Input handling rework for SI
Tom Stellard
tom at stellard.net
Tue Mar 5 10:53:52 PST 2013
On Tue, Mar 05, 2013 at 03:23:33PM +0100, Christian König wrote:
> Hi guys,
>
> while working on indirect addressing for radeonsi (and YES I have
> indirect addressing of consts working and now on the way of getting
> source-indirect addressing of temps/inputs working) I stumbled over
> a whole bunch of problems with the way we currently handle vertex
> and pixel shader input registers. Especially the handling of M0 was
> sucking badly, and was easily broken by my new code.
>
> So please take a look at the attached patches. I have to confess
> that it gotten much further than I expected, but I think the way
> those patches now handle function parameters is much much more
> cleaner than we did before.
>
> The matching mesa patches should hit the maillinglist in few minutes.
>
Hi Christian,
This series has my r-b, I would like to test on R600 before you push it.
Thanks,
Tom
> Cheers,
> Christian.
> From 8f252afc3c18ce86bd7b0feebfc514957cffcc8d Mon Sep 17 00:00:00 2001
> From: =?UTF-8?q?Christian=20K=C3=B6nig?= <christian.koenig at amd.com>
> Date: Sat, 2 Mar 2013 18:48:52 +0100
> Subject: [PATCH 1/7] R600/SI: fix unused variable warning
> MIME-Version: 1.0
> Content-Type: text/plain; charset=UTF-8
> Content-Transfer-Encoding: 8bit
>
> Signed-off-by: Christian K??nig <christian.koenig at amd.com>
> ---
> lib/Target/R600/AMDGPUIndirectAddressing.cpp | 1 -
> 1 file changed, 1 deletion(-)
>
> diff --git a/lib/Target/R600/AMDGPUIndirectAddressing.cpp b/lib/Target/R600/AMDGPUIndirectAddressing.cpp
> index 15840b3..ed6c8ec 100644
> --- a/lib/Target/R600/AMDGPUIndirectAddressing.cpp
> +++ b/lib/Target/R600/AMDGPUIndirectAddressing.cpp
> @@ -289,7 +289,6 @@ bool AMDGPUIndirectAddressingPass::runOnMachineFunction(MachineFunction &MF) {
>
> // We only need to use REG_SEQUENCE for explicit defs, since the
> // register coalescer won't do anything with the implicit defs.
> - MachineInstr *DefInstr = MRI.getVRegDef(Reg);
> if (!regHasExplicitDef(MRI, Reg)) {
> continue;
> }
> --
> 1.7.10.4
>
> From 27c534082d61ce2f330034f5a3b5787787e63988 Mon Sep 17 00:00:00 2001
> From: =?UTF-8?q?Christian=20K=C3=B6nig?= <christian.koenig at amd.com>
> Date: Fri, 1 Mar 2013 10:50:55 +0100
> Subject: [PATCH 2/7] R600/SI: switch types of SGPRs to v*i8
> MIME-Version: 1.0
> Content-Type: text/plain; charset=UTF-8
> Content-Transfer-Encoding: 8bit
>
> Signed-off-by: Christian K??nig <christian.koenig at amd.com>
> ---
> lib/Target/R600/SIISelLowering.cpp | 21 +++++++++++++++++----
> lib/Target/R600/SIInstructions.td | 4 ++--
> lib/Target/R600/SIIntrinsics.td | 4 ++--
> lib/Target/R600/SIRegisterInfo.td | 6 +++---
> 4 files changed, 24 insertions(+), 11 deletions(-)
>
> diff --git a/lib/Target/R600/SIISelLowering.cpp b/lib/Target/R600/SIISelLowering.cpp
> index 0a0fbd9..7aa1037 100644
> --- a/lib/Target/R600/SIISelLowering.cpp
> +++ b/lib/Target/R600/SIISelLowering.cpp
> @@ -28,17 +28,30 @@ SITargetLowering::SITargetLowering(TargetMachine &TM) :
> AMDGPUTargetLowering(TM),
> TII(static_cast<const SIInstrInfo*>(TM.getInstrInfo())),
> TRI(TM.getRegisterInfo()) {
> - addRegisterClass(MVT::v4f32, &AMDGPU::VReg_128RegClass);
> - addRegisterClass(MVT::f32, &AMDGPU::VReg_32RegClass);
> - addRegisterClass(MVT::i32, &AMDGPU::VReg_32RegClass);
> - addRegisterClass(MVT::i64, &AMDGPU::SReg_64RegClass);
> +
> addRegisterClass(MVT::i1, &AMDGPU::SReg_64RegClass);
> + addRegisterClass(MVT::i64, &AMDGPU::SReg_64RegClass);
> +
> + addRegisterClass(MVT::v16i8, &AMDGPU::SReg_128RegClass);
> + addRegisterClass(MVT::v32i8, &AMDGPU::SReg_256RegClass);
> + addRegisterClass(MVT::v64i8, &AMDGPU::SReg_512RegClass);
> +
> + addRegisterClass(MVT::i32, &AMDGPU::VReg_32RegClass);
> + addRegisterClass(MVT::f32, &AMDGPU::VReg_32RegClass);
>
> addRegisterClass(MVT::v1i32, &AMDGPU::VReg_32RegClass);
> +
> addRegisterClass(MVT::v2i32, &AMDGPU::VReg_64RegClass);
> + addRegisterClass(MVT::v2f32, &AMDGPU::VReg_64RegClass);
> +
> addRegisterClass(MVT::v4i32, &AMDGPU::VReg_128RegClass);
> + addRegisterClass(MVT::v4f32, &AMDGPU::VReg_128RegClass);
> +
> addRegisterClass(MVT::v8i32, &AMDGPU::VReg_256RegClass);
> + addRegisterClass(MVT::v8f32, &AMDGPU::VReg_256RegClass);
> +
> addRegisterClass(MVT::v16i32, &AMDGPU::VReg_512RegClass);
> + addRegisterClass(MVT::v16f32, &AMDGPU::VReg_512RegClass);
>
> computeRegisterProperties();
>
> diff --git a/lib/Target/R600/SIInstructions.td b/lib/Target/R600/SIInstructions.td
> index d9d7b4c..bc6d604 100644
> --- a/lib/Target/R600/SIInstructions.td
> +++ b/lib/Target/R600/SIInstructions.td
> @@ -1479,7 +1479,7 @@ multiclass SMRD_Pattern <SMRD Instr_IMM, SMRD Instr_SGPR, ValueType vt> {
>
> defm : SMRD_Pattern <S_LOAD_DWORD_IMM, S_LOAD_DWORD_SGPR, f32>;
> defm : SMRD_Pattern <S_LOAD_DWORD_IMM, S_LOAD_DWORD_SGPR, i32>;
> -defm : SMRD_Pattern <S_LOAD_DWORDX4_IMM, S_LOAD_DWORDX4_SGPR, v4i32>;
> -defm : SMRD_Pattern <S_LOAD_DWORDX8_IMM, S_LOAD_DWORDX8_SGPR, v8i32>;
> +defm : SMRD_Pattern <S_LOAD_DWORDX4_IMM, S_LOAD_DWORDX4_SGPR, v16i8>;
> +defm : SMRD_Pattern <S_LOAD_DWORDX8_IMM, S_LOAD_DWORDX8_SGPR, v32i8>;
>
> } // End isSI predicate
> diff --git a/lib/Target/R600/SIIntrinsics.td b/lib/Target/R600/SIIntrinsics.td
> index 611b9c4..04308d8 100644
> --- a/lib/Target/R600/SIIntrinsics.td
> +++ b/lib/Target/R600/SIIntrinsics.td
> @@ -19,10 +19,10 @@ let TargetPrefix = "SI", isTarget = 1 in {
> /* XXX: We may need a seperate intrinsic here for loading integer values */
> def int_SI_load_const : Intrinsic <[llvm_float_ty], [llvm_i64_ty, llvm_i32_ty], []>;
> def int_SI_vs_load_buffer_index : Intrinsic <[llvm_i32_ty], [], [IntrNoMem]>;
> - def int_SI_vs_load_input : Intrinsic <[llvm_v4f32_ty], [llvm_v4i32_ty, llvm_i16_ty, llvm_i32_ty], [IntrReadMem]> ;
> + def int_SI_vs_load_input : Intrinsic <[llvm_v4f32_ty], [llvm_v16i8_ty, llvm_i16_ty, llvm_i32_ty], [IntrReadMem]> ;
> def int_SI_wqm : Intrinsic <[], [], []>;
>
> - class Sample : Intrinsic <[llvm_v4f32_ty], [llvm_i32_ty, llvm_anyvector_ty, llvm_v8i32_ty, llvm_v4i32_ty, llvm_i32_ty], [IntrReadMem]>;
> + class Sample : Intrinsic <[llvm_v4f32_ty], [llvm_i32_ty, llvm_anyvector_ty, llvm_v32i8_ty, llvm_v16i8_ty, llvm_i32_ty], [IntrReadMem]>;
>
> def int_SI_sample : Sample;
> def int_SI_sampleb : Sample;
> diff --git a/lib/Target/R600/SIRegisterInfo.td b/lib/Target/R600/SIRegisterInfo.td
> index 9e04e24..c4bca2c 100644
> --- a/lib/Target/R600/SIRegisterInfo.td
> +++ b/lib/Target/R600/SIRegisterInfo.td
> @@ -177,11 +177,11 @@ def SReg_64 : RegisterClass<"AMDGPU", [i64, i1], 64,
> (add SGPR_64, VCCReg, EXECReg)
> >;
>
> -def SReg_128 : RegisterClass<"AMDGPU", [v4f32, v4i32], 128, (add SGPR_128)>;
> +def SReg_128 : RegisterClass<"AMDGPU", [v16i8], 128, (add SGPR_128)>;
>
> -def SReg_256 : RegisterClass<"AMDGPU", [v8i32], 256, (add SGPR_256)>;
> +def SReg_256 : RegisterClass<"AMDGPU", [v32i8], 256, (add SGPR_256)>;
>
> -def SReg_512 : RegisterClass<"AMDGPU", [v16i32], 512, (add SGPR_512)>;
> +def SReg_512 : RegisterClass<"AMDGPU", [v64i8], 512, (add SGPR_512)>;
>
> // Register class for all vector registers (VGPRs + Interploation Registers)
> def VReg_32 : RegisterClass<"AMDGPU", [f32, i32, v1i32], 32, (add VGPR_32)>;
> --
> 1.7.10.4
>
> From b0aa2b61d7b86be0cdffd25b6361a9d4fc5da234 Mon Sep 17 00:00:00 2001
> From: =?UTF-8?q?Christian=20K=C3=B6nig?= <christian.koenig at amd.com>
> Date: Sat, 2 Mar 2013 16:33:30 +0100
> Subject: [PATCH 3/7] R600/SI: remove shader type intrinsic
> MIME-Version: 1.0
> Content-Type: text/plain; charset=UTF-8
> Content-Transfer-Encoding: 8bit
>
> Just encode the type as target specific attribute.
