From: Matt Arsenault <Matthew.Arsenault@amd.com>
Date: Thu, 22 May 2014 18:09:07 +0000 (+0000)
Subject: R600: Add dag combine for BFE
X-Git-Url: http://review.tizen.org/git/?a=commitdiff_plain;h=5565f65e1398ff7633ba7e52d80f6b02066b8f33;p=platform%2Fupstream%2Fllvm.git

R600: Add dag combine for BFE

llvm-svn: 209461
---

diff --git a/llvm/lib/Target/R600/AMDGPUISelLowering.cpp b/llvm/lib/Target/R600/AMDGPUISelLowering.cpp
index 78e79bc..a4cb4f5 100644
--- a/llvm/lib/Target/R600/AMDGPUISelLowering.cpp
+++ b/llvm/lib/Target/R600/AMDGPUISelLowering.cpp
@@ -1292,6 +1292,17 @@ static void simplifyI24(SDValue Op, TargetLowering::DAGCombinerInfo &DCI) {
     DCI.CommitTargetLoweringOpt(TLO);
 }
 
+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);
+    return DAG.getConstant(Result, MVT::i32);
+  }
+
+  return DAG.getConstant(Src0 >> Offset, MVT::i32);
+}
+
 SDValue AMDGPUTargetLowering::PerformDAGCombine(SDNode *N,
                                             DAGCombinerInfo &DCI) const {
   SelectionDAG &DAG = DCI.DAG;
@@ -1338,6 +1349,64 @@ SDValue AMDGPUTargetLowering::PerformDAGCombine(SDNode *N,
     case ISD::SELECT_CC: {
       return CombineMinMax(N, DAG);
     }
+  case AMDGPUISD::BFE_I32:
+  case AMDGPUISD::BFE_U32: {
+    assert(!N->getValueType(0).isVector() &&
+           "Vector handling of BFE not implemented");
+    ConstantSDNode *Width = dyn_cast<ConstantSDNode>(N->getOperand(2));
+    if (!Width)
+      break;
+
+    uint32_t WidthVal = Width->getZExtValue() & 0x1f;
+    if (WidthVal == 0)
+      return DAG.getConstant(0, MVT::i32);
+
+    ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
+    if (!Offset)
+      break;
+
+    SDValue BitsFrom = N->getOperand(0);
+    uint32_t OffsetVal = Offset->getZExtValue() & 0x1f;
+
+    bool Signed = N->getOpcode() == AMDGPUISD::BFE_I32;
+
+    if (OffsetVal == 0) {
+      // This is already sign / zero extended, so try to fold away extra BFEs.
+      unsigned SignBits =  Signed ? (32 - WidthVal + 1) : (32 - WidthVal);
+
+      unsigned OpSignBits = DAG.ComputeNumSignBits(BitsFrom);
+      if (OpSignBits >= SignBits)
+        return BitsFrom;
+    }
+
+    if (ConstantSDNode *Val = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
+      if (Signed) {
+        return constantFoldBFE<int32_t>(DAG,
+                                        Val->getSExtValue(),
+                                        OffsetVal,
+                                        WidthVal);
+      }
+
+      return constantFoldBFE<uint32_t>(DAG,
+                                       Val->getZExtValue(),
+                                       OffsetVal,
+                                       WidthVal);
+    }
+
+    APInt Demanded = APInt::getBitsSet(32,
+                                       OffsetVal,
+                                       OffsetVal + WidthVal);
+    APInt KnownZero, KnownOne;
+    TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
+                                          !DCI.isBeforeLegalizeOps());
+    const TargetLowering &TLI = DAG.getTargetLoweringInfo();
+    if (TLO.ShrinkDemandedConstant(BitsFrom, Demanded) ||
+        TLI.SimplifyDemandedBits(BitsFrom, Demanded, KnownZero, KnownOne, TLO)) {
+      DCI.CommitTargetLoweringOpt(TLO);
+    }
+
+    break;
+  }
   }
   return SDValue();
 }
@@ -1560,6 +1629,11 @@ unsigned AMDGPUTargetLowering::ComputeNumSignBitsForTargetNode(
     return std::max(SignBits, Op0SignBits);
   }
 
+  case AMDGPUISD::BFE_U32: {
+    ConstantSDNode *Width = dyn_cast<ConstantSDNode>(Op.getOperand(2));
+    return Width ? 32 - (Width->getZExtValue() & 0x1f) : 1;
+  }
+
   default:
     return 1;
   }
diff --git a/llvm/lib/Target/R600/R600ISelLowering.cpp b/llvm/lib/Target/R600/R600ISelLowering.cpp
