[llvm] [MemCpyOpt] Drop dead `memmove` calls on `memset`'d source data (PR #101930)

Antonio Frighetto via llvm-commits llvm-commits at lists.llvm.org
Wed Nov 27 10:41:11 PST 2024


https://github.com/antoniofrighetto updated https://github.com/llvm/llvm-project/pull/101930

>From 2c338b2ef1c7c67edce917b491309452dbfc788c Mon Sep 17 00:00:00 2001
From: Antonio Frighetto <me at antoniofrighetto.com>
Date: Fri, 20 Sep 2024 17:41:06 +0200
Subject: [PATCH 1/3] [MemCpyOpt] Introduce test for PR101930 (NFC)

---
 .../memset-memmove-redundant-memmove.ll       | 109 ++++++++++++++++++
 1 file changed, 109 insertions(+)
 create mode 100644 llvm/test/Transforms/MemCpyOpt/memset-memmove-redundant-memmove.ll

diff --git a/llvm/test/Transforms/MemCpyOpt/memset-memmove-redundant-memmove.ll b/llvm/test/Transforms/MemCpyOpt/memset-memmove-redundant-memmove.ll
new file mode 100644
index 00000000000000..5b323f119f0950
--- /dev/null
+++ b/llvm/test/Transforms/MemCpyOpt/memset-memmove-redundant-memmove.ll
@@ -0,0 +1,109 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt -passes=memcpyopt -S %s -verify-memoryssa | FileCheck %s
+
+; Redundant memmove.
+define i32 @redundant_memmove() {
+; CHECK-LABEL: @redundant_memmove(
+; CHECK-NEXT:    [[ARRAY:%.*]] = alloca [26 x i32], align 16
+; CHECK-NEXT:    call void @llvm.memset.p0.i64(ptr align 16 [[ARRAY]], i8 0, i64 104, i1 false)
+; CHECK-NEXT:    [[ARRAY_IDX:%.*]] = getelementptr inbounds i8, ptr [[ARRAY]], i64 4
+; CHECK-NEXT:    call void @llvm.memmove.p0.p0.i64(ptr align 16 [[ARRAY]], ptr align 4 [[ARRAY_IDX]], i64 100, i1 false)
+; CHECK-NEXT:    [[VAL:%.*]] = load i32, ptr [[ARRAY]], align 16
+; CHECK-NEXT:    ret i32 [[VAL]]
+;
+  %array = alloca [26 x i32], align 16
+  call void @llvm.memset.p0.i64(ptr align 16 %array, i8 0, i64 104, i1 false)
+  %array.idx = getelementptr inbounds i8, ptr %array, i64 4
+  call void @llvm.memmove.p0.p0.i64(ptr align 16 %array, ptr align 4 %array.idx, i64 100, i1 false)
+  %val = load i32, ptr %array, align 16
+  ret i32 %val
+}
+
+; Used memmove, buffer is reset to zero.
+define i32 @used_memmove_1() {
+; CHECK-LABEL: @used_memmove_1(
+; CHECK-NEXT:    [[ARRAY:%.*]] = alloca [26 x i32], align 16
+; CHECK-NEXT:    call void @llvm.memset.p0.i64(ptr align 16 [[ARRAY]], i8 0, i64 104, i1 false)
+; CHECK-NEXT:    [[ARRAY_IDX:%.*]] = getelementptr inbounds i8, ptr [[ARRAY]], i64 4
+; CHECK-NEXT:    store i32 1, ptr [[ARRAY_IDX]], align 4
+; CHECK-NEXT:    call void @llvm.memmove.p0.p0.i64(ptr align 16 [[ARRAY]], ptr align 4 [[ARRAY_IDX]], i64 100, i1 false)
+; CHECK-NEXT:    [[VAL:%.*]] = load i32, ptr [[ARRAY_IDX]], align 4
+; CHECK-NEXT:    ret i32 [[VAL]]
+;
+  %array = alloca [26 x i32], align 16
+  call void @llvm.memset.p0.i64(ptr align 16 %array, i8 0, i64 104, i1 false)
+  %array.idx = getelementptr inbounds i8, ptr %array, i64 4
+  store i32 1, ptr %array.idx
+  call void @llvm.memmove.p0.p0.i64(ptr align 16 %array, ptr align 4 %array.idx, i64 100, i1 false)
+  %val = load i32, ptr %array.idx, align 4
+  ret i32 %val
+}
+
+; Used memmove.
+define i32 @used_memmove_2() {
+; CHECK-LABEL: @used_memmove_2(
+; CHECK-NEXT:    [[ARRAY:%.*]] = alloca [26 x i32], align 16