>
> Signed-off-by: Christian K??nig <christian.koenig at amd.com>
> ---
> lib/Target/R600/AMDGPUInstructions.td | 7 -------
> lib/Target/R600/AMDGPUIntrinsics.td | 2 --
> lib/Target/R600/R600ISelLowering.cpp | 1 -
> lib/Target/R600/SIISelLowering.cpp | 6 ------
> lib/Target/R600/SIMachineFunctionInfo.cpp | 18 ++++++++++++++++--
> lib/Target/R600/SIMachineFunctionInfo.h | 2 ++
> 6 files changed, 18 insertions(+), 18 deletions(-)
>
> diff --git a/lib/Target/R600/AMDGPUInstructions.td b/lib/Target/R600/AMDGPUInstructions.td
> index 960f108..a59c775 100644
> --- a/lib/Target/R600/AMDGPUInstructions.td
> +++ b/lib/Target/R600/AMDGPUInstructions.td
> @@ -132,13 +132,6 @@ class FNEG <RegisterClass rc> : AMDGPUShaderInst <
> [(set rc:$dst, (fneg rc:$src0))]
> >;
>
> -def SHADER_TYPE : AMDGPUShaderInst <
> - (outs),
> - (ins i32imm:$type),
> - "SHADER_TYPE $type",
> - [(int_AMDGPU_shader_type imm:$type)]
> ->;
> -
> } // usesCustomInserter = 1
>
> multiclass RegisterLoadStore <RegisterClass dstClass, Operand addrClass,
> diff --git a/lib/Target/R600/AMDGPUIntrinsics.td b/lib/Target/R600/AMDGPUIntrinsics.td
> index 2ba2d4b..eecb25b 100644
> --- a/lib/Target/R600/AMDGPUIntrinsics.td
> +++ b/lib/Target/R600/AMDGPUIntrinsics.td
> @@ -50,8 +50,6 @@ let TargetPrefix = "AMDGPU", isTarget = 1 in {
> def int_AMDGPU_umax : Intrinsic<[llvm_i32_ty], [llvm_i32_ty, llvm_i32_ty], [IntrNoMem]>;
> def int_AMDGPU_umin : Intrinsic<[llvm_i32_ty], [llvm_i32_ty, llvm_i32_ty], [IntrNoMem]>;
> def int_AMDGPU_cube : Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty], [IntrNoMem]>;
> -
> - def int_AMDGPU_shader_type : Intrinsic<[], [llvm_i32_ty], []>;
> }
>
> let TargetPrefix = "TGSI", isTarget = 1 in {
> diff --git a/lib/Target/R600/R600ISelLowering.cpp b/lib/Target/R600/R600ISelLowering.cpp
> index b5c2a93..b0f2727 100644
> --- a/lib/Target/R600/R600ISelLowering.cpp
> +++ b/lib/Target/R600/R600ISelLowering.cpp
> @@ -105,7 +105,6 @@ MachineBasicBlock * R600TargetLowering::EmitInstrWithCustomInserter(
>
> switch (MI->getOpcode()) {
> default: return AMDGPUTargetLowering::EmitInstrWithCustomInserter(MI, BB);
> - case AMDGPU::SHADER_TYPE: break;
> case AMDGPU::CLAMP_R600: {
> MachineInstr *NewMI = TII->buildDefaultInstruction(*BB, I,
> AMDGPU::MOV,
> diff --git a/lib/Target/R600/SIISelLowering.cpp b/lib/Target/R600/SIISelLowering.cpp
> index 7aa1037..a9eca31 100644
> --- a/lib/Target/R600/SIISelLowering.cpp
> +++ b/lib/Target/R600/SIISelLowering.cpp
> @@ -83,12 +83,6 @@ MachineBasicBlock * SITargetLowering::EmitInstrWithCustomInserter(
> default:
> return AMDGPUTargetLowering::EmitInstrWithCustomInserter(MI, BB);
> case AMDGPU::BRANCH: return BB;
> - case AMDGPU::SHADER_TYPE:
> - BB->getParent()->getInfo<SIMachineFunctionInfo>()->ShaderType =
> - MI->getOperand(0).getImm();
> - MI->eraseFromParent();
> - break;
> -
> case AMDGPU::SI_INTERP:
> LowerSI_INTERP(MI, *BB, I, MRI);
> break;
> diff --git a/lib/Target/R600/SIMachineFunctionInfo.cpp b/lib/Target/R600/SIMachineFunctionInfo.cpp
> index 7e59b42..e5fbf05 100644
> --- a/lib/Target/R600/SIMachineFunctionInfo.cpp
> +++ b/lib/Target/R600/SIMachineFunctionInfo.cpp
> @@ -10,11 +10,25 @@
>
>
> #include "SIMachineFunctionInfo.h"
> +#include "llvm/IR/Attributes.h"
> +#include "llvm/IR/Function.h"
>
> using namespace llvm;
>
> +const char *SIMachineFunctionInfo::ShaderTypeAttribute = "ShaderType";
> +
> SIMachineFunctionInfo::SIMachineFunctionInfo(const MachineFunction &MF)
> : MachineFunctionInfo(),
> SPIPSInputAddr(0),
> - ShaderType(0)
> - { }
> + ShaderType(0) {
> +
> + AttributeSet Set = MF.getFunction()->getAttributes();
> + Attribute A = Set.getAttribute(AttributeSet::FunctionIndex,
> + ShaderTypeAttribute);
> +
> + if (A.isStringAttribute()) {
> + StringRef Str = A.getValueAsString();
> + if (Str.getAsInteger(0, ShaderType))
> + llvm_unreachable("Can't parse shader type!");
> + }
> +}
> diff --git a/lib/Target/R600/SIMachineFunctionInfo.h b/lib/Target/R600/SIMachineFunctionInfo.h
> index 47271f5..e2f97a4 100644
> --- a/lib/Target/R600/SIMachineFunctionInfo.h
> +++ b/lib/Target/R600/SIMachineFunctionInfo.h
> @@ -23,6 +23,8 @@ namespace llvm {
> /// tells the hardware which interpolation parameters to load.
> class SIMachineFunctionInfo : public MachineFunctionInfo {
> public:
> + static const char *ShaderTypeAttribute;
> +
> SIMachineFunctionInfo(const MachineFunction &MF);
> unsigned SPIPSInputAddr;
> unsigned ShaderType;
> --
> 1.7.10.4
>
> From 92bc687ec5bc549b8d473dade20e56868386415c Mon Sep 17 00:00:00 2001
> From: =?UTF-8?q?Christian=20K=C3=B6nig?= <christian.koenig at amd.com>
> Date: Mon, 4 Mar 2013 12:18:03 +0100
> Subject: [PATCH 4/7] R600/SI: add proper formal parameter handling for SI
> MIME-Version: 1.0
> Content-Type: text/plain; charset=UTF-8
> Content-Transfer-Encoding: 8bit
>
> Signed-off-by: Christian K??nig <christian.koenig at amd.com>
> ---
> lib/Target/R600/AMDGPU.td | 1 +
> lib/Target/R600/AMDGPUCallingConv.td | 42 +++++++++++++
> lib/Target/R600/AMDGPUISelLowering.cpp | 20 +++----
> lib/Target/R600/AMDGPUISelLowering.h | 9 +--
> lib/Target/R600/AMDILISelLowering.cpp | 5 --
> lib/Target/R600/SIISelLowering.cpp | 101 ++++++++++++++++++++++++++++++++
> lib/Target/R600/SIISelLowering.h | 7 +++
> 7 files changed, 163 insertions(+), 22 deletions(-)
> create mode 100644 lib/Target/R600/AMDGPUCallingConv.td
>
> diff --git a/lib/Target/R600/AMDGPU.td b/lib/Target/R600/AMDGPU.td
> index 40f4741..1a26c77 100644
> --- a/lib/Target/R600/AMDGPU.td
> +++ b/lib/Target/R600/AMDGPU.td
> @@ -38,3 +38,4 @@ include "AMDGPUInstrInfo.td"
> include "AMDGPUIntrinsics.td"
> include "AMDGPURegisterInfo.td"
> include "AMDGPUInstructions.td"
> +include "AMDGPUCallingConv.td"
> diff --git a/lib/Target/R600/AMDGPUCallingConv.td b/lib/Target/R600/AMDGPUCallingConv.td
> new file mode 100644
> index 0000000..45ae37e
> --- /dev/null
> +++ b/lib/Target/R600/AMDGPUCallingConv.td
> @@ -0,0 +1,42 @@
> +//===---- AMDCallingConv.td - Calling Conventions for Radeon GPUs ---------===//
> +//
> +// The LLVM Compiler Infrastructure
> +//
> +// This file is distributed under the University of Illinois Open Source
> +// License. See LICENSE.TXT for details.
