summaryrefslogtreecommitdiffstats
path: root/contrib/llvm/lib/Target/AMDGPU/SIFixSGPRCopies.cpp
diff options
context:
space:
mode:
Diffstat (limited to 'contrib/llvm/lib/Target/AMDGPU/SIFixSGPRCopies.cpp')
-rw-r--r--contrib/llvm/lib/Target/AMDGPU/SIFixSGPRCopies.cpp229
1 files changed, 132 insertions, 97 deletions
diff --git a/contrib/llvm/lib/Target/AMDGPU/SIFixSGPRCopies.cpp b/contrib/llvm/lib/Target/AMDGPU/SIFixSGPRCopies.cpp
index 23502b4..96e37c5 100644
--- a/contrib/llvm/lib/Target/AMDGPU/SIFixSGPRCopies.cpp
+++ b/contrib/llvm/lib/Target/AMDGPU/SIFixSGPRCopies.cpp
@@ -82,22 +82,10 @@ using namespace llvm;
namespace {
class SIFixSGPRCopies : public MachineFunctionPass {
-
-private:
+public:
static char ID;
- const TargetRegisterClass *inferRegClassFromUses(const SIRegisterInfo *TRI,
- const MachineRegisterInfo &MRI,
- unsigned Reg,
- unsigned SubReg) const;
- const TargetRegisterClass *inferRegClassFromDef(const SIRegisterInfo *TRI,
- const MachineRegisterInfo &MRI,
- unsigned Reg,
- unsigned SubReg) const;
- bool isVGPRToSGPRCopy(const MachineInstr &Copy, const SIRegisterInfo *TRI,
- const MachineRegisterInfo &MRI) const;
-public:
- SIFixSGPRCopies(TargetMachine &tm) : MachineFunctionPass(ID) { }
+ SIFixSGPRCopies() : MachineFunctionPass(ID) { }
bool runOnMachineFunction(MachineFunction &MF) override;
@@ -105,14 +93,23 @@ public:
return "SI Fix SGPR copies";
}
+ void getAnalysisUsage(AnalysisUsage &AU) const override {
+ AU.setPreservesCFG();
+ MachineFunctionPass::getAnalysisUsage(AU);
+ }
};
} // End anonymous namespace
+INITIALIZE_PASS(SIFixSGPRCopies, DEBUG_TYPE,
+ "SI Fix SGPR copies", false, false)
+
char SIFixSGPRCopies::ID = 0;
-FunctionPass *llvm::createSIFixSGPRCopiesPass(TargetMachine &tm) {
- return new SIFixSGPRCopies(tm);
+char &llvm::SIFixSGPRCopiesID = SIFixSGPRCopies::ID;
+
+FunctionPass *llvm::createSIFixSGPRCopiesPass() {
+ return new SIFixSGPRCopies();
}
static bool hasVGPROperands(const MachineInstr &MI, const SIRegisterInfo *TRI) {
@@ -128,77 +125,115 @@ static bool hasVGPROperands(const MachineInstr &MI, const SIRegisterInfo *TRI) {
return false;
}
-/// This functions walks the use list of Reg until it finds an Instruction
-/// that isn't a COPY returns the register class of that instruction.
-/// \return The register defined by the first non-COPY instruction.
-const TargetRegisterClass *SIFixSGPRCopies::inferRegClassFromUses(
- const SIRegisterInfo *TRI,
- const MachineRegisterInfo &MRI,
- unsigned Reg,
- unsigned SubReg) const {
-
- const TargetRegisterClass *RC
- = TargetRegisterInfo::isVirtualRegister(Reg) ?
- MRI.getRegClass(Reg) :
- TRI->getPhysRegClass(Reg);
-
- RC = TRI->getSubRegClass(RC, SubReg);
- for (MachineRegisterInfo::use_instr_iterator
- I = MRI.use_instr_begin(Reg), E = MRI.use_instr_end(); I != E; ++I) {
- switch (I->getOpcode()) {
- case AMDGPU::COPY:
- RC = TRI->getCommonSubClass(RC, inferRegClassFromUses(TRI, MRI,
- I->getOperand(0).getReg(),
- I->getOperand(0).getSubReg()));
- break;
- }
- }
+static std::pair<const TargetRegisterClass *, const TargetRegisterClass *>
+getCopyRegClasses(const MachineInstr &Copy,
+ const SIRegisterInfo &TRI,
+ const MachineRegisterInfo &MRI) {
+ unsigned DstReg = Copy.getOperand(0).getReg();
+ unsigned SrcReg = Copy.getOperand(1).getReg();
+
+ const TargetRegisterClass *SrcRC =
+ TargetRegisterInfo::isVirtualRegister(SrcReg) ?
+ MRI.getRegClass(SrcReg) :
+ TRI.getPhysRegClass(SrcReg);
- return RC;
+ // We don't really care about the subregister here.
