R600/SI: Use v_cvt_f32_ubyte* instructions

[arsenm]

Summary

This eliminates extra extract instructions when loading an i8 vector to a float vector

[arsenm]

Ping

[arsenm]

r210666

lib/Target/R600/AMDGPUISelLowering.h

   SDValue LowerUINT_TO_FP(SDValue Op, SelectionDAG &DAG) const;
 
 protected:
+  static EVT getEquivalentMemType(LLVMContext &Ctx, EVT VT);
+  static EVT getEquivalentLoadRegType(LLVMContext &Ctx, EVT VT);
 
   /// \brief Helper function that adds Reg to the LiveIn list of the DAG's
   /// MachineFunction.

   SAMPLEB,
   SAMPLED,
   SAMPLEL,
+  CVT_F32_UBYTE0,
+  CVT_F32_UBYTE1,
+  CVT_F32_UBYTE2,
+  CVT_F32_UBYTE3,
   FIRST_MEM_OPCODE_NUMBER = ISD::FIRST_TARGET_MEMORY_OPCODE,
   STORE_MSKOR,
   LOAD_CONSTANT,

lib/Target/R600/AMDGPUISelLowering.cpp

 #include "AMDGPUGenCallingConv.inc"
 
 // Find a larger type to do a load / store of a vector with.
-static MVT getEquivalentMemType(MVT VT) {
+EVT AMDGPUTargetLowering::getEquivalentMemType(LLVMContext &Ctx, EVT VT) {
   unsigned StoreSize = VT.getStoreSizeInBits();
   if (StoreSize <= 32)
-    return MVT::getIntegerVT(StoreSize);
+    return EVT::getIntegerVT(Ctx, StoreSize);
 
   assert(StoreSize % 32 == 0 && "Store size not a multiple of 32");
 
-  return MVT::getVectorVT(MVT::i32, StoreSize / 32);
+  return EVT::getVectorVT(Ctx, MVT::i32, StoreSize / 32);
 }
 
 // Type for a vector that will be loaded to.
-static MVT getEquivalentLoadRegType(MVT VT) {
+EVT AMDGPUTargetLowering::getEquivalentLoadRegType(LLVMContext &Ctx, EVT VT) {
   unsigned StoreSize = VT.getStoreSizeInBits();
   if (StoreSize <= 32)
     return MVT::getIntegerVT(32);

     if (MemEltBits != 64)
       return SDValue();
 
-    MVT StoreVT = getEquivalentMemType(VT.getSimpleVT());
+    EVT StoreVT = getEquivalentMemType(*DAG.getContext(), VT);
     SDValue Cast = DAG.getNode(ISD::BITCAST, DL, StoreVT, Value);
 
     return DAG.getStore(Store->getChain(), DL, Cast,

       EVT PackedVT = EVT::getIntegerVT(*DAG.getContext(), MemVT.getSizeInBits());
 
       // Doing a byte or short store.
-      MVT StoreVT = getEquivalentMemType(PackedVT.getSimpleVT());
+      EVT StoreVT = getEquivalentMemType(*DAG.getContext(), PackedVT);
       Packed = DAG.getNode(ISD::TRUNCATE, DL, StoreVT, Packed);
     }
 

     assert(!VT.isFloatingPoint() && "FP ext loads not yet handled");
 
     // Figure out what type to load as.
-    MVT LoadVT = getEquivalentMemType(MemVT);
-    MVT DestLoadVT = getEquivalentLoadRegType(MemVT);
+    EVT LoadVT = getEquivalentMemType(*DAG.getContext(), MemVT);
+    EVT DestLoadVT = getEquivalentLoadRegType(*DAG.getContext(), MemVT);
 
     // If we're loading a v2i8 as an i16, we need to do an extload of the i16 to
     // i32.

 
   if (ExtType == ISD::NON_EXTLOAD && VT.isVector() &&
       MemVT.getScalarSizeInBits() == 64) {
-    MVT LoadVT = getEquivalentMemType(MemVT);
+    EVT LoadVT = getEquivalentMemType(*DAG.getContext(), MemVT);
     assert(LoadVT != MemVT); // Avoid infinite loop.
 
