More than just loads can read from memory: readonly calls like strlen

also need to be checked for memory modifying instructions before we
can sink them.  THis fixes the second half of PR2297.


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50860 91177308-0d34-0410-b5e6-96231b3b80d8

2 files changed, 32 insertions, 4 deletions

diff --git a/lib/Transforms/Scalar/InstructionCombining.cpp b/lib/Transforms/Scalar/InstructionCombining.cpp
index b0ea9f5..772ef63 100644
--- a/lib/Transforms/Scalar/InstructionCombining.cpp
+++ b/lib/Transforms/Scalar/InstructionCombining.cpp
@@ -11229,8 +11229,8 @@ static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
 
   // We can only sink load instructions if there is nothing between the load and
   // the end of block that could change the value.
-  if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
-    for (BasicBlock::iterator Scan = LI, E = LI->getParent()->end();
+  if (I->mayReadFromMemory()) {
+    for (BasicBlock::iterator Scan = I, E = I->getParent()->end();
          Scan != E; ++Scan)
       if (Scan->mayWriteToMemory())
         return false;
@@ -11388,8 +11388,8 @@ bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) {
     }
 
     // See if we can trivially sink this instruction to a successor basic block.
-    // FIXME: Remove GetResultInst test when first class support for aggregates is
-    // implemented.
+    // FIXME: Remove GetResultInst test when first class support for aggregates
+    // is implemented.
     if (I->hasOneUse() && !isa<GetResultInst>(I)) {
       BasicBlock *BB = I->getParent();
       BasicBlock *UserParent = cast<Instruction>(I->use_back())->getParent();
diff --git a/test/Transforms/InstCombine/2008-05-08-StrLenSink.ll b/test/Transforms/InstCombine/2008-05-08-StrLenSink.ll
new file mode 100644
index 0000000..e81e828
--- /dev/null
+++ b/test/Transforms/InstCombine/2008-05-08-StrLenSink.ll
@@ -0,0 +1,28 @@
+; RUN: llvm-as < %s | opt -instcombine | llvm-dis | %prcontext strlen 1 | not grep ret
+; PR2297
+target datalayout = "e-p:32:32:32-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:32:64-f32:32:32-f64:32:64-v64:64:64-v128:128:128-a0:0:64-f80:128:128"
+target triple = "i386-apple-darwin8"
+
+define i32 @a() nounwind  {
+entry:
+	%"alloca point" = bitcast i32 0 to i32		; <i32> [#uses=0]
+	%tmp1 = call i8* @malloc( i32 10 ) nounwind 		; <i8*> [#uses=5]
+	%tmp3 = getelementptr i8* %tmp1, i32 1		; <i8*> [#uses=1]
+	store i8 0, i8* %tmp3, align 1
+	%tmp5 = getelementptr i8* %tmp1, i32 0		; <i8*> [#uses=1]
+	store i8 1, i8* %tmp5, align 1
+	%tmp7 = call i32 @strlen( i8* %tmp1 ) nounwind readonly 		; <i32> [#uses=1]
+	%tmp9 = getelementptr i8* %tmp1, i32 0		; <i8*> [#uses=1]
+	store i8 0, i8* %tmp9, align 1
+	%tmp11 = call i32 (...)* @b( i8* %tmp1 ) nounwind 		; <i32> [#uses=0]
+	br label %return
+
+return:		; preds = %entry
+	ret i32 %tmp7
+}
+
+declare i8* @malloc(i32) nounwind 
+
+declare i32 @strlen(i8*) nounwind readonly 
+
+declare i32 @b(...)