+; CHECK-NEXT:    call void @llvm.memset.p0.i64(ptr align 16 [[ARRAY]], i8 0, i64 104, i1 false)
+; CHECK-NEXT:    [[ARRAY_IDX:%.*]] = getelementptr inbounds i8, ptr [[ARRAY]], i64 4
+; CHECK-NEXT:    store i32 1, ptr [[ARRAY]], align 4
+; CHECK-NEXT:    call void @llvm.memmove.p0.p0.i64(ptr align 16 [[ARRAY]], ptr align 4 [[ARRAY_IDX]], i64 100, i1 false)
+; CHECK-NEXT:    [[VAL:%.*]] = load i32, ptr [[ARRAY_IDX]], align 4
+; CHECK-NEXT:    ret i32 [[VAL]]
+;
+  %array = alloca [26 x i32], align 16
+  call void @llvm.memset.p0.i64(ptr align 16 %array, i8 0, i64 104, i1 false)
+  %array.idx = getelementptr inbounds i8, ptr %array, i64 4
+  store i32 1, ptr %array
+  call void @llvm.memmove.p0.p0.i64(ptr align 16 %array, ptr align 4 %array.idx, i64 100, i1 false)
+  %val = load i32, ptr %array.idx, align 4
+  ret i32 %val
+}
+
+; Used memmove, buffer clobbered by opaque.
+define i32 @used_memmove_3() {
+; CHECK-LABEL: @used_memmove_3(
+; CHECK-NEXT:    [[ARRAY:%.*]] = alloca [25 x i32], align 16
+; CHECK-NEXT:    call void @llvm.memset.p0.i64(ptr align 16 [[ARRAY]], i8 0, i64 100, i1 false)
+; CHECK-NEXT:    call void @opaque(ptr [[ARRAY]])
+; CHECK-NEXT:    [[ARRAY_IDX:%.*]] = getelementptr inbounds i8, ptr [[ARRAY]], i64 4
+; CHECK-NEXT:    call void @llvm.memmove.p0.p0.i64(ptr align 16 [[ARRAY]], ptr align 4 [[ARRAY_IDX]], i64 96, i1 false)
+; CHECK-NEXT:    [[VAL:%.*]] = load i32, ptr [[ARRAY]], align 16
+; CHECK-NEXT:    ret i32 [[VAL]]
+;
+  %array = alloca [25 x i32], align 16
+  call void @llvm.memset.p0.i64(ptr align 16 %array, i8 0, i64 100, i1 false)
+  call void @opaque(ptr %array)
+  %array.idx = getelementptr inbounds i8, ptr %array, i64 4
+  call void @llvm.memmove.p0.p0.i64(ptr align 16 %array, ptr align 4 %array.idx, i64 96, i1 false)
+  %val = load i32, ptr %array, align 16
+  ret i32 %val
+}
+
+; Redundant memmove, not within the same basic block.
+define i32 @redundant_memmove_different_bbs() {
+; CHECK-LABEL: @redundant_memmove_different_bbs(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[ARRAY:%.*]] = alloca [26 x i32], align 16
+; CHECK-NEXT:    call void @llvm.memset.p0.i64(ptr align 16 [[ARRAY]], i8 0, i64 104, i1 false)
+; CHECK-NEXT:    [[ARRAY_IDX:%.*]] = getelementptr inbounds i8, ptr [[ARRAY]], i64 4
+; CHECK-NEXT:    br label [[USE:%.*]]
+; CHECK:       use:
+; CHECK-NEXT:    call void @llvm.memmove.p0.p0.i64(ptr align 16 [[ARRAY]], ptr align 4 [[ARRAY_IDX]], i64 100, i1 false)
+; CHECK-NEXT:    [[VAL:%.*]] = load i32, ptr [[ARRAY]], align 16
+; CHECK-NEXT:    ret i32 [[VAL]]
+;
+entry:
+  %array = alloca [26 x i32], align 16
+  call void @llvm.memset.p0.i64(ptr align 16 %array, i8 0, i64 104, i1 false)
+  %array.idx = getelementptr inbounds i8, ptr %array, i64 4
+  br label %use
+
+use:                                              ; preds = %entry
+  call void @llvm.memmove.p0.p0.i64(ptr align 16 %array, ptr align 4 %array.idx, i64 100, i1 false)
+  %val = load i32, ptr %array, align 16
+  ret i32 %val
+}
+
+declare void @opaque(ptr)
+declare void @llvm.memset.p0.i64(ptr nocapture, i8, i64, i1)
+declare void @llvm.memmove.p0.p0.i64(ptr nocapture, ptr nocapture, i64, i1) nounwind

