<div dir="ltr"><div class="gmail_quote">Hey David,</div><div class="gmail_quote"><br></div><div class="gmail_quote">Thanks a lot for creating a test case and taking care of the revert. I'll look into what's going wrong here...<br></div><div class="gmail_quote"><br></div><div class="gmail_quote">Nikita<br></div><div class="gmail_quote"><div dir="ltr"><br></div><div dir="ltr">On Thu, Dec 13, 2018 at 4:04 AM David Jones <<a href="mailto:dlj@google.com">dlj@google.com</a>> wrote:<br></div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex"><div dir="ltr"><div dir="ltr"><div dir="ltr"><div dir="ltr">Here is a test case that demonstrates the incorrectly-truncated memset:<div><br></div><div>=====</div><div><div>%"struct.T1" = type { i32, i32, i8, i8, i8, i32, i32, double, double, %"class.T3", %"class.T3" }</div><div>%"class.T2" = type { i8* }</div><div>%"class.T3" = type { i8* }</div><div><br></div><div>; Function Attrs: inlinehint nounwind uwtable</div><div>define linkonce_odr dso_local void @f(%"class.T2"* noalias sret) comdat {</div><div> %2 = alloca %"struct.T1", align 8</div><div> %3 = bitcast %"struct.T1"* %2 to i8*</div><div><br></div><div> %4 = getelementptr inbounds %"struct.T1", %"struct.T1"* %2, i64 0, i32 5</div><div> store i32 0, i32* %4, align 4</div><div> %5 = getelementptr inbounds %"struct.T1", %"struct.T1"* %2, i64 0, i32 6</div><div> store i32 0, i32* %5, align 8</div><div><br></div><div> %6 = getelementptr inbounds %"struct.T1", %"struct.T1"* %2, i64 0, i32 7</div><div> %7 = bitcast double* %6 to i8*</div><div> call void @llvm.memset.p0i8.i64(i8* nonnull align 8 %7, i8 0, i64 64, i1 false) #3</div><div><br></div><div> %8 = bitcast %"struct.T1"* %2 to i8*</div><div> call void @llvm.memset.p0i8.i64(i8* nonnull align 8 %8, i8 0, i64 11, i1 false)</div><div><br></div><div> %9 = tail call i8* @_Znwm(i64 88) #19</div><div> call void @llvm.memcpy.p0i8.p0i8.i64(i8* nonnull align 8 %9, i8* nonnull align 8 %3, i64 40, i1 false) #3</div><div><br></div><div> %10 = getelementptr inbounds i8, i8* %9, i64 40</div><div> %11 = bitcast i8* %10 to %"class.T3"*</div><div> %12 = getelementptr inbounds %"struct.T1", %"struct.T1"* %2, i64 0, i32 9</div><div> call void @ext_f1(%"class.T3"* nonnull %11, %"class.T3"* nonnull dereferenceable(24) %12) #3</div><div><br></div><div> %13 = ptrtoint i8* %9 to i64</div><div> %14 = bitcast %"class.T2"* %0 to i64*</div><div> store i64 %13, i64* %14, align 8</div><div><br></div><div> ret void</div><div>}</div><div><br></div><div>declare void @llvm.memcpy.p0i8.p0i8.i64(i8* nocapture writeonly, i8* nocapture readonly, i64, i1)</div><div>declare void @llvm.memset.p0i8.i64(i8* nocapture writeonly, i8, i64, i1)</div><div>declare noalias nonnull i8* @_Znwm(i64)</div><div>declare void @ext_f1(%"class.T3"*, %"class.T3"* dereferenceable(24)) unnamed_addr</div><div><br></div><div>$f = comdat any</div><div><br></div><div>@keep = dso_local unnamed_addr constant { i8* } { i8* bitcast (void (%"class.T2"*)* @f to i8*) }, align 8</div></div><div>=====</div><div><br></div><div>Here is the problematic output:</div><div><br></div><div>=====</div><div><div>$ diff -u --label before <( $OPT_r348577 /tmp/memset-struct.ll -O2 -f | ./llvm-dis -f - ) --label after <( $OPT_r348661 /tmp/memset-struct.ll -O2 -f | ./llvm-dis -f - ) </div><div>--- before</div><div>+++ after</div><div>@@ -21,7 +21,7 @@</div><div> call void @llvm.memset.p0i8.i64(i8* nonnull align 8 %7, i8 0, i64 64, i1 false)</div><div> call void @llvm.memset.p0i8.i64(i8* nonnull align 8 %3, i8 0, i64 11, i1 false)</div><div> %8 = tail call i8* @_Znwm(i64 88)</div><div>- call void @llvm.memcpy.p0i8.p0i8.i64(i8* nonnull align 8 %8, i8* nonnull align 8 %3, i64 40, i1 false)</div><div>+ call void @llvm.memset.p0i8.i64(i8* nonnull align 8 %8, i8 0, i64 11, i1 false)</div><div> %9 = getelementptr inbounds i8, i8* %8, i64 40</div><div> %10 = bitcast i8* %9 to %class.T3*</div><div> %11 = getelementptr inbounds %struct.T1, %struct.T1* %2, i64 0, i32 9</div><div>@@ -33,9 +33,6 @@</div><div> }</div><div> </div><div> ; Function Attrs: argmemonly nounwind</div><div>-declare void @llvm.memcpy.p0i8.p0i8.i64(i8* nocapture writeonly, i8* nocapture readonly, i64, i1) #0</div><div>-</div><div>-; Function Attrs: argmemonly nounwind</div><div> declare void @llvm.memset.p0i8.i64(i8* nocapture writeonly, i8, i64, i1) #0</div><div> </div><div> declare noalias nonnull i8* @_Znwm(i64) local_unnamed_addr</div><div>=====<br></div></div><div><br></div><div>Since I haven't heard anything back since my last message, I will go ahead and revert this patch. (Sorry about that... :-/ )</div></div></div></div><br><div class="gmail_quote"><div dir="ltr">On Wed, Dec 12, 2018 at 4:21 PM David Jones <<a href="mailto:dlj@google.com" class="gmail-m_-6877082536712387039cremed" target="_blank">dlj@google.com</a>> wrote:<br></div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex"><div dir="ltr">Hello,<div><br></div><div>I'm seeing miscompiles at -O2 that bisect to this revision. I'm working on reducing a test case now... it looks to my (admittedly unfamiliar) eyes like this revision may not correctly handle structs correctly if they are alloca'ed.</div><div><br></div><div>I'll update when I have more (or a test case).</div><div><br></div><div>--dlj</div></div><br><div class="gmail_quote"><div dir="ltr">On Fri, Dec 7, 2018 at 1:19 PM Nikita Popov via llvm-commits <<a href="mailto:llvm-commits@lists.llvm.org" class="gmail-m_-6877082536712387039cremed" target="_blank">llvm-commits@lists.llvm.org</a>> wrote:<br></div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex">Author: nikic<br>
Date: Fri Dec 7 13:16:58 2018<br>
New Revision: 348645<br>
<br>
URL: <a href="http://llvm.org/viewvc/llvm-project?rev=348645&view=rev" rel="noreferrer" class="gmail-m_-6877082536712387039cremed" target="_blank">http://llvm.org/viewvc/llvm-project?rev=348645&view=rev</a><br>
Log:<br>
[MemCpyOpt] memset->memcpy forwarding with undef tail<br>
<br>
Currently memcpyopt optimizes cases like<br>
<br>
memset(a, byte, N);<br>
memcpy(b, a, M);<br>
<br>
to<br>
<br>
memset(a, byte, N);<br>
memset(b, byte, M);<br>
<br>
if M <= N. Often this allows further simplifications down the line,<br>
which drop the first memset entirely.<br>
<br>
This patch extends this optimization for the case where M > N, but we<br>
know that the bytes a[N..M] are undef due to alloca/lifetime.start.<br>
<br>
This situation arises relatively often for Rust code, because Rust does<br>
not initialize trailing structure padding and loves to insert redundant<br>
