R600: Better fix for bug 20982

Just do the left shift as unsigned to avoid the UB.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@218092 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/R600/AMDGPUISelLowering.cpp b/lib/Target/R600/AMDGPUISelLowering.cpp
index 89dab098710558c2c8adaeaa90a201d8ee532801..bf8daf96604b018b5622f3800323d49f75360817 100644 (file)
--- a/lib/Target/R600/AMDGPUISelLowering.cpp
+++ b/lib/Target/R600/AMDGPUISelLowering.cpp
@@ -1896,7 +1896,8 @@ template <typename IntTy>
 static SDValue constantFoldBFE(SelectionDAG &DAG, IntTy Src0,
                                uint32_t Offset, uint32_t Width) {
   if (Width + Offset < 32) {
-    IntTy Result = (Src0 << (32 - Offset - Width)) >> (32 - Width);
+    uint32_t Shl = static_cast<uint32_t>(Src0) << (32 - Offset - Width);
+    IntTy Result = static_cast<IntTy>(Shl) >> (32 - Width);
     return DAG.getConstant(Result, MVT::i32);
   }
 
@@ -2053,12 +2054,8 @@ SDValue AMDGPUTargetLowering::PerformDAGCombine(SDNode *N,
         // Avoid undefined left shift of a negative in the constant fold.
         // TODO: I'm not sure what the behavior of the hardware is, this should
         // probably follow that instead.
-        int32_t Val = CVal->getSExtValue();
-        if (Val < 0)
-          return SDValue();
-
         return constantFoldBFE<int32_t>(DAG,
-                                        Val,
+                                        CVal->getSExtValue(),
                                         OffsetVal,
                                         WidthVal);
       }

diff --git a/test/CodeGen/R600/llvm.AMDGPU.bfe.i32.ll b/test/CodeGen/R600/llvm.AMDGPU.bfe.i32.ll
index 16a140cefa3fc29d6a8b02d0e1b13784d0706afe..eb5094232825e95d60aea935b3bbfc80458e4d7a 100644 (file)
--- a/test/CodeGen/R600/llvm.AMDGPU.bfe.i32.ll
+++ b/test/CodeGen/R600/llvm.AMDGPU.bfe.i32.ll
@@ -370,13 +370,12 @@ define void @bfe_i32_constant_fold_test_15(i32 addrspace(1)* %out) nounwind {
   ret void
 }
 
-; FIXME: This should fold to something
 ; FUNC-LABEL: @bfe_i32_constant_fold_test_16
-; SI: S_BFE_I32 [[SREG:s[0-9]+]], -1, 0x70001
-; SI: V_MOV_B32_e32 [[VREG:v[0-9]+]], [[SREG]]
+; SI-NOT: BFE
+; SI: V_MOV_B32_e32 [[VREG:v[0-9]+]], -1
 ; SI: BUFFER_STORE_DWORD [[VREG]],
 ; SI: S_ENDPGM
-
+; EG-NOT: BFE
 define void @bfe_i32_constant_fold_test_16(i32 addrspace(1)* %out) nounwind {
   %bfe_i32 = call i32 @llvm.AMDGPU.bfe.i32(i32 4294967295, i32 1, i32 7) nounwind readnone
   store i32 %bfe_i32, i32 addrspace(1)* %out, align 4