AMDGPU/GlobalISel: AMDGPURegBankSelect

Assign register banks to virtual registers. Does not use generic
RegBankSelect. After register bank selection all register operand of
G_ instructions have LLT and register banks exclusively. If they had
register class, reassign appropriate register bank.

Assign register banks using machine uniformity analysis:
Sgpr - uniform values and some lane masks
Vgpr - divergent, non S1, values
Vcc  - divergent S1 values(lane masks)

AMDGPURegBankSelect does not consider available instructions and, in
some cases, G_ instructions with some register bank assignment can't be
inst-selected. This is solved in RegBankLegalize.

Exceptions when uniformity analysis does not work:
S32/S64 lane masks:
- need to end up with sgpr register class after instruction selection
- In most cases Uniformity analysis declares them as uniform
  (forced by tablegen) resulting in sgpr S32/S64 reg bank
- When Uniformity analysis declares them as divergent (some phis),
  use intrinsic lane mask analyzer to still assign sgpr register bank
temporal divergence copy:
- COPY to vgpr with implicit use of $exec inside of the cycle
- this copy is declared as uniform by uniformity analysis
- make sure that assigned bank is vgpr
Note: uniformity analysis does not consider that registers with vgpr def
are divergent (you can have uniform value in vgpr).
- TODO: implicit use of $exec could be implemented as indicator
  that instruction is divergent

Author: petar-avramovic

llvm/lib/Target/AMDGPU/AMDGPUGlobalISelUtils.cpp

@@ -9,11 +9,14 @@
 #include "AMDGPUGlobalISelUtils.h"
 #include "GCNSubtarget.h"
 #include "llvm/CodeGen/GlobalISel/GISelKnownBits.h"
+#include "llvm/CodeGen/GlobalISel/GenericMachineInstrs.h"
 #include "llvm/CodeGen/GlobalISel/MIPatternMatch.h"
 #include "llvm/CodeGenTypes/LowLevelType.h"
 #include "llvm/IR/Constants.h"
+#include "llvm/IR/IntrinsicsAMDGPU.h"
 
 using namespace llvm;
+using namespace AMDGPU;
 using namespace MIPatternMatch;
 
 std::pair<Register, unsigned>
@@ -69,3 +72,37 @@ AMDGPU::getBaseWithConstantOffset(MachineRegisterInfo &MRI, Register Reg,
 
   return std::pair(Reg, 0);
 }
+
+IntrinsicLaneMaskAnalyzer::IntrinsicLaneMaskAnalyzer(MachineFunction &MF)
+    : MRI(MF.getRegInfo()) {
+  initLaneMaskIntrinsics(MF);
+}
+
+bool IntrinsicLaneMaskAnalyzer::isS32S64LaneMask(Register Reg) const {
+  return S32S64LaneMask.contains(Reg);
+}
+
+void IntrinsicLaneMaskAnalyzer::initLaneMaskIntrinsics(MachineFunction &MF) {
+  for (auto &MBB : MF) {
+    for (auto &MI : MBB) {
+      GIntrinsic *GI = dyn_cast<GIntrinsic>(&MI);
+      if (GI && GI->is(Intrinsic::amdgcn_if_break)) {
+        S32S64LaneMask.insert(MI.getOperand(3).getReg());
+        findLCSSAPhi(MI.getOperand(0).getReg());
+      }
+
+      if (MI.getOpcode() == AMDGPU::SI_IF ||
+          MI.getOpcode() == AMDGPU::SI_ELSE) {
+        findLCSSAPhi(MI.getOperand(0).getReg());
+      }
+    }
+  }
+}
+
+void IntrinsicLaneMaskAnalyzer::findLCSSAPhi(Register Reg) {
+  S32S64LaneMask.insert(Reg);
+  for (const MachineInstr &LCSSAPhi : MRI.use_instructions(Reg)) {
+    if (LCSSAPhi.isPHI())
+      S32S64LaneMask.insert(LCSSAPhi.getOperand(0).getReg());
+  }
+}

llvm/lib/Target/AMDGPU/AMDGPUGlobalISelUtils.h

@@ -9,6 +9,8 @@
 #ifndef LLVM_LIB_TARGET_AMDGPU_AMDGPUGLOBALISELUTILS_H
 #define LLVM_LIB_TARGET_AMDGPU_AMDGPUGLOBALISELUTILS_H
 
