don't eliminate load from volatile value on paths where the load is dead.

This fixes the second half of PR2262


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@50430 91177308-0d34-0410-b5e6-96231b3b80d8

2 files changed, 34 insertions, 0 deletions

diff --git a/lib/Transforms/Scalar/InstructionCombining.cpp b/lib/Transforms/Scalar/InstructionCombining.cpp
index 86aca073ab..82c67d03d9 100644
--- a/lib/Transforms/Scalar/InstructionCombining.cpp
+++ b/lib/Transforms/Scalar/InstructionCombining.cpp
@@ -9396,6 +9396,15 @@ Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
           LI->getParent() != PN.getIncomingBlock(i) ||
           !isSafeToSinkLoad(LI))
         return 0;
+      
+      // If the PHI is volatile and its block has multiple successors, sinking
+      // it would remove a load of the volatile value from the path through the
+      // other successor.
+      if (isVolatile &&
+          LI->getParent()->getTerminator()->getNumSuccessors() != 1)
+        return 0;
+
+      
     } else if (I->getOperand(1) != ConstantOp) {
       return 0;
     }
diff --git a/test/Transforms/InstCombine/2008-04-29-VolatileLoadDontMerge.ll b/test/Transforms/InstCombine/2008-04-29-VolatileLoadDontMerge.ll
new file mode 100644
index 0000000000..e43c22daa4
--- /dev/null
+++ b/test/Transforms/InstCombine/2008-04-29-VolatileLoadDontMerge.ll
@@ -0,0 +1,25 @@
+; RUN: llvm-as < %s | opt -instcombine | llvm-dis | grep {volatile load} | count 2
+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"
+@g_1 = internal global i32 0		; <i32*> [#uses=3]
+
+define i32 @main() nounwind  {
+entry:
+	%tmp93 = icmp slt i32 0, 10		; <i1> [#uses=0]
+	%tmp34 = volatile load i32* @g_1, align 4		; <i32> [#uses=1]
+	br label %bb
+
+bb:		; preds = %bb, %entry
+	%b.0.reg2mem.0 = phi i32 [ 0, %entry ], [ %tmp6, %bb ]		; <i32> [#uses=1]
+	%tmp3.reg2mem.0 = phi i32 [ %tmp34, %entry ], [ %tmp3, %bb ]		; <i32> [#uses=1]
+	%tmp4 = add i32 %tmp3.reg2mem.0, 5		; <i32> [#uses=1]
+	volatile store i32 %tmp4, i32* @g_1, align 4
+	%tmp6 = add i32 %b.0.reg2mem.0, 1		; <i32> [#uses=2]
+	%tmp9 = icmp slt i32 %tmp6, 10		; <i1> [#uses=1]
+	%tmp3 = volatile load i32* @g_1, align 4		; <i32> [#uses=1]
+	br i1 %tmp9, label %bb, label %bb11
+
+bb11:		; preds = %bb
+	ret i32 0
+}
+