PR4737: Fix a nasty bug in load narrowing with non-power-of-two types.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@79415 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
index 3c4f9de1cd9873ae74a2c77002f4f714d5a46dc0..4138e98f18e7d61abd471835c9cd74d740b2970d 100644 (file)
--- a/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -3466,7 +3466,8 @@ SDValue DAGCombiner::ReduceLoadWidth(SDNode *N) {
       // Is the shift amount a multiple of size of VT?
       if ((ShAmt & (EVTBits-1)) == 0) {
         N0 = N0.getOperand(0);
-        if (N0.getValueType().getSizeInBits() <= EVTBits)
+        // Is the load width a multiple of size of VT?
+        if ((N0.getValueType().getSizeInBits() & (EVTBits-1)) != 0)
           return SDValue();
       }
     }

diff --git a/test/CodeGen/X86/2009-08-19-LoadNarrowingMiscompile.ll b/test/CodeGen/X86/2009-08-19-LoadNarrowingMiscompile.ll
new file mode 100644 (file)
index 0000000..447b064
--- /dev/null
+++ b/test/CodeGen/X86/2009-08-19-LoadNarrowingMiscompile.ll
@@ -0,0 +1,15 @@
+; RUN: llvm-as < %s | llc -march=x86 | FileCheck %s
+
+@a = external global i96, align 4
+@b = external global i64, align 8
+
+define void @c() nounwind {
+; CHECK: movl a+8, %eax
+  %srcval1 = load i96* @a, align 4
+  %sroa.store.elt2 = lshr i96 %srcval1, 64
+  %tmp = trunc i96 %sroa.store.elt2 to i64
+; CHECK: movl %eax, b
+; CHECK: movl $0, b+4
+  store i64 %tmp, i64* @b, align 8
+  ret void
+}