>From 26e1089fbc9144385fb8a94d8dc0098565800ba2 Mon Sep 17 00:00:00 2001
From: Antonio Frighetto <me at antoniofrighetto.com>
Date: Fri, 20 Sep 2024 17:42:17 +0200
Subject: [PATCH 2/3] [MemCpyOpt] Drop dead `memmove` calls on `memset`'d
 source data

When a memmove happens to clobber source data, and such data have
been previously memset'd, the memmove may be redundant.
---
 .../llvm/Transforms/Scalar/MemCpyOptimizer.h  |  3 +-
 .../lib/Transforms/Scalar/MemCpyOptimizer.cpp | 72 ++++++++++++++++++-
 .../memset-memmove-redundant-memmove.ll       |  2 -
 3 files changed, 71 insertions(+), 6 deletions(-)

diff --git a/llvm/include/llvm/Transforms/Scalar/MemCpyOptimizer.h b/llvm/include/llvm/Transforms/Scalar/MemCpyOptimizer.h
index 8ed03d7f3ddbff..63e6fa8e805eca 100644
--- a/llvm/include/llvm/Transforms/Scalar/MemCpyOptimizer.h
+++ b/llvm/include/llvm/Transforms/Scalar/MemCpyOptimizer.h
@@ -66,7 +66,7 @@ class MemCpyOptPass : public PassInfoMixin<MemCpyOptPass> {
                           BasicBlock::iterator &BBI);
   bool processMemSet(MemSetInst *SI, BasicBlock::iterator &BBI);
   bool processMemCpy(MemCpyInst *M, BasicBlock::iterator &BBI);
-  bool processMemMove(MemMoveInst *M);
+  bool processMemMove(MemMoveInst *M, BasicBlock::iterator &BBI);
   bool performCallSlotOptzn(Instruction *cpyLoad, Instruction *cpyStore,
                             Value *cpyDst, Value *cpySrc, TypeSize cpyLen,
                             Align cpyAlign, BatchAAResults &BAA,
@@ -85,6 +85,7 @@ class MemCpyOptPass : public PassInfoMixin<MemCpyOptPass> {
   bool performStackMoveOptzn(Instruction *Load, Instruction *Store,
                              AllocaInst *DestAlloca, AllocaInst *SrcAlloca,
                              TypeSize Size, BatchAAResults &BAA);
+  bool isMemMoveMemSetDependency(MemMoveInst *M);
 
   void eraseInstruction(Instruction *I);
   bool iterateOnFunction(Function &F);
diff --git a/llvm/lib/Transforms/Scalar/MemCpyOptimizer.cpp b/llvm/lib/Transforms/Scalar/MemCpyOptimizer.cpp
index 1d67773585d593..fe5f7c2284c7d0 100644
--- a/llvm/lib/Transforms/Scalar/MemCpyOptimizer.cpp
+++ b/llvm/lib/Transforms/Scalar/MemCpyOptimizer.cpp
@@ -69,6 +69,7 @@ static cl::opt<bool> EnableMemCpyOptWithoutLibcalls(
     cl::desc("Enable memcpyopt even when libcalls are disabled"));
 
