[InstCombine] mempcpy(d,s,n) to memcpy(d,s,n) + n

Summary:
Back-end currently expands mempcpy, but middle-end should work with memcpy instead of mempcpy to enable more memcpy-optimization.

GCC backend emits mempcpy, so LLVM backend could form it too, if we know mempcpy libcall is better than memcpy + n.
https://godbolt.org/z/dOCG96

Reviewers: efriedma, spatel, craig.topper, RKSimon, jdoerfert

Reviewed By: efriedma

Subscribers: hjl.tools, llvm-commits

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D65737

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@370593 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/include/llvm/Transforms/Utils/SimplifyLibCalls.h b/include/llvm/Transforms/Utils/SimplifyLibCalls.h
index b7e6224f8556b74b6d34df778bf5038fcdd6c5da..7fff3e63b31fe0ccc2d2af6c64f6647ab58afaf3 100644 (file)
--- a/include/llvm/Transforms/Utils/SimplifyLibCalls.h
+++ b/include/llvm/Transforms/Utils/SimplifyLibCalls.h
@@ -167,6 +167,7 @@ private:
   Value *optimizeMemCmp(CallInst *CI, IRBuilder<> &B);
   Value *optimizeBCmp(CallInst *CI, IRBuilder<> &B);
   Value *optimizeMemCmpBCmpCommon(CallInst *CI, IRBuilder<> &B);
+  Value *optimizeMemPCpy(CallInst *CI, IRBuilder<> &B);
   Value *optimizeMemCpy(CallInst *CI, IRBuilder<> &B, bool isIntrinsic = false);
   Value *optimizeMemMove(CallInst *CI, IRBuilder<> &B, bool isIntrinsic = false);
   Value *optimizeMemSet(CallInst *CI, IRBuilder<> &B, bool isIntrinsic = false);

diff --git a/lib/Transforms/Utils/SimplifyLibCalls.cpp b/lib/Transforms/Utils/SimplifyLibCalls.cpp
index 7ff1f4aef237215d2345661cd2ace4890a089b1a..1f397c0dedbfbf3a17485a166708d5d595465585 100644 (file)
--- a/lib/Transforms/Utils/SimplifyLibCalls.cpp
+++ b/lib/Transforms/Utils/SimplifyLibCalls.cpp
@@ -998,6 +998,15 @@ Value *LibCallSimplifier::optimizeMemCpy(CallInst *CI, IRBuilder<> &B,
   return CI->getArgOperand(0);
 }
 
+Value *LibCallSimplifier::optimizeMemPCpy(CallInst *CI, IRBuilder<> &B) {
+  Value *Dst = CI->getArgOperand(0);
+  Value *N = CI->getArgOperand(2);
+  // mempcpy(x, y, n) -> llvm.memcpy(align 1 x, align 1 y, n), x + n
+  CallInst *NewCI = B.CreateMemCpy(Dst, 1, CI->getArgOperand(1), 1, N);
+  NewCI->setAttributes(CI->getAttributes());
+  return B.CreateInBoundsGEP(B.getInt8Ty(), Dst, N);
+}
+
 Value *LibCallSimplifier::optimizeMemMove(CallInst *CI, IRBuilder<> &B, bool isIntrinsic) {
   Value *Size = CI->getArgOperand(2);
   if (ConstantInt *LenC = dyn_cast<ConstantInt>(Size))
@@ -2624,6 +2633,8 @@ Value *LibCallSimplifier::optimizeStringMemoryLibCall(CallInst *CI,
       return optimizeMemCmp(CI, Builder);
     case LibFunc_memcpy:
       return optimizeMemCpy(CI, Builder);
+    case LibFunc_mempcpy:
+      return optimizeMemPCpy(CI, Builder);
     case LibFunc_memmove:
       return optimizeMemMove(CI, Builder);
     case LibFunc_memset:

diff --git a/test/Transforms/InstCombine/mempcpy.ll b/test/Transforms/InstCombine/mempcpy.ll
index ce372d8f37d7dcf7721ea126fb1308b89e78f890..7aa6a31a61ac3a09f293e67f40fc15f141a53941 100644 (file)
--- a/test/Transforms/InstCombine/mempcpy.ll
+++ b/test/Transforms/InstCombine/mempcpy.ll
@@ -3,17 +3,41 @@
 
