[WebAssembly] Restore defaults for stores per memop

[tlively]

Summary

Large slowdowns were observed in Rust due to many small, constant
sized copies in conjunction with poorly-optimized memory.copy
implementations. Since memory.copy cannot be expected to be inlined
efficiently by engines at this time, stop using it for the smallest
copies. We continue to lower all memcpy intrinsics to memory.copy,
though.

[aheejin]

What are the default values for those if we unset them?

[alexcrichton]

I've tested this patch compiling the motivating test case for the original performance issue for Firefox and performance is back to what it was in LLVM 8, so I can at least say from my perspective that this does the trick!

[tlively]

In D67639#1671921, @aheejin wrote:

What are the default values for those if we unset them?

8 stores normally or 4 stores when optimizing for size. If SIMD is enabled that will be 128 and 64 bytes, respectively, or else it will be 64 and 32 bytes.

llvm/trunk/lib/Target/WebAssembly/WebAssemblyISelLowering.cpp

WebAssemblyTargetLowering::WebAssemblyTargetLowering(
   setOperationAction(ISD::TRAP, MVT::Other, Legal);
 
   // Exception handling intrinsics
   setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
   setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom);
 
   setMaxAtomicSizeInBitsSupported(64);
 
-  if (Subtarget->hasBulkMemory()) {
-    // Use memory.copy and friends over multiple loads and stores
-    MaxStoresPerMemcpy = 1;
-    MaxStoresPerMemcpyOptSize = 1;
-    MaxStoresPerMemmove = 1;
-    MaxStoresPerMemmoveOptSize = 1;
-    MaxStoresPerMemset = 1;
-    MaxStoresPerMemsetOptSize = 1;
-  }
-
   // Override the __gnu_f2h_ieee/__gnu_h2f_ieee names so that the f32 name is
   // consistent with the f64 and f128 names.
   setLibcallName(RTLIB::FPEXT_F16_F32, "__extendhfsf2");
   setLibcallName(RTLIB::FPROUND_F32_F16, "__truncsfhf2");
 
   // Define the emscripten name for return address helper.
   // TODO: when implementing other WASM backends, make this generic or only do
   // this on emscripten depending on what they end up doing.

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

 ; BULK-MEM-NEXT: memory.fill 0, $0, $1, $pop[[L0]]
 ; BULK-MEM-NEXT: return
 define void @memset_1024(i8* %dest, i8 %val) {
   call void @llvm.memset.p0i8.i32(i8* %dest, i8 %val, i32 1024, i1 0)
   ret void
 }
 
 ; The following tests check that frame index elimination works for
-; bulk memory instructions. The stack pointer is bumped by 16 instead
-; of 10 because the stack pointer in WebAssembly is currently always
+; bulk memory instructions. The stack pointer is bumped by 112 instead
+; of 100 because the stack pointer in WebAssembly is currently always
 ; 16-byte aligned, even in leaf functions, although it is not written
 ; back to the global in this case.
 
 ; TODO: Change TransientStackAlignment to 1 to avoid this extra
 ; arithmetic. This will require forcing the use of StackAlignment in
 ; PrologEpilogEmitter.cpp when
 ; WebAssemblyFrameLowering::needsSPWriteback would be true.
 
 ; CHECK-LABEL: memcpy_alloca_src:
 ; NO-BULK-MEM-NOT: memory.copy
 ; BULK-MEM-NEXT: .functype memcpy_alloca_src (i32) -> ()
 ; BULK-MEM-NEXT: global.get $push[[L0:[0-9]+]]=, __stack_pointer
-; BULK-MEM-NEXT: i32.const $push[[L1:[0-9]+]]=, 16
+; BULK-MEM-NEXT: i32.const $push[[L1:[0-9]+]]=, 112
 ; BULK-MEM-NEXT: i32.sub $push[[L2:[0-9]+]]=, $pop[[L0]], $pop[[L1]]
-; BULK-MEM-NEXT: i32.const $push[[L3:[0-9]+]]=, 6
+; BULK-MEM-NEXT: i32.const $push[[L3:[0-9]+]]=, 12
 ; BULK-MEM-NEXT: i32.add $push[[L4:[0-9]+]]=, $pop[[L2]], $pop[[L3]]
-; BULK-MEM-NEXT: i32.const $push[[L5:[0-9]+]]=, 10
+; BULK-MEM-NEXT: i32.const $push[[L5:[0-9]+]]=, 100
 ; BULK-MEM-NEXT: memory.copy 0, 0, $0, $pop[[L4]], $pop[[L5]]
 ; BULK-MEM-NEXT: return
 define void @memcpy_alloca_src(i8* %dst) {
-  %a = alloca [10 x i8]
-  %p = bitcast [10 x i8]* %a to i8*
-  call void @llvm.memcpy.p0i8.p0i8.i32(i8* %dst, i8* %p, i32 10, i1 false)
+  %a = alloca [100 x i8]
+  %p = bitcast [100 x i8]* %a to i8*
+  call void @llvm.memcpy.p0i8.p0i8.i32(i8* %dst, i8* %p, i32 100, i1 false)
   ret void
 }
 