memcpys. This also fixes <a href="https://bugs.llvm.org/show_bug.cgi?id=39844" rel="noreferrer" class="gmail-m_-6877082536712387039cremed" target="_blank">https://bugs.llvm.org/show_bug.cgi?id=39844</a>.<br>
<br>
For the implementation, I'm reusing a bit of code for a similar existing<br>
optimization (direct memcpy of undef). I've also added memset support to<br>
MemDepAnalysis GetLocation -- Instead, getPointerDependencyFrom could be<br>
used, but it seems to make more sense to add this to GetLocation and thus<br>
make the computation cachable.<br>
<br>
Differential Revision: <a href="https://reviews.llvm.org/D55120" rel="noreferrer" class="gmail-m_-6877082536712387039cremed" target="_blank">https://reviews.llvm.org/D55120</a><br>
<br>
Modified:<br>
llvm/trunk/lib/Analysis/MemoryDependenceAnalysis.cpp<br>
llvm/trunk/lib/Transforms/Scalar/MemCpyOptimizer.cpp<br>
llvm/trunk/test/Transforms/MemCpyOpt/memset-memcpy-oversized.ll<br>
<br>
Modified: llvm/trunk/lib/Analysis/MemoryDependenceAnalysis.cpp<br>
URL: <a href="http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Analysis/MemoryDependenceAnalysis.cpp?rev=348645&r1=348644&r2=348645&view=diff" rel="noreferrer" class="gmail-m_-6877082536712387039cremed" target="_blank">http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Analysis/MemoryDependenceAnalysis.cpp?rev=348645&r1=348644&r2=348645&view=diff</a><br>
==============================================================================<br>
--- llvm/trunk/lib/Analysis/MemoryDependenceAnalysis.cpp (original)<br>
+++ llvm/trunk/lib/Analysis/MemoryDependenceAnalysis.cpp Fri Dec 7 13:16:58 2018<br>
@@ -154,6 +154,12 @@ static ModRefInfo GetLocation(const Inst<br>
return ModRefInfo::Mod;<br>
}<br>
<br>
+ if (const MemSetInst *MI = dyn_cast<MemSetInst>(Inst)) {<br>
+ Loc = MemoryLocation::getForDest(MI);<br>
+ // Conversatively assume ModRef for volatile memset.<br>
+ return MI->isVolatile() ? ModRefInfo::ModRef : ModRefInfo::Mod;<br>
+ }<br>
+<br>
if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {<br>
switch (II->getIntrinsicID()) {<br>
case Intrinsic::lifetime_start:<br>
<br>
Modified: llvm/trunk/lib/Transforms/Scalar/MemCpyOptimizer.cpp<br>
URL: <a href="http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Transforms/Scalar/MemCpyOptimizer.cpp?rev=348645&r1=348644&r2=348645&view=diff" rel="noreferrer" class="gmail-m_-6877082536712387039cremed" target="_blank">http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Transforms/Scalar/MemCpyOptimizer.cpp?rev=348645&r1=348644&r2=348645&view=diff</a><br>
==============================================================================<br>
--- llvm/trunk/lib/Transforms/Scalar/MemCpyOptimizer.cpp (original)<br>
+++ llvm/trunk/lib/Transforms/Scalar/MemCpyOptimizer.cpp Fri Dec 7 13:16:58 2018<br>
@@ -1144,6 +1144,21 @@ bool MemCpyOptPass::processMemSetMemCpyD<br>
return true;<br>
}<br>
<br>
+/// Determine whether the instruction has undefined content for the given Size,<br>
+/// either because it was freshly alloca'd or started its lifetime.<br>
+static bool hasUndefContents(Instruction *I, ConstantInt *Size) {<br>
+ if (isa<AllocaInst>(I))<br>
+ return true;<br>
+<br>
+ if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))<br>
+ if (II->getIntrinsicID() == Intrinsic::lifetime_start)<br>
+ if (ConstantInt *LTSize = dyn_cast<ConstantInt>(II->getArgOperand(0)))<br>
