xor_reversed_v4i32:
 ; SIMD128-FAST:         .functype bitselect_xor_reversed_v4i32 (v128, v128, v128) -> (v128)
 ; SIMD128-FAST-NEXT:  # %bb.0:
-; SIMD128-FAST-NEXT:    v128.xor $push1=, $1, $2
-; SIMD128-FAST-NEXT:    v128.not $push2=, $0
-; SIMD128-FAST-NEXT:    v128.and $push3=, $pop1, $pop2
-; SIMD128-FAST-NEXT:    v128.xor $push0=, $pop3, $2
+; SIMD128-FAST-NEXT:    v128.xor $push2=, $1, $2
+; SIMD128-FAST-NEXT:    v128.and $push1=, $pop2, $0
+; SIMD128-FAST-NEXT:    v128.xor $push0=, $pop1, $1
 ; SIMD128-FAST-NEXT:    return $pop0
 ;
 ; NO-SIMD128-LABEL: bitselect_xor_reversed_v4i32:
 ; NO-SIMD128:         .functype bitselect_xor_reversed_v4i32 (i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32) -> ()
 ; NO-SIMD128-NEXT:  # %bb.0:
-; NO-SIMD128-NEXT:    i32.const $push5=, 12
-; NO-SIMD128-NEXT:    i32.add $push6=, $0, $pop5
-; NO-SIMD128-NEXT:    i32.xor $push2=, $8, $12
-; NO-SIMD128-NEXT:    i32.const $push0=, -1
-; NO-SIMD128-NEXT:    i32.xor $push1=, $4, $pop0
-; NO-SIMD128-NEXT:    i32.and $push3=, $pop2, $pop1
-; NO-SIMD128-NEXT:    i32.xor $push4=, $pop3, $12
-; NO-SIMD128-NEXT:    i32.store 0($pop6), $pop4
-; NO-SIMD128-NEXT:    i32.xor $push8=, $7, $11
-; NO-SIMD128-NEXT:    i32.const $push21=, -1
-; NO-SIMD128-NEXT:    i32.xor $push7=, $3, $pop21
-; NO-SIMD128-NEXT:    i32.and $push9=, $pop8, $pop7
-; NO-SIMD128-NEXT:    i32.xor $push10=, $pop9, $11
-; NO-SIMD128-NEXT:    i32.store 8($0), $pop10
-; NO-SIMD128-NEXT:    i32.xor $push12=, $6, $10
-; NO-SIMD128-NEXT:    i32.const $push20=, -1
-; NO-SIMD128-NEXT:    i32.xor $push11=, $2, $pop20
-; NO-SIMD128-NEXT:    i32.and $push13=, $pop12, $pop11
-; NO-SIMD128-NEXT:    i32.xor $push14=, $pop13, $10
-; NO-SIMD128-NEXT:    i32.store 4($0), $pop14
-; NO-SIMD128-NEXT:    i32.xor $push16=, $5, $9
-; NO-SIMD128-NEXT:    i32.const $push19=, -1
-; NO-SIMD128-NEXT:    i32.xor $push15=, $1, $pop19
-; NO-SIMD128-NEXT:    i32.and $push17=, $pop16, $pop15
-; NO-SIMD128-NEXT:    i32.xor $push18=, $pop17, $9
-; NO-SIMD128-NEXT:    i32.store 0($0), $pop18
+; NO-SIMD128-NEXT:    i32.const $push3=, 12
+; NO-SIMD128-NEXT:    i32.add $push4=, $0, $pop3
+; NO-SIMD128-NEXT:    i32.xor $push0=, $8, $12
+; NO-SIMD128-NEXT:    i32.and $push1=, $pop0, $4
+; NO-SIMD128-NEXT:    i32.xor $push2=, $pop1, $8
+; NO-SIMD128-NEXT:    i32.store 0($pop4), $pop2
+; NO-SIMD128-NEXT:    i32.xor $push5=, $7, $11
+; NO-SIMD128-NEXT:    i32.and $push6=, $pop5, $3
+; NO-SIMD128-NEXT:    i32.xor $push7=, $pop6, $7
+; NO-SIMD128-NEXT:    i32.store 8($0), $pop7
+; NO-SIMD128-NEXT:    i32.xor $push8=, $6, $10
+; NO-SIMD128-NEXT:    i32.and $push9=, $pop8, $2
+; NO-SIMD128-NEXT:    i32.xor $push10=, $pop9, $6
+; NO-SIMD128-NEXT:    i32.store 4($0), $pop10
+; NO-SIMD128-NEXT:    i32.xor $push11=, $5, $9
+; NO-SIMD128-NEXT:    i32.and $push12=, $pop11, $1
+; NO-SIMD128-NEXT:    i32.xor $push13=, $pop12, $5
+; NO-SIMD128-NEXT:    i32.store 0($0), $pop13
 ; NO-SIMD128-NEXT:    return
 ;
 ; NO-SIMD128-FAST-LABEL: bitselect_xor_reversed_v4i32:
 ; NO-SIMD128-FAST:         .functype bitselect_xor_reversed_v4i32 (i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32) -> ()
 ; NO-SIMD128-FAST-NEXT:  # %bb.0:
-; NO-SIMD128-FAST-NEXT:    i32.xor $push2=, $5, $9
-; NO-SIMD128-FAST-NEXT:    i32.const $push0=, -1
-; NO-SIMD128-FAST-NEXT:    i32.xor $push1=, $1, $pop0
-; NO-SIMD128-FAST-NEXT:    i32.and $push3=, $pop2, $pop1
-; NO-SIMD128-FAST-NEXT:    i32.xor $push4=, $pop3, $9
-; NO-SIMD128-FAST-NEXT:    i32.store 0($0), $pop4
-; NO-SIMD128-FAST-NEXT:    i32.xor $push6=, $6, $10
-; NO-SIMD128-FAST-NEXT:    i32.const $push21=, -1
-; NO-SIMD128-FAST-NEXT:    i32.xor $push5=, $2, $pop21
-; NO-SIMD128-FAST-NEXT:    i32.and $push7=, $pop6, $pop5
-; NO-SIMD128-FAST-NEXT:    i32.xor $push8=, $pop7, $10
-; NO-SIMD128-FAST-NEXT:    i32.store 4($0), $pop8
-; NO-SIMD128-FAST-NEXT:    i32.xor $push10=, $7, $11
-; NO-SIMD128-FAST-NEXT:    i32.const $push20=, -1
-; NO-SIMD128-FAST-NEXT:    i32.xor $push9=, $3, $pop20
-; NO-SIMD128-FAST-NEXT:    i32.and $push11=, $pop10, $pop9
-; NO-SIMD128-FAST-NEXT:    i32.xor $push12=, $pop11, $11
-; NO-SIMD128-FAST-NEXT:    i32.store 8($0), $pop12
-; NO-SIMD128-FAST-NEXT:    i32.const $push17=, 12
-; NO-SIMD128-FAST-NEXT:    i32.add $push18=, $0, $pop17
-; NO-SIMD128-FAST-NEXT:    i32.xor $push14=, $8, $12
-; NO-SIMD128-FAST-NEXT:    i32.const $push19=, -1
-; NO-SIMD128-FAST-NEXT:    i32.xor $push13=, $4, $pop19
-; NO-SIMD128-FAST-NEXT:    i32.and $push15=, $pop14, $pop13
-; NO-SIMD128-FAST-NEXT:    i32.xor $push16=, $pop15, $12
-; NO-SIMD128-FAST-NEXT:    i32.store 0($pop18), $pop16
+; NO-SIMD128-FAST-NEXT:    i32.xor $push0=, $5, $9
+; NO-SIMD128-FAST-NEXT:    i32.and $push1=, $pop0, $1
+; NO-SIMD128-FAST-NEXT:    i32.xor $push2=, $pop1, $5
+; NO-SIMD128-FAST-NEXT:    i32.store 0($0), $pop2
+; NO-SIMD128-FAST-NEXT:    i32.xor $push3=, $6, $10
+; NO-SIMD128-FAST-NEXT:    i32.and $push4=, $pop3, $2
+; NO-SIMD128-FAST-NEXT:    i32.xor $push5=, $pop4, $6
+; NO-SIMD128-FAST-NEXT:    i32.store 4($0), $pop5
+; NO-SIMD128-FAST-NEXT:    i32.xor $push6=, $7, $11
+; NO-SIMD128-FAST-NEXT:    i32.and $push7=, $pop6, $3
+; NO-SIMD128-FAST-NEXT:    i32.xor $push8=, $pop7, $7
+; NO-SIMD128-FAST-NEXT:    i32.store 8($0), $pop8
+; NO-SIMD128-FAST-NEXT:    i32.const $push9=, 12
+; NO-SIMD128-FAST-NEXT:    i32.add $push10=, $0, $pop9
+; NO-SIMD128-FAST-NEXT:    i32.xor $push11=, $8, $12
+; NO-SIMD128-FAST-NEXT:    i32.and $push12=, $pop11, $4
+; NO-SIMD128-FAST-NEXT:    i32.xor $push13=, $pop12, $8
+; NO-SIMD128-FAST-NEXT:    i32.store 0($pop10), $pop13
 ; NO-SIMD128-FAST-NEXT:    return
  %xor1 = xor <4 x i32> %v1, %v2
  %notc = xor <4 x i32> %c, <i32 -1, i32 -1, i32 -1, i32 -1>
  %and = and <4 x i32> %xor1, %notc
  %a = xor <4 x i32> %and, %v2
  ret <4 x i32> %a
 }
 