index 489565e6..d6c6830 100644
--- a/llvm/lib/Target/R600/R600ISelLowering.cpp
+++ b/llvm/lib/Target/R600/R600ISelLowering.cpp
@@ -1762,7 +1762,8 @@ SDValue R600TargetLowering::PerformDAGCombine(SDNode *N,
         NewArgs);
   }
   }
-  return SDValue();
+
+  return AMDGPUTargetLowering::PerformDAGCombine(N, DCI);
 }
 
 static bool
diff --git a/llvm/lib/Target/R600/SIISelLowering.cpp b/llvm/lib/Target/R600/SIISelLowering.cpp
index b51c46c..c9e247c 100644
--- a/llvm/lib/Target/R600/SIISelLowering.cpp
+++ b/llvm/lib/Target/R600/SIISelLowering.cpp
@@ -1075,7 +1075,8 @@ SDValue SITargetLowering::PerformDAGCombine(SDNode *N,
       break;
     }
   }
-  return SDValue();
+
+  return AMDGPUTargetLowering::PerformDAGCombine(N, DCI);
 }
 
 /// \brief Test if RegClass is one of the VSrc classes
diff --git a/llvm/test/CodeGen/R600/llvm.AMDGPU.bfe.i32.ll b/llvm/test/CodeGen/R600/llvm.AMDGPU.bfe.i32.ll
index b3fec06..71d2b6e 100644
--- a/llvm/test/CodeGen/R600/llvm.AMDGPU.bfe.i32.ll
+++ b/llvm/test/CodeGen/R600/llvm.AMDGPU.bfe.i32.ll
@@ -48,3 +48,251 @@ define void @v_bfe_print_arg(i32 addrspace(1)* %out, i32 addrspace(1)* %src0) no
   store i32 %bfe_i32, i32 addrspace(1)* %out, align 4
   ret void
 }
+
+; FUNC-LABEL: @bfe_i32_arg_0_width_reg_offset
+; SI-NOT: BFE
+; SI: S_ENDPGM
+; EG-NOT: BFE
+define void @bfe_i32_arg_0_width_reg_offset(i32 addrspace(1)* %out, i32 %src0, i32 %src1) nounwind {
+  %bfe_u32 = call i32 @llvm.AMDGPU.bfe.i32(i32 %src0, i32 %src1, i32 0) nounwind readnone
+  store i32 %bfe_u32, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: @bfe_i32_arg_0_width_imm_offset
+; SI-NOT: BFE
+; SI: S_ENDPGM
+; EG-NOT: BFE
+define void @bfe_i32_arg_0_width_imm_offset(i32 addrspace(1)* %out, i32 %src0, i32 %src1) nounwind {
+  %bfe_u32 = call i32 @llvm.AMDGPU.bfe.i32(i32 %src0, i32 8, i32 0) nounwind readnone
+  store i32 %bfe_u32, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: @bfe_i32_constant_fold_test_0
+; SI-NOT: BFE
+; SI: V_MOV_B32_e32 [[VREG:v[0-9]+]], 0
+; SI: BUFFER_STORE_DWORD [[VREG]],
+; SI: S_ENDPGM
+; EG-NOT: BFE
+define void @bfe_i32_constant_fold_test_0(i32 addrspace(1)* %out) nounwind {
+  %bfe_i32 = call i32 @llvm.AMDGPU.bfe.i32(i32 0, i32 0, i32 0) nounwind readnone
+  store i32 %bfe_i32, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: @bfe_i32_constant_fold_test_1
+; SI-NOT: BFE
+; SI: V_MOV_B32_e32 [[VREG:v[0-9]+]], 0
+; SI: BUFFER_STORE_DWORD [[VREG]],
+; SI: S_ENDPGM
+; EG-NOT: BFE
+define void @bfe_i32_constant_fold_test_1(i32 addrspace(1)* %out) nounwind {
+  %bfe_i32 = call i32 @llvm.AMDGPU.bfe.i32(i32 12334, i32 0, i32 0) nounwind readnone
+  store i32 %bfe_i32, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: @bfe_i32_constant_fold_test_2
+; SI-NOT: BFE
+; SI: V_MOV_B32_e32 [[VREG:v[0-9]+]], 0
+; SI: BUFFER_STORE_DWORD [[VREG]],
+; SI: S_ENDPGM
+; EG-NOT: BFE
+define void @bfe_i32_constant_fold_test_2(i32 addrspace(1)* %out) nounwind {
+  %bfe_i32 = call i32 @llvm.AMDGPU.bfe.i32(i32 0, i32 0, i32 1) nounwind readnone
+  store i32 %bfe_i32, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: @bfe_i32_constant_fold_test_3
+; 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_3(i32 addrspace(1)* %out) nounwind {
+  %bfe_i32 = call i32 @llvm.AMDGPU.bfe.i32(i32 1, i32 0, i32 1) nounwind readnone
+  store i32 %bfe_i32, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: @bfe_i32_constant_fold_test_4
+; 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_4(i32 addrspace(1)* %out) nounwind {
+  %bfe_i32 = call i32 @llvm.AMDGPU.bfe.i32(i32 4294967295, i32 0, i32 1) nounwind readnone
+  store i32 %bfe_i32, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: @bfe_i32_constant_fold_test_5
+; 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_5(i32 addrspace(1)* %out) nounwind {