+ if (LTSize->getZExtValue() >= Size->getZExtValue())<br>
+ return true;<br>
+<br>
+ return false;<br>
+}<br>
+<br>
/// Transform memcpy to memset when its source was just memset.<br>
/// In other words, turn:<br>
/// \code<br>
@@ -1167,12 +1182,23 @@ bool MemCpyOptPass::performMemCpyToMemSe<br>
if (!AA.isMustAlias(MemSet->getRawDest(), MemCpy->getRawSource()))<br>
return false;<br>
<br>
- ConstantInt *CopySize = cast<ConstantInt>(MemCpy->getLength());<br>
+ // A known memset size is required.<br>
ConstantInt *MemSetSize = dyn_cast<ConstantInt>(MemSet->getLength());<br>
+ if (!MemSetSize)<br>
+ return false;<br>
+<br>
// Make sure the memcpy doesn't read any more than what the memset wrote.<br>
// Don't worry about sizes larger than i64.<br>
- if (!MemSetSize || CopySize->getZExtValue() > MemSetSize->getZExtValue())<br>
- return false;<br>
+ ConstantInt *CopySize = cast<ConstantInt>(MemCpy->getLength());<br>
+ if (CopySize->getZExtValue() > MemSetSize->getZExtValue()) {<br>
+ // If the memcpy is larger than the memset, but the memory was undef prior<br>
+ // to the memset, we can just ignore the tail.<br>
+ MemDepResult DepInfo = MD->getDependency(MemSet);<br>
+ if (DepInfo.isDef() && hasUndefContents(DepInfo.getInst(), CopySize))<br>
+ CopySize = MemSetSize;<br>
+ else<br>
+ return false;<br>
+ }<br>
<br>
IRBuilder<> Builder(MemCpy);<br>
Builder.CreateMemSet(MemCpy->getRawDest(), MemSet->getOperand(1),<br>
@@ -1252,19 +1278,7 @@ bool MemCpyOptPass::processMemCpy(MemCpy<br>
if (MemCpyInst *MDep = dyn_cast<MemCpyInst>(SrcDepInfo.getInst()))<br>
return processMemCpyMemCpyDependence(M, MDep);<br>
} else if (SrcDepInfo.isDef()) {<br>
- Instruction *I = SrcDepInfo.getInst();<br>
- bool hasUndefContents = false;<br>
-<br>
- if (isa<AllocaInst>(I)) {<br>
- hasUndefContents = true;<br>
- } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {<br>
- if (II->getIntrinsicID() == Intrinsic::lifetime_start)<br>
- if (ConstantInt *LTSize = dyn_cast<ConstantInt>(II->getArgOperand(0)))<br>
- if (LTSize->getZExtValue() >= CopySize->getZExtValue())<br>
- hasUndefContents = true;<br>
- }<br>
-<br>
- if (hasUndefContents) {<br>
+ if (hasUndefContents(SrcDepInfo.getInst(), CopySize)) {<br>
MD->removeInstruction(M);<br>
M->eraseFromParent();<br>
++NumMemCpyInstr;<br>
<br>
Modified: llvm/trunk/test/Transforms/MemCpyOpt/memset-memcpy-oversized.ll<br>
URL: <a href="http://llvm.org/viewvc/llvm-project/llvm/trunk/test/Transforms/MemCpyOpt/memset-memcpy-oversized.ll?rev=348645&r1=348644&r2=348645&view=diff" rel="noreferrer" class="gmail-m_-6877082536712387039cremed" target="_blank">http://llvm.org/viewvc/llvm-project/llvm/trunk/test/Transforms/MemCpyOpt/memset-memcpy-oversized.ll?rev=348645&r1=348644&r2=348645&view=diff</a><br>
==============================================================================<br>
--- llvm/trunk/test/Transforms/MemCpyOpt/memset-memcpy-oversized.ll (original)<br>
+++ llvm/trunk/test/Transforms/MemCpyOpt/memset-memcpy-oversized.ll Fri Dec 7 13:16:58 2018<br>
@@ -12,7 +12,7 @@ define void @test_alloca(i8* %result) {<br>
; CHECK-NEXT: [[A:%.*]] = alloca [[T:%.*]], align 8<br>
; CHECK-NEXT: [[B:%.*]] = bitcast %T* [[A]] to i8*<br>
; CHECK-NEXT: call void @llvm.memset.p0i8.i64(i8* align 8 [[B]], i8 0, i64 12, i1 false)<br>