+ // SrcRC = TRI.getSubRegClass(SrcRC, Copy.getOperand(1).getSubReg());
+
+ const TargetRegisterClass *DstRC =
+ TargetRegisterInfo::isVirtualRegister(DstReg) ?
+ MRI.getRegClass(DstReg) :
+ TRI.getPhysRegClass(DstReg);
+
+ return std::make_pair(SrcRC, DstRC);
}
-const TargetRegisterClass *SIFixSGPRCopies::inferRegClassFromDef(
- const SIRegisterInfo *TRI,
- const MachineRegisterInfo &MRI,
- unsigned Reg,
- unsigned SubReg) const {
- if (!TargetRegisterInfo::isVirtualRegister(Reg)) {
- const TargetRegisterClass *RC = TRI->getPhysRegClass(Reg);
- return TRI->getSubRegClass(RC, SubReg);
- }
- MachineInstr *Def = MRI.getVRegDef(Reg);
- if (Def->getOpcode() != AMDGPU::COPY) {
- return TRI->getSubRegClass(MRI.getRegClass(Reg), SubReg);
- }
+static bool isVGPRToSGPRCopy(const TargetRegisterClass *SrcRC,
+ const TargetRegisterClass *DstRC,
+ const SIRegisterInfo &TRI) {
+ return TRI.isSGPRClass(DstRC) && TRI.hasVGPRs(SrcRC);
+}
- return inferRegClassFromDef(TRI, MRI, Def->getOperand(1).getReg(),
- Def->getOperand(1).getSubReg());
+static bool isSGPRToVGPRCopy(const TargetRegisterClass *SrcRC,
+ const TargetRegisterClass *DstRC,
+ const SIRegisterInfo &TRI) {
+ return TRI.isSGPRClass(SrcRC) && TRI.hasVGPRs(DstRC);
}
-bool SIFixSGPRCopies::isVGPRToSGPRCopy(const MachineInstr &Copy,
- const SIRegisterInfo *TRI,
- const MachineRegisterInfo &MRI) const {
+// Distribute an SGPR->VGPR copy of a REG_SEQUENCE into a VGPR REG_SEQUENCE.
+//
+// SGPRx = ...
+// SGPRy = REG_SEQUENCE SGPRx, sub0 ...
+// VGPRz = COPY SGPRy
+//
+// ==>
+//
+// VGPRx = COPY SGPRx
+// VGPRz = REG_SEQUENCE VGPRx, sub0
+//
+// This exposes immediate folding opportunities when materializing 64-bit
+// immediates.
+static bool foldVGPRCopyIntoRegSequence(MachineInstr &MI,
+ const SIRegisterInfo *TRI,
+ const SIInstrInfo *TII,
+ MachineRegisterInfo &MRI) {
+ assert(MI.isRegSequence());
+
+ unsigned DstReg = MI.getOperand(0).getReg();
+ if (!TRI->isSGPRClass(MRI.getRegClass(DstReg)))
+ return false;
- unsigned DstReg = Copy.getOperand(0).getReg();
- unsigned SrcReg = Copy.getOperand(1).getReg();
- unsigned SrcSubReg = Copy.getOperand(1).getSubReg();
+ if (!MRI.hasOneUse(DstReg))
+ return false;
- if (!TargetRegisterInfo::isVirtualRegister(DstReg)) {
- // If the destination register is a physical register there isn't really
- // much we can do to fix this.
+ MachineInstr &CopyUse = *MRI.use_instr_begin(DstReg);
+ if (!CopyUse.isCopy())
return false;
- }
- const TargetRegisterClass *DstRC = MRI.getRegClass(DstReg);
+ const TargetRegisterClass *SrcRC, *DstRC;
+ std::tie(SrcRC, DstRC) = getCopyRegClasses(CopyUse, *TRI, MRI);
- const TargetRegisterClass *SrcRC;
+ if (!isSGPRToVGPRCopy(SrcRC, DstRC, *TRI))
+ return false;
- if (!TargetRegisterInfo::isVirtualRegister(SrcReg) ||
- MRI.getRegClass(SrcReg) == &AMDGPU::VReg_1RegClass)
+ // TODO: Could have multiple extracts?
+ unsigned SubReg = CopyUse.getOperand(1).getSubReg();
+ if (SubReg != AMDGPU::NoSubRegister)
return false;
- SrcRC = TRI->getSubRegClass(MRI.getRegClass(SrcReg), SrcSubReg);
- return TRI->isSGPRClass(DstRC) && TRI->hasVGPRs(SrcRC);
+ MRI.setRegClass(DstReg, DstRC);
+
+ // SGPRx = ...
+ // SGPRy = REG_SEQUENCE SGPRx, sub0 ...