+  %bfe_i32 = call i32 @llvm.AMDGPU.bfe.i32(i32 128, i32 7, i32 1) nounwind readnone
+  store i32 %bfe_i32, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: @bfe_i32_constant_fold_test_6
+; SI-NOT: BFE
+; SI: V_MOV_B32_e32 [[VREG:v[0-9]+]], 0xffffff80
+; SI: BUFFER_STORE_DWORD [[VREG]],
+; SI: S_ENDPGM
+; EG-NOT: BFE
+define void @bfe_i32_constant_fold_test_6(i32 addrspace(1)* %out) nounwind {
+  %bfe_i32 = call i32 @llvm.AMDGPU.bfe.i32(i32 128, i32 0, i32 8) nounwind readnone
+  store i32 %bfe_i32, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: @bfe_i32_constant_fold_test_7
+; SI-NOT: BFE
+; SI: V_MOV_B32_e32 [[VREG:v[0-9]+]], 0x7f
+; SI: BUFFER_STORE_DWORD [[VREG]],
+; SI: S_ENDPGM
+; EG-NOT: BFE
+define void @bfe_i32_constant_fold_test_7(i32 addrspace(1)* %out) nounwind {
+  %bfe_i32 = call i32 @llvm.AMDGPU.bfe.i32(i32 127, i32 0, i32 8) nounwind readnone
+  store i32 %bfe_i32, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: @bfe_i32_constant_fold_test_8
+; 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_8(i32 addrspace(1)* %out) nounwind {
+  %bfe_i32 = call i32 @llvm.AMDGPU.bfe.i32(i32 127, i32 6, i32 8) nounwind readnone
+  store i32 %bfe_i32, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: @bfe_i32_constant_fold_test_9
+; 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_9(i32 addrspace(1)* %out) nounwind {
+  %bfe_i32 = call i32 @llvm.AMDGPU.bfe.i32(i32 65536, i32 16, i32 8) nounwind readnone
+  store i32 %bfe_i32, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: @bfe_i32_constant_fold_test_10
+; SI-NOT: BFE
+; SI: V_MOV_B32_e32 [[VREG:v[0-9]+]], 0
+; SI: BUFFER_STORE_DWORD [[VREG]],
+; SI: S_ENDPGM
+; EG-NOT: BFE
+define void @bfe_i32_constant_fold_test_10(i32 addrspace(1)* %out) nounwind {
+  %bfe_i32 = call i32 @llvm.AMDGPU.bfe.i32(i32 65535, i32 16, i32 16) nounwind readnone
+  store i32 %bfe_i32, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: @bfe_i32_constant_fold_test_11
+; SI-NOT: BFE
+; SI: V_MOV_B32_e32 [[VREG:v[0-9]+]], -6
+; SI: BUFFER_STORE_DWORD [[VREG]],
+; SI: S_ENDPGM
+; EG-NOT: BFE
+define void @bfe_i32_constant_fold_test_11(i32 addrspace(1)* %out) nounwind {
+  %bfe_i32 = call i32 @llvm.AMDGPU.bfe.i32(i32 160, i32 4, i32 4) nounwind readnone
+  store i32 %bfe_i32, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: @bfe_i32_constant_fold_test_12
+; SI-NOT: BFE
+; SI: V_MOV_B32_e32 [[VREG:v[0-9]+]], 0
+; SI: BUFFER_STORE_DWORD [[VREG]],
+; SI: S_ENDPGM
+; EG-NOT: BFE
+define void @bfe_i32_constant_fold_test_12(i32 addrspace(1)* %out) nounwind {
+  %bfe_i32 = call i32 @llvm.AMDGPU.bfe.i32(i32 160, i32 31, i32 1) nounwind readnone
+  store i32 %bfe_i32, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: @bfe_i32_constant_fold_test_13
+; 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_13(i32 addrspace(1)* %out) nounwind {
+  %bfe_i32 = call i32 @llvm.AMDGPU.bfe.i32(i32 131070, i32 16, i32 16) nounwind readnone
+  store i32 %bfe_i32, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: @bfe_i32_constant_fold_test_14
+; SI-NOT: BFE
+; SI: V_MOV_B32_e32 [[VREG:v[0-9]+]], 40
+; SI: BUFFER_STORE_DWORD [[VREG]],
+; SI: S_ENDPGM
+; EG-NOT: BFE
+define void @bfe_i32_constant_fold_test_14(i32 addrspace(1)* %out) nounwind {
+  %bfe_i32 = call i32 @llvm.AMDGPU.bfe.i32(i32 160, i32 2, i32 30) nounwind readnone
+  store i32 %bfe_i32, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: @bfe_i32_constant_fold_test_15
+; SI-NOT: BFE
+; SI: V_MOV_B32_e32 [[VREG:v[0-9]+]], 10
+; SI: BUFFER_STORE_DWORD [[VREG]],
+; SI: S_ENDPGM
+; EG-NOT: BFE
+define void @bfe_i32_constant_fold_test_15(i32 addrspace(1)* %out) nounwind {
+  %bfe_i32 = call i32 @llvm.AMDGPU.bfe.i32(i32 160, i32 4, i32 28) nounwind readnone