     SDValue NewLoad = DAG.getLoad(Load->getAddressingMode(),

   unsigned BitWidth = PackedVT.getScalarSizeInBits();
   unsigned NElts = Elts.size();
 
-  MVT StoreVT = getEquivalentMemType(PackedVT.getSimpleVT());
+  EVT StoreVT = getEquivalentMemType(*DAG.getContext(), PackedVT);
 
   assert(NElts > 0);
 

   NODE_NAME_CASE(SAMPLEB)
   NODE_NAME_CASE(SAMPLED)
   NODE_NAME_CASE(SAMPLEL)
+  NODE_NAME_CASE(CVT_F32_UBYTE0)
+  NODE_NAME_CASE(CVT_F32_UBYTE1)
+  NODE_NAME_CASE(CVT_F32_UBYTE2)
+  NODE_NAME_CASE(CVT_F32_UBYTE3)
   NODE_NAME_CASE(STORE_MSKOR)
   NODE_NAME_CASE(TBUFFER_STORE_FORMAT)
   }

lib/Target/R600/AMDGPUInstrInfo.td

   [SDNPCommutative, SDNPAssociative]
 >;
 
+
+def AMDGPUcvt_f32_ubyte0 : SDNode<"AMDGPUISD::CVT_F32_UBYTE0",
+  SDTIntToFPOp, []>;
+def AMDGPUcvt_f32_ubyte1 : SDNode<"AMDGPUISD::CVT_F32_UBYTE1",
+  SDTIntToFPOp, []>;
+def AMDGPUcvt_f32_ubyte2 : SDNode<"AMDGPUISD::CVT_F32_UBYTE2",
+  SDTIntToFPOp, []>;
+def AMDGPUcvt_f32_ubyte3 : SDNode<"AMDGPUISD::CVT_F32_UBYTE3",
+  SDTIntToFPOp, []>;
+
+
 // urecip - This operation is a helper for integer division, it returns the
 // result of 1 / a as a fractional unsigned integer.
 // out = (2^32 / a) + e

lib/Target/R600/SIISelLowering.h

   void adjustWritemask(MachineSDNode *&N, SelectionDAG &DAG) const;
   MachineSDNode *AdjustRegClass(MachineSDNode *N, SelectionDAG &DAG) const;
 
+  static SDValue performUCharToFloatCombine(SDNode *N,
+                                            DAGCombinerInfo &DCI);
+
 public:
   SITargetLowering(TargetMachine &tm);
   bool allowsUnalignedMemoryAccesses(EVT VT, unsigned AS,

lib/Target/R600/SIISelLowering.cpp

 #include "llvm/CodeGen/MachineRegisterInfo.h"
 #include "llvm/CodeGen/SelectionDAG.h"
 #include "llvm/IR/Function.h"
+#include "llvm/ADT/SmallString.h"
 
 using namespace llvm;
 

   setTargetDAGCombine(ISD::SELECT_CC);
   setTargetDAGCombine(ISD::SETCC);
 
+  setTargetDAGCombine(ISD::UINT_TO_FP);
+
   setSchedulingPreference(Sched::RegPressure);
 }
 

 // Custom DAG optimizations
 //===----------------------------------------------------------------------===//
 
