[WebAssembly] Restore defaults for stores per memop

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.

Reviewers: aheejin, alexcrichton

Subscribers: dschuff, sbc100, jgravelle-google, hiraditya, JDevlieghere, sunfish, llvm-commits

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D67639

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@372275 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/WebAssembly/WebAssemblyISelLowering.cpp b/lib/Target/WebAssembly/WebAssemblyISelLowering.cpp
index 78dfa88b510b02889a488e7a41a67eb6d0752f7b..49a9fb0c6190ab52e9a2894efc07c48b772fdc0b 100644 (file)
--- a/lib/Target/WebAssembly/WebAssemblyISelLowering.cpp
+++ b/lib/Target/WebAssembly/WebAssemblyISelLowering.cpp
@@ -259,16 +259,6 @@ WebAssemblyTargetLowering::WebAssemblyTargetLowering(
 
   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");

diff --git a/test/CodeGen/WebAssembly/bulk-memory.ll b/test/CodeGen/WebAssembly/bulk-memory.ll
index 6574511089b74c84103dd4914e88d2cc77e46ddf..89ea7e8f826e2c6a29651edceda56832dee12394 100644 (file)
--- a/test/CodeGen/WebAssembly/bulk-memory.ll
+++ b/test/CodeGen/WebAssembly/bulk-memory.ll
@@ -142,8 +142,8 @@ define void @memset_1024(i8* %dest, i8 %val) {
 }
 
 ; 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.
 
@@ -156,17 +156,17 @@ define void @memset_1024(i8* %dest, i8 %val) {
 ; 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
 }
 
@@ -174,17 +174,17 @@ define void @memcpy_alloca_src(i8* %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
 }
 
@@ -192,16 +192,16 @@ define void @memcpy_alloca_dst(i8* %src) {
 ; 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
 }