 STATISTIC(NumMemCpyInstr, "Number of memcpy instructions deleted");
+STATISTIC(NumMemMoveInstr, "Number of memmove instructions deleted");
 STATISTIC(NumMemSetInfer, "Number of memsets inferred");
 STATISTIC(NumMoveToCpy, "Number of memmoves converted to memcpy");
 STATISTIC(NumCpyToSet, "Number of memcpys converted to memset");
@@ -1843,12 +1844,77 @@ bool MemCpyOptPass::processMemCpy(MemCpyInst *M, BasicBlock::iterator &BBI) {
   return false;
 }
 
+/// Memmove calls with overlapping src/dest buffers that come after a memset may
+/// be removed.
+bool MemCpyOptPass::isMemMoveMemSetDependency(MemMoveInst *M) {
+  const auto &DL = M->getDataLayout();
+  MemSetInst *MS = nullptr;
+  MemoryUseOrDef *MemMoveAccess = MSSA->getMemoryAccess(M);
+  if (!MemMoveAccess)
+    return false;
+
+  BatchAAResults BAA(*AA);
+  MemoryAccess *FirstDef = MemMoveAccess->getDefiningAccess();
+  MemoryLocation SourceLoc = MemoryLocation::getForSource(M);
+  MemoryLocation DestLoc = MemoryLocation::getForDest(M);
+
+  // The first dominating clobbering MemoryAccess for the source location
+  // needs to be the memset.
+  MemoryAccess *SourceClobber =
+      MSSA->getWalker()->getClobberingMemoryAccess(FirstDef, SourceLoc, BAA);
+  if (auto *Def = dyn_cast<MemoryDef>(SourceClobber))
+    MS = dyn_cast_or_null<MemSetInst>(Def->getMemoryInst());
+  if (!MS)
+    return false;
+
+  // The destination buffer must have been memset'd.
+  if (!BAA.isMustAlias(MS->getDest(), M->getDest()))
+    return false;
+
+  // The memmove is of form memmove(x, x + A, B).
+  auto *Source = dyn_cast<GetElementPtrInst>(M->getSource());
+  if (!Source)
+    return false;
+  APInt Offset(DL.getIndexTypeSizeInBits(Source->getType()), 0);
+  auto MemMoveSize = MemoryLocation::getForSource(M).Size;
+  if (!Source->accumulateConstantOffset(DL, Offset) || Offset.isNegative() ||
+      Source->getPointerOperand() != M->getDest() || !MemMoveSize.hasValue())
+    return false;
+
+  LocationSize TotalSize =
+      LocationSize::precise(Offset.getZExtValue() + MemMoveSize.getValue());
+  MemoryLocation CombinedSourceLoc(M->getSource(), TotalSize);
+  MemoryLocation CombinedDestLoc(M->getDest(), TotalSize);
+  if (!isModOrRefSet(BAA.getModRefInfo(MS, CombinedSourceLoc)) ||
+      !isModOrRefSet(BAA.getModRefInfo(MS, CombinedDestLoc)))
+    return false;
+
+  // The first dominating clobbering MemoryAccess for the destination location
+  // needs to be the memset as well.
+  MemoryAccess *DestClobber =
+      MSSA->getWalker()->getClobberingMemoryAccess(FirstDef, DestLoc, BAA);
+  auto *Def = dyn_cast<MemoryDef>(DestClobber);
+  if (Def->getMemoryInst() != MS)
+    return false;
+  return true;
+}
+
 /// Transforms memmove calls to memcpy calls when the src/dst are guaranteed
 /// not to alias.
-bool MemCpyOptPass::processMemMove(MemMoveInst *M) {
+bool MemCpyOptPass::processMemMove(MemMoveInst *M, BasicBlock::iterator &BBI) {
   // See if the source could be modified by this memmove potentially.
-  if (isModSet(AA->getModRefInfo(M, MemoryLocation::getForSource(M))))
+  if (isModSet(AA->getModRefInfo(M, MemoryLocation::getForSource(M)))) {
+    // On the off-chance the memmove clobbers src with previously memset'd
+    // bytes, the memmove may be redundant.
+    if (!M->isVolatile() && isMemMoveMemSetDependency(M)) {
+      LLVM_DEBUG(dbgs() << "Removed redundant memmove.\n");
+      ++BBI;
+      eraseInstruction(M);
+      ++NumMemMoveInstr;
+      return true;
+    }
     return false;
+  }
 
