From 883a0af53a2a4ef612e31b61a22fa4443121a2b8 Mon Sep 17 00:00:00 2001 From: Tom Stellard Date: Thu, 31 May 2012 15:58:17 -0400 Subject: [PATCH] radeon/llvm: Remove AMDIL GLOBALSTORE* instructions --- src/gallium/drivers/radeon/AMDILInstructions.td | 2 - src/gallium/drivers/radeon/R600ISelLowering.cpp | 23 ++++++++ src/gallium/drivers/radeon/R600Instructions.td | 21 ++++--- .../drivers/radeon/R600LowerInstructions.cpp | 67 ---------------------- 4 files changed, 36 insertions(+), 77 deletions(-) diff --git a/src/gallium/drivers/radeon/AMDILInstructions.td b/src/gallium/drivers/radeon/AMDILInstructions.td index 09b651c..0b26ac3 100644 --- a/src/gallium/drivers/radeon/AMDILInstructions.td +++ b/src/gallium/drivers/radeon/AMDILInstructions.td @@ -779,7 +779,6 @@ let Predicates = [Has32BitPtr] in { // Store Memory Operations //===----------------------------------------------------------------------===// defm GLOBALTRUNCSTORE : GTRUNCSTORE<"!global trunc store">; - defm GLOBALSTORE : STORE<"!global store" , global_store>; defm LOCALTRUNCSTORE : LTRUNCSTORE<"!local trunc store">; defm LOCALSTORE : STORE<"!local store" , local_store>; defm PRIVATETRUNCSTORE : PTRUNCSTORE<"!private trunc store">; @@ -1069,7 +1068,6 @@ let Predicates = [Has64BitPtr] in { // Store Memory Operations //===----------------------------------------------------------------------===// defm GLOBALTRUNCSTORE64 : GTRUNCSTORE64<"!global trunc store">; - defm GLOBALSTORE64 : STORE64<"!global store" , global_store>; defm LOCALTRUNCSTORE64 : LTRUNCSTORE64<"!local trunc store">; defm LOCALSTORE64 : STORE64<"!local store" , local_store>; defm PRIVATETRUNCSTORE64 : PTRUNCSTORE64<"!private trunc store">; diff --git a/src/gallium/drivers/radeon/R600ISelLowering.cpp b/src/gallium/drivers/radeon/R600ISelLowering.cpp index 1d4747f..9f9f348 100644 --- a/src/gallium/drivers/radeon/R600ISelLowering.cpp +++ b/src/gallium/drivers/radeon/R600ISelLowering.cpp @@ -146,6 +146,29 @@ MachineBasicBlock * R600TargetLowering::EmitInstrWithCustomInserter( return BB; } + case AMDIL::RAT_WRITE_CACHELESS_eg: + { + // Convert to DWORD address + unsigned NewAddr = MRI.createVirtualRegister( + AMDIL::R600_TReg32_XRegisterClass); + unsigned ShiftValue = MRI.createVirtualRegister( + AMDIL::R600_TReg32RegisterClass); + + // XXX In theory, we should be able to pass ShiftValue directly to + // the LSHR_eg instruction as an inline literal, but I tried doing it + // this way and it didn't produce the correct results. + BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDIL::MOV), ShiftValue) + .addReg(AMDIL::ALU_LITERAL_X) + .addImm(2); + BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDIL::LSHR_eg), NewAddr) + .addOperand(MI->getOperand(1)) + .addReg(ShiftValue); + BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(MI->getOpcode())) + .addOperand(MI->getOperand(0)) + .addReg(NewAddr); + break; + } + case AMDIL::STORE_OUTPUT: { int64_t OutputIndex = MI->getOperand(1).getImm(); diff --git a/src/gallium/drivers/radeon/R600Instructions.td b/src/gallium/drivers/radeon/R600Instructions.td index 1b2768d..e12ebab 100644 --- a/src/gallium/drivers/radeon/R600Instructions.td +++ b/src/gallium/drivers/radeon/R600Instructions.td @@ -50,6 +50,7 @@ def MEMrr : Operand { } def ADDRParam : ComplexPattern; +def ADDRDWord : ComplexPattern; class R600_ALU { @@ -164,13 +165,12 @@ def COND_LE : PatLeaf < case ISD::SETLE: return true;}}}] >; -class EG_CF_RAT cf_inst, bits <6> rat_inst, dag outs, dag ins, - string asm> : - InstR600ISA +class EG_CF_RAT cf_inst, bits <6> rat_inst, bits<4> rat_id, dag outs, + dag ins, string asm, list pattern> : + InstR600ISA { bits<7> RW_GPR; bits<7> INDEX_GPR; - bits<4> RAT_ID; bits<2> RIM; bits<2> TYPE; @@ -186,7 +186,7 @@ class