 ; SIMD128-NEXT:  # %bb.0:
 ; SIMD128-NEXT:    v128.andnot $push0=, $0, $1
 ; SIMD128-NEXT:    return $pop0
 ;
 ; SIMD128-FAST-LABEL: andnot_v2i64:
 ; SIMD128-FAST:         .functype andnot_v2i64 (v128, v128) -> (v128)
 ; SIMD128-FAST-NEXT:  # %bb.0:
 ; SIMD128-FAST-NEXT:    v128.not $push0=, $1
-; SIMD128-FAST-NEXT:    v128.and $push1=, $0, $pop0
+; SIMD128-FAST-NEXT:    v128.and $push1=, $pop0, $0
 ; SIMD128-FAST-NEXT:    return $pop1
 ;
 ; NO-SIMD128-LABEL: andnot_v2i64:
 ; NO-SIMD128:         .functype andnot_v2i64 (i32, i64, i64, i64, i64) -> ()
 ; NO-SIMD128-NEXT:  # %bb.0:
 ; NO-SIMD128-NEXT:    i64.const $push0=, -1
 ; NO-SIMD128-NEXT:    i64.xor $push1=, $4, $pop0
-; NO-SIMD128-NEXT:    i64.and $push2=, $2, $pop1
+; NO-SIMD128-NEXT:    i64.and $push2=, $pop1, $2
 ; NO-SIMD128-NEXT:    i64.store 8($0), $pop2
 ; NO-SIMD128-NEXT:    i64.const $push5=, -1
 ; NO-SIMD128-NEXT:    i64.xor $push3=, $3, $pop5
-; NO-SIMD128-NEXT:    i64.and $push4=, $1, $pop3
+; NO-SIMD128-NEXT:    i64.and $push4=, $pop3, $1
 ; NO-SIMD128-NEXT:    i64.store 0($0), $pop4
 ; NO-SIMD128-NEXT:    return
 ;
 ; NO-SIMD128-FAST-LABEL: andnot_v2i64:
 ; NO-SIMD128-FAST:         .functype andnot_v2i64 (i32, i64, i64, i64, i64) -> ()
 ; NO-SIMD128-FAST-NEXT:  # %bb.0:
 ; NO-SIMD128-FAST-NEXT:    i64.const $push0=, -1
 ; NO-SIMD128-FAST-NEXT:    i64.xor $push1=, $3, $pop0
-; NO-SIMD128-FAST-NEXT:    i64.and $push2=, $1, $pop1
+; NO-SIMD128-FAST-NEXT:    i64.and $push2=, $pop1, $1
 ; NO-SIMD128-FAST-NEXT:    i64.store 0($0), $pop2
 ; NO-SIMD128-FAST-NEXT:    i64.const $push5=, -1
 ; NO-SIMD128-FAST-NEXT:    i64.xor $push3=, $4, $pop5
-; NO-SIMD128-FAST-NEXT:    i64.and $push4=, $2, $pop3
+; NO-SIMD128-FAST-NEXT:    i64.and $push4=, $pop3, $2
 ; NO-SIMD128-FAST-NEXT:    i64.store 8($0), $pop4
 ; NO-SIMD128-FAST-NEXT:    return
  %inv_y = xor <2 x i64> %y, <i64 -1, i64 -1>
  %a = and <2 x i64> %x, %inv_y
  ret <2 x i64> %a
 }
 
 define <2 x i64> @bitselect_v2i64(<2 x i64> %c, <2 x i64> %v1, <2 x i64> %v2) {
 ; SIMD128-LABEL: bitselect_v2i64:
 ; SIMD128:         .functype bitselect_v2i64 (v128, v128, v128) -> (v128)
 ; SIMD128-NEXT:  # %bb.0:
 ; SIMD128-NEXT:    v128.bitselect $push0=, $1, $2, $0
 ; SIMD128-NEXT:    return $pop0
 ;
 ; SIMD128-FAST-LABEL: bitselect_v2i64:
 ; SIMD128-FAST:         .functype bitselect_v2i64 (v128, v128, v128) -> (v128)
 ; SIMD128-FAST-NEXT:  # %bb.0:
 ; SIMD128-FAST-NEXT:    v128.not $push2=, $0
-; SIMD128-FAST-NEXT:    v128.and $push3=, $2, $pop2
+; SIMD128-FAST-NEXT:    v128.and $push3=, $pop2, $2
 ; SIMD128-FAST-NEXT:    v128.and $push0=, $1, $0
 ; SIMD128-FAST-NEXT:    v128.or $push1=, $pop3, $pop0
 ; SIMD128-FAST-NEXT:    return $pop1
 ;
 ; NO-SIMD128-LABEL: bitselect_v2i64:
 ; NO-SIMD128:         .functype bitselect_v2i64 (i32, i64, i64, i64, i64, i64, i64) -> ()
 ; NO-SIMD128-NEXT:  # %bb.0:
 ; NO-SIMD128-NEXT:    i64.const $push1=, -1
 ; NO-SIMD128-NEXT:    i64.xor $push2=, $2, $pop1
-; NO-SIMD128-NEXT:    i64.and $push3=, $6, $pop2
+; NO-SIMD128-NEXT:    i64.and $push3=, $pop2, $6
 ; NO-SIMD128-NEXT:    i64.and $push0=, $4, $2
 ; NO-SIMD128-NEXT:    i64.or $push4=, $pop3, $pop0
 ; NO-SIMD128-NEXT:    i64.store 8($0), $pop4
 ; NO-SIMD128-NEXT:    i64.const $push9=, -1
 ; NO-SIMD128-NEXT:    i64.xor $push6=, $1, $pop9
-; NO-SIMD128-NEXT:    i64.and $push7=, $5, $pop6
+; NO-SIMD128-NEXT:    i64.and $push7=, $pop6, $5
 ; NO-SIMD128-NEXT:    i64.and $push5=, $3, $1
 ; NO-SIMD128-NEXT:    i64.or $push8=, $pop7, $pop5
 ; NO-SIMD128-NEXT:    i64.store 0($0), $pop8
 ; NO-SIMD128-NEXT:    return
 ;
 ; NO-SIMD128-FAST-LABEL: bitselect_v2i64:
 ; NO-SIMD128-FAST:         .functype bitselect_v2i64 (i32, i64, i64, i64, i64, i64, i64) -> ()
 ; NO-SIMD128-FAST-NEXT:  # %bb.0:
 ; NO-SIMD128-FAST-NEXT:    i64.const $push1=, -1
 ; NO-SIMD128-FAST-NEXT:    i64.xor $push2=, $1, $pop1
-; NO-SIMD128-FAST-NEXT:    i64.and $push3=, $5, $pop2
+; NO-SIMD128-FAST-NEXT:    i64.and $push3=, $pop2, $5
 ; NO-SIMD128-FAST-NEXT:    i64.and $push0=, $3, $1
 ; NO-SIMD128-FAST-NEXT:    i64.or $push4=, $pop3, $pop0
 ; NO-SIMD128-FAST-NEXT:    i64.store 0($0), $pop4
 ; NO-SIMD128-FAST-NEXT:    i64.const $push9=, -1
 ; NO-SIMD128-FAST-NEXT:    i64.xor $push6=, $2, $pop9
-; NO-SIMD128-FAST-NEXT:    i64.and $push7=, $6, $pop6
+; NO-SIMD128-FAST-NEXT:    i64.and $push7=, $pop6, $6
 ; NO-SIMD128-FAST-NEXT:    i64.and $push5=, $4, $2
 ; NO-SIMD128-FAST-NEXT:    i64.or $push8=, $pop7, $pop5
 ; NO-SIMD128-FAST-NEXT:    i64.store 8($0), $pop8
 ; NO-SIMD128-FAST-NEXT:    return
   %masked_v1 = and <2 x i64> %v1, %c
   %inv_mask = xor <2 x i64> <i64 -1, i64 -1>, %c
   %masked_v2 = and <2 x i64> %v2, %inv_mask
   %a = or <2 x i64> %masked_v2, %masked_v1

