From: Matt Arsenault <Matthew.Arsenault@amd.com> Date: Tue, 5 Jul 2016 17:09:01 +0000 (+0000) Subject: AMDGPU: Fix folding SGPRs into madak/madmk src0 X-Git-Url: https://granicus.if.org/sourcecode?a=commitdiff_plain;h=6de48c50fea935049fbed2c4027f36bcdd462dc5;p=llvm AMDGPU: Fix folding SGPRs into madak/madmk src0 Because of the special immediate operand, the constant bus is already used so SGPRs are never useful. r263212 changed the name of the immediate operand, which broke the verifier check for the restriction. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@274564 91177308-0d34-0410-b5e6-96231b3b80d8 --- diff --git a/lib/Target/AMDGPU/AMDGPUInstructions.td b/lib/Target/AMDGPU/AMDGPUInstructions.td index 560db41eff6..d77f608fb22 100644 --- a/lib/Target/AMDGPU/AMDGPUInstructions.td +++ b/lib/Target/AMDGPU/AMDGPUInstructions.td @@ -49,6 +49,13 @@ def UnsafeFPMath : Predicate<"TM.Options.UnsafeFPMath">; def InstFlag : OperandWithDefaultOps <i32, (ops (i32 0))>; def ADDRIndirect : ComplexPattern<iPTR, 2, "SelectADDRIndirect", [], []>; +// 32-bit VALU immediate operand that uses the constant bus. +def u32kimm : Operand<i32> { + let OperandNamespace = "AMDGPU"; + let OperandType = "OPERAND_KIMM32"; + let PrintMethod = "printU32ImmOperand"; +} + let OperandType = "OPERAND_IMMEDIATE" in { def u32imm : Operand<i32> { diff --git a/lib/Target/AMDGPU/SIDefines.h b/lib/Target/AMDGPU/SIDefines.h index 23cde1c4bc1..7a5ec05c26c 100644 --- a/lib/Target/AMDGPU/SIDefines.h +++ b/lib/Target/AMDGPU/SIDefines.h @@ -50,7 +50,12 @@ namespace AMDGPU { /// Operand with register or 32-bit immediate OPERAND_REG_IMM32 = MCOI::OPERAND_FIRST_TARGET, /// Operand with register or inline constant - OPERAND_REG_INLINE_C + OPERAND_REG_INLINE_C, + + /// Operand with 32-bit immediate that uses the constant bus. The standard + /// OPERAND_IMMEDIATE should be used for special immediates such as source + /// modifiers. + OPERAND_KIMM32 }; } } diff --git a/lib/Target/AMDGPU/SIInstrInfo.cpp b/lib/Target/AMDGPU/SIInstrInfo.cpp index 159fb72f89f..fc7aca22cef 100644 --- a/lib/Target/AMDGPU/SIInstrInfo.cpp +++ b/lib/Target/AMDGPU/SIInstrInfo.cpp @@ -1695,6 +1695,7 @@ bool SIInstrInfo::verifyInstruction(const MachineInstr &MI, } break; case MCOI::OPERAND_IMMEDIATE: + case AMDGPU::OPERAND_KIMM32: // Check if this operand is an immediate. // FrameIndex operands will be replaced by immediates, so they are // allowed. @@ -1731,6 +1732,10 @@ bool SIInstrInfo::verifyInstruction(const MachineInstr &MI, const int OpIndices[] = { Src0Idx, Src1Idx, Src2Idx }; unsigned ConstantBusCount = 0; + + if (AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::imm) != -1) + ++ConstantBusCount; + unsigned SGPRUsed = findImplicitSGPRRead(MI); if (SGPRUsed != AMDGPU::NoRegister) ++ConstantBusCount; @@ -2020,9 +2025,12 @@ bool SIInstrInfo::isOperandLegal(const MachineInstr &MI, unsigned OpIdx, if (i == OpIdx) continue; const MachineOperand &Op = MI.getOperand(i); - if (Op.isReg() && - (Op.getReg() != SGPRUsed.Reg || Op.getSubReg() != SGPRUsed.SubReg) && - usesConstantBus(MRI, Op, getOpSize(MI, i))) { + if (Op.isReg()) { + if ((Op.getReg() != SGPRUsed.Reg || Op.getSubReg() != SGPRUsed.SubReg) && + usesConstantBus(MRI, Op, getOpSize(MI, i))) { + return false; + } + } else if (InstDesc.OpInfo[i].OperandType == AMDGPU::OPERAND_KIMM32) { return false; } } diff --git a/lib/Target/AMDGPU/SIInstrInfo.td b/lib/Target/AMDGPU/SIInstrInfo.td index f9697057a1d..708d97e9773 100644 --- a/lib/Target/AMDGPU/SIInstrInfo.td +++ b/lib/Target/AMDGPU/SIInstrInfo.td @@ -1582,12 +1582,12 @@ def VOP_I64_I64_I64 : VOPProfile <[i64, i64, i64, untyped]>; def VOP_F32_F32_F32_F32 : VOPProfile <[f32, f32, f32, f32]>; def VOP_MADAK : VOPProfile <[f32, f32, f32, f32]> { - field dag Ins32 = (ins VCSrc_32:$src0, VGPR_32:$src1, u32imm:$imm); + field dag Ins32 = (ins VCSrc_32:$src0, VGPR_32:$src1, u32kimm:$imm); field string Asm32 = "$vdst, $src0, $src1, $imm"; field bit HasExt = 0; } def VOP_MADMK : VOPProfile <[f32, f32, f32, f32]> { - field dag Ins32 = (ins VCSrc_32:$src0, u32imm:$imm, VGPR_32:$src1); + field dag Ins32 = (ins VCSrc_32:$src0, u32kimm:$imm, VGPR_32:$src1); field string Asm32 = "$vdst, $src0, $imm, $src1"; field bit HasExt = 0; } diff --git a/test/CodeGen/AMDGPU/madak.ll b/test/CodeGen/AMDGPU/madak.ll index a394db61bfb..6ea1202ac50 100644 --- a/test/CodeGen/AMDGPU/madak.ll +++ b/test/CodeGen/AMDGPU/madak.ll @@ -191,3 +191,32 @@ define void @no_madak_src1_modifier_f32(float addrspace(1)* noalias %out, float store float %madak, float addrspace(1)* %out.gep, align 4 ret void } + +; SIFoldOperands should not fold the SGPR copy into the instruction +; because the implicit immediate already uses the constant bus. +; GCN-LABEL: {{^}}madak_constant_bus_violation: +; GCN: s_load_dword [[SGPR0:s[0-9]+]], s{{\[[0-9]+:[0-9]+\]}}, {{0xa|0x28}} +; GCN: v_mov_b32_e32 [[SGPR0_VCOPY:v[0-9]+]], [[SGPR0]] +; GCN: buffer_load_dword [[VGPR:v[0-9]+]] +; GCN: v_madak_f32_e32 [[MADAK:v[0-9]+]], 0.5, [[SGPR0_VCOPY]], 0x42280000 +; GCN: v_mul_f32_e32 [[MUL:v[0-9]+]], [[VGPR]], [[MADAK]] +; GCN: buffer_store_dword [[MUL]] +define void @madak_constant_bus_violation(i32 %arg1, float %sgpr0, float %sgpr1) #0 { +bb: + %tmp = icmp eq i32 %arg1, 0 + br i1 %tmp, label %bb3, label %bb4 + +bb3: + store volatile float 0.0, float addrspace(1)* undef + br label %bb4 + +bb4: + %vgpr = load volatile float, float addrspace(1)* undef + %tmp0 = fmul float %sgpr0, 0.5 + %tmp1 = fadd float %tmp0, 42.0 + %tmp2 = fmul float %tmp1, %vgpr + store volatile float %tmp2, float addrspace(1)* undef, align 4 + ret void +} + +attributes #0 = { nounwind}