+  store i32 %bfe_i32, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: @bfe_i32_constant_fold_test_16
+; 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
+  ret void
+}
+
+; FUNC-LABEL: @bfe_i32_constant_fold_test_17
+; SI-NOT: BFE
+; SI: V_MOV_B32_e32 [[VREG:v[0-9]+]], 0x7f
+; SI: BUFFER_STORE_DWORD [[VREG]],
+; SI: S_ENDPGM
+; EG-NOT: BFE
+define void @bfe_i32_constant_fold_test_17(i32 addrspace(1)* %out) nounwind {
+  %bfe_i32 = call i32 @llvm.AMDGPU.bfe.i32(i32 255, i32 1, i32 31) nounwind readnone
+  store i32 %bfe_i32, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: @bfe_i32_constant_fold_test_18
+; SI-NOT: BFE
+; SI: V_MOV_B32_e32 [[VREG:v[0-9]+]], 0
+; SI: BUFFER_STORE_DWORD [[VREG]],
+; SI: S_ENDPGM
+; EG-NOT: BFE
+define void @bfe_i32_constant_fold_test_18(i32 addrspace(1)* %out) nounwind {
+  %bfe_i32 = call i32 @llvm.AMDGPU.bfe.i32(i32 255, i32 31, i32 1) nounwind readnone
+  store i32 %bfe_i32, i32 addrspace(1)* %out, align 4
+  ret void
+}
diff --git a/llvm/test/CodeGen/R600/llvm.AMDGPU.bfe.u32.ll b/llvm/test/CodeGen/R600/llvm.AMDGPU.bfe.u32.ll
index 0d47863..6ed1ad5 100644
--- a/llvm/test/CodeGen/R600/llvm.AMDGPU.bfe.u32.ll
+++ b/llvm/test/CodeGen/R600/llvm.AMDGPU.bfe.u32.ll
@@ -38,3 +38,422 @@ define void @bfe_u32_imm_arg_arg(i32 addrspace(1)* %out, i32 %src1, i32 %src2) n
   store i32 %bfe_u32, i32 addrspace(1)* %out, align 4
   ret void
 }
+
+; FUNC-LABEL: @bfe_u32_arg_0_width_reg_offset
+; SI-NOT: BFE
+; SI: S_ENDPGM
+; EG-NOT: BFE
+define void @bfe_u32_arg_0_width_reg_offset(i32 addrspace(1)* %out, i32 %src0, i32 %src1) nounwind {
+  %bfe_u32 = call i32 @llvm.AMDGPU.bfe.u32(i32 %src0, i32 %src1, i32 0) nounwind readnone
+  store i32 %bfe_u32, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: @bfe_u32_arg_0_width_imm_offset
+; SI-NOT: BFE
+; SI: S_ENDPGM
+; EG-NOT: BFE
+define void @bfe_u32_arg_0_width_imm_offset(i32 addrspace(1)* %out, i32 %src0, i32 %src1) nounwind {
+  %bfe_u32 = call i32 @llvm.AMDGPU.bfe.u32(i32 %src0, i32 8, i32 0) nounwind readnone
+  store i32 %bfe_u32, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: @bfe_u32_zextload_i8
+; SI: BUFFER_LOAD_UBYTE
+; SI-NOT: BFE
+; SI: S_ENDPGM
+define void @bfe_u32_zextload_i8(i32 addrspace(1)* %out, i8 addrspace(1)* %in) nounwind {
+  %load = load i8 addrspace(1)* %in
+  %ext = zext i8 %load to i32
+  %bfe = call i32 @llvm.AMDGPU.bfe.u32(i32 %ext, i32 0, i32 8)
+  store i32 %bfe, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: @bfe_u32_zext_in_reg_i8
+; SI: BUFFER_LOAD_DWORD
+; SI: V_ADD_I32
+; SI-NEXT: V_AND_B32_e32
+; SI-NOT: BFE
+; SI: S_ENDPGM
+define void @bfe_u32_zext_in_reg_i8(i32 addrspace(1)* %out, i32 addrspace(1)* %in) nounwind {
+  %load = load i32 addrspace(1)* %in, align 4
+  %add = add i32 %load, 1
+  %ext = and i32 %add, 255
+  %bfe = call i32 @llvm.AMDGPU.bfe.u32(i32 %ext, i32 0, i32 8)
+  store i32 %bfe, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: @bfe_u32_zext_in_reg_i16
+; SI: BUFFER_LOAD_DWORD
+; SI: V_ADD_I32
+; SI-NEXT: V_AND_B32_e32
+; SI-NOT: BFE
+; SI: S_ENDPGM
+define void @bfe_u32_zext_in_reg_i16(i32 addrspace(1)* %out, i32 addrspace(1)* %in) nounwind {
+  %load = load i32 addrspace(1)* %in, align 4
+  %add = add i32 %load, 1
+  %ext = and i32 %add, 65535
+  %bfe = call i32 @llvm.AMDGPU.bfe.u32(i32 %ext, i32 0, i32 16)
+  store i32 %bfe, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: @bfe_u32_zext_in_reg_i8_offset_1
+; SI: BUFFER_LOAD_DWORD
+; SI: V_ADD_I32
+; SI: BFE
+; SI: S_ENDPGM
+define void @bfe_u32_zext_in_reg_i8_offset_1(i32 addrspace(1)* %out, i32 addrspace(1)* %in) nounwind {
+  %load = load i32 addrspace(1)* %in, align 4
+  %add = add i32 %load, 1
+  %ext = and i32 %add, 255
+  %bfe = call i32 @llvm.AMDGPU.bfe.u32(i32 %ext, i32 1, i32 8)