 ; SIMD128:         .functype bitselect_xor_reversed_v2i64 (v128, v128, v128) -> (v128)
 ; SIMD128-NEXT:  # %bb.0:
 ; SIMD128-NEXT:    v128.bitselect $push0=, $2, $1, $0
 ; SIMD128-NEXT:    return $pop0
 ;
 ; SIMD128-FAST-LABEL: bitselect_xor_reversed_v2i64:
 ; SIMD128-FAST:         .functype bitselect_xor_reversed_v2i64 (v128, v128, v128) -> (v128)
 ; SIMD128-FAST-NEXT:  # %bb.0:
-; SIMD128-FAST-NEXT:    v128.xor $push1=, $1, $2
-; SIMD128-FAST-NEXT:    v128.not $push2=, $0
-; SIMD128-FAST-NEXT:    v128.and $push3=, $pop1, $pop2
-; SIMD128-FAST-NEXT:    v128.xor $push0=, $pop3, $2
+; SIMD128-FAST-NEXT:    v128.xor $push2=, $1, $2
+; SIMD128-FAST-NEXT:    v128.and $push1=, $pop2, $0
+; SIMD128-FAST-NEXT:    v128.xor $push0=, $pop1, $1
 ; SIMD128-FAST-NEXT:    return $pop0
 ;
 ; NO-SIMD128-LABEL: bitselect_xor_reversed_v2i64:
 ; NO-SIMD128:         .functype bitselect_xor_reversed_v2i64 (i32, i64, i64, i64, i64, i64, i64) -> ()
 ; NO-SIMD128-NEXT:  # %bb.0:
-; NO-SIMD128-NEXT:    i64.xor $push2=, $4, $6
-; NO-SIMD128-NEXT:    i64.const $push0=, -1
-; NO-SIMD128-NEXT:    i64.xor $push1=, $2, $pop0
-; NO-SIMD128-NEXT:    i64.and $push3=, $pop2, $pop1
-; NO-SIMD128-NEXT:    i64.xor $push4=, $pop3, $6
-; NO-SIMD128-NEXT:    i64.store 8($0), $pop4
-; NO-SIMD128-NEXT:    i64.xor $push6=, $3, $5
-; NO-SIMD128-NEXT:    i64.const $push9=, -1
-; NO-SIMD128-NEXT:    i64.xor $push5=, $1, $pop9
-; NO-SIMD128-NEXT:    i64.and $push7=, $pop6, $pop5
-; NO-SIMD128-NEXT:    i64.xor $push8=, $pop7, $5
-; NO-SIMD128-NEXT:    i64.store 0($0), $pop8
+; NO-SIMD128-NEXT:    i64.xor $push0=, $4, $6
+; NO-SIMD128-NEXT:    i64.and $push1=, $pop0, $2
+; NO-SIMD128-NEXT:    i64.xor $push2=, $pop1, $4
+; NO-SIMD128-NEXT:    i64.store 8($0), $pop2
+; NO-SIMD128-NEXT:    i64.xor $push3=, $3, $5
+; NO-SIMD128-NEXT:    i64.and $push4=, $pop3, $1
+; NO-SIMD128-NEXT:    i64.xor $push5=, $pop4, $3
+; NO-SIMD128-NEXT:    i64.store 0($0), $pop5
 ; NO-SIMD128-NEXT:    return
 ;
 ; NO-SIMD128-FAST-LABEL: bitselect_xor_reversed_v2i64:
 ; NO-SIMD128-FAST:         .functype bitselect_xor_reversed_v2i64 (i32, i64, i64, i64, i64, i64, i64) -> ()
 ; NO-SIMD128-FAST-NEXT:  # %bb.0:
-; NO-SIMD128-FAST-NEXT:    i64.xor $push2=, $3, $5
-; NO-SIMD128-FAST-NEXT:    i64.const $push0=, -1
-; NO-SIMD128-FAST-NEXT:    i64.xor $push1=, $1, $pop0
-; NO-SIMD128-FAST-NEXT:    i64.and $push3=, $pop2, $pop1
-; NO-SIMD128-FAST-NEXT:    i64.xor $push4=, $pop3, $5
-; NO-SIMD128-FAST-NEXT:    i64.store 0($0), $pop4
-; NO-SIMD128-FAST-NEXT:    i64.xor $push6=, $4, $6
-; NO-SIMD128-FAST-NEXT:    i64.const $push9=, -1
-; NO-SIMD128-FAST-NEXT:    i64.xor $push5=, $2, $pop9
-; NO-SIMD128-FAST-NEXT:    i64.and $push7=, $pop6, $pop5
-; NO-SIMD128-FAST-NEXT:    i64.xor $push8=, $pop7, $6
-; NO-SIMD128-FAST-NEXT:    i64.store 8($0), $pop8
+; NO-SIMD128-FAST-NEXT:    i64.xor $push0=, $3, $5
+; NO-SIMD128-FAST-NEXT:    i64.and $push1=, $pop0, $1
+; NO-SIMD128-FAST-NEXT:    i64.xor $push2=, $pop1, $3
+; NO-SIMD128-FAST-NEXT:    i64.store 0($0), $pop2
+; NO-SIMD128-FAST-NEXT:    i64.xor $push3=, $4, $6
+; NO-SIMD128-FAST-NEXT:    i64.and $push4=, $pop3, $2
+; NO-SIMD128-FAST-NEXT:    i64.xor $push5=, $pop4, $4
+; NO-SIMD128-FAST-NEXT:    i64.store 8($0), $pop5
 ; NO-SIMD128-FAST-NEXT:    return
  %xor1 = xor <2 x i64> %v1, %v2
  %notc = xor <2 x i64> %c, <i64 -1, i64 -1>
  %and = and <2 x i64> %xor1, %notc
  %a = xor <2 x i64> %and, %v2
  ret <2 x i64> %a
 }
 

llvm/test/CodeGen/WebAssembly/simd-build-pair.ll

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc < %s -verify-machineinstrs -disable-wasm-fallthrough-return-opt -wasm-disable-explicit-locals -wasm-keep-registers -mattr=simd128 | FileCheck %s
 
 target triple = "wasm32-unknown-unknown"
 
 ; Test that BUILD_PAIR dag nodes are correctly lowered.
 ; This code produces a selection DAG like the following:
 
 ;    t0: ch = EntryToken
 ;  t3: v4i32,ch = load<(load 16 from `ptr undef`)> t0, undef:i32, undef:i32
 ;        t30: i32 = extract_vector_elt t3, Constant:i32<2>
 ;        t28: i32 = extract_vector_elt t3, Constant:i32<3>
 ;      t24: i64 = build_pair t30, t28
 ;    t8: ch = store<(store 8 into `ptr undef`, align 1)> t3:1, t24, undef:i32, undef:i32
 ;  t9: ch = WebAssemblyISD::RETURN t8
 
-define void @build_pair_i32s() {
+define void @build_pair_i32s(ptr %in, ptr %out) {
 ; CHECK-LABEL: build_pair_i32s:
-; CHECK:         .functype build_pair_i32s () -> ()
+; CHECK:         .functype build_pair_i32s (i32, i32) -> ()
 ; CHECK-NEXT:  # %bb.0: # %entry
-; CHECK-NEXT:    v128.load $push0=
-; CHECK-NEXT:    i8x16.shuffle $push1=
-; CHECK-NEXT:    v128.store64_lane
+; CHECK-NEXT:    v128.load $push0=, 0($0)
+; CHECK-NEXT:    i8x16.shuffle $push1=, $pop0, $2, 8, 9, 10, 11, 12, 13, 14, 15, 0, 1, 2, 3, 0, 1, 2, 3
+; CHECK-NEXT:    v128.store64_lane 0($1):p2align=0, $pop1, 0
 ; CHECK-NEXT:    return
 entry:
-  %0 = load <4 x i32>, ptr undef, align 16
-  %shuffle.i184 = shufflevector <4 x i32> %0, <4 x i32> undef, <4 x i32> <i32 2, i32 3, i32 undef, i32 undef>
+  %load = load <4 x i32>, ptr %in, align 16
+  %shuffle.i184 = shufflevector <4 x i32> %load, <4 x i32> undef, <4 x i32> <i32 2, i32 3, i32 undef, i32 undef>
   %bc357 = bitcast <4 x i32> %shuffle.i184 to <2 x i64>
-  %1 = extractelement <2 x i64> %bc357, i32 0
-  store i64 %1, ptr undef, align 1
+  %extract = extractelement <2 x i64> %bc357, i32 0
+  store i64 %extract, ptr %out, align 1
   ret void
 }

llvm/test/CodeGen/WebAssembly/simd-build-vector.ll

; CHECK-NEXT:    return $pop6
   %v0 = insertelement <8 x i16> undef, i16 %s0, i32 0
   ret <8 x i16> %v0
 }
 
 define <4 x i32> @half_shuffle_i32x4(<4 x i32> %src) {
 ; CHECK-LABEL: half_shuffle_i32x4:
 ; CHECK:         .functype half_shuffle_i32x4 (v128) -> (v128)
 ; CHECK-NEXT:  # %bb.0:
-; CHECK-NEXT:    i8x16.shuffle $push0=, $0, $0, 0, 1, 2, 3, 8, 9, 10, 11, 0, 1, 2, 3, 0, 1, 2, 3
-; CHECK-NEXT:    i32.const $push1=, 0
-; CHECK-NEXT:    i32x4.replace_lane $push2=, $pop0, 0, $pop1
-; CHECK-NEXT:    i32.const $push3=, 3
-; CHECK-NEXT:    i32x4.replace_lane $push4=, $pop2, 3, $pop3
-; CHECK-NEXT:    return $pop4
+; CHECK-NEXT:    v128.const $push0=, 0, 0, 0, 3
+; CHECK-NEXT:    i8x16.shuffle $push1=, $pop0, $0, 0, 1, 2, 3, 24, 25, 26, 27, 16, 17, 18, 19, 12, 13, 14, 15
+; CHECK-NEXT:    return $pop1
   %s0 = extractelement <4 x i32> %src, i32 0
   %s2 = extractelement <4 x i32> %src, i32 2
   %v0 = insertelement <4 x i32> undef, i32 0, i32 0
   %v1 = insertelement <4 x i32> %v0, i32 %s2, i32 1
   %v2 = insertelement <4 x i32> %v1, i32 %s0, i32 2
   %v3 = insertelement <4 x i32> %v2, i32 3, i32 3
   ret <4 x i32> %v3
 }

llvm/test/CodeGen/WebAssembly/simd-comparisons.ll

 define <4 x i1> @compare_ord_v4f32 (<4 x float> %x, <4 x float> %y) {
   %res = fcmp ord <4 x float> %x, %y
   ret <4 x i1> %res
 }
 