+SDValue SITargetLowering::performUCharToFloatCombine(SDNode *N,
+                                                     DAGCombinerInfo &DCI) {
+  EVT VT = N->getValueType(0);
+  EVT ScalarVT = VT.getScalarType();
+  if (ScalarVT != MVT::f32)
+    return SDValue();
+
+  SelectionDAG &DAG = DCI.DAG;
+  SDLoc DL(N);
+
+  SDValue Src = N->getOperand(0);
+  EVT SrcVT = Src.getValueType();
+
+  // TODO: We could try to match extracting the higher bytes, which would be
+  // easier if i8 vectors weren't promoted to i32 vectors, particularly after
+  // types are legalized. v4i8 -> v4f32 is probably the only case to worry
+  // about in practice.
+  if (DCI.isAfterLegalizeVectorOps() && SrcVT == MVT::i32) {
+    if (DAG.MaskedValueIsZero(Src, APInt::getHighBitsSet(32, 24))) {
+      SDValue Cvt = DAG.getNode(AMDGPUISD::CVT_F32_UBYTE0, DL, VT, Src);
+      DCI.AddToWorklist(Cvt.getNode());
+      return Cvt;
+    }
+  }
+
+  // We are primarily trying to catch operations on illegal vector types
+  // before they are expanded.
+  // For scalars, we can use the more flexible method of checking masked bits
+  // after legalization.
+  if (!DCI.isBeforeLegalize() ||
+      !SrcVT.isVector() ||
+      SrcVT.getVectorElementType() != MVT::i8) {
+    return SDValue();
+  }
+
+  assert(DCI.isBeforeLegalize() && "Unexpected legal type");
+
+  // Weird sized vectors are a pain to handle, but we know 3 is really the same
+  // size as 4.
+  unsigned NElts = SrcVT.getVectorNumElements();
+  if (!SrcVT.isSimple() && NElts != 3)
+    return SDValue();
+
+  // Handle v4i8 -> v4f32 extload. Replace the v4i8 with a legal i32 load to
+  // prevent a mess from expanding to v4i32 and repacking.
+  if (ISD::isNormalLoad(Src.getNode()) && Src.hasOneUse()) {
+    EVT LoadVT = getEquivalentMemType(*DAG.getContext(), SrcVT);
+    EVT RegVT = getEquivalentLoadRegType(*DAG.getContext(), SrcVT);
+    EVT FloatVT = EVT::getVectorVT(*DAG.getContext(), MVT::f32, NElts);
+
+    LoadSDNode *Load = cast<LoadSDNode>(Src);
+    SDValue NewLoad = DAG.getExtLoad(ISD::ZEXTLOAD, DL, RegVT,
+                                     Load->getChain(),
+                                     Load->getBasePtr(),
+                                     LoadVT,
+                                     Load->getMemOperand());
+
+    // Make sure successors of the original load stay after it by updating
+    // them to use the new Chain.
+    DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 1), NewLoad.getValue(1));
+
+    SmallVector<SDValue, 4> Elts;
+    if (RegVT.isVector())
+      DAG.ExtractVectorElements(NewLoad, Elts);
+    else
+      Elts.push_back(NewLoad);
+
+    SmallVector<SDValue, 4> Ops;
+
+    unsigned EltIdx = 0;
+    for (SDValue Elt : Elts) {
+      unsigned ComponentsInElt = std::min(4u, NElts - 4 * EltIdx);
+      for (unsigned I = 0; I < ComponentsInElt; ++I) {
+        unsigned Opc = AMDGPUISD::CVT_F32_UBYTE0 + I;
+        SDValue Cvt = DAG.getNode(Opc, DL, MVT::f32, Elt);
+        DCI.AddToWorklist(Cvt.getNode());
+        Ops.push_back(Cvt);
+      }
+
+      ++EltIdx;
+    }
+
+    assert(Ops.size() == NElts);
+
+    return DAG.getNode(ISD::BUILD_VECTOR, DL, FloatVT, Ops);
+  }
+
+  return SDValue();
+}
+
 SDValue SITargetLowering::PerformDAGCombine(SDNode *N,
                                             DAGCombinerInfo &DCI) const {
   SelectionDAG &DAG = DCI.DAG;

       }
       break;
     }
+
+  case AMDGPUISD::CVT_F32_UBYTE0:
+  case AMDGPUISD::CVT_F32_UBYTE1:
+  case AMDGPUISD::CVT_F32_UBYTE2:
+  case AMDGPUISD::CVT_F32_UBYTE3: {
+    unsigned Offset = N->getOpcode() - AMDGPUISD::CVT_F32_UBYTE0;
+
+    SDValue Src = N->getOperand(0);
+    APInt Demanded = APInt::getBitsSet(32, 8 * Offset, 8 * Offset + 8);
+
+    APInt KnownZero, KnownOne;
+    TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
+                                          !DCI.isBeforeLegalizeOps());
+    const TargetLowering &TLI = DAG.getTargetLoweringInfo();
+    if (TLO.ShrinkDemandedConstant(Src, Demanded) ||
+        TLI.SimplifyDemandedBits(Src, Demanded, KnownZero, KnownOne, TLO)) {
+      DCI.CommitTargetLoweringOpt(TLO);
+    }
+
+    break;
+  }
+
+  case ISD::UINT_TO_FP: {
+    return performUCharToFloatCombine(N, DCI);
+  }
   }
 
   return AMDGPUTargetLowering::PerformDAGCombine(N, DCI);

