Move SeparateConstOffsetFromGEPPass() before LSR() and enable EnableGEPOpt by default.

[gsocshubham]

Summary

GEP's across basic blocks were not getting splitted due to EnableGEPOpt which was turned off by default. Hence, EarlyCSE missed the opportunity to eliminate common part of GEP's. This can be achieved by simply turning GEP pass on.

1. This patch moves SeparateConstOffsetFromGEPPass() just before LSR().
2. It enables EnableGEPOpt by default.

Resolves - https://github.com/llvm/llvm-project/issues/50528

Added an unit test.

[gsocshubham]

Below are SPECS intrate benchmark results obtained on commit f6c79c6ae49f3a642bebe32a2346186c38bb83d7

Peak tuning(-O3)

Benchmark          %change

500.perlbench_r    2.173
505.mcf_r          -0.472
520.omnetpp_r      0.055
523.xalancbmk_r    -0.939
525.x264_r         -4.048
531.deepsjeng_r    4.944
541.leela_r        0.604
557.xz_r           0.209

Base tuning(-O2)

Benchmark          %change

500.perlbench_r    1.341
505.mcf_r          -0.526
520.omnetpp_r      0.961
523.xalancbmk_r    1.257
525.x264_r         -6.739
531.deepsjeng_r    0
541.leela_r        -0.462
557.xz_r           0.522

The only benchmark with most regression is 525.x264_r. Most of the benchmarks shows improved results.

[gsocshubham]

llvm/lib/Transforms/Scalar/SeparateConstOffsetFromGEP.cpp

1050 ↗

(On Diff #439975)

Let me know best course of action on lowerToSingleIndexGEPs().

I have temporarily removed it because I was running SPEC benchmarks for AArch64 and GEP's are being splitted using lowerToSingleIndexGEPs() when called using clang.

But when called using opt, lowerToArithmetics() gets called. I want to have lowerToArithmetics() as default.

Also, this pass was no longer used.

[gsocshubham]

Please review and suggest changes.

[xbolva00]

Missing context
Missing test
x264_r regression is big one..
Many tests fail in precommit CI - please update them

[dmgreen]

Hello. Unfortunately, I doubt that people will be in favour of this approach, especially if it is introducing ptr2int and int2ptr's so early in the pass pipeline. It looks like a pass that needs to be run as part of the backend.

There is a run of the pass in the AArch64 backend already, but it is disabled by default: https://github.com/llvm/llvm-project/blob/main/llvm/lib/Target/AArch64/AArch64TargetMachine.cpp#L581
Enabling it didn't seem to help with the original case in #50528. Would it be possible to teach it what it needs to for that case, and from there enable the EnableGEPOpt flag?

[gsocshubham]

In D128582#3624787, @dmgreen wrote:

Hello. Unfortunately, I doubt that people will be in favour of this approach, especially if it is introducing ptr2int and int2ptr's so early in the pass pipeline. It looks like a pass that needs to be run as part of the backend.

There is a run of the pass in the AArch64 backend already, but it is disabled by default: https://github.com/llvm/llvm-project/blob/main/llvm/lib/Target/AArch64/AArch64TargetMachine.cpp#L581

Hello David,

Passing -aarch64-enable-gep-opt=true -O3 would do the job - https://clang.godbolt.org/z/hcjsh1vex what this patch was proposing. WDYT?

Enabling it didn't seem to help with the original case in #50528. Would it be possible to teach it what it needs to for that case, and from there enable the EnableGEPOpt flag?

Sure. For #50528, enabling EnableGEPOpt reduces GEP instructions by half but is done by splitting GEP into ptr2int and int2ptr.

[nikic]

In D128582#3624787, @dmgreen wrote:

Hello. Unfortunately, I doubt that people will be in favour of this approach, especially if it is introducing ptr2int and int2ptr's so early in the pass pipeline. It looks like a pass that needs to be run as part of the backend.

Indeed. This pass, especially in lowerToArithmetics() mode, is only usable in the backend.

[dmgreen]

Passing -aarch64-enable-gep-opt=true -O3 would do the job - https://clang.godbolt.org/z/hcjsh1vex what this patch was proposing. WDYT?

..

Sure. For #50528, enabling EnableGEPOpt reduces GEP instructions by half but is done by splitting GEP into ptr2int and int2ptr.

Oh yeah I see it does. I must have missed the -O3 off the time I tried it. That's good. I wonder why it was never enabled in the past.

From the look of it, it runs after LSR, which I think it would need to run before. Otherwise it is likely to mess up what LSR has tried to do. That would be before the call to TargetPassConfig::addIRPasses(). I'm not sure if the LICM run is necessary either, but I see it is used in other backends. We would need to gather some benchmark to see how it behaves. Like do the issues in x264 still occur, and what happens across more benchmark cases.