 ; CHECK-LABEL: compare_ord_nnan_v4f32:
 ; NO-SIMD128-NOT: f32x4
 ; SIMD128-NEXT: .functype compare_ord_nnan_v4f32 (v128, v128) -> (v128){{$}}
-; SIMD128-NEXT: f32x4.eq $push[[T0:[0-9]+]]=, $0, $0{{$}}
-; SIMD128-NEXT: f32x4.eq $push[[T1:[0-9]+]]=, $1, $1{{$}}
-; SIMD128-NEXT: v128.and $push[[R:[0-9]+]]=, $pop[[T0]], $pop[[T1]]{{$}}
-; SIMD128-NEXT: return   $pop[[R]]{{$}}
+; SIMD128-NEXT: v128.const $push0=, 1, 1, 1, 1
+; SIMD128-NEXT: return   $pop0
 define <4 x i1> @compare_ord_nnan_v4f32 (<4 x float> %x, <4 x float> %y) {
   %res = fcmp nnan ord <4 x float> %x, %y
   ret <4 x i1> %res
 }
 
 ; CHECK-LABEL: compare_sext_ord_v4f32:
 ; NO-SIMD128-NOT: f32x4
 ; SIMD128-NEXT: .functype compare_sext_ord_v4f32 (v128, v128) -> (v128){{$}}
 ; SIMD128-NEXT: f32x4.eq $push[[T0:[0-9]+]]=, $0, $0{{$}}
 ; SIMD128-NEXT: f32x4.eq $push[[T1:[0-9]+]]=, $1, $1{{$}}
 ; SIMD128-NEXT: v128.and $push[[R:[0-9]+]]=, $pop[[T0]], $pop[[T1]]{{$}}
 ; SIMD128-NEXT: return   $pop[[R]]{{$}}
 define <4 x i32> @compare_sext_ord_v4f32 (<4 x float> %x, <4 x float> %y) {
   %cmp = fcmp ord <4 x float> %x, %y
   %res = sext <4 x i1> %cmp to <4 x i32>
   ret <4 x i32> %res
 }
 
 ; CHECK-LABEL: compare_sext_ord_nnan_v4f32:
 ; NO-SIMD128-NOT: f32x4
 ; SIMD128-NEXT: .functype compare_sext_ord_nnan_v4f32 (v128, v128) -> (v128){{$}}
-; SIMD128-NEXT: f32x4.eq $push[[T0:[0-9]+]]=, $0, $0{{$}}
-; SIMD128-NEXT: f32x4.eq $push[[T1:[0-9]+]]=, $1, $1{{$}}
-; SIMD128-NEXT: v128.and $push[[R:[0-9]+]]=, $pop[[T0]], $pop[[T1]]{{$}}
-; SIMD128-NEXT: return   $pop[[R]]{{$}}
+; SIMD128-NEXT: v128.const      $push0=, -1, -1, -1, -1
+; SIMD128-NEXT: return   $pop0
 define <4 x i32> @compare_sext_ord_nnan_v4f32 (<4 x float> %x, <4 x float> %y) {
   %cmp = fcmp nnan ord <4 x float> %x, %y
   %res = sext <4 x i1> %cmp to <4 x i32>
   ret <4 x i32> %res
 }
 
 ; CHECK-LABEL: compare_ueq_v4f32:
 ; NO-SIMD128-NOT: f32x4

 define <4 x i1> @compare_uno_v4f32 (<4 x float> %x, <4 x float> %y) {
   %res = fcmp uno <4 x float> %x, %y
   ret <4 x i1> %res
 }
 
 ; CHECK-LABEL: compare_uno_nnan_v4f32:
 ; NO-SIMD128-NOT: f32x4
 ; SIMD128-NEXT: .functype compare_uno_nnan_v4f32 (v128, v128) -> (v128){{$}}
-; SIMD128-NEXT: f32x4.ne $push[[T0:[0-9]+]]=, $0, $0{{$}}
-; SIMD128-NEXT: f32x4.ne $push[[T1:[0-9]+]]=, $1, $1{{$}}
-; SIMD128-NEXT: v128.or $push[[R:[0-9]+]]=, $pop[[T0]], $pop[[T1]]{{$}}
-; SIMD128-NEXT: return $pop[[R]]{{$}}
+; SIMD128-NEXT: v128.const $push0=, 0, 0, 0, 0
+; SIMD128-NEXT: return $pop0
 define <4 x i1> @compare_uno_nnan_v4f32 (<4 x float> %x, <4 x float> %y) {
   %res = fcmp nnan uno <4 x float> %x, %y
   ret <4 x i1> %res
 }
 
 ; CHECK-LABEL: compare_sext_uno_v4f32:
 ; NO-SIMD128-NOT: f32x4
 ; SIMD128-NEXT: .functype compare_sext_uno_v4f32 (v128, v128) -> (v128){{$}}
 ; SIMD128-NEXT: f32x4.ne $push[[T0:[0-9]+]]=, $0, $0{{$}}
 ; SIMD128-NEXT: f32x4.ne $push[[T1:[0-9]+]]=, $1, $1{{$}}
 ; SIMD128-NEXT: v128.or $push[[R:[0-9]+]]=, $pop[[T0]], $pop[[T1]]{{$}}
 ; SIMD128-NEXT: return $pop[[R]]{{$}}
 define <4 x i32> @compare_sext_uno_v4f32 (<4 x float> %x, <4 x float> %y) {
   %cmp = fcmp uno <4 x float> %x, %y
   %res = sext <4 x i1> %cmp to <4 x i32>
   ret <4 x i32> %res
 }
 
 ; CHECK-LABEL: compare_sext_uno_nnan_v4f32:
 ; NO-SIMD128-NOT: f32x4
 ; SIMD128-NEXT: .functype compare_sext_uno_nnan_v4f32 (v128, v128) -> (v128){{$}}
-; SIMD128-NEXT: f32x4.ne $push[[T0:[0-9]+]]=, $0, $0{{$}}
-; SIMD128-NEXT: f32x4.ne $push[[T1:[0-9]+]]=, $1, $1{{$}}
-; SIMD128-NEXT: v128.or $push[[R:[0-9]+]]=, $pop[[T0]], $pop[[T1]]{{$}}
-; SIMD128-NEXT: return $pop[[R]]{{$}}
+; SIMD128-NEXT: v128.const $push0=, 0, 0, 0, 0
+; SIMD128-NEXT: return $pop0
 define <4 x i32> @compare_sext_uno_nnan_v4f32 (<4 x float> %x, <4 x float> %y) {
   %cmp = fcmp nnan uno <4 x float> %x, %y
   %res = sext <4 x i1> %cmp to <4 x i32>
   ret <4 x i32> %res
 }
 
 ; CHECK-LABEL: compare_oeq_v2f64:
 ; NO-SIMD128-NOT: f64x2

 define <2 x i1> @compare_ord_v2f64 (<2 x double> %x, <2 x double> %y) {
   %res = fcmp ord <2 x double> %x, %y
   ret <2 x i1> %res
 }
 
 ; CHECK-LABEL: compare_ord_nnan_v2f64:
 ; NO-SIMD128-NOT: f64x2
 ; SIMD128-NEXT: .functype compare_ord_nnan_v2f64 (v128, v128) -> (v128){{$}}
-; SIMD128-NEXT: f64x2.eq $push[[T0:[0-9]+]]=, $0, $0{{$}}
-; SIMD128-NEXT: f64x2.eq $push[[T1:[0-9]+]]=, $1, $1{{$}}
-; SIMD128-NEXT: v128.and $push[[R:[0-9]+]]=, $pop[[T0]], $pop[[T1]]{{$}}
-; SIMD128-NEXT: return   $pop[[R]]{{$}}
+; SIMD128-NEXT: v128.const $push0=, 1, 1
+; SIMD128-NEXT: return $pop0
 define <2 x i1> @compare_ord_nnan_v2f64 (<2 x double> %x, <2 x double> %y) {
   %res = fcmp nnan ord <2 x double> %x, %y
   ret <2 x i1> %res
 }
 
 ; CHECK-LABEL: compare_sext_ord_v2f64:
 ; NO-SIMD128-NOT: f64x2
 ; SIMD128-NEXT: .functype compare_sext_ord_v2f64 (v128, v128) -> (v128){{$}}
 ; SIMD128-NEXT: f64x2.eq $push[[T0:[0-9]+]]=, $0, $0{{$}}
 ; SIMD128-NEXT: f64x2.eq $push[[T1:[0-9]+]]=, $1, $1{{$}}
 ; SIMD128-NEXT: v128.and $push[[R:[0-9]+]]=, $pop[[T0]], $pop[[T1]]{{$}}
 ; SIMD128-NEXT: return   $pop[[R]]{{$}}
 define <2 x i64> @compare_sext_ord_v2f64 (<2 x double> %x, <2 x double> %y) {
   %cmp = fcmp ord <2 x double> %x, %y
   %res = sext <2 x i1> %cmp to <2 x i64>
   ret <2 x i64> %res
 }
 
