From 364a6747aa06ead940783e99f09e4759d271bdb3 Mon Sep 17 00:00:00 2001 From: Matt Arsenault Date: Wed, 11 Jun 2014 17:50:44 +0000 Subject: [PATCH] R600/SI: Use v_cvt_f32_ubyte* instructions This eliminates extra extract instructions when loading an i8 vector to a float vector. llvm-svn: 210666 --- llvm/lib/Target/R600/AMDGPUISelLowering.cpp | 17 +- llvm/lib/Target/R600/AMDGPUISelLowering.h | 6 + llvm/lib/Target/R600/AMDGPUInstrInfo.td | 11 ++ llvm/lib/Target/R600/AMDGPUIntrinsics.td | 4 + llvm/lib/Target/R600/SIISelLowering.cpp | 118 ++++++++++++++ llvm/lib/Target/R600/SIISelLowering.h | 3 + llvm/lib/Target/R600/SIInstructions.td | 16 +- llvm/test/CodeGen/R600/bitcast.ll | 14 ++ llvm/test/CodeGen/R600/cvt_f32_ubyte.ll | 171 +++++++++++++++++++++ .../test/CodeGen/R600/llvm.AMDGPU.cvt_f32_ubyte.ll | 42 +++++ 10 files changed, 397 insertions(+), 5 deletions(-) create mode 100644 llvm/test/CodeGen/R600/cvt_f32_ubyte.ll create mode 100644 llvm/test/CodeGen/R600/llvm.AMDGPU.cvt_f32_ubyte.ll diff --git a/llvm/lib/Target/R600/AMDGPUISelLowering.cpp b/llvm/lib/Target/R600/AMDGPUISelLowering.cpp index 0f1a77a..92ab174 100644 --- a/llvm/lib/Target/R600/AMDGPUISelLowering.cpp +++ b/llvm/lib/Target/R600/AMDGPUISelLowering.cpp @@ -786,6 +786,18 @@ SDValue AMDGPUTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, return DAG.getNode(AMDGPUISD::MAD_I24, DL, VT, Op.getOperand(1), Op.getOperand(2), Op.getOperand(3)); + case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte0: + return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE0, DL, VT, Op.getOperand(1)); + + case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte1: + return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE1, DL, VT, Op.getOperand(1)); + + case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte2: + return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE2, DL, VT, Op.getOperand(1)); + + case AMDGPUIntrinsic::AMDGPU_cvt_f32_ubyte3: + return DAG.getNode(AMDGPUISD::CVT_F32_UBYTE3, DL, VT, Op.getOperand(1)); + case AMDGPUIntrinsic::AMDGPU_bfe_i32: return DAG.getNode(AMDGPUISD::BFE_I32, DL, VT, Op.getOperand(1), @@ -1256,7 +1268,6 @@ SDValue AMDGPUTargetLowering::LowerUINT_TO_FP(SDValue Op, FloatHi = DAG.getNode(ISD::FMUL, DL, MVT::f32, FloatHi, DAG.getConstantFP(4294967296.0f, MVT::f32)); // 2^32 return DAG.getNode(ISD::FADD, DL, MVT::f32, FloatLo, FloatHi); - } SDValue AMDGPUTargetLowering::ExpandSIGN_EXTEND_INREG(SDValue Op, @@ -1582,6 +1593,10 @@ const char* AMDGPUTargetLowering::getTargetNodeName(unsigned Opcode) const { 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) } diff --git a/llvm/lib/Target/R600/AMDGPUISelLowering.h b/llvm/lib/Target/R600/AMDGPUISelLowering.h index ee98b3b..f50afab 100644 --- a/llvm/lib/Target/R600/AMDGPUISelLowering.h +++ b/llvm/lib/Target/R600/AMDGPUISelLowering.h @@ -205,6 +205,12 @@ enum { SAMPLEB, SAMPLED, SAMPLEL, + + // These cvt_f32_ubyte* nodes need to remain consecutive and in order. + 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, diff --git a/llvm/lib/Target/R600/AMDGPUInstrInfo.td b/llvm/lib/Target/R600/AMDGPUInstrInfo.td index f96dbb4..03137800 100644 --- a/llvm/lib/Target/R600/AMDGPUInstrInfo.td +++ b/llvm/lib/Target/R600/AMDGPUInstrInfo.td @@ -59,6 +59,17 @@ def AMDGPUumin : SDNode<"AMDGPUISD::UMIN", SDTIntBinOp, [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 diff --git a/llvm/lib/Target/R600/AMDGPUIntrinsics.td b/llvm/lib/Target/R600/AMDGPUIntrinsics.td index 9ad5e72..5790357 100644 --- a/llvm/lib/Target/R600/AMDGPUIntrinsics.td +++ b/llvm/lib/Target/R600/AMDGPUIntrinsics.td @@ -53,6 +53,10 @@ let TargetPrefix = "AMDGPU", isTarget = 1 in { def int_AMDGPU_imul24 : Intrinsic<[llvm_i32_ty], [llvm_i32_ty, llvm_i32_ty], [IntrNoMem]>; def int_AMDGPU_imad24 : Intrinsic<[llvm_i32_ty], [llvm_i32_ty, llvm_i32_ty, llvm_i32_ty], [IntrNoMem]>; def int_AMDGPU_umad24 : Intrinsic<[llvm_i32_ty], [llvm_i32_ty, llvm_i32_ty, llvm_i32_ty], [IntrNoMem]>; + def int_AMDGPU_cvt_f32_ubyte0 : Intrinsic<[llvm_float_ty], [llvm_i32_ty], [IntrNoMem]>; + def int_AMDGPU_cvt_f32_ubyte1 : Intrinsic<[llvm_float_ty], [llvm_i32_ty], [IntrNoMem]>; + def int_AMDGPU_cvt_f32_ubyte2 : Intrinsic<[llvm_float_ty], [llvm_i32_ty], [IntrNoMem]>; + def int_AMDGPU_cvt_f32_ubyte3 : Intrinsic<[llvm_float_ty], [llvm_i32_ty], [IntrNoMem]>; def int_AMDGPU_cube : Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty], [IntrNoMem]>; def int_AMDGPU_bfi : Intrinsic<[llvm_i32_ty], [llvm_i32_ty, llvm_i32_ty, llvm_i32_ty], [IntrNoMem]>; def int_AMDGPU_bfe_i32 : Intrinsic<[llvm_i32_ty], [llvm_i32_ty, llvm_i32_ty, llvm_i32_ty], [IntrNoMem]>; diff --git a/llvm/lib/Target/R600/SIISelLowering.cpp b/llvm/lib/Target/R600/SIISelLowering.cpp index 608aad2..846aeb6 100644 --- a/llvm/lib/Target/R600/SIISelLowering.cpp +++ b/llvm/lib/Target/R600/SIISelLowering.cpp @@ -24,6 +24,7 @@ #include "llvm/CodeGen/MachineRegisterInfo.h" #include "llvm/CodeGen/SelectionDAG.h" #include "llvm/IR/Function.h" +#include "llvm/ADT/SmallString.h" using namespace llvm; @@ -214,6 +215,8 @@ SITargetLowering::SITargetLowering(TargetMachine &TM) : setTargetDAGCombine(ISD::SELECT_CC); setTargetDAGCombine(ISD::SETCC); + setTargetDAGCombine(ISD::UINT_TO_FP); + setSchedulingPreference(Sched::RegPressure); } @@ -979,6 +982,96 @@ SDValue SITargetLowering::LowerSTORE(SDValue Op, SelectionDAG &DAG) const { // 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(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 Elts; + if (RegVT.isVector()) + DAG.ExtractVectorElements(NewLoad, Elts); + else + Elts.push_back(NewLoad); + + SmallVector 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; @@ -1020,6 +1113,31 @@ SDValue SITargetLowering::PerformDAGCombine(SDNode *N, } 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); diff --git a/llvm/lib/Target/R600/SIISelLowering.h b/llvm/lib/Target/R600/SIISelLowering.h index d9bad4e..2f97a9a 100644 --- a/llvm/lib/Target/R600/SIISelLowering.h +++ b/llvm/lib/Target/R600/SIISelLowering.h @@ -43,6 +43,9 @@ class SITargetLowering : public AMDGPUTargetLowering { 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, diff --git a/llvm/lib/Target/R600/SIInstructions.td b/llvm/lib/Target/R600/SIInstructions.td index 7d37d10..9af1a70 100644 --- a/llvm/lib/Target/R600/SIInstructions.td +++ b/llvm/lib/Target/R600/SIInstructions.td @@ -978,10 +978,18 @@ defm V_CVT_F32_F64 : VOP1_32_64 <0x0000000f, "V_CVT_F32_F64", 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))] >; diff --git a/llvm/test/CodeGen/R600/bitcast.ll b/llvm/test/CodeGen/R600/bitcast.ll index 539c597..a1a98f2 100644 --- a/llvm/test/CodeGen/R600/bitcast.ll +++ b/llvm/test/CodeGen/R600/bitcast.ll @@ -42,3 +42,17 @@ define void @v2i16_to_f32(float addrspace(1)* %out, <2 x i16> addrspace(1)* %in) store float %bc, float addrspace(1)* %out, align 4 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 +} diff --git a/llvm/test/CodeGen/R600/cvt_f32_ubyte.ll b/llvm/test/CodeGen/R600/cvt_f32_ubyte.ll new file mode 100644 index 0000000..fe97a44 --- /dev/null +++ b/llvm/test/CodeGen/R600/cvt_f32_ubyte.ll @@ -0,0 +1,171 @@ +; 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_UBYTE +; SI: V_CVT_F32_UBYTE0_e32 +; SI: BUFFER_LOAD_UBYTE +; SI: V_CVT_F32_UBYTE0_e32 +; SI: BUFFER_LOAD_UBYTE +; SI: V_CVT_F32_UBYTE0_e32 +; SI: BUFFER_LOAD_UBYTE +; SI: V_CVT_F32_UBYTE0_e32 + +; XXX - replace with this when v4i8 loads aren't scalarized anymore. +; XSI: BUFFER_LOAD_DWORD +; XSI: V_CVT_F32_U32_e32 +; XSI: V_CVT_F32_U32_e32 +; XSI: V_CVT_F32_U32_e32 +; XSI: 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, ; 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_DWORD +; SI: BUFFER_STORE_DWORD +; SI: BUFFER_STORE_DWORD +; SI: BUFFER_STORE_DWORD +; SI: BUFFER_STORE_DWORD +; SI: BUFFER_STORE_DWORD +; SI: BUFFER_STORE_DWORD +; SI: BUFFER_STORE_DWORD +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 +} diff --git a/llvm/test/CodeGen/R600/llvm.AMDGPU.cvt_f32_ubyte.ll b/llvm/test/CodeGen/R600/llvm.AMDGPU.cvt_f32_ubyte.ll new file mode 100644 index 0000000..6facb47 --- /dev/null +++ b/llvm/test/CodeGen/R600/llvm.AMDGPU.cvt_f32_ubyte.ll @@ -0,0 +1,42 @@ +; RUN: llc -march=r600 -mcpu=SI < %s | FileCheck -check-prefix=SI %s + +declare float @llvm.AMDGPU.cvt.f32.ubyte0(i32) nounwind readnone +declare float @llvm.AMDGPU.cvt.f32.ubyte1(i32) nounwind readnone +declare float @llvm.AMDGPU.cvt.f32.ubyte2(i32) nounwind readnone +declare float @llvm.AMDGPU.cvt.f32.ubyte3(i32) nounwind readnone + +; SI-LABEL: @test_unpack_byte0_to_float: +; SI: V_CVT_F32_UBYTE0 +define void @test_unpack_byte0_to_float(float addrspace(1)* %out, i32 addrspace(1)* %in) nounwind { + %val = load i32 addrspace(1)* %in, align 4 + %cvt = call float @llvm.AMDGPU.cvt.f32.ubyte0(i32 %val) nounwind readnone + store float %cvt, float addrspace(1)* %out, align 4 + ret void +} + +; SI-LABEL: @test_unpack_byte1_to_float: +; SI: V_CVT_F32_UBYTE1 +define void @test_unpack_byte1_to_float(float addrspace(1)* %out, i32 addrspace(1)* %in) nounwind { + %val = load i32 addrspace(1)* %in, align 4 + %cvt = call float @llvm.AMDGPU.cvt.f32.ubyte1(i32 %val) nounwind readnone + store float %cvt, float addrspace(1)* %out, align 4 + ret void +} + +; SI-LABEL: @test_unpack_byte2_to_float: +; SI: V_CVT_F32_UBYTE2 +define void @test_unpack_byte2_to_float(float addrspace(1)* %out, i32 addrspace(1)* %in) nounwind { + %val = load i32 addrspace(1)* %in, align 4 + %cvt = call float @llvm.AMDGPU.cvt.f32.ubyte2(i32 %val) nounwind readnone + store float %cvt, float addrspace(1)* %out, align 4 + ret void +} + +; SI-LABEL: @test_unpack_byte3_to_float: +; SI: V_CVT_F32_UBYTE3 +define void @test_unpack_byte3_to_float(float addrspace(1)* %out, i32 addrspace(1)* %in) nounwind { + %val = load i32 addrspace(1)* %in, align 4 + %cvt = call float @llvm.AMDGPU.cvt.f32.ubyte3(i32 %val) nounwind readnone + store float %cvt, float addrspace(1)* %out, align 4 + ret void +} -- 2.7.4