[gsocshubham]

In D128582#3628937, @dmgreen wrote:

Passing -aarch64-enable-gep-opt=true -O3 would do the job - https://clang.godbolt.org/z/hcjsh1vex what this patch was proposing. WDYT?

..

Sure. For #50528, enabling EnableGEPOpt reduces GEP instructions by half but is done by splitting GEP into ptr2int and int2ptr.

Oh yeah I see it does. I must have missed the -O3 off the time I tried it. That's good. I wonder why it was never enabled in the past.

From the look of it, it runs after LSR, which I think it would need to run before. Otherwise it is likely to mess up what LSR has tried to do. That would be before the call to TargetPassConfig::addIRPasses(). I'm not sure if the LICM run is necessary either, but I see it is used in other backends. We would need to gather some benchmark to see how it behaves. Like do the issues in x264 still occur, and what happens across more benchmark cases.

Will running it before LSR make any difference compared to current location?

The issue in x264 was occuring due to pass registeration way early at IR level. It is irrelevant now since it is already enabled in AArch64 if passed relevant flags. As we have been given suggestion to move GEP pass just before Selection DAG, I have registered/moved GEP pass from addIRPasses() to at the end of AArch64PassConfig::addCodeGenPrepare() but there is no regression in x264 benchmark result as compared to master.

[dmgreen]

In D128582#3637881, @gsocshubham wrote:

In D128582#3628937, @dmgreen wrote:

Passing -aarch64-enable-gep-opt=true -O3 would do the job - https://clang.godbolt.org/z/hcjsh1vex what this patch was proposing. WDYT?

..

Sure. For #50528, enabling EnableGEPOpt reduces GEP instructions by half but is done by splitting GEP into ptr2int and int2ptr.

Oh yeah I see it does. I must have missed the -O3 off the time I tried it. That's good. I wonder why it was never enabled in the past.

From the look of it, it runs after LSR, which I think it would need to run before. Otherwise it is likely to mess up what LSR has tried to do. That would be before the call to TargetPassConfig::addIRPasses(). I'm not sure if the LICM run is necessary either, but I see it is used in other backends. We would need to gather some benchmark to see how it behaves. Like do the issues in x264 still occur, and what happens across more benchmark cases.

Will running it before LSR make any difference compared to current location?

The issue in x264 was occuring due to pass registeration way early at IR level. It is irrelevant now since it is already enabled in AArch64 if passed relevant flags. As we have been given suggestion to move GEP pass just before Selection DAG, I have registered/moved GEP pass from addIRPasses() to at the end of AArch64PassConfig::addCodeGenPrepare() but there is no regression in x264 benchmark result as compared to master.

Everything in AArch64PassConfig::addIRPasses counts as the "backend" from the point of view of LLVM. They are still llvm-ir passes, but ran as part of the backend prior to ISel lowering to perform certain optimizations and help with lowering that are target-specific, but easier to perform on IR than MIR. I think the pass really needs to run before the call to LSR, as LSR will be very opinionated about the geps in loops and we don't want to mess that up. LSR is added by TargetPassConfig::addIRPasses(), so moving the SeparateConstOffsetFromGEP passes anywhere before that call in AArch64PassConfig::addIRPasses should be OK.

[gsocshubham]

Move GEP pass before LSR pass.

[gsocshubham]

In D128582#3640458, @dmgreen wrote:

In D128582#3637881, @gsocshubham wrote:

In D128582#3628937, @dmgreen wrote:

Passing -aarch64-enable-gep-opt=true -O3 would do the job - https://clang.godbolt.org/z/hcjsh1vex what this patch was proposing. WDYT?

..

Sure. For #50528, enabling EnableGEPOpt reduces GEP instructions by half but is done by splitting GEP into ptr2int and int2ptr.

Oh yeah I see it does. I must have missed the -O3 off the time I tried it. That's good. I wonder why it was never enabled in the past.

From the look of it, it runs after LSR, which I think it would need to run before. Otherwise it is likely to mess up what LSR has tried to do. That would be before the call to TargetPassConfig::addIRPasses(). I'm not sure if the LICM run is necessary either, but I see it is used in other backends. We would need to gather some benchmark to see how it behaves. Like do the issues in x264 still occur, and what happens across more benchmark cases.

Will running it before LSR make any difference compared to current location?

The issue in x264 was occuring due to pass registeration way early at IR level. It is irrelevant now since it is already enabled in AArch64 if passed relevant flags. As we have been given suggestion to move GEP pass just before Selection DAG, I have registered/moved GEP pass from addIRPasses() to at the end of AArch64PassConfig::addCodeGenPrepare() but there is no regression in x264 benchmark result as compared to master.

Everything in AArch64PassConfig::addIRPasses counts as the "backend" from the point of view of LLVM. They are still llvm-ir passes, but ran as part of the backend prior to ISel lowering to perform certain optimizations and help with lowering that are target-specific, but easier to perform on IR than MIR. I think the pass really needs to run before the call to LSR, as LSR will be very opinionated about the geps in loops and we don't want to mess that up. LSR is added by TargetPassConfig::addIRPasses(), so moving the SeparateConstOffsetFromGEP passes anywhere before that call in AArch64PassConfig::addIRPasses should be OK.