+  store i32 %bfe, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: @bfe_u32_zext_in_reg_i8_offset_3
+; SI: BUFFER_LOAD_DWORD
+; SI: V_ADD_I32
+; SI-NEXT: V_AND_B32_e32 {{v[0-9]+}}, 0xf8
+; SI-NEXT: BFE
+; SI: S_ENDPGM
+define void @bfe_u32_zext_in_reg_i8_offset_3(i32 addrspace(1)* %out, i32 addrspace(1)* %in) nounwind {
+  %load = load i32 addrspace(1)* %in, align 4
+  %add = add i32 %load, 1
+  %ext = and i32 %add, 255
+  %bfe = call i32 @llvm.AMDGPU.bfe.u32(i32 %ext, i32 3, i32 8)
+  store i32 %bfe, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: @bfe_u32_zext_in_reg_i8_offset_7
+; SI: BUFFER_LOAD_DWORD
+; SI: V_ADD_I32
+; SI-NEXT: V_AND_B32_e32 {{v[0-9]+}}, 0x80
+; SI-NEXT: BFE
+; SI: S_ENDPGM
+define void @bfe_u32_zext_in_reg_i8_offset_7(i32 addrspace(1)* %out, i32 addrspace(1)* %in) nounwind {
+  %load = load i32 addrspace(1)* %in, align 4
+  %add = add i32 %load, 1
+  %ext = and i32 %add, 255
+  %bfe = call i32 @llvm.AMDGPU.bfe.u32(i32 %ext, i32 7, i32 8)
+  store i32 %bfe, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: @bfe_u32_zext_in_reg_i16_offset_8
+; SI: BUFFER_LOAD_DWORD
+; SI: V_ADD_I32
+; SI-NEXT: BFE
+; SI: S_ENDPGM
+define void @bfe_u32_zext_in_reg_i16_offset_8(i32 addrspace(1)* %out, i32 addrspace(1)* %in) nounwind {
+  %load = load i32 addrspace(1)* %in, align 4
+  %add = add i32 %load, 1
+  %ext = and i32 %add, 65535
+  %bfe = call i32 @llvm.AMDGPU.bfe.u32(i32 %ext, i32 8, i32 8)
+  store i32 %bfe, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: @bfe_u32_test_1
+; SI: BUFFER_LOAD_DWORD
+; SI: V_BFE_U32 {{v[0-9]+}}, {{v[0-9]+}}, 0, 1
+; SI: S_ENDPGM
+; EG: BFE_UINT
+define void @bfe_u32_test_1(i32 addrspace(1)* %out, i32 addrspace(1)* %in) nounwind {
+  %x = load i32 addrspace(1)* %in, align 4
+  %bfe = call i32 @llvm.AMDGPU.bfe.u32(i32 %x, i32 0, i32 1)
+  store i32 %bfe, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define void @bfe_u32_test_2(i32 addrspace(1)* %out, i32 addrspace(1)* %in) nounwind {
+  %x = load i32 addrspace(1)* %in, align 4
+  %shl = shl i32 %x, 31
+  %bfe = call i32 @llvm.AMDGPU.bfe.u32(i32 %shl, i32 0, i32 8)
+  store i32 %bfe, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define void @bfe_u32_test_3(i32 addrspace(1)* %out, i32 addrspace(1)* %in) nounwind {
+  %x = load i32 addrspace(1)* %in, align 4
+  %shl = shl i32 %x, 31
+  %bfe = call i32 @llvm.AMDGPU.bfe.u32(i32 %shl, i32 0, i32 1)
+  store i32 %bfe, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: @bfe_u32_test_4
+; SI-NOT: LSHL
+; SI-NOT: SHR
+; SI-NOT: BFE
+; SI: V_MOV_B32_e32 [[VREG:v[0-9]+]], 0
+; SI: BUFFER_STORE_DWORD [[VREG]],
+; SI: S_ENDPGM
+define void @bfe_u32_test_4(i32 addrspace(1)* %out, i32 addrspace(1)* %in) nounwind {
+  %x = load i32 addrspace(1)* %in, align 4
+  %shl = shl i32 %x, 31
+  %shr = lshr i32 %shl, 31
+  %bfe = call i32 @llvm.AMDGPU.bfe.u32(i32 %shr, i32 31, i32 1)
+  store i32 %bfe, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: @bfe_u32_test_5
+; SI: BUFFER_LOAD_DWORD
+; SI-NOT: LSHL
+; SI-NOT: SHR
+; SI: V_BFE_I32 {{v[0-9]+}}, {{v[0-9]+}}, 0, 1
+; SI: S_ENDPGM
+define void @bfe_u32_test_5(i32 addrspace(1)* %out, i32 addrspace(1)* %in) nounwind {
+  %x = load i32 addrspace(1)* %in, align 4
+  %shl = shl i32 %x, 31
+  %shr = ashr i32 %shl, 31
+  %bfe = call i32 @llvm.AMDGPU.bfe.u32(i32 %shr, i32 0, i32 1)
+  store i32 %bfe, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: @bfe_u32_test_6
+; SI: V_LSHLREV_B32_e32 v{{[0-9]+}}, 31, v{{[0-9]+}}
+; SI: V_BFE_U32 {{v[0-9]+}}, {{v[0-9]+}}, 1, 31
+; SI: S_ENDPGM
+define void @bfe_u32_test_6(i32 addrspace(1)* %out, i32 addrspace(1)* %in) nounwind {
+  %x = load i32 addrspace(1)* %in, align 4
+  %shl = shl i32 %x, 31
+  %bfe = call i32 @llvm.AMDGPU.bfe.u32(i32 %shl, i32 1, i32 31)