-; CHECK-NEXT: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[RESULT:%.*]], i8* align 8 [[B]], i64 16, i1 false)<br>
+; CHECK-NEXT: call void @llvm.memset.p0i8.i64(i8* [[RESULT:%.*]], i8 0, i64 12, i1 false)<br>
; CHECK-NEXT: ret void<br>
;<br>
%a = alloca %T, align 8<br>
@@ -28,7 +28,7 @@ define void @test_alloca_with_lifetimes(<br>
; CHECK-NEXT: [[B:%.*]] = bitcast %T* [[A]] to i8*<br>
; CHECK-NEXT: call void @llvm.lifetime.start.p0i8(i64 16, i8* [[B]])<br>
; CHECK-NEXT: call void @llvm.memset.p0i8.i64(i8* align 8 [[B]], i8 0, i64 12, i1 false)<br>
-; CHECK-NEXT: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[RESULT:%.*]], i8* align 8 [[B]], i64 16, i1 false)<br>
+; CHECK-NEXT: call void @llvm.memset.p0i8.i64(i8* [[RESULT:%.*]], i8 0, i64 12, i1 false)<br>
; CHECK-NEXT: call void @llvm.lifetime.end.p0i8(i64 16, i8* [[B]])<br>
; CHECK-NEXT: ret void<br>
;<br>
@@ -46,7 +46,7 @@ define void @test_malloc_with_lifetimes(<br>
; CHECK-NEXT: [[A:%.*]] = call i8* @malloc(i64 16)<br>
; CHECK-NEXT: call void @llvm.lifetime.start.p0i8(i64 16, i8* [[A]])<br>
; CHECK-NEXT: call void @llvm.memset.p0i8.i64(i8* align 8 [[A]], i8 0, i64 12, i1 false)<br>
-; CHECK-NEXT: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[RESULT:%.*]], i8* align 8 [[A]], i64 16, i1 false)<br>
+; CHECK-NEXT: call void @llvm.memset.p0i8.i64(i8* [[RESULT:%.*]], i8 0, i64 12, i1 false)<br>
; CHECK-NEXT: call void @llvm.lifetime.end.p0i8(i64 16, i8* [[A]])<br>
; CHECK-NEXT: call void @free(i8* [[A]])<br>
; CHECK-NEXT: ret void<br>
@@ -98,7 +98,7 @@ define void @test_volatile_memset(i8* %r<br>
; CHECK-NEXT: [[A:%.*]] = alloca [[T:%.*]], align 8<br>
; CHECK-NEXT: [[B:%.*]] = bitcast %T* [[A]] to i8*<br>
; CHECK-NEXT: call void @llvm.memset.p0i8.i64(i8* align 8 [[B]], i8 0, i64 12, i1 true)<br>
-; CHECK-NEXT: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[RESULT:%.*]], i8* align 8 [[B]], i64 16, i1 false)<br>
+; CHECK-NEXT: call void @llvm.memset.p0i8.i64(i8* [[RESULT:%.*]], i8 0, i64 12, i1 false)<br>
; CHECK-NEXT: ret void<br>
;<br>
%a = alloca %T, align 8<br>
<br>
<br>
_______________________________________________<br>
llvm-commits mailing list<br>
<a href="mailto:llvm-commits@lists.llvm.org" class="gmail-m_-6877082536712387039cremed" target="_blank">llvm-commits@lists.llvm.org</a><br>
<a href="http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-commits" rel="noreferrer" class="gmail-m_-6877082536712387039cremed" target="_blank">http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-commits</a><br>
</blockquote></div>
</blockquote></div></div><br><div class="gmail_quote"><div dir="ltr">On Wed, Dec 12, 2018 at 4:21 PM David Jones <<a href="mailto:dlj@google.com" target="_blank">dlj@google.com</a>> wrote:<br></div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex"><div dir="ltr">Hello,<div><br></div><div>I'm seeing miscompiles at -O2 that bisect to this revision. I'm working on reducing a test case now... it looks to my (admittedly unfamiliar) eyes like this revision may not correctly handle structs correctly if they are alloca'ed.</div><div><br></div><div>I'll update when I have more (or a test case).</div><div><br></div><div>--dlj</div></div><br><div class="gmail_quote"><div dir="ltr">On Fri, Dec 7, 2018 at 1:19 PM Nikita Popov via llvm-commits <<a href="mailto:llvm-commits@lists.llvm.org" target="_blank">llvm-commits@lists.llvm.org</a>> wrote:<br></div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex">Author: nikic<br>