I have moved GEP pass just before LSR pass and below are benchmarks results at peak (O3) -

Benchmark              %change w.r.t master
500.perlbench_r ->     2.085
505.mcf_r       ->     0.466
520.omnetpp_r   ->     0.607
523.xalancbmk_r ->     -0.515
531.deepsjeng_r ->     0.326
541.leela_r     ->     -0.980
557.xz_r        ->     -0.0571

[gsocshubham]

In D128582#3624787, @dmgreen wrote:

Hello. Unfortunately, I doubt that people will be in favour of this approach, especially if it is introducing ptr2int and int2ptr's so early in the pass pipeline. It looks like a pass that needs to be run as part of the backend.

There is a run of the pass in the AArch64 backend already, but it is disabled by default: https://github.com/llvm/llvm-project/blob/main/llvm/lib/Target/AArch64/AArch64TargetMachine.cpp#L581
Enabling it didn't seem to help with the original case in #50528. Would it be possible to teach it what it needs to for that case, and from there enable the EnableGEPOpt flag?

Regarding teaching SeparateConstOffsetFromGEPPass pass to handle constant GEP's -

From #50528, there is only one case where there is gep with all constant indices as below -

1. store i32 %inc, i32* getelementptr inbounds (%struct.state_t, %struct.state_t* @s, i64 0, i32 2)

-> At first, this was not being considered for gep as it is store instruction and hence I split above into 2 instructions in testcase as below -

%temp = getelementptr inbounds %struct.state_t, %struct.state_t* @s, i64 0, i32 2
  store i32 %inc, i32* %temp

2. %temp is not being considered by the pass since it has all constant indices -

// The backend can already nicely handle the case where all indices are
// constant.
if (GEP->hasAllConstantIndices())
  return false;

I tried forcefully splitting it by passing the checks but it does not seem right to split GEP with all constant indices at IR level. From the comment backend passes should take care of it but I still see repeated instructions -

madd    x8, x8, x10, x9
madd    x0, x11, x12, x8

I see 2 occurrencea of above set of instruction in assembly. Ideally it should occur only once.

Reference - https://clang.godbolt.org/z/KfrfT97hn

[dmgreen]

It sounds like this would be a good first step, and we can look into the other geps in the issue in a separate patch if needed. Can you:

• Change the EnableGEPOpt from false to true
• Add a test case from the bug, run through llc to show the codegen improvements.
• Update the patch with full context

[nikic]

(Patch no longer modifies the middle-end pipeline in unacceptable ways)

[gsocshubham]

a. Updated patch with context.
b. Updated title and summary.
c. Moved SeparateConstOffsetFromGEPPass() before LSR()
d. Added an unit test.

[gsocshubham]

This patch does not enable splitting of GEP with all constant indices - SeparateConstOffsetFromGEP.cpp:969

// The backend can already nicely handle the case where all indices are
// constant.
if (GEP->hasAllConstantIndices())
  return false;

[dmgreen]

Thanks - Please reread this and see what "with full context" means: https://llvm.org/docs/Phabricator.html#requesting-a-review-via-the-web-interface. The patch should be created with -U9999999.

I was thinking of the original longer test from the issue, but it is a bit long. This one might be OK too.

llvm/test/CodeGen/AArch64/cond-br-tuning.ll

33

Change the store here to store 10 not 0. That should keep this testing what it did previously.

llvm/test/Transforms/SeparateConstOffsetFromGEP/split-gep.ll

3

Remove -aarch64-enable-gep-opt - it is on by default now.

31

Remove extra line.

[gsocshubham]

Updated patch using -U9999999 and fixed review comments.

llvm/test/CodeGen/AArch64/cond-br-tuning.ll

33

Done.

[gsocshubham]

In D128582#3657201, @dmgreen wrote:

Thanks - Please reread this and see what "with full context" means: https://llvm.org/docs/Phabricator.html#requesting-a-review-via-the-web-interface. The patch should be created with -U9999999.

Understood. Thanks. I have updated patch using -U9999999. Is it fine now?

I was thinking of the original longer test from the issue, but it is a bit long. This one might be OK too.

[dmgreen]

Thanks. This looks good, but did the test go missing from the previous revision?

llvm/test/CodeGen/AArch64/O3-pipeline.ll

12

This should be removed, as the file doesn't use the script.

[gsocshubham]

Fixed review comments.

llvm/test/Transforms/SeparateConstOffsetFromGEP/split-gep.ll

3

Done.

31

Done.

[gsocshubham]

In D128582#3665002, @dmgreen wrote:

Thanks. This looks good, but did the test go missing from the previous revision?

Yes. It got removed. I have added it.

[gsocshubham]

On my latest revision, build is failed. Is it due to pre merge test failures?

