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

[xbolva00]

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

[xbolva00]

Ping

[RKSimon]

Shouldn't this wait until we have confirmed the backend can create mempcpy ?

[xbolva00]

I dont think so.

Nowdays middle end does nothing with mempcpy, backend expands it to memcpy + n and then possibly does some memcpy opts. Middle end loses opportunities to do something better.

To sum up, with this patch nothing should regress, only improve due to middle-end opt.

[jdoerfert]

I think this makes sense, do you have verified that it doesn't regress anything?

[xbolva00]

mempcpy is quite rare, so hard to measure it on codebases.

I wrote small C “benchmarks” and no perf changes (backend produced memcpy/inlined anyway).

Another motivation is our recent attribute annotation work so we would get all good things from memcpy. Otherwise, I would have to annotate it individually.

It seems (some gcc discussion) only x86-64 has fast mempcpy libc impl. - so I really think backend should create mempcpy under target hook.

Not sure, maybe @craig.topper has more info about mempcpy perf on x86-64.

[jdoerfert]

I like that we get the attributes, let's wait and see if anyone objects.

[craig.topper]

@hjl.tools can you speak to mempcpy perf?

[xbolva00]

As side note, there are a few cases in real word code
https://codesearch.isocpp.org/cgi-bin/cgi_ppsearch?q=mempcpy&search=Search

where mempcpy's return value is not used.. so it is pessiimizaton to use mempcpy instead of memcpy..

[xbolva00]

Ping.

[xbolva00]

ping

[efriedma]

LGTM

[xbolva00]

Rebased + more tests.

llvm/trunk/include/llvm/Transforms/Utils/SimplifyLibCalls.h

private:
   Value *optimizeStrTo(CallInst *CI, IRBuilder<> &B);
   Value *optimizeStrSpn(CallInst *CI, IRBuilder<> &B);
   Value *optimizeStrCSpn(CallInst *CI, IRBuilder<> &B);
   Value *optimizeStrStr(CallInst *CI, IRBuilder<> &B);
   Value *optimizeMemChr(CallInst *CI, IRBuilder<> &B);
   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);
   Value *optimizeRealloc(CallInst *CI, IRBuilder<> &B);
   Value *optimizeWcslen(CallInst *CI, IRBuilder<> &B);
   // Wrapper for all String/Memory Library Call Optimizations
   Value *optimizeStringMemoryLibCall(CallInst *CI, IRBuilder<> &B);
 

llvm/trunk/lib/Transforms/Utils/SimplifyLibCalls.cpp

Value *LibCallSimplifier::optimizeMemCpy(CallInst *CI, IRBuilder<> &B,
   if (isIntrinsic)
     return nullptr;
 
   // memcpy(x, y, n) -> llvm.memcpy(align 1 x, align 1 y, n)
   B.CreateMemCpy(CI->getArgOperand(0), 1, CI->getArgOperand(1), 1, Size);
   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))
     annotateDereferenceableBytes(CI, {0, 1}, LenC->getZExtValue());
 
   if (isIntrinsic)
     return nullptr;
 
if (TLI->getLibFunc(*Callee, Func) && TLI->has(Func)) {
     case LibFunc_memchr:
       return optimizeMemChr(CI, Builder);
     case LibFunc_bcmp:
       return optimizeBCmp(CI, Builder);
     case LibFunc_memcmp:
       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:
       return optimizeMemSet(CI, Builder);
     case LibFunc_realloc:
       return optimizeRealloc(CI, Builder);
     case LibFunc_wcslen:
       return optimizeWcslen(CI, Builder);

llvm/trunk/test/Transforms/InstCombine/mempcpy.ll

 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
 ; RUN: opt -instcombine -S < %s | FileCheck %s
 
 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
 }
 
 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
 }
 
 declare i8* @mempcpy(i8*, i8* nocapture readonly, i64)