+#include "llvm/ADT/DenseSet.h"
+#include "llvm/CodeGen/MachineFunction.h"
 #include "llvm/CodeGen/Register.h"
 #include <utility>
 
@@ -26,6 +28,26 @@ std::pair<Register, unsigned>
 getBaseWithConstantOffset(MachineRegisterInfo &MRI, Register Reg,
                           GISelKnownBits *KnownBits = nullptr,
                           bool CheckNUW = false);
+
+// Currently finds S32/S64 lane masks that can be declared as divergent by
+// uniformity analysis (all are phis at the moment).
+// These are defined as i32/i64 in some IR intrinsics (not as i1).
+// Tablegen forces(via telling that lane mask IR intrinsics are uniform) most of
+// S32/S64 lane masks to be uniform, as this results in them ending up with sgpr
+// reg class after instruction-select don't search for all of them.
+class IntrinsicLaneMaskAnalyzer {
+  SmallDenseSet<Register, 8> S32S64LaneMask;
+  MachineRegisterInfo &MRI;
+
+public:
+  IntrinsicLaneMaskAnalyzer(MachineFunction &MF);
+  bool isS32S64LaneMask(Register Reg) const;
+
+private:
+  void initLaneMaskIntrinsics(MachineFunction &MF);
+  // This will not be needed when we turn off LCSSA for global-isel.
+  void findLCSSAPhi(Register Reg);
+};
 }
 }
 

llvm/lib/Target/AMDGPU/AMDGPURegBankSelect.cpp

@@ -16,12 +16,21 @@
 //===----------------------------------------------------------------------===//
 
 #include "AMDGPU.h"
+#include "AMDGPUGlobalISelUtils.h"
+#include "GCNSubtarget.h"
+#include "llvm/CodeGen/GlobalISel/CSEInfo.h"
+#include "llvm/CodeGen/GlobalISel/CSEMIRBuilder.h"
+#include "llvm/CodeGen/GlobalISel/MachineIRBuilder.h"
 #include "llvm/CodeGen/MachineFunctionPass.h"
+#include "llvm/CodeGen/MachineInstr.h"
+#include "llvm/CodeGen/MachineUniformityAnalysis.h"
+#include "llvm/CodeGen/TargetPassConfig.h"
 #include "llvm/InitializePasses.h"
 
 #define DEBUG_TYPE "amdgpu-regbankselect"
 
 using namespace llvm;
+using namespace AMDGPU;
 
 namespace {
 
@@ -40,6 +49,9 @@ class AMDGPURegBankSelect : public MachineFunctionPass {
   }
 
   void getAnalysisUsage(AnalysisUsage &AU) const override {
+    AU.addRequired<TargetPassConfig>();
+    AU.addRequired<GISelCSEAnalysisWrapperPass>();
+    AU.addRequired<MachineUniformityAnalysisPass>();
     MachineFunctionPass::getAnalysisUsage(AU);
   }
 
@@ -55,6 +67,9 @@ class AMDGPURegBankSelect : public MachineFunctionPass {
 
 INITIALIZE_PASS_BEGIN(AMDGPURegBankSelect, DEBUG_TYPE,
                       "AMDGPU Register Bank Select", false, false)
+INITIALIZE_PASS_DEPENDENCY(TargetPassConfig)
+INITIALIZE_PASS_DEPENDENCY(GISelCSEAnalysisWrapperPass)
+INITIALIZE_PASS_DEPENDENCY(MachineUniformityAnalysisPass)
 INITIALIZE_PASS_END(AMDGPURegBankSelect, DEBUG_TYPE,
                     "AMDGPU Register Bank Select", false, false)
 
@@ -66,9 +81,232 @@ FunctionPass *llvm::createAMDGPURegBankSelectPass() {
   return new AMDGPURegBankSelect();
 }
 