Build Status
    Buildable 176471	
    Build 265103: pre-merge checks	x64 windows failed · x64 debian failed

What is the script used for building? Can someone please point to it? In my local, build is passing. I want to replicate current build at my local with same config options.

[dmgreen]

Thanks. The buildbot failure seems unrelated.

This LGTM if you can alter the test slightly and things are passing locally.

llvm/test/CodeGen/AArch64/cond-br-tuning.ll

36

%d = icmp ne i32 %c, 10

39

store i32 10, i32* %ptr, align 4

[gsocshubham]

llvm/test/CodeGen/AArch64/cond-br-tuning.ll

36

Do you mean it to change it to %d = icmp ne i32 %c, 0?

It is already %d = icmp ne i32 %c, 10

[dmgreen]

llvm/test/CodeGen/AArch64/cond-br-tuning.ll

36

Yeah sorry, that is what I meant. The store should be storing a value that isn't 0, so the csel isn't optimized away. The cmp should still not be present (it does get optimized, as it can re-use the adds flags).

[gsocshubham]

Fixed review comments.

There is only one LIT test which is failing now - CodeGen/AArch64/GlobalISel/arm64-irtranslator.ll.

[gsocshubham]

llvm/test/CodeGen/AArch64/cond-br-tuning.ll

36

Done.

39

Done.

[gsocshubham]

In D128582#3668948, @gsocshubham wrote:

Fixed review comments.

There is only one LIT test which is failing now - CodeGen/AArch64/GlobalISel/arm64-irtranslator.ll.

Regarding CodeGen/AArch64/GlobalISel/arm64-irtranslator.ll, it is failing due to change in GEP representation. There are -O0 and -O3 checks from which -O3 checks are failing due to change in GEP representation. I tried to update testcase using - ../../../../utils/update_mir_test_checks.py --llc-binary=../../../../../install/bin/llc arm64-irtranslator.ll due to which checks had been updated but now -O0 checks are failing. The testcase contains 6k lines and is difficult to update manually.

I used this option to just update O3 checks but that is not helping -
../../../../utils/update_mir_test_checks.py --llc-binary=../../../../../install/bin/llc arm64-irtranslator.ll --filter O3

How do I just update O3 checks for testcase CodeGen/AArch64/GlobalISel/arm64-irtranslator.l? I have updated CodeGen/AArch64/GlobalISel/arm64-irtranslator-gep.ll using utils/update_mir_test_checks.py successfully.

[dmgreen]

In D128582#3668948, @gsocshubham wrote:

There is only one LIT test which is failing now - CodeGen/AArch64/GlobalISel/arm64-irtranslator.ll.

It looks like it is removing a unused alloca. Try this:

index ef559652380e..3fb33ecbb2c7 100644
--- a/llvm/test/CodeGen/AArch64/GlobalISel/arm64-irtranslator.ll
+++ b/llvm/test/CodeGen/AArch64/GlobalISel/arm64-irtranslator.ll
@@ -1458,10 +1458,12 @@ define void @test_lifetime_intrin() {
 ; O3-LABEL: name: test_lifetime_intrin
 ; O3: {{%[0-9]+}}:_(p0) = G_FRAME_INDEX %stack.0.slot
 ; O3-NEXT: LIFETIME_START %stack.0.slot
+; O3-NEXT: G_STORE
 ; O3-NEXT: LIFETIME_END %stack.0.slot
 ; O3-NEXT: RET_ReallyLR
   %slot = alloca i8, i32 4
   call void @llvm.lifetime.start.p0i8(i64 0, i8* %slot)
+  store volatile i8 10, i8* %slot
   call void @llvm.lifetime.end.p0i8(i64 0, i8* %slot)
   ret void
 }

llvm/test/CodeGen/AArch64/GlobalISel/arm64-irtranslator-gep.ll

54

Make this %v1, %v2, I think is better. One of the loads is otherwise unused.

llvm/test/Transforms/SeparateConstOffsetFromGEP/split-gep.ll

1

Oh - If this test is using the AArch64 target then it needs to be moved into an AArch64 subdirectory, which is only ran when AArch64 is a registered target. If it doesn't exist already, make sure there is a lit.local.cfg.

[gsocshubham]

Fix review comments and lit test failures - AArch64/GlobalISel/arm64-irtranslator.ll and AArch64/GlobalISel/arm64-irtranslator-gep.ll

llvm/test/CodeGen/AArch64/GlobalISel/arm64-irtranslator-gep.ll

54

Done.

[gsocshubham]

In D128582#3669192, @dmgreen wrote:

In D128582#3668948, @gsocshubham wrote:

There is only one LIT test which is failing now - CodeGen/AArch64/GlobalISel/arm64-irtranslator.ll.