> +//
> +//===----------------------------------------------------------------------===//
> +//
> +// This describes the calling conventions for the AMD Radeon GPUs.
> +//
> +//===----------------------------------------------------------------------===//
> +
> +// Inversion of CCIfInReg
> +class CCIfNotInReg<CCAction A> : CCIf<"!ArgFlags.isInReg()", A> {}
> +
> +// Calling convention for SI
> +def CC_SI : CallingConv<[
> +
> + CCIfInReg<CCIfType<[f32, i32] , CCAssignToReg<[
> + SGPR0, SGPR1, SGPR2, SGPR3, SGPR4, SGPR5, SGPR6, SGPR7,
> + SGPR8, SGPR9, SGPR10, SGPR11, SGPR12, SGPR13, SGPR14, SGPR15
> + ]>>>,
> +
> + CCIfInReg<CCIfType<[i64] , CCAssignToRegWithShadow<
> + [ SGPR0, SGPR2, SGPR4, SGPR6, SGPR8, SGPR10, SGPR12, SGPR14 ],
> + [ SGPR1, SGPR3, SGPR5, SGPR7, SGPR9, SGPR11, SGPR12, SGPR15 ]
> + >>>,
> +
> + CCIfNotInReg<CCIfType<[f32, i32] , CCAssignToReg<[
> + VGPR0, VGPR1, VGPR2, VGPR3, VGPR4, VGPR5, VGPR6, VGPR7,
> + VGPR8, VGPR9, VGPR10, VGPR11, VGPR12, VGPR13, VGPR14, VGPR15,
> + VGPR16, VGPR17, VGPR18, VGPR19, VGPR20, VGPR21, VGPR22, VGPR23,
> + VGPR24, VGPR25, VGPR26, VGPR27, VGPR28, VGPR29, VGPR30, VGPR31
> + ]>>>
> +
> +]>;
> +
> +def CC_AMDGPU : CallingConv<[
> + CCIf<"State.getTarget().getSubtarget<AMDGPUSubtarget>().device()"#
> + "->getGeneration() == AMDGPUDeviceInfo::HD7XXX", CCDelegateTo<CC_SI>>
> +]>;
> diff --git a/lib/Target/R600/AMDGPUISelLowering.cpp b/lib/Target/R600/AMDGPUISelLowering.cpp
> index 0a33264..5995b6f 100644
> --- a/lib/Target/R600/AMDGPUISelLowering.cpp
> +++ b/lib/Target/R600/AMDGPUISelLowering.cpp
> @@ -14,7 +14,10 @@
> //===----------------------------------------------------------------------===//
>
> #include "AMDGPUISelLowering.h"
> +#include "AMDGPURegisterInfo.h"
> #include "AMDILIntrinsicInfo.h"
> +#include "AMDGPUSubtarget.h"
> +#include "llvm/CodeGen/CallingConvLower.h"
> #include "llvm/CodeGen/MachineFunction.h"
> #include "llvm/CodeGen/MachineRegisterInfo.h"
> #include "llvm/CodeGen/SelectionDAG.h"
> @@ -22,6 +25,8 @@
>
> using namespace llvm;
>
> +#include "AMDGPUGenCallingConv.inc"
> +
> AMDGPUTargetLowering::AMDGPUTargetLowering(TargetMachine &TM) :
> TargetLowering(TM, new TargetLoweringObjectFileELF()) {
>
> @@ -64,17 +69,10 @@ AMDGPUTargetLowering::AMDGPUTargetLowering(TargetMachine &TM) :
> // TargetLowering Callbacks
> //===---------------------------------------------------------------------===//
>
> -SDValue AMDGPUTargetLowering::LowerFormalArguments(
> - SDValue Chain,
> - CallingConv::ID CallConv,
> - bool isVarArg,
> - const SmallVectorImpl<ISD::InputArg> &Ins,
> - DebugLoc DL, SelectionDAG &DAG,
> - SmallVectorImpl<SDValue> &InVals) const {
> - for (unsigned i = 0, e = Ins.size(); i < e; ++i) {
> - InVals.push_back(SDValue());
> - }
> - return Chain;
> +void AMDGPUTargetLowering::AnalyzeFormalArguments(CCState &State,
> + const SmallVectorImpl<ISD::InputArg> &Ins) const {
> +
> + State.AnalyzeFormalArguments(Ins, CC_AMDGPU);
> }
>
> SDValue AMDGPUTargetLowering::LowerReturn(
> diff --git a/lib/Target/R600/AMDGPUISelLowering.h b/lib/Target/R600/AMDGPUISelLowering.h
> index 9e7d997..f31b646 100644
> --- a/lib/Target/R600/AMDGPUISelLowering.h
> +++ b/lib/Target/R600/AMDGPUISelLowering.h
> @@ -39,15 +39,12 @@ protected:
> bool isHWTrueValue(SDValue Op) const;
> bool isHWFalseValue(SDValue Op) const;
>
> + void AnalyzeFormalArguments(CCState &State,
> + const SmallVectorImpl<ISD::InputArg> &Ins) const;
> +
> public:
> AMDGPUTargetLowering(TargetMachine &TM);
>
> - virtual SDValue LowerFormalArguments(SDValue Chain, CallingConv::ID CallConv,
> - bool isVarArg,
> - const SmallVectorImpl<ISD::InputArg> &Ins,
> - DebugLoc DL, SelectionDAG &DAG,
> - SmallVectorImpl<SDValue> &InVals) const;
> -
> virtual SDValue LowerReturn(SDValue Chain, CallingConv::ID CallConv,
> bool isVarArg,
> const SmallVectorImpl<ISD::OutputArg> &Outs,
> diff --git a/lib/Target/R600/AMDILISelLowering.cpp b/lib/Target/R600/AMDILISelLowering.cpp
> index f65e1f3..922cac1 100644
> --- a/lib/Target/R600/AMDILISelLowering.cpp
> +++ b/lib/Target/R600/AMDILISelLowering.cpp
> @@ -33,11 +33,6 @@
>
> using namespace llvm;
> //===----------------------------------------------------------------------===//
> -// Calling Convention Implementation
> -//===----------------------------------------------------------------------===//
> -#include "AMDGPUGenCallingConv.inc"
> -
> -//===----------------------------------------------------------------------===//
> // TargetLowering Implementation Help Functions End
> //===----------------------------------------------------------------------===//
>
> diff --git a/lib/Target/R600/SIISelLowering.cpp b/lib/Target/R600/SIISelLowering.cpp
> index a9eca31..8b46cb0 100644
> --- a/lib/Target/R600/SIISelLowering.cpp
> +++ b/lib/Target/R600/SIISelLowering.cpp
> @@ -18,6 +18,8 @@
> #include "SIInstrInfo.h"
> #include "SIMachineFunctionInfo.h"
> #include "SIRegisterInfo.h"
> +#include "llvm/IR/Function.h"
> +#include "llvm/CodeGen/CallingConvLower.h"
> #include "llvm/CodeGen/MachineInstrBuilder.h"
> #include "llvm/CodeGen/MachineRegisterInfo.h"
> #include "llvm/CodeGen/SelectionDAG.h"
> @@ -74,6 +76,105 @@ SITargetLowering::SITargetLowering(TargetMachine &TM) :
> setTargetDAGCombine(ISD::SETCC);
> }
>
> +SDValue SITargetLowering::LowerFormalArguments(
> + SDValue Chain,
> + CallingConv::ID CallConv,
> + bool isVarArg,
> + const SmallVectorImpl<ISD::InputArg> &Ins,
> + DebugLoc DL, SelectionDAG &DAG,
> + SmallVectorImpl<SDValue> &InVals) const {
> +
> + const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
> +
> + MachineFunction &MF = DAG.getMachineFunction();
> + FunctionType *FType = MF.getFunction()->getFunctionType();
> +
> + assert(CallConv == CallingConv::C);
> +
> + SmallVector<ISD::InputArg, 16> Splits;
> + for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
> + const ISD::InputArg &Arg = Ins[i];
> +
> + // Split vertices into their elements
> + if (Arg.VT.isVector()) {
> + ISD::InputArg NewArg = Arg;
> + NewArg.Flags.setSplit();
> + NewArg.VT = Arg.VT.getVectorElementType();
> +
> + // We REALLY want the ORIGINAL number of vertex elements here, e.g. a
> + // three or five element vertex only needs three or five registers,
> + // NOT four or eigth.