   LLVM_DEBUG(dbgs() << "MemCpyOptPass: Optimizing memmove -> memcpy: " << *M
                     << "\n");
@@ -2066,7 +2132,7 @@ bool MemCpyOptPass::iterateOnFunction(Function &F) {
       else if (auto *M = dyn_cast<MemCpyInst>(I))
         RepeatInstruction = processMemCpy(M, BI);
       else if (auto *M = dyn_cast<MemMoveInst>(I))
-        RepeatInstruction = processMemMove(M);
+        RepeatInstruction = processMemMove(M, BI);
       else if (auto *CB = dyn_cast<CallBase>(I)) {
         for (unsigned i = 0, e = CB->arg_size(); i != e; ++i) {
           if (CB->isByValArgument(i))
diff --git a/llvm/test/Transforms/MemCpyOpt/memset-memmove-redundant-memmove.ll b/llvm/test/Transforms/MemCpyOpt/memset-memmove-redundant-memmove.ll
index 5b323f119f0950..5fcd8c80606a67 100644
--- a/llvm/test/Transforms/MemCpyOpt/memset-memmove-redundant-memmove.ll
+++ b/llvm/test/Transforms/MemCpyOpt/memset-memmove-redundant-memmove.ll
@@ -7,7 +7,6 @@ define i32 @redundant_memmove() {
 ; CHECK-NEXT:    [[ARRAY:%.*]] = alloca [26 x i32], align 16
 ; CHECK-NEXT:    call void @llvm.memset.p0.i64(ptr align 16 [[ARRAY]], i8 0, i64 104, i1 false)
 ; CHECK-NEXT:    [[ARRAY_IDX:%.*]] = getelementptr inbounds i8, ptr [[ARRAY]], i64 4
-; CHECK-NEXT:    call void @llvm.memmove.p0.p0.i64(ptr align 16 [[ARRAY]], ptr align 4 [[ARRAY_IDX]], i64 100, i1 false)
 ; CHECK-NEXT:    [[VAL:%.*]] = load i32, ptr [[ARRAY]], align 16
 ; CHECK-NEXT:    ret i32 [[VAL]]
 ;
@@ -88,7 +87,6 @@ define i32 @redundant_memmove_different_bbs() {
 ; CHECK-NEXT:    [[ARRAY_IDX:%.*]] = getelementptr inbounds i8, ptr [[ARRAY]], i64 4
 ; CHECK-NEXT:    br label [[USE:%.*]]
 ; CHECK:       use:
-; CHECK-NEXT:    call void @llvm.memmove.p0.p0.i64(ptr align 16 [[ARRAY]], ptr align 4 [[ARRAY_IDX]], i64 100, i1 false)
 ; CHECK-NEXT:    [[VAL:%.*]] = load i32, ptr [[ARRAY]], align 16
 ; CHECK-NEXT:    ret i32 [[VAL]]
 ;

>From 693081c95e31222586f96ca369267e750b11dcd9 Mon Sep 17 00:00:00 2001
From: Antonio Frighetto <me at antoniofrighetto.com>
Date: Wed, 27 Nov 2024 19:39:31 +0100
Subject: [PATCH 3/3] !fixup single mssa query over the combined dest location

---
 .../lib/Transforms/Scalar/MemCpyOptimizer.cpp | 47 ++++++++-----------
 1 file changed, 19 insertions(+), 28 deletions(-)

diff --git a/llvm/lib/Transforms/Scalar/MemCpyOptimizer.cpp b/llvm/lib/Transforms/Scalar/MemCpyOptimizer.cpp
index fe5f7c2284c7d0..2a7bce4512bffc 100644
--- a/llvm/lib/Transforms/Scalar/MemCpyOptimizer.cpp
+++ b/llvm/lib/Transforms/Scalar/MemCpyOptimizer.cpp
@@ -1848,35 +1848,17 @@ bool MemCpyOptPass::processMemCpy(MemCpyInst *M, BasicBlock::iterator &BBI) {
 /// be removed.
 bool MemCpyOptPass::isMemMoveMemSetDependency(MemMoveInst *M) {
   const auto &DL = M->getDataLayout();
-  MemSetInst *MS = nullptr;
   MemoryUseOrDef *MemMoveAccess = MSSA->getMemoryAccess(M);
   if (!MemMoveAccess)
     return false;
 