It looks like it is removing a unused alloca. Try this:

index ef559652380e..3fb33ecbb2c7 100644
--- a/llvm/test/CodeGen/AArch64/GlobalISel/arm64-irtranslator.ll
+++ b/llvm/test/CodeGen/AArch64/GlobalISel/arm64-irtranslator.ll
@@ -1458,10 +1458,12 @@ define void @test_lifetime_intrin() {
 ; O3-LABEL: name: test_lifetime_intrin
 ; O3: {{%[0-9]+}}:_(p0) = G_FRAME_INDEX %stack.0.slot
 ; O3-NEXT: LIFETIME_START %stack.0.slot
+; O3-NEXT: G_STORE
 ; O3-NEXT: LIFETIME_END %stack.0.slot
 ; O3-NEXT: RET_ReallyLR
   %slot = alloca i8, i32 4
   call void @llvm.lifetime.start.p0i8(i64 0, i8* %slot)
+  store volatile i8 10, i8* %slot
   call void @llvm.lifetime.end.p0i8(i64 0, i8* %slot)
   ret void
 }

Thanks a lot for this.

Are there any other suggestions on this patch?

[dmgreen]

Yeah - the patch LGTM. Thanks.

Do you have commit access? If not, I can commit it for you, I just need a "name <[e-mail]>" to attribute it to.

[gsocshubham]

In D128582#3671454, @dmgreen wrote:

Yeah - the patch LGTM. Thanks.

Do you have commit access? If not, I can commit it for you, I just need a "name <[e-mail]>" to attribute it to.

I do not have commit access. Please commit for me.

Here are the details -

Name - Shubham Narlawar
Email - [e-mail]

"Shubham Narlawar <[e-mail]>"

llvm/lib/Target/AArch64/AArch64TargetMachine.cpp

 static cl::opt<bool>
     EnableCondOpt("aarch64-enable-condopt",
                   cl::desc("Enable the condition optimizer pass"),
                   cl::init(true), cl::Hidden);
 
 static cl::opt<bool>
     EnableGEPOpt("aarch64-enable-gep-opt", cl::Hidden,
                  cl::desc("Enable optimizations on complex GEPs"),
-                 cl::init(false));
+                 cl::init(true));
 
 static cl::opt<bool>
     BranchRelaxation("aarch64-enable-branch-relax", cl::Hidden, cl::init(true),
                      cl::desc("Relax out of range conditional branches"));
 
 static cl::opt<bool> EnableCompressJumpTables(
     "aarch64-enable-compress-jump-tables", cl::Hidden, cl::init(true),
     cl::desc("Use smallest entry possible for jump tables"));
void AArch64PassConfig::addIRPasses() {
   // pointer values N iterations ahead.
   if (TM->getOptLevel() != CodeGenOpt::None) {
     if (EnableLoopDataPrefetch)
       addPass(createLoopDataPrefetchPass());
     if (EnableFalkorHWPFFix)
       addPass(createFalkorMarkStridedAccessesPass());
   }
 
-  TargetPassConfig::addIRPasses();
-
-  addPass(createAArch64StackTaggingPass(
-      /*IsOptNone=*/TM->getOptLevel() == CodeGenOpt::None));
-
-  // Match interleaved memory accesses to ldN/stN intrinsics.
-  if (TM->getOptLevel() != CodeGenOpt::None) {
-    addPass(createInterleavedLoadCombinePass());
-    addPass(createInterleavedAccessPass());
-  }
-
   if (TM->getOptLevel() == CodeGenOpt::Aggressive && EnableGEPOpt) {
     // Call SeparateConstOffsetFromGEP pass to extract constants within indices
     // and lower a GEP with multiple indices to either arithmetic operations or
     // multiple GEPs with single index.
     addPass(createSeparateConstOffsetFromGEPPass(true));
     // Call EarlyCSE pass to find and remove subexpressions in the lowered
     // result.
     addPass(createEarlyCSEPass());
     // Do loop invariant code motion in case part of the lowered result is
     // invariant.
     addPass(createLICMPass());
   }
 
+  TargetPassConfig::addIRPasses();
+
+  addPass(createAArch64StackTaggingPass(
+      /*IsOptNone=*/TM->getOptLevel() == CodeGenOpt::None));
+
+  // Match interleaved memory accesses to ldN/stN intrinsics.
+  if (TM->getOptLevel() != CodeGenOpt::None) {
+    addPass(createInterleavedLoadCombinePass());
+    addPass(createInterleavedAccessPass());
+  }
+
   // Add Control Flow Guard checks.
   if (TM->getTargetTriple().isOSWindows())
     addPass(createCFGuardCheckPass());
 
   if (TM->Options.JMCInstrument)
     addPass(createJMCInstrumenterPass());
 }
 

llvm/test/CodeGen/AArch64/GlobalISel/arm64-irtranslator-gep.ll

 ; NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
 ; RUN: llc -O0 -stop-after=irtranslator -global-isel -verify-machineinstrs %s -o - 2>&1 | FileCheck %s --check-prefix=O0
 ; RUN: llc -O3 -stop-after=irtranslator -global-isel -verify-machineinstrs %s -o - 2>&1 | FileCheck %s --check-prefix=O3
 target datalayout = "e-m:o-i64:64-i128:128-n32:64-S128"
 target triple = "aarch64--"
 