> + Type *ParamType = FType->getParamType(Arg.OrigArgIndex);
> + unsigned NumElements = ParamType->getVectorNumElements();
> +
> + for (unsigned j = 0; j != NumElements; ++j) {
> + Splits.push_back(NewArg);
> + NewArg.PartOffset += NewArg.VT.getStoreSize();
> + }
> +
> + } else {
> + Splits.push_back(Arg);
> + }
> + }
> +
> + SmallVector<CCValAssign, 16> ArgLocs;
> + CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
> + getTargetMachine(), ArgLocs, *DAG.getContext());
> +
> + AnalyzeFormalArguments(CCInfo, Splits);
> +
> + for (unsigned i = 0, e = Ins.size(), ArgIdx = 0; i != e; ++i) {
> +
> + CCValAssign &VA = ArgLocs[ArgIdx++];
> + assert(VA.isRegLoc() && "Parameter must be in a register!");
> +
> + unsigned Reg = VA.getLocReg();
> + MVT VT = VA.getLocVT();
> +
> + if (VT == MVT::i64) {
> + // For now assume it is a pointer
> + Reg = TRI->getMatchingSuperReg(Reg, AMDGPU::sub0,
> + &AMDGPU::SReg_64RegClass);
> + Reg = MF.addLiveIn(Reg, &AMDGPU::SReg_64RegClass);
> + InVals.push_back(DAG.getCopyFromReg(Chain, DL, Reg, VT));
> + continue;
> + }
> +
> + const TargetRegisterClass *RC = TRI->getMinimalPhysRegClass(Reg, VT);
> +
> + Reg = MF.addLiveIn(Reg, RC);
> + SDValue Val = DAG.getCopyFromReg(Chain, DL, Reg, VT);
> +
> + const ISD::InputArg &Arg = Ins[i];
> + if (Arg.VT.isVector()) {
> +
> + // Build a vector from the registers
> + Type *ParamType = FType->getParamType(Arg.OrigArgIndex);
> + unsigned NumElements = ParamType->getVectorNumElements();
> +
> + SmallVector<SDValue, 4> Regs;
> + Regs.push_back(Val);
> + for (unsigned j = 1; j != NumElements; ++j) {
> + Reg = ArgLocs[ArgIdx++].getLocReg();
> + Reg = MF.addLiveIn(Reg, RC);
> + Regs.push_back(DAG.getCopyFromReg(Chain, DL, Reg, VT));
> + }
> +
> + // Fill up the missing vector elements
> + NumElements = Arg.VT.getVectorNumElements() - NumElements;
> + for (unsigned j = 0; j != NumElements; ++j)
> + Regs.push_back(DAG.getUNDEF(VT));
> +
> + InVals.push_back(DAG.getNode(ISD::BUILD_VECTOR, DL, Arg.VT,
> + Regs.data(), Regs.size()));
> + continue;
> + }
> +
> + InVals.push_back(Val);
> + }
> + return Chain;
> +}
> +
> MachineBasicBlock * SITargetLowering::EmitInstrWithCustomInserter(
> MachineInstr * MI, MachineBasicBlock * BB) const {
> MachineRegisterInfo & MRI = BB->getParent()->getRegInfo();
> diff --git a/lib/Target/R600/SIISelLowering.h b/lib/Target/R600/SIISelLowering.h
> index 737162f..6d77d88 100644
> --- a/lib/Target/R600/SIISelLowering.h
> +++ b/lib/Target/R600/SIISelLowering.h
> @@ -43,6 +43,13 @@ class SITargetLowering : public AMDGPUTargetLowering {
>
> public:
> SITargetLowering(TargetMachine &tm);
> +
> + SDValue LowerFormalArguments(SDValue Chain, CallingConv::ID CallConv,
> + bool isVarArg,
> + const SmallVectorImpl<ISD::InputArg> &Ins,
> + DebugLoc DL, SelectionDAG &DAG,
> + SmallVectorImpl<SDValue> &InVals) const;
> +
> virtual MachineBasicBlock * EmitInstrWithCustomInserter(MachineInstr * MI,
> MachineBasicBlock * BB) const;
> virtual EVT getSetCCResultType(EVT VT) const;
> --
> 1.7.10.4
>
> From 942c4da4c81157f21c37a031231faa809647470c Mon Sep 17 00:00:00 2001
> From: =?UTF-8?q?Christian=20K=C3=B6nig?= <christian.koenig at amd.com>
> Date: Sat, 2 Mar 2013 16:40:22 +0100
> Subject: [PATCH 5/7] R600/SI: remove SGPR address space
> MIME-Version: 1.0
> Content-Type: text/plain; charset=UTF-8
> Content-Transfer-Encoding: 8bit
>
> Signed-off-by: Christian K??nig <christian.koenig at amd.com>
> ---
> lib/Target/R600/AMDIL.h | 35 +++++++++++++--------------
> lib/Target/R600/SIISelLowering.cpp | 47 ------------------------------------
> lib/Target/R600/SIISelLowering.h | 1 -
> lib/Target/R600/SIInstructions.td | 5 ++--
> 4 files changed, 19 insertions(+), 69 deletions(-)
>
> diff --git a/lib/Target/R600/AMDIL.h b/lib/Target/R600/AMDIL.h
> index b39fbdb..39ab664 100644
> --- a/lib/Target/R600/AMDIL.h
> +++ b/lib/Target/R600/AMDIL.h
> @@ -96,24 +96,23 @@ enum AddressSpaces {
> ADDRESS_NONE = 5, ///< Address space for unknown memory.
> PARAM_D_ADDRESS = 6, ///< Address space for direct addressible parameter memory (CONST0)
> PARAM_I_ADDRESS = 7, ///< Address space for indirect addressible parameter memory (VTX1)
> - USER_SGPR_ADDRESS = 8, ///< Address space for USER_SGPRS on SI
> - CONSTANT_BUFFER_0 = 9,
> - CONSTANT_BUFFER_1 = 10,
> - CONSTANT_BUFFER_2 = 11,
> - CONSTANT_BUFFER_3 = 12,
> - CONSTANT_BUFFER_4 = 13,
> - CONSTANT_BUFFER_5 = 14,
> - CONSTANT_BUFFER_6 = 15,
> - CONSTANT_BUFFER_7 = 16,
> - CONSTANT_BUFFER_8 = 17,
> - CONSTANT_BUFFER_9 = 18,
> - CONSTANT_BUFFER_10 = 19,
> - CONSTANT_BUFFER_11 = 20,
> - CONSTANT_BUFFER_12 = 21,
> - CONSTANT_BUFFER_13 = 22,
> - CONSTANT_BUFFER_14 = 23,
> - CONSTANT_BUFFER_15 = 24,
> - LAST_ADDRESS = 25
> + CONSTANT_BUFFER_0 = 8,
> + CONSTANT_BUFFER_1 = 9,
> + CONSTANT_BUFFER_2 = 10,
> + CONSTANT_BUFFER_3 = 11,
> + CONSTANT_BUFFER_4 = 12,
> + CONSTANT_BUFFER_5 = 13,
> + CONSTANT_BUFFER_6 = 14,
> + CONSTANT_BUFFER_7 = 15,
> + CONSTANT_BUFFER_8 = 16,
> + CONSTANT_BUFFER_9 = 17,
> + CONSTANT_BUFFER_10 = 18,
> + CONSTANT_BUFFER_11 = 19,
> + CONSTANT_BUFFER_12 = 20,
> + CONSTANT_BUFFER_13 = 21,
> + CONSTANT_BUFFER_14 = 22,
> + CONSTANT_BUFFER_15 = 23,
> + LAST_ADDRESS = 24
> };
>
> } // namespace AMDGPUAS
> diff --git a/lib/Target/R600/SIISelLowering.cpp b/lib/Target/R600/SIISelLowering.cpp
> index 8b46cb0..a65e06f 100644
> --- a/lib/Target/R600/SIISelLowering.cpp
> +++ b/lib/Target/R600/SIISelLowering.cpp
> @@ -62,11 +62,6 @@ SITargetLowering::SITargetLowering(TargetMachine &TM) :
>
> setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
>
> - // We need to custom lower loads from the USER_SGPR address space, so we can
> - // add the SGPRs as livein registers.