 ; CHECK-LABEL: compare_sext_ord_nnan_v2f64:
 ; NO-SIMD128-NOT: f64x2
 ; SIMD128-NEXT: .functype compare_sext_ord_nnan_v2f64 (v128, v128) -> (v128){{$}}
-; SIMD128-NEXT: f64x2.eq $push[[T0:[0-9]+]]=, $0, $0{{$}}
-; SIMD128-NEXT: f64x2.eq $push[[T1:[0-9]+]]=, $1, $1{{$}}
-; SIMD128-NEXT: v128.and $push[[R:[0-9]+]]=, $pop[[T0]], $pop[[T1]]{{$}}
-; SIMD128-NEXT: return   $pop[[R]]{{$}}
+; SIMD128-NEXT: v128.const $push0=, -1, -1
+; SIMD128-NEXT: return $pop0
 define <2 x i64> @compare_sext_ord_nnan_v2f64 (<2 x double> %x, <2 x double> %y) {
   %cmp = fcmp nnan ord <2 x double> %x, %y
   %res = sext <2 x i1> %cmp to <2 x i64>
   ret <2 x i64> %res
 }
 
 ; CHECK-LABEL: compare_ueq_v2f64:
 ; NO-SIMD128-NOT: f64x2

 define <2 x i1> @compare_uno_v2f64 (<2 x double> %x, <2 x double> %y) {
   %res = fcmp uno <2 x double> %x, %y
   ret <2 x i1> %res
 }
 
 ; CHECK-LABEL: compare_uno_nnan_v2f64:
 ; NO-SIMD128-NOT: f64x2
 ; SIMD128-NEXT: .functype compare_uno_nnan_v2f64 (v128, v128) -> (v128){{$}}
-; SIMD128-NEXT: f64x2.ne $push[[T0:[0-9]+]]=, $0, $0{{$}}
-; SIMD128-NEXT: f64x2.ne $push[[T1:[0-9]+]]=, $1, $1{{$}}
-; SIMD128-NEXT: v128.or $push[[R:[0-9]+]]=, $pop[[T0]], $pop[[T1]]{{$}}
-; SIMD128-NEXT: return $pop[[R]]{{$}}
+; SIMD128-NEXT: v128.const $push0=, 0, 0
+; SIMD128-NEXT: return $pop0
 define <2 x i1> @compare_uno_nnan_v2f64 (<2 x double> %x, <2 x double> %y) {
   %res = fcmp nnan uno <2 x double> %x, %y
   ret <2 x i1> %res
 }
 
 ; CHECK-LABEL: compare_sext_uno_v2f64:
 ; NO-SIMD128-NOT: f64x2
 ; SIMD128-NEXT: .functype compare_sext_uno_v2f64 (v128, v128) -> (v128){{$}}
 ; SIMD128-NEXT: f64x2.ne $push[[T0:[0-9]+]]=, $0, $0{{$}}
 ; SIMD128-NEXT: f64x2.ne $push[[T1:[0-9]+]]=, $1, $1{{$}}
 ; SIMD128-NEXT: v128.or $push[[R:[0-9]+]]=, $pop[[T0]], $pop[[T1]]{{$}}
 ; SIMD128-NEXT: return $pop[[R]]{{$}}
 define <2 x i64> @compare_sext_uno_v2f64 (<2 x double> %x, <2 x double> %y) {
   %cmp = fcmp uno <2 x double> %x, %y
   %res = sext <2 x i1> %cmp to <2 x i64>
   ret <2 x i64> %res
 }
 
 ; CHECK-LABEL: compare_sext_uno_nnan_v2f64:
 ; NO-SIMD128-NOT: f64x2
 ; SIMD128-NEXT: .functype compare_sext_uno_nnan_v2f64 (v128, v128) -> (v128){{$}}
-; SIMD128-NEXT: f64x2.ne $push[[T0:[0-9]+]]=, $0, $0{{$}}
-; SIMD128-NEXT: f64x2.ne $push[[T1:[0-9]+]]=, $1, $1{{$}}
-; SIMD128-NEXT: v128.or $push[[R:[0-9]+]]=, $pop[[T0]], $pop[[T1]]{{$}}
-; SIMD128-NEXT: return $pop[[R]]{{$}}
+; SIMD128-NEXT: v128.const $push0=, 0, 0
+; SIMD128-NEXT: return $pop0
 define <2 x i64> @compare_sext_uno_nnan_v2f64 (<2 x double> %x, <2 x double> %y) {
   %cmp = fcmp nnan uno <2 x double> %x, %y
   %res = sext <2 x i1> %cmp to <2 x i64>
   ret <2 x i64> %res
 }

llvm/test/CodeGen/WebAssembly/simd-illegal-signext.ll

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc < %s -mattr=+simd128 | FileCheck %s
 
 ; Regression test for a crash caused by
 ; WebAssemblyTargetLowering::ReplaceNodeResults not being
 ; implemented. Since SIMD is enabled, sign_ext_inreg is custom lowered
 ; but the result is i16, an illegal value. This requires
 ; ReplaceNodeResults to resolve, but the default implementation is to
 ; abort.
 
 target triple = "wasm32-unknown-emscripten"
 
-define void @foo() {
+define void @foo(ptr %arg) {
 ; CHECK-LABEL: foo:
-; CHECK:         .functype foo () -> ()
-; CHECK-NEXT:    .local i32
+; CHECK:         .functype foo (i32) -> ()
 ; CHECK-NEXT:  # %bb.0: # %entry
-; CHECK-NEXT:    i32.const 0
+; CHECK-NEXT:    local.get 0
+; CHECK-NEXT:    i32.load 0
+; CHECK-NEXT:    local.tee 0
 ; CHECK-NEXT:    local.get 0
 ; CHECK-NEXT:    i32.load8_s 0
 ; CHECK-NEXT:    i32.store16 0
 ; CHECK-NEXT:    # fallthrough-return
 entry:
-  %0 = load ptr, ptr undef, align 4
-  %1 = load i32, ptr %0, align 4
-  %2 = load i32, ptr undef, align 4
-  %conv67 = trunc i32 %2 to i8
+  %ptr.1 = load ptr, ptr %arg, align 4
+  %data = load i32, ptr %ptr.1, align 4
+  %conv67 = trunc i32 %data to i8
   %conv68 = sext i8 %conv67 to i16
-  store i16 %conv68, ptr null, align 2
+  store i16 %conv68, ptr %ptr.1, align 2
   ret void
 }

llvm/test/CodeGen/WebAssembly/simd-offset.ll

; CHECK-NEXT:    # fallthrough-return
   ret <8 x i16> %v2
 }
 
 define <8 x i8> @load_ext_v8i16_from_numeric_address() {
 ; CHECK-LABEL: load_ext_v8i16_from_numeric_address:
 ; CHECK:         .functype load_ext_v8i16_from_numeric_address () -> (v128)
 ; CHECK-NEXT:  # %bb.0:
 ; CHECK-NEXT:    i32.const 0
-; CHECK-NEXT:    v128.load64_zero 32
+; CHECK-NEXT:    v128.load 32
 ; CHECK-NEXT:    # fallthrough-return
   %s = inttoptr i32 32 to ptr
   %v = load <8 x i8>, ptr %s
   ret <8 x i8> %v
 }
 
 @gv_v8i16 = global <8 x i16> <i16 42, i16 42, i16 42, i16 42, i16 42, i16 42, i16 42, i16 42>
 define <8 x i16> @load_v8i16_from_global_address() {

 }
 
 define <4 x i32> @load_sext_v4i8_to_v4i32_from_numeric_address() {
 ; CHECK-LABEL: load_sext_v4i8_to_v4i32_from_numeric_address:
 ; CHECK:         .functype load_sext_v4i8_to_v4i32_from_numeric_address () -> (v128)
 ; CHECK-NEXT:    .local v128
 ; CHECK-NEXT:  # %bb.0:
 ; CHECK-NEXT:    i32.const 0
-; CHECK-NEXT:    v128.load32_zero 32
+; CHECK-NEXT:    v128.load 32
 ; CHECK-NEXT:    local.get 0
 ; CHECK-NEXT:    i8x16.shuffle 0, 0, 0, 0, 1, 0, 0, 0, 2, 0, 0, 0, 3, 0, 0, 0
 ; CHECK-NEXT:    i32.const 24
 ; CHECK-NEXT:    i32x4.shl
 ; CHECK-NEXT:    i32.const 24
 ; CHECK-NEXT:    i32x4.shr_s
 ; CHECK-NEXT:    # fallthrough-return
   %s = inttoptr i32 32 to ptr
   %v = load <4 x i8>, ptr %s
   %v2 = sext <4 x i8> %v to <4 x i32>
   ret <4 x i32> %v2
 }
 
 define <4 x i32> @load_zext_v4i8_to_v4i32_from_numeric_address() {
 ; CHECK-LABEL: load_zext_v4i8_to_v4i32_from_numeric_address:
 ; CHECK:         .functype load_zext_v4i8_to_v4i32_from_numeric_address () -> (v128)
 ; CHECK-NEXT:  # %bb.0:
 ; CHECK-NEXT:    v128.const 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0
 ; CHECK-NEXT:    i32.const 0
-; CHECK-NEXT:    v128.load32_zero 32
+; CHECK-NEXT:    v128.load 32
 ; CHECK-NEXT:    i8x16.shuffle 16, 1, 2, 3, 17, 5, 6, 7, 18, 9, 10, 11, 19, 13, 14, 15
 ; CHECK-NEXT:    # fallthrough-return
   %s = inttoptr i32 32 to ptr
   %v = load <4 x i8>, ptr %s
   %v2 = zext <4 x i8> %v to <4 x i32>
   ret <4 x i32> %v2
 }
 
 define <4 x i16> @load_ext_v4i32_from_numeric_address() {
 ; CHECK-LABEL: load_ext_v4i32_from_numeric_address:
 ; CHECK:         .functype load_ext_v4i32_from_numeric_address () -> (v128)
 ; CHECK-NEXT:  # %bb.0:
 ; CHECK-NEXT:    i32.const 0
-; CHECK-NEXT:    v128.load64_zero 32
+; CHECK-NEXT:    v128.load 32
 ; CHECK-NEXT:    # fallthrough-return
   %s = inttoptr i32 32 to ptr
   %v = load <4 x i16>, ptr %s
   ret <4 x i16> %v
 }
 