 define i32 @cse_gep([4 x i32]* %ptr, i32 %idx) {
   ; O0-LABEL: name: cse_gep
   ; O0: bb.1 (%ir-block.0):
-  ; O0:   liveins: $w1, $x0
-  ; O0:   [[COPY:%[0-9]+]]:_(p0) = COPY $x0
-  ; O0:   [[COPY1:%[0-9]+]]:_(s32) = COPY $w1
-  ; O0:   [[SEXT:%[0-9]+]]:_(s64) = G_SEXT [[COPY1]](s32)
-  ; O0:   [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 16
-  ; O0:   [[MUL:%[0-9]+]]:_(s64) = G_MUL [[SEXT]], [[C]]
-  ; O0:   [[PTR_ADD:%[0-9]+]]:_(p0) = G_PTR_ADD [[COPY]], [[MUL]](s64)
-  ; O0:   [[COPY2:%[0-9]+]]:_(p0) = COPY [[PTR_ADD]](p0)
-  ; O0:   [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY2]](p0) :: (load (s32) from %ir.gep1)
-  ; O0:   [[MUL1:%[0-9]+]]:_(s64) = G_MUL [[SEXT]], [[C]]
-  ; O0:   [[PTR_ADD1:%[0-9]+]]:_(p0) = G_PTR_ADD [[COPY]], [[MUL1]](s64)
-  ; O0:   [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
-  ; O0:   [[PTR_ADD2:%[0-9]+]]:_(p0) = G_PTR_ADD [[PTR_ADD1]], [[C1]](s64)
-  ; O0:   [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p0) :: (load (s32) from %ir.gep2)
-  ; O0:   [[ADD:%[0-9]+]]:_(s32) = G_ADD [[LOAD1]], [[LOAD1]]
-  ; O0:   $w0 = COPY [[ADD]](s32)
-  ; O0:   RET_ReallyLR implicit $w0
+  ; O0-NEXT:   liveins: $w1, $x0
+  ; O0-NEXT: {{  $}}
+  ; O0-NEXT:   [[COPY:%[0-9]+]]:_(p0) = COPY $x0
+  ; O0-NEXT:   [[COPY1:%[0-9]+]]:_(s32) = COPY $w1
+  ; O0-NEXT:   [[SEXT:%[0-9]+]]:_(s64) = G_SEXT [[COPY1]](s32)
+  ; O0-NEXT:   [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 16
+  ; O0-NEXT:   [[MUL:%[0-9]+]]:_(s64) = G_MUL [[SEXT]], [[C]]
+  ; O0-NEXT:   [[PTR_ADD:%[0-9]+]]:_(p0) = G_PTR_ADD [[COPY]], [[MUL]](s64)
+  ; O0-NEXT:   [[COPY2:%[0-9]+]]:_(p0) = COPY [[PTR_ADD]](p0)
+  ; O0-NEXT:   [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY2]](p0) :: (load (s32) from %ir.gep1)
+  ; O0-NEXT:   [[MUL1:%[0-9]+]]:_(s64) = G_MUL [[SEXT]], [[C]]
+  ; O0-NEXT:   [[PTR_ADD1:%[0-9]+]]:_(p0) = G_PTR_ADD [[COPY]], [[MUL1]](s64)
+  ; O0-NEXT:   [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+  ; O0-NEXT:   [[PTR_ADD2:%[0-9]+]]:_(p0) = G_PTR_ADD [[PTR_ADD1]], [[C1]](s64)
+  ; O0-NEXT:   [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p0) :: (load (s32) from %ir.gep2)
+  ; O0-NEXT:   [[ADD:%[0-9]+]]:_(s32) = G_ADD [[LOAD1]], [[LOAD1]]
+  ; O0-NEXT:   $w0 = COPY [[ADD]](s32)
+  ; O0-NEXT:   RET_ReallyLR implicit $w0
   ; O3-LABEL: name: cse_gep
   ; O3: bb.1 (%ir-block.0):