lib/Target/R600/SIInstructions.td

 defm V_CVT_F64_F32 : VOP1_64_32 <0x00000010, "V_CVT_F64_F32",
   [(set f64:$dst, (fextend f32:$src0))]
 >;
-//defm V_CVT_F32_UBYTE0 : VOP1_32 <0x00000011, "V_CVT_F32_UBYTE0", []>;
-//defm V_CVT_F32_UBYTE1 : VOP1_32 <0x00000012, "V_CVT_F32_UBYTE1", []>;
-//defm V_CVT_F32_UBYTE2 : VOP1_32 <0x00000013, "V_CVT_F32_UBYTE2", []>;
-//defm V_CVT_F32_UBYTE3 : VOP1_32 <0x00000014, "V_CVT_F32_UBYTE3", []>;
+defm V_CVT_F32_UBYTE0 : VOP1_32 <0x00000011, "V_CVT_F32_UBYTE0",
+  [(set f32:$dst, (AMDGPUcvt_f32_ubyte0 i32:$src0))]
+>;
+defm V_CVT_F32_UBYTE1 : VOP1_32 <0x00000012, "V_CVT_F32_UBYTE1",
+  [(set f32:$dst, (AMDGPUcvt_f32_ubyte1 i32:$src0))]
+>;
+defm V_CVT_F32_UBYTE2 : VOP1_32 <0x00000013, "V_CVT_F32_UBYTE2",
+  [(set f32:$dst, (AMDGPUcvt_f32_ubyte2 i32:$src0))]
+>;
+defm V_CVT_F32_UBYTE3 : VOP1_32 <0x00000014, "V_CVT_F32_UBYTE3",
+  [(set f32:$dst, (AMDGPUcvt_f32_ubyte3 i32:$src0))]
+>;
 defm V_CVT_U32_F64 : VOP1_32_64 <0x00000015, "V_CVT_U32_F64",
   [(set i32:$dst, (fp_to_uint f64:$src0))]
 >;

test/CodeGen/R600/bitcast.ll

-; RUN: llc < %s -march=r600 -mcpu=SI -verify-machineinstrs | FileCheck %s
+; RUN: llc -march=r600 -mcpu=SI -verify-machineinstrs < %s | FileCheck %s
 
 ; This test just checks that the compiler doesn't crash.
 ; CHECK-LABEL: @v32i8_to_v8i32

   store <16 x i8> %1, <16 x i8> addrspace(1)* %out
   ret void
 }
+
+define void @v4i8_to_i32(i32 addrspace(1)* %out, <4 x i8> addrspace(1)* %in) nounwind {
+  %load = load <4 x i8> addrspace(1)* %in, align 4
+  %bc = bitcast <4 x i8> %load to i32
+  store i32 %bc, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+define void @i32_to_v4i8(<4 x i8> addrspace(1)* %out, i32 addrspace(1)* %in) nounwind {
+  %load = load i32 addrspace(1)* %in, align 4
+  %bc = bitcast i32 %load to <4 x i8>
+  store <4 x i8> %bc, <4 x i8> addrspace(1)* %out, align 4
+  ret void
+}