 @gv_v4i32 = global <4 x i32> <i32 42, i32 42, i32 42, i32 42>
 define <4 x i32> @load_v4i32_from_global_address() {
; CHECK-NEXT:    # fallthrough-return
   ret <2 x i64> %v2
 }
 
 define <2 x i32> @load_ext_v2i64_from_numeric_address() {
 ; CHECK-LABEL: load_ext_v2i64_from_numeric_address:
 ; CHECK:         .functype load_ext_v2i64_from_numeric_address () -> (v128)
 ; CHECK-NEXT:  # %bb.0:
 ; CHECK-NEXT:    i32.const 0
-; CHECK-NEXT:    v128.load64_zero 32
+; CHECK-NEXT:    v128.load 32
 ; CHECK-NEXT:    # fallthrough-return
   %s = inttoptr i32 32 to ptr
   %v = load <2 x i32>, ptr %s
   ret <2 x i32> %v
 }
 
 @gv_v2i64 = global <2 x i64> <i64 42, i64 42>
 define <2 x i64> @load_v2i64_from_global_address() {

llvm/test/CodeGen/WebAssembly/simd.ll

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
-; RUN: llc < %s -verify-machineinstrs -disable-wasm-fallthrough-return-opt -wasm-disable-explicit-locals -wasm-keep-registers -mattr=+simd128,+sign-ext | FileCheck %s --check-prefix=SIMD128
-; RUN: llc < %s -verify-machineinstrs -disable-wasm-fallthrough-return-opt -wasm-disable-explicit-locals -wasm-keep-registers | FileCheck %s --check-prefix=NO-SIMD128
+; RUN: llc < %s -O1 -verify-machineinstrs -disable-wasm-fallthrough-return-opt -wasm-disable-explicit-locals -wasm-keep-registers -mattr=+simd128,+sign-ext | FileCheck %s --check-prefix=SIMD128
+; RUN: llc < %s -O1 -verify-machineinstrs -disable-wasm-fallthrough-return-opt -wasm-disable-explicit-locals -wasm-keep-registers | FileCheck %s --check-prefix=NO-SIMD128
 
 ; Test that basic SIMD128 vector manipulation operations assemble as expected.
 
 target triple = "wasm32-unknown-unknown"
 
 ; ==============================================================================
 ; 16 x i8
 ; ==============================================================================

llvm/test/CodeGen/WebAssembly/stack-protector.ll

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc -verify-machineinstrs -mtriple=wasm32-unknown-unknown < %s | FileCheck -check-prefix=WASM32 %s
 
 @"\01LC" = internal constant [11 x i8] c"buf == %s\0A\00"    ; <ptr> [#uses=1]
 
 define void @test(ptr %a) nounwind ssp {
 ; WASM32-LABEL: test:
 ; WASM32:         .functype test (i32) -> ()
 ; WASM32-NEXT:    .local i32
 ; WASM32-NEXT:  # %bb.0: # %entry
 ; WASM32-NEXT:    global.get __stack_pointer
-; WASM32-NEXT:    i32.const 32
+; WASM32-NEXT:    i32.const 16
 ; WASM32-NEXT:    i32.sub
 ; WASM32-NEXT:    local.tee 1
 ; WASM32-NEXT:    global.set __stack_pointer
 ; WASM32-NEXT:    local.get 1
-; WASM32-NEXT:    local.get 0
-; WASM32-NEXT:    i32.store 16
-; WASM32-NEXT:    local.get 1
 ; WASM32-NEXT:    i32.const 0
 ; WASM32-NEXT:    i32.load __stack_chk_guard
-; WASM32-NEXT:    i32.store 28
+; WASM32-NEXT:    i32.store 12
 ; WASM32-NEXT:    local.get 1
-; WASM32-NEXT:    i32.const 20
+; WASM32-NEXT:    i32.const 4
 ; WASM32-NEXT:    i32.add
 ; WASM32-NEXT:    local.get 0
 ; WASM32-NEXT:    call strcpy
 ; WASM32-NEXT:    drop
 ; WASM32-NEXT:    local.get 1
 ; WASM32-NEXT:    local.get 1
-; WASM32-NEXT:    i32.const 20
+; WASM32-NEXT:    i32.const 4
 ; WASM32-NEXT:    i32.add
 ; WASM32-NEXT:    i32.store 0
 ; WASM32-NEXT:    i32.const LC
 ; WASM32-NEXT:    local.get 1
 ; WASM32-NEXT:    call printf
 ; WASM32-NEXT:    drop
 ; WASM32-NEXT:    block
 ; WASM32-NEXT:    i32.const 0
 ; WASM32-NEXT:    i32.load __stack_chk_guard
 ; WASM32-NEXT:    local.get 1
-; WASM32-NEXT:    i32.load 28
+; WASM32-NEXT:    i32.load 12
 ; WASM32-NEXT:    i32.eq
 ; WASM32-NEXT:    br_if 0 # 0: down to label0
 ; WASM32-NEXT:  # %bb.1: # %return
 ; WASM32-NEXT:    call __stack_chk_fail
 ; WASM32-NEXT:    unreachable
 ; WASM32-NEXT:  .LBB0_2: # %return
 ; WASM32-NEXT:    end_block # label0:
 ; WASM32-NEXT:    local.get 1
-; WASM32-NEXT:    i32.const 32
+; WASM32-NEXT:    i32.const 16
 ; WASM32-NEXT:    i32.add
 ; WASM32-NEXT:    global.set __stack_pointer
 ; WASM32-NEXT:    # fallthrough-return
 entry:
   %a_addr = alloca ptr    ; <ptr> [#uses=2]
   %buf = alloca [8 x i8]    ; <ptr> [#uses=2]
   %"alloca point" = bitcast i32 0 to i32    ; <i32> [#uses=0]
   store ptr %a, ptr %a_addr
   %0 = load ptr, ptr %a_addr, align 4    ; <ptr> [#uses=1]
   %1 = call ptr @strcpy(ptr %buf, ptr %0) nounwind    ; <ptr> [#uses=0]
   %2 = call i32 (ptr, ...) @printf(ptr @"\01LC", ptr %buf) nounwind    ; <i32> [#uses=0]
   br label %return
 
 return:   ; preds = %entry
   ret void
 }
 
 define i32 @test_return_i32(ptr %a) nounwind ssp {
 ; WASM32-LABEL: test_return_i32:
 ; WASM32:         .functype test_return_i32 (i32) -> (i32)
 ; WASM32-NEXT:    .local i32
 ; WASM32-NEXT:  # %bb.0: # %entry
 ; WASM32-NEXT:    global.get __stack_pointer
-; WASM32-NEXT:    i32.const 32
+; WASM32-NEXT:    i32.const 16
 ; WASM32-NEXT:    i32.sub
 ; WASM32-NEXT:    local.tee 1
 ; WASM32-NEXT:    global.set __stack_pointer
 ; WASM32-NEXT:    local.get 1
-; WASM32-NEXT:    local.get 0
-; WASM32-NEXT:    i32.store 16
-; WASM32-NEXT:    local.get 1
 ; WASM32-NEXT:    i32.const 0
 ; WASM32-NEXT:    i32.load __stack_chk_guard
-; WASM32-NEXT:    i32.store 28
+; WASM32-NEXT:    i32.store 12
 ; WASM32-NEXT:    local.get 1
-; WASM32-NEXT:    i32.const 20
+; WASM32-NEXT:    i32.const 4
 ; WASM32-NEXT:    i32.add
 ; WASM32-NEXT:    local.get 0
 ; WASM32-NEXT:    call strcpy
 ; WASM32-NEXT:    drop
 ; WASM32-NEXT:    local.get 1
 ; WASM32-NEXT:    local.get 1
-; WASM32-NEXT:    i32.const 20
+; WASM32-NEXT:    i32.const 4
 ; WASM32-NEXT:    i32.add
 ; WASM32-NEXT:    i32.store 0
 ; WASM32-NEXT:    i32.const LC
 ; WASM32-NEXT:    local.get 1
 ; WASM32-NEXT:    call printf
 ; WASM32-NEXT:    drop
 ; WASM32-NEXT:    block
 ; WASM32-NEXT:    i32.const 0
 ; WASM32-NEXT:    i32.load __stack_chk_guard
 ; WASM32-NEXT:    local.get 1
-; WASM32-NEXT:    i32.load 28
+; WASM32-NEXT:    i32.load 12
 ; WASM32-NEXT:    i32.eq
 ; WASM32-NEXT:    br_if 0 # 0: down to label1
 ; WASM32-NEXT:  # %bb.1: # %return
 ; WASM32-NEXT:    call __stack_chk_fail
 ; WASM32-NEXT:    unreachable
 ; WASM32-NEXT:  .LBB1_2: # %return
 ; WASM32-NEXT:    end_block # label1:
 ; WASM32-NEXT:    local.get 1
-; WASM32-NEXT:    i32.const 32
+; WASM32-NEXT:    i32.const 16
 ; WASM32-NEXT:    i32.add
 ; WASM32-NEXT:    global.set __stack_pointer
 ; WASM32-NEXT:    i32.const 0
 ; WASM32-NEXT:    # fallthrough-return
 entry:
   %a_addr = alloca ptr    ; <ptr> [#uses=2]
   %buf = alloca [8 x i8]    ; <ptr> [#uses=2]
   %"alloca point" = bitcast i32 0 to i32    ; <i32> [#uses=0]

llvm/test/CodeGen/WebAssembly/umulo-i64.ll

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc < %s -wasm-disable-explicit-locals -wasm-keep-registers | FileCheck %s
 ; Test that UMULO works correctly on 64-bit operands.
 target triple = "wasm32-unknown-unknown"
 