+ // VGPRz = COPY SGPRy
+
+ // =>
+ // VGPRx = COPY SGPRx
+ // VGPRz = REG_SEQUENCE VGPRx, sub0
+
+ MI.getOperand(0).setReg(CopyUse.getOperand(0).getReg());
+
+ for (unsigned I = 1, N = MI.getNumOperands(); I != N; I += 2) {
+ unsigned SrcReg = MI.getOperand(I).getReg();
+ unsigned SrcSubReg = MI.getOperand(I).getReg();
+
+ const TargetRegisterClass *SrcRC = MRI.getRegClass(SrcReg);
+ assert(TRI->isSGPRClass(SrcRC) &&
+ "Expected SGPR REG_SEQUENCE to only have SGPR inputs");
+
+ SrcRC = TRI->getSubRegClass(SrcRC, SrcSubReg);
+ const TargetRegisterClass *NewSrcRC = TRI->getEquivalentVGPRClass(SrcRC);
+
+ unsigned TmpReg = MRI.createVirtualRegister(NewSrcRC);
+
+ BuildMI(*MI.getParent(), &MI, MI.getDebugLoc(), TII->get(AMDGPU::COPY), TmpReg)
+ .addOperand(MI.getOperand(I));
+
+ MI.getOperand(I).setReg(TmpReg);
+ }
+
+ CopyUse.eraseFromParent();
+ return true;
}
bool SIFixSGPRCopies::runOnMachineFunction(MachineFunction &MF) {
@@ -207,40 +242,38 @@ bool SIFixSGPRCopies::runOnMachineFunction(MachineFunction &MF) {
static_cast<const SIRegisterInfo *>(MF.getSubtarget().getRegisterInfo());
const SIInstrInfo *TII =
static_cast<const SIInstrInfo *>(MF.getSubtarget().getInstrInfo());
+
+ SmallVector<MachineInstr *, 16> Worklist;
+
for (MachineFunction::iterator BI = MF.begin(), BE = MF.end();
BI != BE; ++BI) {
MachineBasicBlock &MBB = *BI;
for (MachineBasicBlock::iterator I = MBB.begin(), E = MBB.end();
- I != E; ++I) {
+ I != E; ++I) {
MachineInstr &MI = *I;
- if (MI.getOpcode() == AMDGPU::COPY && isVGPRToSGPRCopy(MI, TRI, MRI)) {
- DEBUG(dbgs() << "Fixing VGPR -> SGPR copy:\n");
- DEBUG(MI.print(dbgs()));
- TII->moveToVALU(MI);
-
- }
switch (MI.getOpcode()) {
- default: continue;
- case AMDGPU::PHI: {
- DEBUG(dbgs() << "Fixing PHI: " << MI);
-
- for (unsigned i = 1; i < MI.getNumOperands(); i += 2) {
- const MachineOperand &Op = MI.getOperand(i);
- unsigned Reg = Op.getReg();
- const TargetRegisterClass *RC
- = inferRegClassFromDef(TRI, MRI, Reg, Op.getSubReg());
+ default:
+ continue;
+ case AMDGPU::COPY: {
+ // If the destination register is a physical register there isn't really
+ // much we can do to fix this.
+ if (!TargetRegisterInfo::isVirtualRegister(MI.getOperand(0).getReg()))
+ continue;
- MRI.constrainRegClass(Op.getReg(), RC);
- }
- unsigned Reg = MI.getOperand(0).getReg();
- const TargetRegisterClass *RC = inferRegClassFromUses(TRI, MRI, Reg,
- MI.getOperand(0).getSubReg());
- if (TRI->getCommonSubClass(RC, &AMDGPU::VGPR_32RegClass)) {
- MRI.constrainRegClass(Reg, &AMDGPU::VGPR_32RegClass);
+ const TargetRegisterClass *SrcRC, *DstRC;
+ std::tie(SrcRC, DstRC) = getCopyRegClasses(MI, *TRI, MRI);
+ if (isVGPRToSGPRCopy(SrcRC, DstRC, *TRI)) {
+ DEBUG(dbgs() << "Fixing VGPR -> SGPR copy: " << MI);
+ TII->moveToVALU(MI);
}
+ break;
+ }
+ case AMDGPU::PHI: {
+ DEBUG(dbgs() << "Fixing PHI: " << MI);
+ unsigned Reg = MI.getOperand(0).getReg();
if (!TRI->isSGPRClass(MRI.getRegClass(Reg)))
break;
@@ -310,8 +343,10 @@ bool SIFixSGPRCopies::runOnMachineFunction(MachineFunction &MF) {
}
case AMDGPU::REG_SEQUENCE: {
if (TRI->hasVGPRs(TII->getOpRegClass(MI, 0)) ||
- !hasVGPROperands(MI, TRI))
+ !hasVGPROperands(MI, TRI)) {
+ foldVGPRCopyIntoRegSequence(MI, TRI, TII, MRI);
continue;
+ }
DEBUG(dbgs() << "Fixing REG_SEQUENCE: " << MI);
OpenPOWER on IntegriCloud