test/CodeGen/R600/cvt_f32_ubyte.ll

+; RUN: llc -march=r600 -mcpu=SI -verify-machineinstrs < %s | FileCheck -check-prefix=SI %s
+
+; SI-LABEL: @load_i8_to_f32:
+; SI: BUFFER_LOAD_UBYTE [[LOADREG:v[0-9]+]],
+; SI-NOT: BFE
+; SI-NOT: LSHR
+; SI: V_CVT_F32_UBYTE0_e32 [[CONV:v[0-9]+]], [[LOADREG]]
+; SI: BUFFER_STORE_DWORD [[CONV]],
+define void @load_i8_to_f32(float addrspace(1)* noalias %out, i8 addrspace(1)* noalias %in) nounwind {
+  %load = load i8 addrspace(1)* %in, align 1
+  %cvt = uitofp i8 %load to float
+  store float %cvt, float addrspace(1)* %out, align 4
+  ret void
+}
+
+; SI-LABEL: @load_v2i8_to_v2f32:
+; SI: BUFFER_LOAD_USHORT [[LOADREG:v[0-9]+]],
+; SI-NOT: BFE
+; SI-NOT: LSHR
+; SI-NOT: AND
+; SI-DAG: V_CVT_F32_UBYTE1_e32 v[[HIRESULT:[0-9]+]], [[LOADREG]]
+; SI-DAG: V_CVT_F32_UBYTE0_e32 v[[LORESULT:[0-9]+]], [[LOADREG]]
+; SI: BUFFER_STORE_DWORDX2 v{{\[}}[[LORESULT]]:[[HIRESULT]]{{\]}},
+define void @load_v2i8_to_v2f32(<2 x float> addrspace(1)* noalias %out, <2 x i8> addrspace(1)* noalias %in) nounwind {
+  %load = load <2 x i8> addrspace(1)* %in, align 1
+  %cvt = uitofp <2 x i8> %load to <2 x float>
+  store <2 x float> %cvt, <2 x float> addrspace(1)* %out, align 16
+  ret void
+}
+
+; SI-LABEL: @load_v3i8_to_v3f32:
+; SI-NOT: BFE
+; SI-NOT: V_CVT_F32_UBYTE3_e32
+; SI-DAG: V_CVT_F32_UBYTE2_e32
+; SI-DAG: V_CVT_F32_UBYTE1_e32
+; SI-DAG: V_CVT_F32_UBYTE0_e32
+; SI: BUFFER_STORE_DWORDX2 v{{\[}}[[LORESULT]]:[[HIRESULT]]{{\]}},
+define void @load_v3i8_to_v3f32(<3 x float> addrspace(1)* noalias %out, <3 x i8> addrspace(1)* noalias %in) nounwind {
+  %load = load <3 x i8> addrspace(1)* %in, align 1
+  %cvt = uitofp <3 x i8> %load to <3 x float>
+  store <3 x float> %cvt, <3 x float> addrspace(1)* %out, align 16
+  ret void
+}
+
+; SI-LABEL: @load_v4i8_to_v4f32:
+; SI: BUFFER_LOAD_DWORD [[LOADREG:v[0-9]+]],
+; SI-NOT: BFE
+; SI-NOT: LSHR
+; SI-DAG: V_CVT_F32_UBYTE3_e32 v[[HIRESULT:[0-9]+]], [[LOADREG]]
+; SI-DAG: V_CVT_F32_UBYTE2_e32 v{{[0-9]+}}, [[LOADREG]]
+; SI-DAG: V_CVT_F32_UBYTE1_e32 v{{[0-9]+}}, [[LOADREG]]
+; SI-DAG: V_CVT_F32_UBYTE0_e32 v[[LORESULT:[0-9]+]], [[LOADREG]]
+; SI: BUFFER_STORE_DWORDX4 v{{\[}}[[LORESULT]]:[[HIRESULT]]{{\]}},
+define void @load_v4i8_to_v4f32(<4 x float> addrspace(1)* noalias %out, <4 x i8> addrspace(1)* noalias %in) nounwind {
+  %load = load <4 x i8> addrspace(1)* %in, align 1
+  %cvt = uitofp <4 x i8> %load to <4 x float>
+  store <4 x float> %cvt, <4 x float> addrspace(1)* %out, align 16
+  ret void
+}
+
+; XXX - This should really still be able to use the V_CVT_F32_UBYTE0
+; for each component, but computeKnownBits doesn't handle vectors very
+; well.
+
+; SI-LABEL: @load_v4i8_to_v4f32_2_uses:
+; SI: BUFFER_LOAD_DWORD
+; SI: V_CVT_F32_U32_e32
+; SI: V_CVT_F32_U32_e32
+; SI: V_CVT_F32_U32_e32
+; SI: V_CVT_F32_U32_e32
+; SI: S_ENDPGM
+define void @load_v4i8_to_v4f32_2_uses(<4 x float> addrspace(1)* noalias %out, <4 x i8> addrspace(1)* noalias %out2, <4 x i8> addrspace(1)* noalias %in) nounwind {
+  %load = load <4 x i8> addrspace(1)* %in, align 4
+  %cvt = uitofp <4 x i8> %load to <4 x float>
+  store <4 x float> %cvt, <4 x float> addrspace(1)* %out, align 16