EG_CF_RAT cf_inst, bits <6> rat_inst, dag outs, dag ins, bits<1> BARRIER; /* CF_ALLOC_EXPORT_WORD0_RAT */ - let Inst{3-0} = RAT_ID; + let Inst{3-0} = rat_id; let Inst{9-4} = rat_inst; let Inst{10} = 0; /* Reserved */ let Inst{12-11} = RIM; @@ -772,9 +772,12 @@ class TRIG_HELPER_r700 : Pat < let Predicates = [isEG] in { -def RAT_WRITE_CACHELESS_eg : - EG_CF_RAT <0x57, 0x2, (outs), (ins R600_TReg32_X:$rw_gpr, - R600_TReg32_X:$index_gpr, i32imm:$rat_id), ""> +let usesCustomInserter = 1 in { + +def RAT_WRITE_CACHELESS_eg : EG_CF_RAT <0x57, 0x2, 0, (outs), + (ins R600_TReg32_X:$rw_gpr, R600_TReg32_X:$index_gpr), + "RAT_WRITE_CACHELESS_eg $rw_gpr, $index_gpr", + [(global_store (i32 R600_TReg32_X:$rw_gpr), R600_TReg32_X:$index_gpr)]> { let RIM = 0; /* XXX: Have a separate instruction for non-indexed writes. */ @@ -791,6 +794,8 @@ def RAT_WRITE_CACHELESS_eg : let BARRIER = 1; } +} // End usesCustomInserter = 1 + class VTX_READ_eg pattern> : InstR600ISA < (outs R600_TReg32_X:$dst), (ins MEMxi:$ptr), diff --git a/src/gallium/drivers/radeon/R600LowerInstructions.cpp b/src/gallium/drivers/radeon/R600LowerInstructions.cpp index 932825a..99ca839 100644 --- a/src/gallium/drivers/radeon/R600LowerInstructions.cpp +++ b/src/gallium/drivers/radeon/R600LowerInstructions.cpp @@ -39,12 +39,6 @@ namespace { void lowerFLT(MachineInstr &MI); - void calcAddress(const MachineOperand &ptrOp, - const MachineOperand &indexOp, - unsigned indexReg, - MachineBasicBlock &MBB, - MachineBasicBlock::iterator I) const; - public: R600LowerInstructionsPass(TargetMachine &tm) : MachineFunctionPass(ID), TM(tm), @@ -99,47 +93,6 @@ bool R600LowerInstructionsPass::runOnMachineFunction(MachineFunction &MF) } */ /* XXX: This is an optimization */ - case AMDIL::GLOBALSTORE_i32: - case AMDIL::GLOBALSTORE_f32: - { - MachineOperand &ptrOperand = MI.getOperand(1); - MachineOperand &indexOperand = MI.getOperand(2); - unsigned rwReg = - MRI->createVirtualRegister(&AMDIL::R600_TReg32_XRegClass); - unsigned byteIndexReg = - MRI->createVirtualRegister(&AMDIL::R600_TReg32RegClass); - unsigned shiftReg = - MRI->createVirtualRegister(&AMDIL::R600_TReg32RegClass); - unsigned indexReg = - MRI->createVirtualRegister(&AMDIL::R600_TReg32_XRegClass); - - /* Move the store value to the correct register class */ - BuildMI(MBB, I, MBB.findDebugLoc(I), TII->get(AMDIL::COPY), rwReg) - .addOperand(MI.getOperand(0)); - - /* Calculate the address in the RAT */ - calcAddress(ptrOperand, indexOperand, byteIndexReg, MBB, I); - - - BuildMI(MBB, I, MBB.findDebugLoc(I), TII->get(AMDIL::MOV), shiftReg) - .addReg(AMDIL::ALU_LITERAL_X) - .addImm(2); - - /* XXX: Check GPU family */ - BuildMI(MBB, I, MBB.findDebugLoc(I), - TII->get(AMDIL::LSHR_eg), indexReg) - .addReg(byteIndexReg) - .addReg(shiftReg); - - /* XXX: Check GPU Family */ - BuildMI(MBB, I, MBB.findDebugLoc(I), - TII->get(AMDIL::RAT_WRITE_CACHELESS_eg)) - .addReg(rwReg) - .addReg(indexReg) - .addImm(0); - break; - } - case AMDIL::ILT: BuildMI(MBB, I, MBB.findDebugLoc(I), TII->get(AMDIL::SETGT_INT)) .addOperand(MI.getOperand(0)) @@ -195,23 +148,3 @@ bool R600LowerInstructionsPass::runOnMachineFunction(MachineFunction &MF) } return false; } - -void R600LowerInstructionsPass::calcAddress(const MachineOperand &ptrOp, - const MachineOperand &indexOp, - unsigned indexReg, - MachineBasicBlock &MBB, - MachineBasicBlock::iterator I) const -{ - /* Optimize the case where the indexOperand is 0 */ - if (indexOp.isImm() && indexOp.getImm() == 0) { - assert(ptrOp.isReg()); - BuildMI(MBB, I, MBB.findDebugLoc(I), - TII->get(AMDIL::COPY), indexReg) - .addOperand(ptrOp); - } else { - BuildMI(MBB, I, MBB.findDebugLoc(I), - TII->get(AMDIL::ADD_INT), indexReg) - .addOperand(indexOp) - .addOperand(ptrOp); - } -} -- 2.7.4