> - setOperationAction(ISD::LOAD, MVT::i32, Custom);
> - setOperationAction(ISD::LOAD, MVT::i64, Custom);
> -
> setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
> setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
>
> @@ -245,7 +240,6 @@ SDValue SITargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
> switch (Op.getOpcode()) {
> default: return AMDGPUTargetLowering::LowerOperation(Op, DAG);
> case ISD::BRCOND: return LowerBRCOND(Op, DAG);
> - case ISD::LOAD: return LowerLOAD(Op, DAG);
> case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
> case ISD::INTRINSIC_WO_CHAIN: {
> unsigned IntrinsicID =
> @@ -357,47 +351,6 @@ SDValue SITargetLowering::LowerBRCOND(SDValue BRCOND,
> return Chain;
> }
>
> -SDValue SITargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const {
> - EVT VT = Op.getValueType();
> - LoadSDNode *Ptr = dyn_cast<LoadSDNode>(Op);
> -
> - assert(Ptr);
> -
> - unsigned AddrSpace = Ptr->getPointerInfo().getAddrSpace();
> -
> - // We only need to lower USER_SGPR address space loads
> - if (AddrSpace != AMDGPUAS::USER_SGPR_ADDRESS) {
> - return SDValue();
> - }
> -
> - // Loads from the USER_SGPR address space can only have constant value
> - // pointers.
> - ConstantSDNode *BasePtr = dyn_cast<ConstantSDNode>(Ptr->getBasePtr());
> - assert(BasePtr);
> -
> - unsigned TypeDwordWidth = VT.getSizeInBits() / 32;
> - const TargetRegisterClass * dstClass;
> - switch (TypeDwordWidth) {
> - default:
> - assert(!"USER_SGPR value size not implemented");
> - return SDValue();
> - case 1:
> - dstClass = &AMDGPU::SReg_32RegClass;
> - break;
> - case 2:
> - dstClass = &AMDGPU::SReg_64RegClass;
> - break;
> - }
> - uint64_t Index = BasePtr->getZExtValue();
> - assert(Index % TypeDwordWidth == 0 && "USER_SGPR not properly aligned");
> - unsigned SGPRIndex = Index / TypeDwordWidth;
> - unsigned Reg = dstClass->getRegister(SGPRIndex);
> -
> - DAG.ReplaceAllUsesOfValueWith(Op, CreateLiveInRegister(DAG, dstClass, Reg,
> - VT));
> - return SDValue();
> -}
> -
> SDValue SITargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
> SDValue LHS = Op.getOperand(0);
> SDValue RHS = Op.getOperand(1);
> diff --git a/lib/Target/R600/SIISelLowering.h b/lib/Target/R600/SIISelLowering.h
> index 6d77d88..8ae1400 100644
> --- a/lib/Target/R600/SIISelLowering.h
> +++ b/lib/Target/R600/SIISelLowering.h
> @@ -31,7 +31,6 @@ class SITargetLowering : public AMDGPUTargetLowering {
> void LowerSI_WQM(MachineInstr *MI, MachineBasicBlock &BB,
> MachineBasicBlock::iterator I, MachineRegisterInfo & MRI) const;
>
> - SDValue LowerLOAD(SDValue Op, SelectionDAG &DAG) const;
> SDValue LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const;
> SDValue LowerBRCOND(SDValue Op, SelectionDAG &DAG) const;
>
> diff --git a/lib/Target/R600/SIInstructions.td b/lib/Target/R600/SIInstructions.td
> index bc6d604..15a3a98 100644
> --- a/lib/Target/R600/SIInstructions.td
> +++ b/lib/Target/R600/SIInstructions.td
> @@ -1337,9 +1337,8 @@ def : Pat <
> /********** ===================== **********/
>
> def : Pat <
> - (int_SI_fs_interp_constant imm:$attr_chan, imm:$attr, SReg_32:$params),
> - (V_INTERP_MOV_F32 INTERP.P0, imm:$attr_chan, imm:$attr,
> - (S_MOV_B32 SReg_32:$params))
> + (int_SI_fs_interp_constant imm:$attr_chan, imm:$attr, M0Reg:$params),
> + (V_INTERP_MOV_F32 INTERP.P0, imm:$attr_chan, imm:$attr, M0Reg:$params)
> >;
>
> def : Pat <
> --
> 1.7.10.4
>
> From fa0ee67efe2f38eef47579fe9a55ce03f08dec4e Mon Sep 17 00:00:00 2001
> From: =?UTF-8?q?Christian=20K=C3=B6nig?= <christian.koenig at amd.com>
> Date: Tue, 5 Mar 2013 11:12:42 +0100
> Subject: [PATCH 6/7] R600/SI: remove SI_vs_load_buffer_index
> MIME-Version: 1.0
> Content-Type: text/plain; charset=UTF-8
> Content-Transfer-Encoding: 8bit
>
> Signed-off-by: Christian K??nig <christian.koenig at amd.com>
> ---
> lib/Target/R600/SIISelLowering.cpp | 14 --------------
> lib/Target/R600/SIIntrinsics.td | 1 -
> 2 files changed, 15 deletions(-)
>
> diff --git a/lib/Target/R600/SIISelLowering.cpp b/lib/Target/R600/SIISelLowering.cpp
> index a65e06f..34ab229 100644
> --- a/lib/Target/R600/SIISelLowering.cpp
> +++ b/lib/Target/R600/SIISelLowering.cpp
> @@ -60,8 +60,6 @@ SITargetLowering::SITargetLowering(TargetMachine &TM) :
> setOperationAction(ISD::ADD, MVT::i64, Legal);
> setOperationAction(ISD::ADD, MVT::i32, Legal);
>
> - setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
> -
> setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
> setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
>
> @@ -241,18 +239,6 @@ SDValue SITargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
> default: return AMDGPUTargetLowering::LowerOperation(Op, DAG);
> case ISD::BRCOND: return LowerBRCOND(Op, DAG);
> case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
> - case ISD::INTRINSIC_WO_CHAIN: {
> - unsigned IntrinsicID =
> - cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
> - EVT VT = Op.getValueType();
> - switch (IntrinsicID) {
> - case AMDGPUIntrinsic::SI_vs_load_buffer_index:
> - return CreateLiveInRegister(DAG, &AMDGPU::VReg_32RegClass,
> - AMDGPU::VGPR0, VT);
> - default: return AMDGPUTargetLowering::LowerOperation(Op, DAG);
> - }
> - break;
> - }
> }
> return SDValue();
> }
> diff --git a/lib/Target/R600/SIIntrinsics.td b/lib/Target/R600/SIIntrinsics.td
> index 04308d8..aa17510 100644
> --- a/lib/Target/R600/SIIntrinsics.td
> +++ b/lib/Target/R600/SIIntrinsics.td
> @@ -18,7 +18,6 @@ let TargetPrefix = "SI", isTarget = 1 in {
> def int_SI_export : Intrinsic <[], [llvm_i32_ty, llvm_i32_ty, llvm_i32_ty, llvm_i32_ty, llvm_i32_ty, llvm_float_ty, llvm_float_ty, llvm_float_ty, llvm_float_ty], []>;
> /* XXX: We may need a seperate intrinsic here for loading integer values */
> def int_SI_load_const : Intrinsic <[llvm_float_ty], [llvm_i64_ty, llvm_i32_ty], []>;
> - def int_SI_vs_load_buffer_index : Intrinsic <[llvm_i32_ty], [], [IntrNoMem]>;
> def int_SI_vs_load_input : Intrinsic <[llvm_v4f32_ty], [llvm_v16i8_ty, llvm_i16_ty, llvm_i32_ty], [IntrReadMem]> ;
> def int_SI_wqm : Intrinsic <[], [], []>;
>
> --
> 1.7.10.4
>
> From 87e1c8a14288c05f435e180d6ab52a37c075d3ef Mon Sep 17 00:00:00 2001
> From: =?UTF-8?q?Christian=20K=C3=B6nig?= <christian.koenig at amd.com>
> Date: Sat, 2 Mar 2013 18:47:20 +0100
> Subject: [PATCH 7/7] R600/SI: rework input interpolation
> MIME-Version: 1.0
> Content-Type: text/plain; charset=UTF-8
> Content-Transfer-Encoding: 8bit
>
> Signed-off-by: Christian K??nig <christian.koenig at amd.com>
> ---
> lib/Target/R600/AMDGPU.h | 1 -
> lib/Target/R600/AMDGPUAsmPrinter.cpp | 2 +-
> lib/Target/R600/AMDGPUTargetMachine.cpp | 5 -
> lib/Target/R600/SIAssignInterpRegs.cpp | 152 -----------------------------
> lib/Target/R600/SIISelLowering.cpp | 69 +++++++------
> lib/Target/R600/SIISelLowering.h | 4 -
> lib/Target/R600/SIInstructions.td | 67 ++-----------
> lib/Target/R600/SIIntrinsics.td | 13 +--
> lib/Target/R600/SIMachineFunctionInfo.cpp | 4 +-
> lib/Target/R600/SIMachineFunctionInfo.h | 2 +-
> lib/Target/R600/SIRegisterInfo.td | 47 +--------
> 11 files changed, 46 insertions(+), 320 deletions(-)
> delete mode 100644 lib/Target/R600/SIAssignInterpRegs.cpp
>
> diff --git a/lib/Target/R600/AMDGPU.h b/lib/Target/R600/AMDGPU.h
> index ba87918..0eb7b94 100644
> --- a/lib/Target/R600/AMDGPU.h
> +++ b/lib/Target/R600/AMDGPU.h
> @@ -27,7 +27,6 @@ FunctionPass *createR600LowerConstCopy(TargetMachine &tm);
>
> // SI Passes
> FunctionPass *createSIAnnotateControlFlowPass();
> -FunctionPass *createSIAssignInterpRegsPass(TargetMachine &tm);
> FunctionPass *createSILowerControlFlowPass(TargetMachine &tm);
> FunctionPass *createSICodeEmitterPass(formatted_raw_ostream &OS);
> FunctionPass *createSIInsertWaits(TargetMachine &tm);
> diff --git a/lib/Target/R600/AMDGPUAsmPrinter.cpp b/lib/Target/R600/AMDGPUAsmPrinter.cpp
> index c30dbe4..f600144 100644
> --- a/lib/Target/R600/AMDGPUAsmPrinter.cpp
> +++ b/lib/Target/R600/AMDGPUAsmPrinter.cpp
> @@ -141,5 +141,5 @@ void AMDGPUAsmPrinter::EmitProgramInfo(MachineFunction &MF) {
> SIMachineFunctionInfo * MFI = MF.getInfo<SIMachineFunctionInfo>();
> OutStreamer.EmitIntValue(MaxSGPR + 1, 4);
> OutStreamer.EmitIntValue(MaxVGPR + 1, 4);
> - OutStreamer.EmitIntValue(MFI->SPIPSInputAddr, 4);
> + OutStreamer.EmitIntValue(MFI->PSInputAddr, 4);
> }
> diff --git a/lib/Target/R600/AMDGPUTargetMachine.cpp b/lib/Target/R600/AMDGPUTargetMachine.cpp
> index e2f00be..31cd9b5 100644
> --- a/lib/Target/R600/AMDGPUTargetMachine.cpp
> +++ b/lib/Target/R600/AMDGPUTargetMachine.cpp
> @@ -112,11 +112,6 @@ bool AMDGPUPassConfig::addInstSelector() {
> }
>
> bool AMDGPUPassConfig::addPreRegAlloc() {
> - const AMDGPUSubtarget &ST = TM->getSubtarget<AMDGPUSubtarget>();
> -
> - if (ST.device()->getGeneration() > AMDGPUDeviceInfo::HD6XXX) {
> - addPass(createSIAssignInterpRegsPass(*TM));
> - }
> addPass(createAMDGPUConvertToISAPass(*TM));
> return false;
> }
> diff --git a/lib/Target/R600/SIAssignInterpRegs.cpp b/lib/Target/R600/SIAssignInterpRegs.cpp
> deleted file mode 100644
> index 832e44d..0000000
> --- a/lib/Target/R600/SIAssignInterpRegs.cpp
> +++ /dev/null
> @@ -1,152 +0,0 @@
> -//===-- SIAssignInterpRegs.cpp - Assign interpolation registers -----------===//
> -//
> -// The LLVM Compiler Infrastructure
> -//
> -// This file is distributed under the University of Illinois Open Source
> -// License. See LICENSE.TXT for details.