 define i8* @memcpy_nonconst_n(i8* %d, i8* nocapture readonly %s, i64 %n) {
 ; CHECK-LABEL: @memcpy_nonconst_n(
-; CHECK-NEXT:    [[R:%.*]] = tail call i8* @mempcpy(i8* [[D:%.*]], i8* [[S:%.*]], i64 [[N:%.*]])
-; CHECK-NEXT:    ret i8* [[R]]
+; CHECK-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 1 [[D:%.*]], i8* align 1 [[S:%.*]], i64 [[N:%.*]], i1 false)
+; CHECK-NEXT:    [[TMP1:%.*]] = getelementptr inbounds i8, i8* [[D]], i64 [[N]]
+; CHECK-NEXT:    ret i8* [[TMP1]]
 ;
   %r = tail call i8* @mempcpy(i8* %d, i8* %s, i64 %n)
   ret i8* %r
 }
 
+define i8* @memcpy_nonconst_n_copy_attrs(i8* %d, i8* nocapture readonly %s, i64 %n) {
+; CHECK-LABEL: @memcpy_nonconst_n_copy_attrs(
+; CHECK-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 1 dereferenceable(16) [[D:%.*]], i8* align 1 [[S:%.*]], i64 [[N:%.*]], i1 false)
+; CHECK-NEXT:    [[TMP1:%.*]] = getelementptr inbounds i8, i8* [[D]], i64 [[N]]
+; CHECK-NEXT:    ret i8* [[TMP1]]
+;
+  %r = tail call i8* @mempcpy(i8* dereferenceable(16) %d, i8* %s, i64 %n)
+  ret i8* %r
+}
+
+define void @memcpy_nonconst_n_unused_retval(i8* %d, i8* nocapture readonly %s, i64 %n) {
+; CHECK-LABEL: @memcpy_nonconst_n_unused_retval(
+; CHECK-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 1 [[D:%.*]], i8* align 1 [[S:%.*]], i64 [[N:%.*]], i1 false)
+; CHECK-NEXT:    ret void
+;
+  call i8* @mempcpy(i8* %d, i8* %s, i64 %n)
+  ret void
+}
+
 define i8* @memcpy_small_const_n(i8* %d, i8* nocapture readonly %s) {
 ; CHECK-LABEL: @memcpy_small_const_n(
-; CHECK-NEXT:    [[R:%.*]] = tail call i8* @mempcpy(i8* [[D:%.*]], i8* [[S:%.*]], i64 8)
-; CHECK-NEXT:    ret i8* [[R]]
+; CHECK-NEXT:    [[TMP1:%.*]] = bitcast i8* [[S:%.*]] to i64*
+; CHECK-NEXT:    [[TMP2:%.*]] = bitcast i8* [[D:%.*]] to i64*
+; CHECK-NEXT:    [[TMP3:%.*]] = load i64, i64* [[TMP1]], align 1
+; CHECK-NEXT:    store i64 [[TMP3]], i64* [[TMP2]], align 1
+; CHECK-NEXT:    [[TMP4:%.*]] = getelementptr inbounds i8, i8* [[D]], i64 8
+; CHECK-NEXT:    ret i8* [[TMP4]]
 ;
   %r = tail call i8* @mempcpy(i8* %d, i8* %s, i64 8)
   ret i8* %r
@@ -21,8 +45,9 @@ define i8* @memcpy_small_const_n(i8* %d, i8* nocapture readonly %s) {
 
 define i8* @memcpy_big_const_n(i8* %d, i8* nocapture readonly %s) {
 ; CHECK-LABEL: @memcpy_big_const_n(
-; CHECK-NEXT:    [[R:%.*]] = tail call i8* @mempcpy(i8* [[D:%.*]], i8* [[S:%.*]], i64 1024)
-; CHECK-NEXT:    ret i8* [[R]]
+; CHECK-NEXT:    call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 1 dereferenceable(1024) [[D:%.*]], i8* align 1 dereferenceable(1024) [[S:%.*]], i64 1024, i1 false)
+; CHECK-NEXT:    [[TMP1:%.*]] = getelementptr inbounds i8, i8* [[D]], i64 1024
+; CHECK-NEXT:    ret i8* [[TMP1]]
 ;
   %r = tail call i8* @mempcpy(i8* %d, i8* %s, i64 1024)
   ret i8* %r