Don't eliminate memcpy's when the address of the pointer may itself b…

…e relevant. Fixes PR18304. Patch by David Wiberg! llvm-svn: 212970
llvm · Jul 14, 2014 · 703e488 · 703e488
1 parent 0e767b1
commit 703e488
Show file tree

Hide file tree

Showing 7 changed files with 35 additions and 7 deletions.
diff --git a/llvm/lib/Transforms/Scalar/MemCpyOptimizer.cpp b/llvm/lib/Transforms/Scalar/MemCpyOptimizer.cpp
@@ -684,6 +684,12 @@ bool MemCpyOpt::performCallSlotOptzn(Instruction *cpy,
     }
   }
 
+  // Check that src isn't captured by the called function since the
+  // transformation can cause aliasing issues in that case.
+  for (unsigned i = 0, e = CS.arg_size(); i != e; ++i)
+    if (CS.getArgument(i) == cpySrc && !CS.doesNotCapture(i))
+      return false;
+
   // Since we're changing the parameter to the callsite, we need to make sure
   // that what would be the new parameter dominates the callsite.
   DominatorTree &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();

diff --git a/llvm/test/Transforms/MemCpyOpt/2008-02-24-MultipleUseofSRet.ll b/llvm/test/Transforms/MemCpyOpt/2008-02-24-MultipleUseofSRet.ll
@@ -6,7 +6,7 @@ target triple = "i386-pc-linux-gnu"
 
 %0 = type { x86_fp80, x86_fp80 }
 
-define internal fastcc void @initialize(%0* noalias sret %agg.result) nounwind {
+define internal fastcc void @initialize(%0* noalias nocapture sret %agg.result) nounwind {
 entry:
   %agg.result.03 = getelementptr %0* %agg.result, i32 0, i32 0
   store x86_fp80 0xK00000000000000000000, x86_fp80* %agg.result.03
@@ -15,7 +15,7 @@ entry:
   ret void
 }
 
-declare fastcc x86_fp80 @passed_uninitialized(%0*) nounwind
+declare fastcc x86_fp80 @passed_uninitialized(%0* nocapture) nounwind
 
 define fastcc void @badly_optimized() nounwind {
 entry:

diff --git a/llvm/test/Transforms/MemCpyOpt/2008-03-13-ReturnSlotBitcast.ll b/llvm/test/Transforms/MemCpyOpt/2008-03-13-ReturnSlotBitcast.ll
@@ -4,7 +4,7 @@ target datalayout = "E-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:32:64-f3
 %a = type { i32 }
 %b = type { float }
 
-declare void @g(%a*)
+declare void @g(%a* nocapture)
 
 define float @f() {
 entry:

diff --git a/llvm/test/Transforms/MemCpyOpt/capturing-func.ll b/llvm/test/Transforms/MemCpyOpt/capturing-func.ll
@@ -0,0 +1,22 @@
+; RUN: opt < %s -basicaa -memcpyopt -S | FileCheck %s
+
+target datalayout = "e"
+
+declare void @foo(i8*)
+declare void @llvm.memcpy.p0i8.p0i8.i32(i8* nocapture, i8* nocapture, i32, i32, i1) nounwind
+
+define void @test() {
+  %ptr1 = alloca i8
+  %ptr2 = alloca i8
+  call void @foo(i8* %ptr2)
+  call void @llvm.memcpy.p0i8.p0i8.i32(i8* %ptr1, i8* %ptr2, i32 1, i32 1, i1 false)
+  call void @foo(i8* %ptr1)
+  ret void
+
+  ; Check that the transformation isn't applied if the called function can
+  ; capture the pointer argument (i.e. the nocapture attribute isn't present)
+  ; CHECK-LABEL: @test(
+  ; CHECK: call void @foo(i8* %ptr2)
+  ; CHECK-NEXT: call void @llvm.memcpy
+  ; CHECK-NEXT: call void @foo(i8* %ptr1)
+}
diff --git a/llvm/test/Transforms/MemCpyOpt/loadstore-sret.ll b/llvm/test/Transforms/MemCpyOpt/loadstore-sret.ll
@@ -22,4 +22,4 @@ _ZNSt8auto_ptrIiED1Ev.exit:
   ret void
 }
 
-declare void @_Z3barv(%"class.std::auto_ptr"* sret)
+declare void @_Z3barv(%"class.std::auto_ptr"* nocapture sret)
diff --git a/llvm/test/Transforms/MemCpyOpt/memcpy.ll b/llvm/test/Transforms/MemCpyOpt/memcpy.ll
@@ -29,7 +29,7 @@ entry:
 ; CHECK: ret void
 }
 
-declare void @ccoshl(%0* sret , x86_fp80, x86_fp80) nounwind 
+declare void @ccoshl(%0* nocapture sret, x86_fp80, x86_fp80) nounwind 
 
 
 ; The intermediate alloca and one of the memcpy's should be eliminated, the
@@ -202,7 +202,7 @@ define void @test10(%opaque* noalias nocapture sret %x, i32 %y) {
   ret void
 }
 
-declare void @f1(%struct.big* sret)
+declare void @f1(%struct.big* nocapture sret)
 declare void @f2(%struct.big*)
 
 ; CHECK: attributes [[NUW]] = { nounwind }

diff --git a/llvm/test/Transforms/MemCpyOpt/sret.ll b/llvm/test/Transforms/MemCpyOpt/sret.ll
@@ -25,6 +25,6 @@ entry:
   ret void
 }
 
-declare void @ccoshl(%0* noalias sret, %0* byval) nounwind
+declare void @ccoshl(%0* noalias nocapture sret, %0* byval) nounwind
 
 declare void @llvm.memcpy.p0i8.p0i8.i32(i8* nocapture, i8* nocapture, i32, i32, i1) nounwind