+class RegBankSelectHelper {
+  MachineIRBuilder &B;
+  MachineRegisterInfo &MRI;
+  AMDGPU::IntrinsicLaneMaskAnalyzer &ILMA;
+  const MachineUniformityInfo &MUI;
+  const SIRegisterInfo &TRI;
+  const RegisterBank *SgprRB;
+  const RegisterBank *VgprRB;
+  const RegisterBank *VccRB;
+
+public:
+  RegBankSelectHelper(MachineIRBuilder &B,
+                      AMDGPU::IntrinsicLaneMaskAnalyzer &ILMA,
+                      const MachineUniformityInfo &MUI,
+                      const SIRegisterInfo &TRI, const RegisterBankInfo &RBI)
+      : B(B), MRI(*B.getMRI()), ILMA(ILMA), MUI(MUI), TRI(TRI),
+        SgprRB(&RBI.getRegBank(AMDGPU::SGPRRegBankID)),
+        VgprRB(&RBI.getRegBank(AMDGPU::VGPRRegBankID)),
+        VccRB(&RBI.getRegBank(AMDGPU::VCCRegBankID)) {}
+
+  bool shouldRegBankSelect(MachineInstr &MI) {
+    return MI.isPreISelOpcode() || MI.isCopy();
+  }
+
+  // Temporal divergence copy: COPY to vgpr with implicit use of $exec inside of
+  // the cycle
+  // Note: uniformity analysis does not consider that registers with vgpr def
+  // are divergent (you can have uniform value in vgpr).
+  // - TODO: implicit use of $exec could be implemented as indicator that
+  //   instruction is divergent
+  bool isTemporalDivergenceCopy(Register Reg) {
+    MachineInstr *MI = MRI.getVRegDef(Reg);
+    if (!MI->isCopy())
+      return false;
+
+    for (auto Op : MI->implicit_operands()) {
+      if (!Op.isReg())
+        continue;
+
+      if (Op.getReg() == TRI.getExec()) {
+        return true;
+      }
+    }
+
+    return false;
+  }
+
+  void setRBDef(MachineInstr &MI, MachineOperand &DefOP,
+                const RegisterBank *RB) {
+    Register Reg = DefOP.getReg();
+
+    if (!MRI.getRegClassOrNull(Reg)) {
+      MRI.setRegBank(Reg, *RB);
+      return;
+    }
+
+    // Register that already has Register class got it during pre-inst selection
+    // of another instruction. Maybe cross bank copy was required so we insert a
+    // copy that can be removed later. This simplifies post regbanklegalize
+    // combiner and avoids need to special case some patterns.
+    LLT Ty = MRI.getType(Reg);
+    Register NewReg = MRI.createVirtualRegister({RB, Ty});
+    DefOP.setReg(NewReg);
+
+    auto &MBB = *MI.getParent();
+    B.setInsertPt(MBB, MBB.SkipPHIsAndLabels(std::next(MI.getIterator())));
+    B.buildCopy(Reg, NewReg);
+
+    // The problem was discovered for uniform S1 that was used as both
+    // lane mask(vcc) and regular sgpr S1.
+    // - lane-mask(vcc) use was by si_if, this use is divergent and requires
+    //   non-trivial sgpr-S1-to-vcc copy. But pre-inst-selection of si_if sets
+    //   sreg_64_xexec(S1) on def of uniform S1 making it lane-mask.
+    // - the regular sgpr S1(uniform) instruction is now broken since
+    //   it uses sreg_64_xexec(S1) which is divergent.
+
+    // Replace virtual registers with register class on generic instructions
+    // uses with virtual registers with register bank.
+    SmallVector<MachineOperand *, 4> RegUsesOnGInstrs;
+    for (auto &UseMI : MRI.use_instructions(Reg)) {
+      if (shouldRegBankSelect(UseMI)) {
+        for (MachineOperand &Op : UseMI.operands()) {
+          if (Op.isReg() && Op.getReg() == Reg)
+            RegUsesOnGInstrs.push_back(&Op);
+        }
+      }
+    }
+    for (MachineOperand *Op : RegUsesOnGInstrs) {
+      Op->setReg(NewReg);
+    }
+  }
+
+  Register tryGetVReg(MachineOperand &Op) {
+    if (!Op.isReg())
+      return {};
+
+    Register Reg = Op.getReg();
+    if (!Reg.isVirtual())
+      return {};
+
+    return Reg;
+  }
+
+  void assignBanksOnDefs(MachineInstr &MI) {
+    if (!shouldRegBankSelect(MI))
+      return;
+
+    for (MachineOperand &DefOP : MI.defs()) {
+      Register DefReg = tryGetVReg(DefOP);
+      if (!DefReg.isValid())
+        continue;