-  ; O3:   liveins: $w1, $x0
-  ; O3:   [[COPY:%[0-9]+]]:_(p0) = COPY $x0
-  ; O3:   [[COPY1:%[0-9]+]]:_(s32) = COPY $w1
-  ; O3:   [[SEXT:%[0-9]+]]:_(s64) = G_SEXT [[COPY1]](s32)
-  ; O3:   [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 16
-  ; O3:   [[MUL:%[0-9]+]]:_(s64) = G_MUL [[SEXT]], [[C]]
-  ; O3:   [[PTR_ADD:%[0-9]+]]:_(p0) = G_PTR_ADD [[COPY]], [[MUL]](s64)
-  ; O3:   [[COPY2:%[0-9]+]]:_(p0) = COPY [[PTR_ADD]](p0)
-  ; O3:   [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY2]](p0) :: (load (s32) from %ir.gep1)
-  ; O3:   [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
-  ; O3:   [[PTR_ADD1:%[0-9]+]]:_(p0) = G_PTR_ADD [[PTR_ADD]], [[C1]](s64)
-  ; O3:   [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p0) :: (load (s32) from %ir.gep2)
-  ; O3:   [[ADD:%[0-9]+]]:_(s32) = G_ADD [[LOAD1]], [[LOAD1]]
-  ; O3:   $w0 = COPY [[ADD]](s32)
-  ; O3:   RET_ReallyLR implicit $w0
+  ; O3-NEXT:   liveins: $w1, $x0
+  ; O3-NEXT: {{  $}}
+  ; O3-NEXT:   [[COPY:%[0-9]+]]:_(p0) = COPY $x0
+  ; O3-NEXT:   [[COPY1:%[0-9]+]]:_(s32) = COPY $w1
+  ; O3-NEXT:   [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+  ; O3-NEXT:   [[SEXT:%[0-9]+]]:_(s64) = G_SEXT [[COPY1]](s32)
+  ; O3-NEXT:   [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 16
+  ; O3-NEXT:   [[MUL:%[0-9]+]]:_(s64) = G_MUL [[SEXT]], [[C1]]
+  ; O3-NEXT:   [[PTR_ADD:%[0-9]+]]:_(p0) = G_PTR_ADD [[COPY]], [[MUL]](s64)
+  ; O3-NEXT:   [[COPY2:%[0-9]+]]:_(p0) = COPY [[PTR_ADD]](p0)
+  ; O3-NEXT:   [[SHL:%[0-9]+]]:_(s64) = G_SHL [[SEXT]], [[C]](s64)
+  ; O3-NEXT:   [[PTR_ADD1:%[0-9]+]]:_(p0) = G_PTR_ADD [[COPY]], [[SHL]](s64)
+  ; O3-NEXT:   [[COPY3:%[0-9]+]]:_(p0) = COPY [[PTR_ADD1]](p0)
+  ; O3-NEXT:   [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 4
+  ; O3-NEXT:   [[PTR_ADD2:%[0-9]+]]:_(p0) = G_PTR_ADD [[COPY3]], [[C2]](s64)
+  ; O3-NEXT:   [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p0) :: (load (s32) from %ir.3)
+  ; O3-NEXT:   [[ADD:%[0-9]+]]:_(s32) = G_ADD [[LOAD]], [[LOAD]]
+  ; O3-NEXT:   $w0 = COPY [[ADD]](s32)
+  ; O3-NEXT:   RET_ReallyLR implicit $w0
   %sidx = sext i32 %idx to i64
   %gep1 = getelementptr inbounds [4 x i32], [4 x i32]* %ptr, i64 %sidx, i64 0
   %v1 = load i32, i32* %gep1
   %gep2 = getelementptr inbounds [4 x i32], [4 x i32]* %ptr, i64 %sidx, i64 1
   %v2 = load i32, i32* %gep2
   %res = add i32 %v2, %v2
   ret i32 %res
 }