-  BatchAAResults BAA(*AA);
-  MemoryAccess *FirstDef = MemMoveAccess->getDefiningAccess();
-  MemoryLocation SourceLoc = MemoryLocation::getForSource(M);
-  MemoryLocation DestLoc = MemoryLocation::getForDest(M);
-
-  // The first dominating clobbering MemoryAccess for the source location
-  // needs to be the memset.
-  MemoryAccess *SourceClobber =
-      MSSA->getWalker()->getClobberingMemoryAccess(FirstDef, SourceLoc, BAA);
-  if (auto *Def = dyn_cast<MemoryDef>(SourceClobber))
-    MS = dyn_cast_or_null<MemSetInst>(Def->getMemoryInst());
-  if (!MS)
-    return false;
-
-  // The destination buffer must have been memset'd.
-  if (!BAA.isMustAlias(MS->getDest(), M->getDest()))
-    return false;
-
   // The memmove is of form memmove(x, x + A, B).
+  MemoryLocation SourceLoc = MemoryLocation::getForSource(M);
   auto *Source = dyn_cast<GetElementPtrInst>(M->getSource());
   if (!Source)
     return false;
   APInt Offset(DL.getIndexTypeSizeInBits(Source->getType()), 0);
-  auto MemMoveSize = MemoryLocation::getForSource(M).Size;
+  auto MemMoveSize = SourceLoc.Size;
   if (!Source->accumulateConstantOffset(DL, Offset) || Offset.isNegative() ||
       Source->getPointerOperand() != M->getDest() || !MemMoveSize.hasValue())
     return false;
@@ -1885,16 +1867,25 @@ bool MemCpyOptPass::isMemMoveMemSetDependency(MemMoveInst *M) {
       LocationSize::precise(Offset.getZExtValue() + MemMoveSize.getValue());
   MemoryLocation CombinedSourceLoc(M->getSource(), TotalSize);
   MemoryLocation CombinedDestLoc(M->getDest(), TotalSize);
-  if (!isModOrRefSet(BAA.getModRefInfo(MS, CombinedSourceLoc)) ||
-      !isModOrRefSet(BAA.getModRefInfo(MS, CombinedDestLoc)))
+
+  // The first dominating clobbering MemoryAccess for the combined location
+  // needs to be a memset.
+  BatchAAResults BAA(*AA);
+  MemSetInst *MS = nullptr;
+  MemoryAccess *FirstDef = MemMoveAccess->getDefiningAccess();
+  MemoryAccess *DestClobber = MSSA->getWalker()->getClobberingMemoryAccess(
+      FirstDef, CombinedDestLoc, BAA);
+  if (auto *Def = dyn_cast<MemoryDef>(DestClobber))
+    MS = dyn_cast_or_null<MemSetInst>(Def->getMemoryInst());
+  if (!MS)
     return false;
 
-  // The first dominating clobbering MemoryAccess for the destination location
-  // needs to be the memset as well.
-  MemoryAccess *DestClobber =
-      MSSA->getWalker()->getClobberingMemoryAccess(FirstDef, DestLoc, BAA);
-  auto *Def = dyn_cast<MemoryDef>(DestClobber);
-  if (Def->getMemoryInst() != MS)
+  // The destination buffer must have been memset'd.
+  if (!BAA.isMustAlias(MS->getDest(), M->getDest()))
+    return false;
+
+  if (!isModOrRefSet(BAA.getModRefInfo(MS, CombinedSourceLoc)) ||
+      !isModOrRefSet(BAA.getModRefInfo(MS, CombinedDestLoc)))
     return false;
   return true;
 }



More information about the llvm-commits mailing list