Date: Fri Dec 7 13:16:58 2018<br>
New Revision: 348645<br>
<br>
URL: <a href="http://llvm.org/viewvc/llvm-project?rev=348645&view=rev" rel="noreferrer" target="_blank">http://llvm.org/viewvc/llvm-project?rev=348645&view=rev</a><br>
Log:<br>
[MemCpyOpt] memset->memcpy forwarding with undef tail<br>
<br>
Currently memcpyopt optimizes cases like<br>
<br>
memset(a, byte, N);<br>
memcpy(b, a, M);<br>
<br>
to<br>
<br>
memset(a, byte, N);<br>
memset(b, byte, M);<br>
<br>
if M <= N. Often this allows further simplifications down the line,<br>
which drop the first memset entirely.<br>
<br>
This patch extends this optimization for the case where M > N, but we<br>
know that the bytes a[N..M] are undef due to alloca/lifetime.start.<br>
<br>
This situation arises relatively often for Rust code, because Rust does<br>
not initialize trailing structure padding and loves to insert redundant<br>
memcpys. This also fixes <a href="https://bugs.llvm.org/show_bug.cgi?id=39844" rel="noreferrer" target="_blank">https://bugs.llvm.org/show_bug.cgi?id=39844</a>.<br>
<br>
For the implementation, I'm reusing a bit of code for a similar existing<br>
optimization (direct memcpy of undef). I've also added memset support to<br>
MemDepAnalysis GetLocation -- Instead, getPointerDependencyFrom could be<br>
used, but it seems to make more sense to add this to GetLocation and thus<br>
make the computation cachable.<br>
<br>
Differential Revision: <a href="https://reviews.llvm.org/D55120" rel="noreferrer" target="_blank">https://reviews.llvm.org/D55120</a><br>
<br>
Modified:<br>
llvm/trunk/lib/Analysis/MemoryDependenceAnalysis.cpp<br>
llvm/trunk/lib/Transforms/Scalar/MemCpyOptimizer.cpp<br>
llvm/trunk/test/Transforms/MemCpyOpt/memset-memcpy-oversized.ll<br>
<br>
Modified: llvm/trunk/lib/Analysis/MemoryDependenceAnalysis.cpp<br>
URL: <a href="http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Analysis/MemoryDependenceAnalysis.cpp?rev=348645&r1=348644&r2=348645&view=diff" rel="noreferrer" target="_blank">http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Analysis/MemoryDependenceAnalysis.cpp?rev=348645&r1=348644&r2=348645&view=diff</a><br>
==============================================================================<br>
--- llvm/trunk/lib/Analysis/MemoryDependenceAnalysis.cpp (original)<br>
+++ llvm/trunk/lib/Analysis/MemoryDependenceAnalysis.cpp Fri Dec 7 13:16:58 2018<br>
@@ -154,6 +154,12 @@ static ModRefInfo GetLocation(const Inst<br>
return ModRefInfo::Mod;<br>
}<br>
<br>
+ if (const MemSetInst *MI = dyn_cast<MemSetInst>(Inst)) {<br>
+ Loc = MemoryLocation::getForDest(MI);<br>
+ // Conversatively assume ModRef for volatile memset.<br>
+ return MI->isVolatile() ? ModRefInfo::ModRef : ModRefInfo::Mod;<br>
+ }<br>
+<br>
if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {<br>
switch (II->getIntrinsicID()) {<br>
case Intrinsic::lifetime_start:<br>
<br>
Modified: llvm/trunk/lib/Transforms/Scalar/MemCpyOptimizer.cpp<br>
URL: <a href="http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Transforms/Scalar/MemCpyOptimizer.cpp?rev=348645&r1=348644&r2=348645&view=diff" rel="noreferrer" target="_blank">http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Transforms/Scalar/MemCpyOptimizer.cpp?rev=348645&r1=348644&r2=348645&view=diff</a><br>