+  %add = add <4 x i8> %load, <i8 9, i8 9, i8 9, i8 9> ; Second use of %load
+  store <4 x i8> %add, <4 x i8> addrspace(1)* %out2, align 4
+  ret void
+}
+
+; Make sure this doesn't crash.
+; SI-LABEL: @load_v7i8_to_v7f32:
+; SI: S_ENDPGM
+define void @load_v7i8_to_v7f32(<7 x float> addrspace(1)* noalias %out, <7 x i8> addrspace(1)* noalias %in) nounwind {
+  %load = load <7 x i8> addrspace(1)* %in, align 1
+  %cvt = uitofp <7 x i8> %load to <7 x float>
+  store <7 x float> %cvt, <7 x float> addrspace(1)* %out, align 16
+  ret void
+}
+
+; SI-LABEL: @load_v8i8_to_v8f32:
+; SI: BUFFER_LOAD_DWORDX2 v{{\[}}[[LOLOAD:[0-9]+]]:[[HILOAD:[0-9]+]]{{\]}},
+; SI-NOT: BFE
+; SI-NOT: LSHR
+; SI-DAG: V_CVT_F32_UBYTE3_e32 v{{[0-9]+}}, v[[LOLOAD]]
+; SI-DAG: V_CVT_F32_UBYTE2_e32 v{{[0-9]+}}, v[[LOLOAD]]
+; SI-DAG: V_CVT_F32_UBYTE1_e32 v{{[0-9]+}}, v[[LOLOAD]]
+; SI-DAG: V_CVT_F32_UBYTE0_e32 v{{[0-9]+}}, v[[LOLOAD]]
+; SI-DAG: V_CVT_F32_UBYTE3_e32 v{{[0-9]+}}, v[[HILOAD]]
+; SI-DAG: V_CVT_F32_UBYTE2_e32 v{{[0-9]+}}, v[[HILOAD]]
+; SI-DAG: V_CVT_F32_UBYTE1_e32 v{{[0-9]+}}, v[[HILOAD]]
+; SI-DAG: V_CVT_F32_UBYTE0_e32 v{{[0-9]+}}, v[[HILOAD]]
+; SI-NOT: BFE
+; SI-NOT: LSHR
+; SI: BUFFER_STORE_DWORDX4
+; SI: BUFFER_STORE_DWORDX4
+define void @load_v8i8_to_v8f32(<8 x float> addrspace(1)* noalias %out, <8 x i8> addrspace(1)* noalias %in) nounwind {
+  %load = load <8 x i8> addrspace(1)* %in, align 1
+  %cvt = uitofp <8 x i8> %load to <8 x float>
+  store <8 x float> %cvt, <8 x float> addrspace(1)* %out, align 16
+  ret void
+}
+
+; SI-LABEL: @i8_zext_inreg_i32_to_f32:
+; SI: BUFFER_LOAD_DWORD [[LOADREG:v[0-9]+]],
+; SI: V_ADD_I32_e32 [[ADD:v[0-9]+]], 2, [[LOADREG]]
+; SI-NEXT: V_CVT_F32_UBYTE0_e32 [[CONV:v[0-9]+]], [[ADD]]
+; SI: BUFFER_STORE_DWORD [[CONV]],
+define void @i8_zext_inreg_i32_to_f32(float addrspace(1)* noalias %out, i32 addrspace(1)* noalias %in) nounwind {
+  %load = load i32 addrspace(1)* %in, align 4
+  %add = add i32 %load, 2
+  %inreg = and i32 %add, 255
+  %cvt = uitofp i32 %inreg to float
+  store float %cvt, float addrspace(1)* %out, align 4
+  ret void
+}
+
+; SI-LABEL: @i8_zext_inreg_hi1_to_f32:
+define void @i8_zext_inreg_hi1_to_f32(float addrspace(1)* noalias %out, i32 addrspace(1)* noalias %in) nounwind {
+  %load = load i32 addrspace(1)* %in, align 4
+  %inreg = and i32 %load, 65280
+  %shr = lshr i32 %inreg, 8
+  %cvt = uitofp i32 %shr to float
+  store float %cvt, float addrspace(1)* %out, align 4
+  ret void
+}
+
+
+; We don't get these ones because of the zext, but instcombine removes
+; them so it shouldn't really matter.
+define void @i8_zext_i32_to_f32(float addrspace(1)* noalias %out, i8 addrspace(1)* noalias %in) nounwind {
+  %load = load i8 addrspace(1)* %in, align 1
+  %ext = zext i8 %load to i32
+  %cvt = uitofp i32 %ext to float
+  store float %cvt, float addrspace(1)* %out, align 4
+  ret void
+}
+
+define void @v4i8_zext_v4i32_to_v4f32(<4 x float> addrspace(1)* noalias %out, <4 x i8> addrspace(1)* noalias %in) nounwind {
+  %load = load <4 x i8> addrspace(1)* %in, align 1
+  %ext = zext <4 x i8> %load to <4 x i32>
+  %cvt = uitofp <4 x i32> %ext to <4 x float>
+  store <4 x float> %cvt, <4 x float> addrspace(1)* %out, align 16
+  ret void
+}