 ; CHECK-LABEL: memcpy_alloca_dst:
 ; NO-BULK-MEM-NOT: memory.copy
 ; BULK-MEM-NEXT: .functype memcpy_alloca_dst (i32) -> ()
 ; BULK-MEM-NEXT: global.get $push[[L0:[0-9]+]]=, __stack_pointer
-; BULK-MEM-NEXT: i32.const $push[[L1:[0-9]+]]=, 16
+; BULK-MEM-NEXT: i32.const $push[[L1:[0-9]+]]=, 112
 ; BULK-MEM-NEXT: i32.sub $push[[L2:[0-9]+]]=, $pop[[L0]], $pop[[L1]]
-; BULK-MEM-NEXT: i32.const $push[[L3:[0-9]+]]=, 6
+; BULK-MEM-NEXT: i32.const $push[[L3:[0-9]+]]=, 12
 ; BULK-MEM-NEXT: i32.add $push[[L4:[0-9]+]]=, $pop[[L2]], $pop[[L3]]
-; BULK-MEM-NEXT: i32.const $push[[L5:[0-9]+]]=, 10
+; BULK-MEM-NEXT: i32.const $push[[L5:[0-9]+]]=, 100
 ; BULK-MEM-NEXT: memory.copy 0, 0, $pop[[L4]], $0, $pop[[L5]]
 ; BULK-MEM-NEXT: return
 define void @memcpy_alloca_dst(i8* %src) {
-  %a = alloca [10 x i8]
-  %p = bitcast [10 x i8]* %a to i8*
-  call void @llvm.memcpy.p0i8.p0i8.i32(i8* %p, i8* %src, i32 10, i1 false)
+  %a = alloca [100 x i8]
+  %p = bitcast [100 x i8]* %a to i8*
+  call void @llvm.memcpy.p0i8.p0i8.i32(i8* %p, i8* %src, i32 100, i1 false)
   ret void
 }
 
 ; CHECK-LABEL: memset_alloca:
 ; NO-BULK-MEM-NOT: memory.fill
 ; BULK-MEM-NEXT: .functype memset_alloca (i32) -> ()
 ; BULK-MEM-NEXT: global.get $push[[L0:[0-9]+]]=, __stack_pointer
-; BULK-MEM-NEXT: i32.const $push[[L1:[0-9]+]]=, 16
+; BULK-MEM-NEXT: i32.const $push[[L1:[0-9]+]]=, 112
 ; BULK-MEM-NEXT: i32.sub $push[[L2:[0-9]+]]=, $pop[[L0]], $pop[[L1]]
-; BULK-MEM-NEXT: i32.const $push[[L3:[0-9]+]]=, 6
+; BULK-MEM-NEXT: i32.const $push[[L3:[0-9]+]]=, 12
 ; BULK-MEM-NEXT: i32.add $push[[L4:[0-9]+]]=, $pop[[L2]], $pop[[L3]]
-; BULK-MEM-NEXT: i32.const $push[[L5:[0-9]+]]=, 10
+; BULK-MEM-NEXT: i32.const $push[[L5:[0-9]+]]=, 100
 ; BULK-MEM-NEXT: memory.fill 0, $pop[[L4]], $0, $pop[[L5]]
 ; BULK-MEM-NEXT: return
 define void @memset_alloca(i8 %val) {
-  %a = alloca [10 x i8]
-  %p = bitcast [10 x i8]* %a to i8*
-  call void @llvm.memset.p0i8.i32(i8* %p, i8 %val, i32 10, i1 false)
+  %a = alloca [100 x i8]
+  %p = bitcast [100 x i8]* %a to i8*
+  call void @llvm.memset.p0i8.i32(i8* %p, i8 %val, i32 100, i1 false)
   ret void
 }