==============================================================================<br>
--- llvm/trunk/lib/Transforms/Scalar/MemCpyOptimizer.cpp (original)<br>
+++ llvm/trunk/lib/Transforms/Scalar/MemCpyOptimizer.cpp Fri Dec 7 13:16:58 2018<br>
@@ -1144,6 +1144,21 @@ bool MemCpyOptPass::processMemSetMemCpyD<br>
return true;<br>
}<br>
<br>
+/// Determine whether the instruction has undefined content for the given Size,<br>
+/// either because it was freshly alloca'd or started its lifetime.<br>
+static bool hasUndefContents(Instruction *I, ConstantInt *Size) {<br>
+ if (isa<AllocaInst>(I))<br>
+ return true;<br>
+<br>
+ if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))<br>
+ if (II->getIntrinsicID() == Intrinsic::lifetime_start)<br>
+ if (ConstantInt *LTSize = dyn_cast<ConstantInt>(II->getArgOperand(0)))<br>
+ if (LTSize->getZExtValue() >= Size->getZExtValue())<br>
+ return true;<br>
+<br>
+ return false;<br>
+}<br>
+<br>
/// Transform memcpy to memset when its source was just memset.<br>
/// In other words, turn:<br>
/// \code<br>
@@ -1167,12 +1182,23 @@ bool MemCpyOptPass::performMemCpyToMemSe<br>
if (!AA.isMustAlias(MemSet->getRawDest(), MemCpy->getRawSource()))<br>
return false;<br>
<br>
- ConstantInt *CopySize = cast<ConstantInt>(MemCpy->getLength());<br>
+ // A known memset size is required.<br>
ConstantInt *MemSetSize = dyn_cast<ConstantInt>(MemSet->getLength());<br>
+ if (!MemSetSize)<br>
+ return false;<br>
+<br>
// Make sure the memcpy doesn't read any more than what the memset wrote.<br>
// Don't worry about sizes larger than i64.<br>
- if (!MemSetSize || CopySize->getZExtValue() > MemSetSize->getZExtValue())<br>
- return false;<br>
+ ConstantInt *CopySize = cast<ConstantInt>(MemCpy->getLength());<br>
+ if (CopySize->getZExtValue() > MemSetSize->getZExtValue()) {<br>
+ // If the memcpy is larger than the memset, but the memory was undef prior<br>
+ // to the memset, we can just ignore the tail.<br>
+ MemDepResult DepInfo = MD->getDependency(MemSet);<br>
+ if (DepInfo.isDef() && hasUndefContents(DepInfo.getInst(), CopySize))<br>
+ CopySize = MemSetSize;<br>
+ else<br>
+ return false;<br>
+ }<br>
<br>
IRBuilder<> Builder(MemCpy);<br>
Builder.CreateMemSet(MemCpy->getRawDest(), MemSet->getOperand(1),<br>
@@ -1252,19 +1278,7 @@ bool MemCpyOptPass::processMemCpy(MemCpy<br>
if (MemCpyInst *MDep = dyn_cast<MemCpyInst>(SrcDepInfo.getInst()))<br>
return processMemCpyMemCpyDependence(M, MDep);<br>
} else if (SrcDepInfo.isDef()) {<br>
- Instruction *I = SrcDepInfo.getInst();<br>
- bool hasUndefContents = false;<br>
-<br>
- if (isa<AllocaInst>(I)) {<br>
- hasUndefContents = true;<br>
- } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {<br>
- if (II->getIntrinsicID() == Intrinsic::lifetime_start)<br>
- if (ConstantInt *LTSize = dyn_cast<ConstantInt>(II->getArgOperand(0)))<br>
- if (LTSize->getZExtValue() >= CopySize->getZExtValue())<br>
- hasUndefContents = true;<br>
- }<br>
-<br>
- if (hasUndefContents) {<br>
+ if (hasUndefContents(SrcDepInfo.getInst(), CopySize)) {<br>
MD->removeInstruction(M);<br>
M->eraseFromParent();<br>
++NumMemCpyInstr;<br>
<br>
Modified: llvm/trunk/test/Transforms/MemCpyOpt/memset-memcpy-oversized.ll<br>