> -//
> -//===----------------------------------------------------------------------===//
> -//
> -/// \file
> -/// \brief This pass maps the pseudo interpolation registers to the correct physical
> -/// registers.
> -//
> -/// Prior to executing a fragment shader, the GPU loads interpolation
> -/// parameters into physical registers. The specific physical register that each
> -/// interpolation parameter ends up in depends on the type of the interpolation
> -/// parameter as well as how many interpolation parameters are used by the
> -/// shader.
> -//
> -//===----------------------------------------------------------------------===//
> -
> -
> -
> -#include "AMDGPU.h"
> -#include "AMDIL.h"
> -#include "SIMachineFunctionInfo.h"
> -#include "llvm/CodeGen/MachineFunctionPass.h"
> -#include "llvm/CodeGen/MachineInstrBuilder.h"
> -#include "llvm/CodeGen/MachineRegisterInfo.h"
> -
> -using namespace llvm;
> -
> -namespace {
> -
> -class SIAssignInterpRegsPass : public MachineFunctionPass {
> -
> -private:
> - static char ID;
> - TargetMachine &TM;
> -
> - void addLiveIn(MachineFunction * MF, MachineRegisterInfo & MRI,
> - unsigned physReg, unsigned virtReg);
> -
> -public:
> - SIAssignInterpRegsPass(TargetMachine &tm) :
> - MachineFunctionPass(ID), TM(tm) { }
> -
> - virtual bool runOnMachineFunction(MachineFunction &MF);
> -
> - const char *getPassName() const { return "SI Assign intrpolation registers"; }
> -};
> -
> -} // End anonymous namespace
> -
> -char SIAssignInterpRegsPass::ID = 0;
> -
> -#define INTERP_VALUES 16
> -#define REQUIRED_VALUE_MAX_INDEX 7
> -
> -struct InterpInfo {
> - bool Enabled;
> - unsigned Regs[3];
> - unsigned RegCount;
> -};
> -
> -
> -FunctionPass *llvm::createSIAssignInterpRegsPass(TargetMachine &tm) {
> - return new SIAssignInterpRegsPass(tm);
> -}
> -
> -bool SIAssignInterpRegsPass::runOnMachineFunction(MachineFunction &MF) {
> -
> - struct InterpInfo InterpUse[INTERP_VALUES] = {
> - {false, {AMDGPU::PERSP_SAMPLE_I, AMDGPU::PERSP_SAMPLE_J}, 2},
> - {false, {AMDGPU::PERSP_CENTER_I, AMDGPU::PERSP_CENTER_J}, 2},
> - {false, {AMDGPU::PERSP_CENTROID_I, AMDGPU::PERSP_CENTROID_J}, 2},
> - {false, {AMDGPU::PERSP_I_W, AMDGPU::PERSP_J_W, AMDGPU::PERSP_1_W}, 3},
> - {false, {AMDGPU::LINEAR_SAMPLE_I, AMDGPU::LINEAR_SAMPLE_J}, 2},
> - {false, {AMDGPU::LINEAR_CENTER_I, AMDGPU::LINEAR_CENTER_J}, 2},
> - {false, {AMDGPU::LINEAR_CENTROID_I, AMDGPU::LINEAR_CENTROID_J}, 2},
> - {false, {AMDGPU::LINE_STIPPLE_TEX_COORD}, 1},
> - {false, {AMDGPU::POS_X_FLOAT}, 1},
> - {false, {AMDGPU::POS_Y_FLOAT}, 1},
> - {false, {AMDGPU::POS_Z_FLOAT}, 1},
> - {false, {AMDGPU::POS_W_FLOAT}, 1},
> - {false, {AMDGPU::FRONT_FACE}, 1},
> - {false, {AMDGPU::ANCILLARY}, 1},
> - {false, {AMDGPU::SAMPLE_COVERAGE}, 1},
> - {false, {AMDGPU::POS_FIXED_PT}, 1}
> - };
> -
> - SIMachineFunctionInfo * MFI = MF.getInfo<SIMachineFunctionInfo>();
> - // This pass is only needed for pixel shaders.
> - if (MFI->ShaderType != ShaderType::PIXEL) {
> - return false;
> - }
> - MachineRegisterInfo &MRI = MF.getRegInfo();
> - bool ForceEnable = true;
> -
> - // First pass, mark the interpolation values that are used.
> - for (unsigned InterpIdx = 0; InterpIdx < INTERP_VALUES; InterpIdx++) {
> - for (unsigned RegIdx = 0; RegIdx < InterpUse[InterpIdx].RegCount;
> - RegIdx++) {
> - InterpUse[InterpIdx].Enabled = InterpUse[InterpIdx].Enabled ||
> - !MRI.use_empty(InterpUse[InterpIdx].Regs[RegIdx]);
> - if (InterpUse[InterpIdx].Enabled &&
> - InterpIdx <= REQUIRED_VALUE_MAX_INDEX) {
> - ForceEnable = false;
> - }
> - }
> - }
> -
> - // At least one interpolation mode must be enabled or else the GPU will hang.
> - if (ForceEnable) {
> - InterpUse[0].Enabled = true;
> - }
> -
> - unsigned UsedVgprs = 0;
> -
> - // Second pass, replace with VGPRs.