+  store i32 %bfe, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: @bfe_u32_test_7
+; SI: V_LSHLREV_B32_e32 v{{[0-9]+}}, 31, v{{[0-9]+}}
+; SI-NOT: BFE
+; SI: S_ENDPGM
+define void @bfe_u32_test_7(i32 addrspace(1)* %out, i32 addrspace(1)* %in) nounwind {
+  %x = load i32 addrspace(1)* %in, align 4
+  %shl = shl i32 %x, 31
+  %bfe = call i32 @llvm.AMDGPU.bfe.u32(i32 %shl, i32 0, i32 31)
+  store i32 %bfe, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: @bfe_u32_test_8
+; SI: V_LSHLREV_B32_e32 v{{[0-9]+}}, 31, v{{[0-9]+}}
+; SI: V_BFE_U32 {{v[0-9]+}}, {{v[0-9]+}}, 31, 1
+; SI: S_ENDPGM
+define void @bfe_u32_test_8(i32 addrspace(1)* %out, i32 addrspace(1)* %in) nounwind {
+  %x = load i32 addrspace(1)* %in, align 4
+  %shl = shl i32 %x, 31
+  %bfe = call i32 @llvm.AMDGPU.bfe.u32(i32 %shl, i32 31, i32 1)
+  store i32 %bfe, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: @bfe_u32_constant_fold_test_0
+; SI-NOT: BFE
+; SI: V_MOV_B32_e32 [[VREG:v[0-9]+]], 0
+; SI: BUFFER_STORE_DWORD [[VREG]],
+; SI: S_ENDPGM
+; EG-NOT: BFE
+define void @bfe_u32_constant_fold_test_0(i32 addrspace(1)* %out) nounwind {
+  %bfe_u32 = call i32 @llvm.AMDGPU.bfe.u32(i32 0, i32 0, i32 0) nounwind readnone
+  store i32 %bfe_u32, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: @bfe_u32_constant_fold_test_1
+; SI-NOT: BFE
+; SI: V_MOV_B32_e32 [[VREG:v[0-9]+]], 0
+; SI: BUFFER_STORE_DWORD [[VREG]],
+; SI: S_ENDPGM
+; EG-NOT: BFE
+define void @bfe_u32_constant_fold_test_1(i32 addrspace(1)* %out) nounwind {
+  %bfe_u32 = call i32 @llvm.AMDGPU.bfe.u32(i32 12334, i32 0, i32 0) nounwind readnone
+  store i32 %bfe_u32, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: @bfe_u32_constant_fold_test_2
+; SI-NOT: BFE
+; SI: V_MOV_B32_e32 [[VREG:v[0-9]+]], 0
+; SI: BUFFER_STORE_DWORD [[VREG]],
+; SI: S_ENDPGM
+; EG-NOT: BFE
+define void @bfe_u32_constant_fold_test_2(i32 addrspace(1)* %out) nounwind {
+  %bfe_u32 = call i32 @llvm.AMDGPU.bfe.u32(i32 0, i32 0, i32 1) nounwind readnone
+  store i32 %bfe_u32, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: @bfe_u32_constant_fold_test_3
+; 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_u32_constant_fold_test_3(i32 addrspace(1)* %out) nounwind {
+  %bfe_u32 = call i32 @llvm.AMDGPU.bfe.u32(i32 1, i32 0, i32 1) nounwind readnone
+  store i32 %bfe_u32, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: @bfe_u32_constant_fold_test_4
+; 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_u32_constant_fold_test_4(i32 addrspace(1)* %out) nounwind {
+  %bfe_u32 = call i32 @llvm.AMDGPU.bfe.u32(i32 4294967295, i32 0, i32 1) nounwind readnone
+  store i32 %bfe_u32, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: @bfe_u32_constant_fold_test_5
+; 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_u32_constant_fold_test_5(i32 addrspace(1)* %out) nounwind {
+  %bfe_u32 = call i32 @llvm.AMDGPU.bfe.u32(i32 128, i32 7, i32 1) nounwind readnone
+  store i32 %bfe_u32, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: @bfe_u32_constant_fold_test_6
+; SI-NOT: BFE
+; SI: V_MOV_B32_e32 [[VREG:v[0-9]+]], 0x80
+; SI: BUFFER_STORE_DWORD [[VREG]],
+; SI: S_ENDPGM
+; EG-NOT: BFE
+define void @bfe_u32_constant_fold_test_6(i32 addrspace(1)* %out) nounwind {
+  %bfe_u32 = call i32 @llvm.AMDGPU.bfe.u32(i32 128, i32 0, i32 8) nounwind readnone
+  store i32 %bfe_u32, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: @bfe_u32_constant_fold_test_7
+; SI-NOT: BFE
+; SI: V_MOV_B32_e32 [[VREG:v[0-9]+]], 0x7f
+; SI: BUFFER_STORE_DWORD [[VREG]],
+; SI: S_ENDPGM
+; EG-NOT: BFE
+define void @bfe_u32_constant_fold_test_7(i32 addrspace(1)* %out) nounwind {
+  %bfe_u32 = call i32 @llvm.AMDGPU.bfe.u32(i32 127, i32 0, i32 8) nounwind readnone