URL: <a href="http://llvm.org/viewvc/llvm-project/llvm/trunk/test/Transforms/MemCpyOpt/memset-memcpy-oversized.ll?rev=348645&r1=348644&r2=348645&view=diff" rel="noreferrer" target="_blank">http://llvm.org/viewvc/llvm-project/llvm/trunk/test/Transforms/MemCpyOpt/memset-memcpy-oversized.ll?rev=348645&r1=348644&r2=348645&view=diff</a><br>
==============================================================================<br>
--- llvm/trunk/test/Transforms/MemCpyOpt/memset-memcpy-oversized.ll (original)<br>
+++ llvm/trunk/test/Transforms/MemCpyOpt/memset-memcpy-oversized.ll Fri Dec 7 13:16:58 2018<br>
@@ -12,7 +12,7 @@ define void @test_alloca(i8* %result) {<br>
; CHECK-NEXT: [[A:%.*]] = alloca [[T:%.*]], align 8<br>
; CHECK-NEXT: [[B:%.*]] = bitcast %T* [[A]] to i8*<br>
; CHECK-NEXT: call void @llvm.memset.p0i8.i64(i8* align 8 [[B]], i8 0, i64 12, i1 false)<br>
-; CHECK-NEXT: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[RESULT:%.*]], i8* align 8 [[B]], i64 16, i1 false)<br>
+; CHECK-NEXT: call void @llvm.memset.p0i8.i64(i8* [[RESULT:%.*]], i8 0, i64 12, i1 false)<br>
; CHECK-NEXT: ret void<br>
;<br>
%a = alloca %T, align 8<br>
@@ -28,7 +28,7 @@ define void @test_alloca_with_lifetimes(<br>
; CHECK-NEXT: [[B:%.*]] = bitcast %T* [[A]] to i8*<br>
; CHECK-NEXT: call void @llvm.lifetime.start.p0i8(i64 16, i8* [[B]])<br>
; CHECK-NEXT: call void @llvm.memset.p0i8.i64(i8* align 8 [[B]], i8 0, i64 12, i1 false)<br>
-; CHECK-NEXT: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[RESULT:%.*]], i8* align 8 [[B]], i64 16, i1 false)<br>
+; CHECK-NEXT: call void @llvm.memset.p0i8.i64(i8* [[RESULT:%.*]], i8 0, i64 12, i1 false)<br>
; CHECK-NEXT: call void @llvm.lifetime.end.p0i8(i64 16, i8* [[B]])<br>
; CHECK-NEXT: ret void<br>
;<br>
@@ -46,7 +46,7 @@ define void @test_malloc_with_lifetimes(<br>
; CHECK-NEXT: [[A:%.*]] = call i8* @malloc(i64 16)<br>
; CHECK-NEXT: call void @llvm.lifetime.start.p0i8(i64 16, i8* [[A]])<br>
; CHECK-NEXT: call void @llvm.memset.p0i8.i64(i8* align 8 [[A]], i8 0, i64 12, i1 false)<br>
-; CHECK-NEXT: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[RESULT:%.*]], i8* align 8 [[A]], i64 16, i1 false)<br>
+; CHECK-NEXT: call void @llvm.memset.p0i8.i64(i8* [[RESULT:%.*]], i8 0, i64 12, i1 false)<br>
; CHECK-NEXT: call void @llvm.lifetime.end.p0i8(i64 16, i8* [[A]])<br>
; CHECK-NEXT: call void @free(i8* [[A]])<br>
; CHECK-NEXT: ret void<br>
@@ -98,7 +98,7 @@ define void @test_volatile_memset(i8* %r<br>
; CHECK-NEXT: [[A:%.*]] = alloca [[T:%.*]], align 8<br>
; CHECK-NEXT: [[B:%.*]] = bitcast %T* [[A]] to i8*<br>
; CHECK-NEXT: call void @llvm.memset.p0i8.i64(i8* align 8 [[B]], i8 0, i64 12, i1 true)<br>
-; CHECK-NEXT: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[RESULT:%.*]], i8* align 8 [[B]], i64 16, i1 false)<br>
+; CHECK-NEXT: call void @llvm.memset.p0i8.i64(i8* [[RESULT:%.*]], i8 0, i64 12, i1 false)<br>
; CHECK-NEXT: ret void<br>
;<br>
%a = alloca %T, align 8<br>
<br>
<br>
_______________________________________________<br>
llvm-commits mailing list<br>
<a href="mailto:llvm-commits@lists.llvm.org" target="_blank">llvm-commits@lists.llvm.org</a><br>
<a href="http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-commits" rel="noreferrer" target="_blank">http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-commits</a><br>
</blockquote></div>
</blockquote></div>
</blockquote></div></div>