 ; Function Attrs: inlinehint
 define void @"_ZN4core3num21_$LT$impl$u20$u64$GT$15overflowing_mul17h07be88b4cbac028fE"(i64, i64) unnamed_addr #0 {
 ; CHECK-LABEL: _ZN4core3num21_$LT$impl$u20$u64$GT$15overflowing_mul17h07be88b4cbac028fE:
 ; CHECK:         .functype _ZN4core3num21_$LT$impl$u20$u64$GT$15overflowing_mul17h07be88b4cbac028fE (i64, i64) -> ()
 ; CHECK-NEXT:  # %bb.0: # %start
-; CHECK-NEXT:    global.get $push2=, __stack_pointer
-; CHECK-NEXT:    i32.const $push3=, 16
-; CHECK-NEXT:    i32.sub $push6=, $pop2, $pop3
-; CHECK-NEXT:    local.tee $push5=, $2=, $pop6
-; CHECK-NEXT:    global.set __stack_pointer, $pop5
-; CHECK-NEXT:    i64.const $push0=, 0
-; CHECK-NEXT:    i64.const $push4=, 0
-; CHECK-NEXT:    call __multi3, $2, $0, $pop0, $1, $pop4
-; CHECK-NEXT:    i64.load $push1=, 0($2)
-; CHECK-NEXT:    i64.store 0($2), $pop1
 ; CHECK-NEXT:    unreachable
 start:
   %2 = call { i64, i1 } @llvm.umul.with.overflow.i64(i64 %0, i64 %1)
   %3 = extractvalue { i64, i1 } %2, 0
   store i64 %3, ptr undef
   unreachable
 }
 

llvm/test/CodeGen/WebAssembly/userstack.ll

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
-; RUN: llc < %s --mtriple=wasm32-unknown-unknown -disable-wasm-fallthrough-return-opt -wasm-keep-registers | FileCheck -DPTR=32 %s --check-prefix=CHECK-32
-; RUN: llc < %s --mtriple=wasm64-unknown-unknown -disable-wasm-fallthrough-return-opt -wasm-keep-registers | FileCheck -DPTR=64 %s --check-prefix=CHECK-64
+; RUN: llc < %s -O1 --mtriple=wasm32-unknown-unknown -asm-verbose=false -disable-wasm-fallthrough-return-opt -wasm-keep-registers | FileCheck -DPTR=32 %s
+; RUN: llc < %s -O1 --mtriple=wasm64-unknown-unknown -asm-verbose=false -disable-wasm-fallthrough-return-opt -wasm-keep-registers | FileCheck -DPTR=64 %s
 
 declare void @ext_func(ptr %ptr)
 declare void @ext_func_i32(ptr %ptr)
 
 ; Check that there is an extra local for the stack pointer.
 define void @alloca32() noredzone {
 ; CHECK-32-LABEL: alloca32:
 ; CHECK-32:         .functype alloca32 () -> ()
; CHECK-64-NEXT:    return $pop5
   ;; Here @str is a global address and not an immediate, so cannot be folded
   %3 = getelementptr [3 x i8], ptr @str, i32 0, i32 %2
   %4 = load i8, ptr %3, align 8
   %5 = and i8 %4, 67
   ret i8 %5
 }
 
 ; TODO: test over-aligned alloca
+;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:
+; CHECK: {{.*}}

llvm/test/DebugInfo/WebAssembly/stackified-debug.ll

 ; RUN: llc -verify-machineinstrs < %s | FileCheck %s
-; RUN: llc -filetype=obj %s -o - | llvm-dwarfdump - | FileCheck %s --check-prefix DWARF
+; RUN: llc -O1 -filetype=obj %s -o - | llvm-dwarfdump - | FileCheck %s --check-prefix DWARF
 
 ; Input C code:
 
 ;  int i = input();  // Nested case
 ;  int j = input();  // Trivial def-use.
 ;  output(i, j);
 
 ; The ll below generates 330 lines of .S, so relevant parts that the

llvm/test/MC/WebAssembly/func-address.ll

-; RUN: llc -O2 -filetype=obj %s -o - | llvm-readobj -S -r --expand-relocs - | FileCheck %s
+; RUN: llc -O1 -filetype=obj %s -o - | llvm-readobj -S -r --expand-relocs - | FileCheck %s
 
 target triple = "wasm32-unknown-unknown"
 
 declare i32 @import1()
 declare i32 @import2()
 declare i32 @import3()
 
 ; call the imports to make sure they are included in the imports section

llvm/test/MC/WebAssembly/stack-ptr-mclower.ll

-; RUN: llc --mtriple=wasm32-unknown-unknown -asm-verbose=false %s -o - | FileCheck --check-prefixes CHECK -DPTR=i32 %s
-; RUN: llc --mtriple=wasm64-unknown-unknown -asm-verbose=false %s -o - | FileCheck --check-prefixes CHECK -DPTR=i64 %s
+; RUN: llc -O1 --mtriple=wasm32-unknown-unknown -asm-verbose=false %s -o - | FileCheck --check-prefixes CHECK -DPTR=i32 %s
+; RUN: llc -O1 --mtriple=wasm64-unknown-unknown -asm-verbose=false %s -o - | FileCheck --check-prefixes CHECK -DPTR=i64 %s
 
 ; empty function that does not use stack, to check WebAssemblyMCLowerPrePass
 ; works correctly.
 define hidden void @bar() #0 {
   ret void
 }
 
 ; Function that uses explict stack, and should generate a reference to

llvm/test/MC/WebAssembly/stack-ptr.ll

-; RUN: llc --mtriple=wasm32-unknown-unknown -filetype=obj %s -o - | obj2yaml | FileCheck --check-prefixes CHECK,CHK32 %s
-; RUN: llc --mtriple=wasm64-unknown-unknown -filetype=obj %s -o - | obj2yaml | FileCheck --check-prefixes CHECK,CHK64 %s
+; RUN: llc -O1 --mtriple=wasm32-unknown-unknown -filetype=obj %s -o - | obj2yaml | FileCheck --check-prefixes CHECK,CHK32 %s
+; RUN: llc -O1 --mtriple=wasm64-unknown-unknown -filetype=obj %s -o - | obj2yaml | FileCheck --check-prefixes CHECK,CHK64 %s
 
 ; Function that uses explict stack, and should generate a reference to
 ; __stack_pointer, along with the corresponding reloction entry.
 define hidden void @foo() #0 {
 entry:
   alloca i32, align 4
   ret void
 }

llvm/test/tools/UpdateTestChecks/update_llc_test_checks/Inputs/wasm_generated_funcs.ll.generated.expected