+  store i32 %bfe_u32, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: @bfe_u32_constant_fold_test_8
+; 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_u32_constant_fold_test_8(i32 addrspace(1)* %out) nounwind {
+  %bfe_u32 = call i32 @llvm.AMDGPU.bfe.u32(i32 127, i32 6, i32 8) nounwind readnone
+  store i32 %bfe_u32, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: @bfe_u32_constant_fold_test_9
+; SI-NOT: BFE
+; SI: V_MOV_B32_e32 [[VREG:v[0-9]+]], 1
+; SI: BUFFER_STORE_DWORD [[VREG]],
+; SI: S_ENDPGM
+; EG-NOT: BFEfppppppppppppp
+define void @bfe_u32_constant_fold_test_9(i32 addrspace(1)* %out) nounwind {
+  %bfe_u32 = call i32 @llvm.AMDGPU.bfe.u32(i32 65536, i32 16, i32 8) nounwind readnone
+  store i32 %bfe_u32, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: @bfe_u32_constant_fold_test_10
+; SI-NOT: BFE
+; SI: V_MOV_B32_e32 [[VREG:v[0-9]+]], 0
+; SI: BUFFER_STORE_DWORD [[VREG]],
+; SI: S_ENDPGM
+; EG-NOT: BFE
+define void @bfe_u32_constant_fold_test_10(i32 addrspace(1)* %out) nounwind {
+  %bfe_u32 = call i32 @llvm.AMDGPU.bfe.u32(i32 65535, i32 16, i32 16) nounwind readnone
+  store i32 %bfe_u32, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: @bfe_u32_constant_fold_test_11
+; SI-NOT: BFE
+; SI: V_MOV_B32_e32 [[VREG:v[0-9]+]], 10
+; SI: BUFFER_STORE_DWORD [[VREG]],
+; SI: S_ENDPGM
+; EG-NOT: BFE
+define void @bfe_u32_constant_fold_test_11(i32 addrspace(1)* %out) nounwind {
+  %bfe_u32 = call i32 @llvm.AMDGPU.bfe.u32(i32 160, i32 4, i32 4) nounwind readnone
+  store i32 %bfe_u32, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: @bfe_u32_constant_fold_test_12
+; SI-NOT: BFE
+; SI: V_MOV_B32_e32 [[VREG:v[0-9]+]], 0
+; SI: BUFFER_STORE_DWORD [[VREG]],
+; SI: S_ENDPGM
+; EG-NOT: BFE
+define void @bfe_u32_constant_fold_test_12(i32 addrspace(1)* %out) nounwind {
+  %bfe_u32 = call i32 @llvm.AMDGPU.bfe.u32(i32 160, i32 31, i32 1) nounwind readnone
+  store i32 %bfe_u32, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: @bfe_u32_constant_fold_test_13
+; 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_u32_constant_fold_test_13(i32 addrspace(1)* %out) nounwind {
+  %bfe_u32 = call i32 @llvm.AMDGPU.bfe.u32(i32 131070, i32 16, i32 16) nounwind readnone
+  store i32 %bfe_u32, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: @bfe_u32_constant_fold_test_14
+; SI-NOT: BFE
+; SI: V_MOV_B32_e32 [[VREG:v[0-9]+]], 40
+; SI: BUFFER_STORE_DWORD [[VREG]],
+; SI: S_ENDPGM
+; EG-NOT: BFE
+define void @bfe_u32_constant_fold_test_14(i32 addrspace(1)* %out) nounwind {
+  %bfe_u32 = call i32 @llvm.AMDGPU.bfe.u32(i32 160, i32 2, i32 30) nounwind readnone
+  store i32 %bfe_u32, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: @bfe_u32_constant_fold_test_15
+; SI-NOT: BFE
+; SI: V_MOV_B32_e32 [[VREG:v[0-9]+]], 10
+; SI: BUFFER_STORE_DWORD [[VREG]],
+; SI: S_ENDPGM
+; EG-NOT: BFE
+define void @bfe_u32_constant_fold_test_15(i32 addrspace(1)* %out) nounwind {
+  %bfe_u32 = call i32 @llvm.AMDGPU.bfe.u32(i32 160, i32 4, i32 28) nounwind readnone
+  store i32 %bfe_u32, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: @bfe_u32_constant_fold_test_16
+; SI-NOT: BFE
+; SI: V_MOV_B32_e32 [[VREG:v[0-9]+]], 0x7f
+; SI: BUFFER_STORE_DWORD [[VREG]],
+; SI: S_ENDPGM
+; EG-NOT: BFE
+define void @bfe_u32_constant_fold_test_16(i32 addrspace(1)* %out) nounwind {
+  %bfe_u32 = call i32 @llvm.AMDGPU.bfe.u32(i32 4294967295, i32 1, i32 7) nounwind readnone
+  store i32 %bfe_u32, i32 addrspace(1)* %out, align 4
+  ret void
+}
diff --git a/llvm/test/CodeGen/R600/sext-in-reg.ll b/llvm/test/CodeGen/R600/sext-in-reg.ll
index d2ab0b9..404c9b8 100644
--- a/llvm/test/CodeGen/R600/sext-in-reg.ll
+++ b/llvm/test/CodeGen/R600/sext-in-reg.ll
@@ -382,10 +382,57 @@ define void @sext_in_reg_to_illegal_type(i16 addrspace(1)* nocapture %out, i8 ad
 