> - for (unsigned InterpIdx = 0; InterpIdx < INTERP_VALUES; InterpIdx++) {
> - if (!InterpUse[InterpIdx].Enabled) {
> - continue;
> - }
> - MFI->SPIPSInputAddr |= (1 << InterpIdx);
> -
> - for (unsigned RegIdx = 0; RegIdx < InterpUse[InterpIdx].RegCount;
> - RegIdx++, UsedVgprs++) {
> - unsigned NewReg = AMDGPU::VReg_32RegClass.getRegister(UsedVgprs);
> - unsigned VirtReg = MRI.createVirtualRegister(&AMDGPU::VReg_32RegClass);
> - MRI.replaceRegWith(InterpUse[InterpIdx].Regs[RegIdx], VirtReg);
> - addLiveIn(&MF, MRI, NewReg, VirtReg);
> - }
> - }
> -
> - return false;
> -}
> -
> -void SIAssignInterpRegsPass::addLiveIn(MachineFunction * MF,
> - MachineRegisterInfo & MRI,
> - unsigned physReg, unsigned virtReg) {
> - const TargetInstrInfo * TII = TM.getInstrInfo();
> - if (!MRI.isLiveIn(physReg)) {
> - MRI.addLiveIn(physReg, virtReg);
> - MF->front().addLiveIn(physReg);
> - BuildMI(MF->front(), MF->front().begin(), DebugLoc(),
> - TII->get(TargetOpcode::COPY), virtReg)
> - .addReg(physReg);
> - } else {
> - MRI.replaceRegWith(virtReg, MRI.getLiveInVirtReg(physReg));
> - }
> -}
> diff --git a/lib/Target/R600/SIISelLowering.cpp b/lib/Target/R600/SIISelLowering.cpp
> index 34ab229..fead115 100644
> --- a/lib/Target/R600/SIISelLowering.cpp
> +++ b/lib/Target/R600/SIISelLowering.cpp
> @@ -14,6 +14,7 @@
>
> #include "SIISelLowering.h"
> #include "AMDIL.h"
> +#include "AMDGPU.h"
> #include "AMDILIntrinsicInfo.h"
> #include "SIInstrInfo.h"
> #include "SIMachineFunctionInfo.h"
> @@ -81,14 +82,32 @@ SDValue SITargetLowering::LowerFormalArguments(
>
> MachineFunction &MF = DAG.getMachineFunction();
> FunctionType *FType = MF.getFunction()->getFunctionType();
> + SIMachineFunctionInfo *Info = MF.getInfo<SIMachineFunctionInfo>();
>
> assert(CallConv == CallingConv::C);
>
> SmallVector<ISD::InputArg, 16> Splits;
> - for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
> + uint32_t Skipped = 0;
> +
> + for (unsigned i = 0, e = Ins.size(), PSInputNum = 0; i != e; ++i) {
> const ISD::InputArg &Arg = Ins[i];
>
> - // Split vertices into their elements
> + // First check if it's a PS input addr
> + if (Info->ShaderType == ShaderType::PIXEL && !Arg.Flags.isInReg()) {
> +
> + assert((PSInputNum <= 15) && "Too many PS inputs!");
> +
> + if (!Arg.Used) {
> + // We can savely skip PS inputs
> + Skipped |= 1 << i;
> + ++PSInputNum;
> + continue;
> + }
> +
> + Info->PSInputAddr |= 1 << PSInputNum++;
> + }
> +
> + // Second split vertices into their elements
> if (Arg.VT.isVector()) {
> ISD::InputArg NewArg = Arg;
> NewArg.Flags.setSplit();
> @@ -114,10 +133,22 @@ SDValue SITargetLowering::LowerFormalArguments(
> CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
> getTargetMachine(), ArgLocs, *DAG.getContext());
>
> + // At least one interpolation mode must be enabled or else the GPU will hang.
> + if (Info->ShaderType == ShaderType::PIXEL && (Info->PSInputAddr & 0x7F) == 0) {
> + Info->PSInputAddr |= 1;
> + CCInfo.AllocateReg(AMDGPU::VGPR0);
> + CCInfo.AllocateReg(AMDGPU::VGPR1);
> + }
> +
> AnalyzeFormalArguments(CCInfo, Splits);
>
> for (unsigned i = 0, e = Ins.size(), ArgIdx = 0; i != e; ++i) {
>
> + if (Skipped & (1 << i)) {
> + InVals.push_back(SDValue());
> + continue;
> + }
> +
> CCValAssign &VA = ArgLocs[ArgIdx++];
> assert(VA.isRegLoc() && "Parameter must be in a register!");
>
> @@ -177,9 +208,6 @@ MachineBasicBlock * SITargetLowering::EmitInstrWithCustomInserter(
> default:
> return AMDGPUTargetLowering::EmitInstrWithCustomInserter(MI, BB);
> case AMDGPU::BRANCH: return BB;
> - case AMDGPU::SI_INTERP:
> - LowerSI_INTERP(MI, *BB, I, MRI);
> - break;
> case AMDGPU::SI_WQM:
> LowerSI_WQM(MI, *BB, I, MRI);
> break;
> @@ -195,37 +223,6 @@ void SITargetLowering::LowerSI_WQM(MachineInstr *MI, MachineBasicBlock &BB,
> MI->eraseFromParent();
> }
>
> -void SITargetLowering::LowerSI_INTERP(MachineInstr *MI, MachineBasicBlock &BB,
> - MachineBasicBlock::iterator I, MachineRegisterInfo & MRI) const {
> - unsigned tmp = MRI.createVirtualRegister(&AMDGPU::VReg_32RegClass);
> - unsigned M0 = MRI.createVirtualRegister(&AMDGPU::M0RegRegClass);
> - MachineOperand dst = MI->getOperand(0);
> - MachineOperand iReg = MI->getOperand(1);
> - MachineOperand jReg = MI->getOperand(2);
> - MachineOperand attr_chan = MI->getOperand(3);
> - MachineOperand attr = MI->getOperand(4);
> - MachineOperand params = MI->getOperand(5);
> -
> - BuildMI(BB, I, BB.findDebugLoc(I), TII->get(AMDGPU::S_MOV_B32), M0)
> - .addOperand(params);
> -
> - BuildMI(BB, I, BB.findDebugLoc(I), TII->get(AMDGPU::V_INTERP_P1_F32), tmp)
> - .addOperand(iReg)
> - .addOperand(attr_chan)
> - .addOperand(attr)
> - .addReg(M0);
> -
> - BuildMI(BB, I, BB.findDebugLoc(I), TII->get(AMDGPU::V_INTERP_P2_F32))
> - .addOperand(dst)
> - .addReg(tmp)
> - .addOperand(jReg)
> - .addOperand(attr_chan)
> - .addOperand(attr)
> - .addReg(M0);
> -
> - MI->eraseFromParent();
> -}
> -
> EVT SITargetLowering::getSetCCResultType(EVT VT) const {
> return MVT::i1;
> }
> diff --git a/lib/Target/R600/SIISelLowering.h b/lib/Target/R600/SIISelLowering.h
> index 8ae1400..0411565 100644
> --- a/lib/Target/R600/SIISelLowering.h
> +++ b/lib/Target/R600/SIISelLowering.h
> @@ -24,10 +24,6 @@ class SITargetLowering : public AMDGPUTargetLowering {
> const SIInstrInfo * TII;
> const TargetRegisterInfo * TRI;
>
> - void LowerMOV_IMM(MachineInstr *MI, MachineBasicBlock &BB,
> - MachineBasicBlock::iterator I, unsigned Opocde) const;
> - void LowerSI_INTERP(MachineInstr *MI, MachineBasicBlock &BB,
> - MachineBasicBlock::iterator I, MachineRegisterInfo & MRI) const;
> void LowerSI_WQM(MachineInstr *MI, MachineBasicBlock &BB,
> MachineBasicBlock::iterator I, MachineRegisterInfo & MRI) const;
>
> diff --git a/lib/Target/R600/SIInstructions.td b/lib/Target/R600/SIInstructions.td
> index 15a3a98..0ab9e4e 100644
> --- a/lib/Target/R600/SIInstructions.td
> +++ b/lib/Target/R600/SIInstructions.td
> @@ -1044,13 +1044,6 @@ def S_ABSDIFF_I32 : SOP2_32 <0x0000002c, "S_ABSDIFF_I32", []>;
>
> let isCodeGenOnly = 1, isPseudo = 1 in {
>
> -def SET_M0 : InstSI <
> - (outs SReg_32:$dst),
> - (ins i32imm:$src0),
> - "SET_M0 $dst, $src0",
> - [(set SReg_32:$dst, (int_SI_set_M0 imm:$src0))]
> ->;
> -
> def LOAD_CONST : AMDGPUShaderInst <
> (outs GPRF32:$dst),
> (ins i32imm:$src),
> @@ -1060,13 +1053,6 @@ def LOAD_CONST : AMDGPUShaderInst <
>
> let usesCustomInserter = 1 in {
>
> -def SI_INTERP : InstSI <
> - (outs VReg_32:$dst),
> - (ins VReg_32:$i, VReg_32:$j, i32imm:$attr_chan, i32imm:$attr, SReg_32:$params),
> - "SI_INTERP $dst, $i, $j, $attr_chan, $attr, $params",
> - []
> ->;
> -
> def SI_WQM : InstSI <
> (outs),
> (ins),
> @@ -1337,57 +1323,16 @@ def : Pat <