+
+      // Copies can have register class on def registers.
+      if (MI.isCopy() && MRI.getRegClassOrNull(DefReg)) {
+        continue;
+      }
+
+      if (MUI.isUniform(DefReg) || ILMA.isS32S64LaneMask(DefReg)) {
+        setRBDef(MI, DefOP, SgprRB);
+      } else {
+        if (MRI.getType(DefReg) == LLT::scalar(1))
+          setRBDef(MI, DefOP, VccRB);
+        else
+          setRBDef(MI, DefOP, VgprRB);
+      }
+    }
+  }
+
+  void constrainRBUse(MachineInstr &MI, MachineOperand &UseOP,
+                      const RegisterBank *RB) {
+    Register Reg = UseOP.getReg();
+
+    LLT Ty = MRI.getType(Reg);
+    Register NewReg = MRI.createVirtualRegister({RB, Ty});
+    UseOP.setReg(NewReg);
+
+    if (MI.isPHI()) {
+      auto DefMI = MRI.getVRegDef(Reg)->getIterator();
+      MachineBasicBlock *DefMBB = DefMI->getParent();
+      B.setInsertPt(*DefMBB, DefMBB->SkipPHIsAndLabels(std::next(DefMI)));
+    } else {
+      B.setInstr(MI);
+    }
+
+    B.buildCopy(NewReg, Reg);
+  }
+
+  void constrainBanksOnUses(MachineInstr &MI) {
+    if (!shouldRegBankSelect(MI))
+      return;
+
+    // Copies can have register class on use registers.
+    if (MI.isCopy())
+      return;
+
+    for (MachineOperand &UseOP : MI.uses()) {
+      auto UseReg = tryGetVReg(UseOP);
+      if (!UseReg.isValid())
+        continue;
+
+      // UseReg already has register bank.
+      if (MRI.getRegBankOrNull(UseReg))
+        continue;
+
+      if (!isTemporalDivergenceCopy(UseReg) &&
+          (MUI.isUniform(UseReg) || ILMA.isS32S64LaneMask(UseReg))) {
+        constrainRBUse(MI, UseOP, SgprRB);
+      } else {
+        if (MRI.getType(UseReg) == LLT::scalar(1))
+          constrainRBUse(MI, UseOP, VccRB);
+        else
+          constrainRBUse(MI, UseOP, VgprRB);
+      }
+    }
+  }
+};
+
 bool AMDGPURegBankSelect::runOnMachineFunction(MachineFunction &MF) {
   if (MF.getProperties().hasProperty(
           MachineFunctionProperties::Property::FailedISel))
     return false;
+
+  // Setup the instruction builder with CSE.
+  const TargetPassConfig &TPC = getAnalysis<TargetPassConfig>();
+  GISelCSEAnalysisWrapper &Wrapper =
+      getAnalysis<GISelCSEAnalysisWrapperPass>().getCSEWrapper();
+  GISelCSEInfo &CSEInfo = Wrapper.get(TPC.getCSEConfig());
+  GISelObserverWrapper Observer;
+  Observer.addObserver(&CSEInfo);
+
+  CSEMIRBuilder B(MF);
+  B.setCSEInfo(&CSEInfo);
+  B.setChangeObserver(Observer);
+
+  RAIIDelegateInstaller DelegateInstaller(MF, &Observer);
+  RAIIMFObserverInstaller MFObserverInstaller(MF, Observer);
+
+  IntrinsicLaneMaskAnalyzer ILMA(MF);
+  MachineUniformityInfo &MUI =
+      getAnalysis<MachineUniformityAnalysisPass>().getUniformityInfo();
+  const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>();
+  RegBankSelectHelper RBSHelper(B, ILMA, MUI, *ST.getRegisterInfo(),
+                                *ST.getRegBankInfo());
+
+  // Assign register banks to ALL def registers on G_ instructions.
+  // Same for copies if they have no register bank or class on def.
+  for (MachineBasicBlock &MBB : MF) {
+    for (MachineInstr &MI : MBB) {
+      RBSHelper.assignBanksOnDefs(MI);
+    }
+  }
+
+  // At this point all virtual registers have register class or bank
+  // - Defs of G_ instructions have register banks.
+  // - Defs and uses of inst-selected instructions have register class.
+  // - Defs and uses of copies can have either register class or bank
+  //   and most notably:
+  // - Uses of G_ instructions can have either register class or bank.
+
+  // Reassign uses of G_ instructions to only have register banks.
+  for (MachineBasicBlock &MBB : MF) {
+    for (MachineInstr &MI : MBB) {
+      RBSHelper.constrainBanksOnUses(MI);
+    }
+  }
+
+  // Defs and uses of G_ instructions have register banks exclusively.
   return true;
 }