 }
 
 attributes #0 = { noredzone nounwind ssp uwtable "frame-pointer"="all" }
 ; CHECK-LABEL: check_boundaries:
 ; CHECK:         .functype check_boundaries () -> (i32)
 ; CHECK-NEXT:    .local i32
 ; CHECK-NEXT:  # %bb.0:
 ; CHECK-NEXT:    global.get __stack_pointer
-; CHECK-NEXT:    i32.const 32
+; CHECK-NEXT:    i32.const 16
 ; CHECK-NEXT:    i32.sub
 ; CHECK-NEXT:    local.tee 0
 ; CHECK-NEXT:    global.set __stack_pointer
 ; CHECK-NEXT:    local.get 0
-; CHECK-NEXT:    i32.const 0
-; CHECK-NEXT:    i32.store 24
-; CHECK-NEXT:    local.get 0
-; CHECK-NEXT:    i32.const 0
-; CHECK-NEXT:    i32.store 28
-; CHECK-NEXT:    block
+; CHECK-NEXT:    i32.const 1
+; CHECK-NEXT:    i32.store 12
 ; CHECK-NEXT:    block
-; CHECK-NEXT:    i32.const 0
-; CHECK-NEXT:    br_if 0 # 0: down to label1
+; CHECK-NEXT:    i32.const 1
+; CHECK-NEXT:    br_if 0 # 0: down to label0
 ; CHECK-NEXT:  # %bb.1:
 ; CHECK-NEXT:    local.get 0
-; CHECK-NEXT:    i32.const 2
-; CHECK-NEXT:    i32.store 20
-; CHECK-NEXT:    local.get 0
 ; CHECK-NEXT:    i32.const 1
-; CHECK-NEXT:    i32.store 24
-; CHECK-NEXT:    local.get 0
-; CHECK-NEXT:    i32.const 3
-; CHECK-NEXT:    i32.store 16
-; CHECK-NEXT:    local.get 0
-; CHECK-NEXT:    i32.const 4
 ; CHECK-NEXT:    i32.store 12
-; CHECK-NEXT:    br 1 # 1: down to label0
 ; CHECK-NEXT:  .LBB0_2:
-; CHECK-NEXT:    end_block # label1:
-; CHECK-NEXT:    local.get 0
-; CHECK-NEXT:    i32.const 1
-; CHECK-NEXT:    i32.store 16
-; CHECK-NEXT:  .LBB0_3:
 ; CHECK-NEXT:    end_block # label0:
-; CHECK-NEXT:    block
-; CHECK-NEXT:    block
-; CHECK-NEXT:    local.get 0
-; CHECK-NEXT:    i32.load 24
-; CHECK-NEXT:    br_if 0 # 0: down to label3
-; CHECK-NEXT:  # %bb.4:
-; CHECK-NEXT:    local.get 0
-; CHECK-NEXT:    i32.const 2
-; CHECK-NEXT:    i32.store 20
-; CHECK-NEXT:    local.get 0
-; CHECK-NEXT:    i32.const 1
-; CHECK-NEXT:    i32.store 24
-; CHECK-NEXT:    local.get 0
-; CHECK-NEXT:    i32.const 3
-; CHECK-NEXT:    i32.store 16
-; CHECK-NEXT:    local.get 0
-; CHECK-NEXT:    i32.const 4
-; CHECK-NEXT:    i32.store 12
-; CHECK-NEXT:    br 1 # 1: down to label2
-; CHECK-NEXT:  .LBB0_5:
-; CHECK-NEXT:    end_block # label3:
-; CHECK-NEXT:    local.get 0
-; CHECK-NEXT:    i32.const 1
-; CHECK-NEXT:    i32.store 16
-; CHECK-NEXT:  .LBB0_6:
-; CHECK-NEXT:    end_block # label2:
 ; CHECK-NEXT:    local.get 0
-; CHECK-NEXT:    i32.const 32
+; CHECK-NEXT:    i32.const 16
 ; CHECK-NEXT:    i32.add
 ; CHECK-NEXT:    global.set __stack_pointer
 ; CHECK-NEXT:    i32.const 0
 ; CHECK-NEXT:    # fallthrough-return
 ;
 ; CHECK-LABEL: __original_main:
 ; CHECK:         .functype __original_main () -> (i32)
-; CHECK-NEXT:    .local i32
 ; CHECK-NEXT:  # %bb.0:
-; CHECK-NEXT:    global.get __stack_pointer
-; CHECK-NEXT:    i32.const 32
-; CHECK-NEXT:    i32.sub
-; CHECK-NEXT:    local.tee 0
-; CHECK-NEXT:    global.set __stack_pointer
 ; CHECK-NEXT:    i32.const 0
 ; CHECK-NEXT:    i32.const 1
 ; CHECK-NEXT:    i32.store x
-; CHECK-NEXT:    local.get 0
-; CHECK-NEXT:    i32.const 0
-; CHECK-NEXT:    i32.store 28
-; CHECK-NEXT:    local.get 0
-; CHECK-NEXT:    i32.const 1
-; CHECK-NEXT:    i32.store 24
-; CHECK-NEXT:    local.get 0
-; CHECK-NEXT:    i32.const 2
-; CHECK-NEXT:    i32.store 20
-; CHECK-NEXT:    local.get 0
-; CHECK-NEXT:    i32.const 3
-; CHECK-NEXT:    i32.store 16
-; CHECK-NEXT:    local.get 0
-; CHECK-NEXT:    i32.const 4
-; CHECK-NEXT:    i32.store 12
 ; CHECK-NEXT:    #APP
 ; CHECK-NEXT:    #NO_APP
-; CHECK-NEXT:    local.get 0
-; CHECK-NEXT:    i32.const 2
-; CHECK-NEXT:    i32.store 20
-; CHECK-NEXT:    local.get 0
-; CHECK-NEXT:    i32.const 1
-; CHECK-NEXT:    i32.store 24
-; CHECK-NEXT:    local.get 0
-; CHECK-NEXT:    i32.const 3
-; CHECK-NEXT:    i32.store 16
-; CHECK-NEXT:    local.get 0
-; CHECK-NEXT:    i32.const 4
-; CHECK-NEXT:    i32.store 12
-; CHECK-NEXT:    local.get 0
-; CHECK-NEXT:    i32.const 32
-; CHECK-NEXT:    i32.add
-; CHECK-NEXT:    global.set __stack_pointer
 ; CHECK-NEXT:    i32.const 0
 ; CHECK-NEXT:    # fallthrough-return
 ;
 ; CHECK-LABEL: main:
 ; CHECK:         .functype main (i32, i32) -> (i32)
 ; CHECK-NEXT:  # %bb.0: # %body
 ; CHECK-NEXT:    call __original_main
 ; CHECK-NEXT:    # fallthrough-return

llvm/test/tools/UpdateTestChecks/update_llc_test_checks/Inputs/wasm_generated_funcs.ll.nogenerated.expected

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc -enable-machine-outliner -mtriple=wasm32-unknown-linux < %s | FileCheck %s
 
 ; NOTE: Machine outliner doesn't run.
 @x = global i32 0, align 4
 
 define dso_local i32 @check_boundaries() #0 {
 ; CHECK-LABEL: check_boundaries:
 ; CHECK:         .functype check_boundaries () -> (i32)
 ; CHECK-NEXT:    .local i32
 ; CHECK-NEXT:  # %bb.0:
 ; CHECK-NEXT:    global.get __stack_pointer
-; CHECK-NEXT:    i32.const 32
+; CHECK-NEXT:    i32.const 16
 ; CHECK-NEXT:    i32.sub
 ; CHECK-NEXT:    local.tee 0
 ; CHECK-NEXT:    global.set __stack_pointer
 ; CHECK-NEXT:    local.get 0
-; CHECK-NEXT:    i32.const 0
-; CHECK-NEXT:    i32.store 24
-; CHECK-NEXT:    local.get 0
-; CHECK-NEXT:    i32.const 0
-; CHECK-NEXT:    i32.store 28
-; CHECK-NEXT:    block
+; CHECK-NEXT:    i32.const 1
+; CHECK-NEXT:    i32.store 12
 ; CHECK-NEXT:    block
-; CHECK-NEXT:    i32.const 0
-; CHECK-NEXT:    br_if 0 # 0: down to label1
+; CHECK-NEXT:    i32.const 1
+; CHECK-NEXT:    br_if 0 # 0: down to label0
 ; CHECK-NEXT:  # %bb.1:
 ; CHECK-NEXT:    local.get 0
-; CHECK-NEXT:    i32.const 2
-; CHECK-NEXT:    i32.store 20
-; CHECK-NEXT:    local.get 0
 ; CHECK-NEXT:    i32.const 1
-; CHECK-NEXT:    i32.store 24
-; CHECK-NEXT:    local.get 0
-; CHECK-NEXT:    i32.const 3
-; CHECK-NEXT:    i32.store 16
-; CHECK-NEXT:    local.get 0
-; CHECK-NEXT:    i32.const 4
 ; CHECK-NEXT:    i32.store 12
-; CHECK-NEXT:    br 1 # 1: down to label0
 ; CHECK-NEXT:  .LBB0_2:
-; CHECK-NEXT:    end_block # label1:
-; CHECK-NEXT:    local.get 0
-; CHECK-NEXT:    i32.const 1
-; CHECK-NEXT:    i32.store 16
-; CHECK-NEXT:  .LBB0_3:
 ; CHECK-NEXT:    end_block # label0:
-; CHECK-NEXT:    block
-; CHECK-NEXT:    block
-; CHECK-NEXT:    local.get 0
-; CHECK-NEXT:    i32.load 24
-; CHECK-NEXT:    br_if 0 # 0: down to label3
-; CHECK-NEXT:  # %bb.4:
-; CHECK-NEXT:    local.get 0
-; CHECK-NEXT:    i32.const 2
-; CHECK-NEXT:    i32.store 20
-; CHECK-NEXT:    local.get 0
-; CHECK-NEXT:    i32.const 1
-; CHECK-NEXT:    i32.store 24
-; CHECK-NEXT:    local.get 0
-; CHECK-NEXT:    i32.const 3
-; CHECK-NEXT:    i32.store 16
-; CHECK-NEXT:    local.get 0
-; CHECK-NEXT:    i32.const 4
-; CHECK-NEXT:    i32.store 12
-; CHECK-NEXT:    br 1 # 1: down to label2
-; CHECK-NEXT:  .LBB0_5:
-; CHECK-NEXT:    end_block # label3:
-; CHECK-NEXT:    local.get 0
-; CHECK-NEXT:    i32.const 1
-; CHECK-NEXT:    i32.store 16
-; CHECK-NEXT:  .LBB0_6:
-; CHECK-NEXT:    end_block # label2:
 ; CHECK-NEXT:    local.get 0
-; CHECK-NEXT:    i32.const 32
+; CHECK-NEXT:    i32.const 16
 ; CHECK-NEXT:    i32.add
 ; CHECK-NEXT:    global.set __stack_pointer
 ; CHECK-NEXT:    i32.const 0
 ; CHECK-NEXT:    # fallthrough-return
   %1 = alloca i32, align 4
   %2 = alloca i32, align 4
   %3 = alloca i32, align 4
   %4 = alloca i32, align 4