> /********** ===================== **********/
>
> def : Pat <
> - (int_SI_fs_interp_constant imm:$attr_chan, imm:$attr, M0Reg:$params),
> + (int_SI_fs_constant imm:$attr_chan, imm:$attr, M0Reg:$params),
> (V_INTERP_MOV_F32 INTERP.P0, imm:$attr_chan, imm:$attr, M0Reg:$params)
> >;
>
> def : Pat <
> - (int_SI_fs_interp_linear_center imm:$attr_chan, imm:$attr, SReg_32:$params),
> - (SI_INTERP (f32 LINEAR_CENTER_I), (f32 LINEAR_CENTER_J), imm:$attr_chan,
> - imm:$attr, SReg_32:$params)
> ->;
> -
> -def : Pat <
> - (int_SI_fs_interp_linear_centroid imm:$attr_chan, imm:$attr, SReg_32:$params),
> - (SI_INTERP (f32 LINEAR_CENTROID_I), (f32 LINEAR_CENTROID_J), imm:$attr_chan,
> - imm:$attr, SReg_32:$params)
> ->;
> -
> -def : Pat <
> - (int_SI_fs_interp_persp_center imm:$attr_chan, imm:$attr, SReg_32:$params),
> - (SI_INTERP (f32 PERSP_CENTER_I), (f32 PERSP_CENTER_J), imm:$attr_chan,
> - imm:$attr, SReg_32:$params)
> ->;
> -
> -def : Pat <
> - (int_SI_fs_interp_persp_centroid imm:$attr_chan, imm:$attr, SReg_32:$params),
> - (SI_INTERP (f32 PERSP_CENTROID_I), (f32 PERSP_CENTROID_J), imm:$attr_chan,
> - imm:$attr, SReg_32:$params)
> ->;
> -
> -def : Pat <
> - (int_SI_fs_read_face),
> - (f32 FRONT_FACE)
> ->;
> -
> -def : Pat <
> - (int_SI_fs_read_pos 0),
> - (f32 POS_X_FLOAT)
> ->;
> -
> -def : Pat <
> - (int_SI_fs_read_pos 1),
> - (f32 POS_Y_FLOAT)
> ->;
> -
> -def : Pat <
> - (int_SI_fs_read_pos 2),
> - (f32 POS_Z_FLOAT)
> ->;
> -
> -def : Pat <
> - (int_SI_fs_read_pos 3),
> - (f32 POS_W_FLOAT)
> + (int_SI_fs_interp imm:$attr_chan, imm:$attr, M0Reg:$params, VReg_64:$ij),
> + (V_INTERP_P2_F32 (V_INTERP_P1_F32 (EXTRACT_SUBREG VReg_64:$ij, sub0),
> + imm:$attr_chan, imm:$attr, M0Reg:$params),
> + (EXTRACT_SUBREG VReg_64:$ij, sub1),
> + imm:$attr_chan, imm:$attr, M0Reg:$params)
> >;
>
> /********** ================== **********/
> diff --git a/lib/Target/R600/SIIntrinsics.td b/lib/Target/R600/SIIntrinsics.td
> index aa17510..7c23d17 100644
> --- a/lib/Target/R600/SIIntrinsics.td
> +++ b/lib/Target/R600/SIIntrinsics.td
> @@ -29,17 +29,8 @@ let TargetPrefix = "SI", isTarget = 1 in {
>
> /* Interpolation Intrinsics */
>
> - def int_SI_set_M0 : Intrinsic <[llvm_i32_ty], [llvm_i32_ty]>;
> - class Interp : Intrinsic <[llvm_float_ty], [llvm_i32_ty, llvm_i32_ty, llvm_i32_ty], [IntrReadMem]>;
> -
> - def int_SI_fs_interp_linear_center : Interp;
> - def int_SI_fs_interp_linear_centroid : Interp;
> - def int_SI_fs_interp_persp_center : Interp;
> - def int_SI_fs_interp_persp_centroid : Interp;
> - def int_SI_fs_interp_constant : Interp;
> -
> - def int_SI_fs_read_face : Intrinsic <[llvm_float_ty], [], [IntrNoMem]>;
> - def int_SI_fs_read_pos : Intrinsic <[llvm_float_ty], [llvm_i32_ty], [IntrNoMem]>;
> + def int_SI_fs_constant : Intrinsic <[llvm_float_ty], [llvm_i32_ty, llvm_i32_ty, llvm_i32_ty], [IntrReadMem]>;
> + def int_SI_fs_interp : Intrinsic <[llvm_float_ty], [llvm_i32_ty, llvm_i32_ty, llvm_i32_ty, llvm_v2i32_ty], [IntrReadMem]>;
>
> /* Control flow Intrinsics */
>
> diff --git a/lib/Target/R600/SIMachineFunctionInfo.cpp b/lib/Target/R600/SIMachineFunctionInfo.cpp
> index e5fbf05..1a4e4cb 100644
> --- a/lib/Target/R600/SIMachineFunctionInfo.cpp
> +++ b/lib/Target/R600/SIMachineFunctionInfo.cpp
> @@ -19,8 +19,8 @@ const char *SIMachineFunctionInfo::ShaderTypeAttribute = "ShaderType";
>
> SIMachineFunctionInfo::SIMachineFunctionInfo(const MachineFunction &MF)
> : MachineFunctionInfo(),
> - SPIPSInputAddr(0),
> - ShaderType(0) {
> + ShaderType(0),
> + PSInputAddr(0) {
>
> AttributeSet Set = MF.getFunction()->getAttributes();
> Attribute A = Set.getAttribute(AttributeSet::FunctionIndex,
> diff --git a/lib/Target/R600/SIMachineFunctionInfo.h b/lib/Target/R600/SIMachineFunctionInfo.h
> index e2f97a4..91a809b 100644
> --- a/lib/Target/R600/SIMachineFunctionInfo.h
> +++ b/lib/Target/R600/SIMachineFunctionInfo.h
> @@ -26,8 +26,8 @@ public:
> static const char *ShaderTypeAttribute;
>
> SIMachineFunctionInfo(const MachineFunction &MF);
> - unsigned SPIPSInputAddr;
> unsigned ShaderType;
> + unsigned PSInputAddr;
> };
>
> } // End namespace llvm
> diff --git a/lib/Target/R600/SIRegisterInfo.td b/lib/Target/R600/SIRegisterInfo.td
> index c4bca2c..3dcad50 100644
> --- a/lib/Target/R600/SIRegisterInfo.td
> +++ b/lib/Target/R600/SIRegisterInfo.td
> @@ -34,32 +34,6 @@ foreach Index = 0-255 in {
> }
> }
>
> -// virtual Interpolation registers
> -def PERSP_SAMPLE_I : SIReg <"PERSP_SAMPLE_I">;
> -def PERSP_SAMPLE_J : SIReg <"PERSP_SAMPLE_J">;
> -def PERSP_CENTER_I : SIReg <"PERSP_CENTER_I">;
> -def PERSP_CENTER_J : SIReg <"PERSP_CENTER_J">;
> -def PERSP_CENTROID_I : SIReg <"PERSP_CENTROID_I">;
> -def PERSP_CENTROID_J : SIReg <"PERP_CENTROID_J">;
> -def PERSP_I_W : SIReg <"PERSP_I_W">;
> -def PERSP_J_W : SIReg <"PERSP_J_W">;
> -def PERSP_1_W : SIReg <"PERSP_1_W">;
> -def LINEAR_SAMPLE_I : SIReg <"LINEAR_SAMPLE_I">;
> -def LINEAR_SAMPLE_J : SIReg <"LINEAR_SAMPLE_J">;
> -def LINEAR_CENTER_I : SIReg <"LINEAR_CENTER_I">;
> -def LINEAR_CENTER_J : SIReg <"LINEAR_CENTER_J">;
> -def LINEAR_CENTROID_I : SIReg <"LINEAR_CENTROID_I">;
> -def LINEAR_CENTROID_J : SIReg <"LINEAR_CENTROID_J">;
> -def LINE_STIPPLE_TEX_COORD : SIReg <"LINE_STIPPLE_TEX_COORD">;
> -def POS_X_FLOAT : SIReg <"POS_X_FLOAT">;
> -def POS_Y_FLOAT : SIReg <"POS_Y_FLOAT">;
> -def POS_Z_FLOAT : SIReg <"POS_Z_FLOAT">;
> -def POS_W_FLOAT : SIReg <"POS_W_FLOAT">;
> -def FRONT_FACE : SIReg <"FRONT_FACE">;
> -def ANCILLARY : SIReg <"ANCILLARY">;
> -def SAMPLE_COVERAGE : SIReg <"SAMPLE_COVERAGE">;
> -def POS_FIXED_PT : SIReg <"POS_FIXED_PT">;
> -
> //===----------------------------------------------------------------------===//
> // Groupings using register classes and tuples
> //===----------------------------------------------------------------------===//
> @@ -202,26 +176,7 @@ def SSrc_32 : RegisterClass<"AMDGPU", [i32, f32], 32, (add SReg_32)>;
>
> def SSrc_64 : RegisterClass<"AMDGPU", [i64, i1], 64, (add SReg_64)>;
>
> -def VSrc_32 : RegisterClass<"AMDGPU", [i32, f32], 32,
> - (add VReg_32, SReg_32,
> - PERSP_SAMPLE_I, PERSP_SAMPLE_J,
> - PERSP_CENTER_I, PERSP_CENTER_J,
> - PERSP_CENTROID_I, PERSP_CENTROID_J,
> - PERSP_I_W, PERSP_J_W, PERSP_1_W,
> - LINEAR_SAMPLE_I, LINEAR_SAMPLE_J,
> - LINEAR_CENTER_I, LINEAR_CENTER_J,
> - LINEAR_CENTROID_I, LINEAR_CENTROID_J,
> - LINE_STIPPLE_TEX_COORD,
> - POS_X_FLOAT,
> - POS_Y_FLOAT,
> - POS_Z_FLOAT,
> - POS_W_FLOAT,
> - FRONT_FACE,
> - ANCILLARY,
> - SAMPLE_COVERAGE,
> - POS_FIXED_PT
> - )
> ->;
> +def VSrc_32 : RegisterClass<"AMDGPU", [i32, f32], 32, (add VReg_32, SReg_32)>;
>
> def VSrc_64 : RegisterClass<"AMDGPU", [i64], 64, (add VReg_64, SReg_64)>;
>
> --
> 1.7.10.4
>
More information about the llvm-commits
mailing list