llvm/test/CodeGen/AArch64/O3-pipeline.ll

 ; RUN: llc --debugify-and-strip-all-safe=0 -mtriple=arm64-- -O3 -debug-pass=Structure < %s -o /dev/null 2>&1 | \
 ; RUN:     grep -v "Verify generated machine code" | FileCheck %s
 
 ; REQUIRES: asserts
 
 ; CHECK-LABEL: Pass Arguments:
 ; CHECK-NEXT: Target Library Information
 ; CHECK-NEXT: Target Pass Configuration
 ; CHECK-NEXT: Machine Module Information
 ; CHECK-NEXT: Target Transform Information
 ; CHECK-NEXT: Assumption Cache Tracker
 ; CHECK-NEXT: Profile summary info
 ; CHECK-NEXT: Type-Based Alias Analysis
 ; CHECK-NEXT: Scoped NoAlias Alias Analysis
 ; CHECK-NEXT: Create Garbage Collector Module Metadata
 ; CHECK-NEXT: Machine Branch Probability Analysis
 ; CHECK-NEXT: Default Regalloc Eviction Advisor
 ; CHECK-NEXT:   ModulePass Manager
 ; CHECK-NEXT:     Pre-ISel Intrinsic Lowering
 ; CHECK-NEXT:     FunctionPass Manager
 ; CHECK-NEXT:       Expand Atomic instructions
 ; CHECK-NEXT:     SVE intrinsics optimizations
 ; CHECK-NEXT:       FunctionPass Manager
 ; CHECK-NEXT:         Dominator Tree Construction
 ; CHECK-NEXT:     FunctionPass Manager
 ; CHECK-NEXT:       Simplify the CFG
 ; CHECK-NEXT:       Dominator Tree Construction
 ; CHECK-NEXT:       Natural Loop Information
 ; CHECK-NEXT:       Canonicalize natural loops
 ; CHECK-NEXT:       Lazy Branch Probability Analysis
 ; CHECK-NEXT:       Lazy Block Frequency Analysis
 ; CHECK-NEXT:       Optimization Remark Emitter
 ; CHECK-NEXT:       Scalar Evolution Analysis
 ; CHECK-NEXT:       Loop Data Prefetch
 ; CHECK-NEXT:       Falkor HW Prefetch Fix
-; CHECK-NEXT:       Module Verifier
+; CHECK-NEXT:       Split GEPs to a variadic base and a constant offset for better CSE
+; CHECK-NEXT:       Early CSE
 ; CHECK-NEXT:       Basic Alias Analysis (stateless AA impl)
+; CHECK-NEXT:       Function Alias Analysis Results
+; CHECK-NEXT:       Memory SSA
 ; CHECK-NEXT:       Canonicalize natural loops
+; CHECK-NEXT:       LCSSA Verifier
+; CHECK-NEXT:       Loop-Closed SSA Form Pass
+; CHECK-NEXT:       Scalar Evolution Analysis
+; CHECK-NEXT:       Lazy Branch Probability Analysis
+; CHECK-NEXT:       Lazy Block Frequency Analysis
+; CHECK-NEXT:       Loop Pass Manager
+; CHECK-NEXT:         Loop Invariant Code Motion
+; CHECK-NEXT:       Module Verifier
 ; CHECK-NEXT:       Loop Pass Manager
 ; CHECK-NEXT:         Canonicalize Freeze Instructions in Loops
 ; CHECK-NEXT:         Induction Variable Users
 ; CHECK-NEXT:         Loop Strength Reduction
 ; CHECK-NEXT:       Basic Alias Analysis (stateless AA impl)
 ; CHECK-NEXT:       Function Alias Analysis Results
 ; CHECK-NEXT:       Merge contiguous icmps into a memcmp
 ; CHECK-NEXT:       Natural Loop Information

llvm/test/CodeGen/AArch64/cond-br-tuning.ll

 L2:
   store i32 1, i32* %ptr, align 4
   ret void
 }
 
 define void @test_add_cbz_multiple_use(i32 %a, i32 %b, i32* %ptr) {
 ; CHECK-LABEL: test_add_cbz_multiple_use:
 ; CHECK:       // %bb.0: // %common.ret
-; CHECK-NEXT:    adds w8, w0, w1
-; CHECK-NEXT:    csel w8, wzr, w8, ne
+; CHECK-NEXT:    mov w8, #10
+; CHECK-NEXT:    adds w9, w0, w1
+; CHECK-NEXT:    csel w8, w8, w9, ne
 ; CHECK-NEXT:    str w8, [x2]
 ; CHECK-NEXT:    ret
   %c = add nsw i32 %a, %b
   %d = icmp ne i32 %c, 0
   br i1 %d, label %L1, label %L2
 L1:
-  store i32 0, i32* %ptr, align 4
+  store i32 10, i32* %ptr, align 4
   ret void
 L2:
   store i32 %c, i32* %ptr, align 4
   ret void
 }
 
 define void @test_add_cbz_64(i64 %a, i64 %b, i64* %ptr) {
 ; CHECK-LABEL: test_add_cbz_64:

llvm/test/Transforms/SeparateConstOffsetFromGEP/split-gep.ll

+; RUN: llc < %s -O3 -mtriple=aarch64-linux-gnu | FileCheck %s
+
+%struct = type { i32, i32, i32 }
+
+define i32 @test1(%struct* %ptr, i64 %idx) {
+; CHECK-LABEL: test1:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    mov w8, #12
+; CHECK-NEXT:    madd x8, x1, x8, x0
+; CHECK-NEXT:    ldr w9, [x8, #4]
+; CHECK-NEXT:    tbnz w9, #31, .LBB0_2
+; CHECK-NEXT:  // %bb.1:
+; CHECK-NEXT:    mov w0, wzr
+; CHECK-NEXT:    ret
+; CHECK-NEXT:  .LBB0_2: // %then
+; CHECK-NEXT:    ldr w8, [x8, #8]
+; CHECK-NEXT:    add w0, w9, w8
+; CHECK-NEXT:    ret
+ %gep.1 = getelementptr %struct, %struct* %ptr, i64 %idx, i32 1
+ %lv.1 = load i32, i32* %gep.1
+ %c = icmp slt i32 %lv.1, 0
+ br i1 %c, label %then, label %else
+
+then:
+ %gep.2 = getelementptr %struct, %struct* %ptr, i64 %idx, i32 2
+ %lv.2 = load i32, i32* %gep.2
+ %res = add i32 %lv.1, %lv.2
+ ret i32 %res
+
+else:
+ ret i32 0
+}