 declare i32 @llvm.AMDGPU.bfe.i32(i32, i32, i32) nounwind readnone
 
+; FUNC-LABEL: @bfe_0_width
+; SI-NOT: BFE
+; SI: S_ENDPGM
+define void @bfe_0_width(i32 addrspace(1)* %out, i32 addrspace(1)* %ptr) nounwind {
+  %load = load i32 addrspace(1)* %ptr, align 4
+  %bfe = call i32 @llvm.AMDGPU.bfe.i32(i32 %load, i32 8, i32 0) nounwind readnone
+  store i32 %bfe, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: @bfe_8_bfe_8
+; SI: V_BFE_I32
+; SI-NOT: BFE
+; SI: S_ENDPGM
+define void @bfe_8_bfe_8(i32 addrspace(1)* %out, i32 addrspace(1)* %ptr) nounwind {
+  %load = load i32 addrspace(1)* %ptr, align 4
+  %bfe0 = call i32 @llvm.AMDGPU.bfe.i32(i32 %load, i32 0, i32 8) nounwind readnone
+  %bfe1 = call i32 @llvm.AMDGPU.bfe.i32(i32 %bfe0, i32 0, i32 8) nounwind readnone
+  store i32 %bfe1, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: @bfe_8_bfe_16
+; SI: V_BFE_I32 v{{[0-9]+}}, v{{[0-9]+}}, 0, 8
+; SI: S_ENDPGM
+define void @bfe_8_bfe_16(i32 addrspace(1)* %out, i32 addrspace(1)* %ptr) nounwind {
+  %load = load i32 addrspace(1)* %ptr, align 4
+  %bfe0 = call i32 @llvm.AMDGPU.bfe.i32(i32 %load, i32 0, i32 8) nounwind readnone
+  %bfe1 = call i32 @llvm.AMDGPU.bfe.i32(i32 %bfe0, i32 0, i32 16) nounwind readnone
+  store i32 %bfe1, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; This really should be folded into 1
+; FUNC-LABEL: @bfe_16_bfe_8
+; SI: V_BFE_I32 v{{[0-9]+}}, v{{[0-9]+}}, 0, 16
+; SI: V_BFE_I32 v{{[0-9]+}}, v{{[0-9]+}}, 0, 8
+; SI: S_ENDPGM
+define void @bfe_16_bfe_8(i32 addrspace(1)* %out, i32 addrspace(1)* %ptr) nounwind {
+  %load = load i32 addrspace(1)* %ptr, align 4
+  %bfe0 = call i32 @llvm.AMDGPU.bfe.i32(i32 %load, i32 0, i32 16) nounwind readnone
+  %bfe1 = call i32 @llvm.AMDGPU.bfe.i32(i32 %bfe0, i32 0, i32 8) nounwind readnone
+  store i32 %bfe1, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
 ; Make sure there isn't a redundant BFE
 ; FUNC-LABEL: @sext_in_reg_i8_to_i32_bfe
 ; SI: S_BFE_I32 s{{[0-9]+}}, s{{[0-9]+}}, 0x80000
 ; SI-NOT: BFE
+; SI: S_ENDPGM
 define void @sext_in_reg_i8_to_i32_bfe(i32 addrspace(1)* %out, i32 %a, i32 %b) nounwind {
   %c = add i32 %a, %b ; add to prevent folding into extload
   %bfe = call i32 @llvm.AMDGPU.bfe.i32(i32 %c, i32 0, i32 8) nounwind readnone
@@ -394,3 +441,40 @@ define void @sext_in_reg_i8_to_i32_bfe(i32 addrspace(1)* %out, i32 %a, i32 %b) n
   store i32 %ashr, i32 addrspace(1)* %out, align 4
   ret void
 }
+
+; FUNC-LABEL: @sext_in_reg_i8_to_i32_bfe_wrong
+define void @sext_in_reg_i8_to_i32_bfe_wrong(i32 addrspace(1)* %out, i32 %a, i32 %b) nounwind {
+  %c = add i32 %a, %b ; add to prevent folding into extload
+  %bfe = call i32 @llvm.AMDGPU.bfe.i32(i32 %c, i32 8, i32 0) nounwind readnone
+  %shl = shl i32 %bfe, 24
+  %ashr = ashr i32 %shl, 24
+  store i32 %ashr, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: @sextload_i8_to_i32_bfe
+; SI: BUFFER_LOAD_SBYTE
+; SI-NOT: BFE
+; SI: S_ENDPGM
+define void @sextload_i8_to_i32_bfe(i32 addrspace(1)* %out, i8 addrspace(1)* %ptr) nounwind {
+  %load = load i8 addrspace(1)* %ptr, align 1
+  %sext = sext i8 %load to i32
+  %bfe = call i32 @llvm.AMDGPU.bfe.i32(i32 %sext, i32 0, i32 8) nounwind readnone
+  %shl = shl i32 %bfe, 24
+  %ashr = ashr i32 %shl, 24
+  store i32 %ashr, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: @sextload_i8_to_i32_bfe_0:
+; SI-NOT: BFE
+; SI: S_ENDPGM
+define void @sextload_i8_to_i32_bfe_0(i32 addrspace(1)* %out, i8 addrspace(1)* %ptr) nounwind {
+  %load = load i8 addrspace(1)* %ptr, align 1
+  %sext = sext i8 %load to i32
+  %bfe = call i32 @llvm.AMDGPU.bfe.i32(i32 %sext, i32 8, i32 0) nounwind readnone
+  %shl = shl i32 %bfe, 24
+  %ashr = ashr i32 %shl, 24
+  store i32 %ashr, i32 addrspace(1)* %out, align 4
+  ret void
+}