[llvm] GlobalISel lane masks merging (PR #73337)

[Pierre van Houtryve via llvm-commits]

================
@@ -0,0 +1,213 @@
+//===-- AMDGPUGlobalISelDivergenceLowering.cpp ----------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+// GlobalISel pass that selects divergent i1 phis as lane mask phis.
+// Lane mask merging uses same algorithm as SDAG in SILowerI1Copies.
+// Handles all cases of temporal divergence.
+//
+// For divergent non-phi i1 and uniform i1 uses outside of the cycle this pass
+// currently depends on LCSSA to insert phis with one incoming.
+//===----------------------------------------------------------------------===//
+
+#include "AMDGPU.h"
+#include "SILowerI1Copies.h"
+#include "llvm/CodeGen/GlobalISel/MachineIRBuilder.h"
+#include "llvm/CodeGen/MachineFunctionPass.h"
+#include "llvm/CodeGen/MachineUniformityAnalysis.h"
+#include "llvm/InitializePasses.h"
+
+#define DEBUG_TYPE "global-isel-divergence-lowering"
+
+using namespace llvm;
+
+namespace {
+
+class AMDGPUGlobalISelDivergenceLowering : public MachineFunctionPass {
+public:
+  static char ID;
+
+public:
+  AMDGPUGlobalISelDivergenceLowering() : MachineFunctionPass(ID) {
+    initializeAMDGPUGlobalISelDivergenceLoweringPass(
+        *PassRegistry::getPassRegistry());
+  }
+
+  bool runOnMachineFunction(MachineFunction &MF) override;
+
+  StringRef getPassName() const override {
+    return "GlobalISel divergence lowering";
+  }
+
+  void getAnalysisUsage(AnalysisUsage &AU) const override {
+    AU.setPreservesCFG();
+    AU.addRequired<MachineCycleInfoWrapperPass>();
+    AU.addRequired<MachineDominatorTree>();
+    AU.addRequired<MachinePostDominatorTree>();
+    MachineFunctionPass::getAnalysisUsage(AU);
+  }
+};
+
+class DivergenceLoweringHelper : public PhiLoweringHelper {
+public:
+  DivergenceLoweringHelper(MachineFunction *MF, MachineDominatorTree *DT,
+                           MachinePostDominatorTree *PDT,
+                           MachineUniformityInfo *MUI);
+
+private:
+  MachineUniformityInfo *MUI = nullptr;
+
+public:
+  void markAsLaneMask(Register DstReg) const override;
+  void getCandidatesForLowering(
+      SmallVectorImpl<MachineInstr *> &Vreg1Phis) const override;
+  void collectIncomingValuesFromPhi(
+      const MachineInstr *MI,
+      SmallVectorImpl<Incoming> &Incomings) const override;
+  void replaceDstReg(Register NewReg, Register OldReg,
+                     MachineBasicBlock *MBB) override;
+  void buildMergeLaneMasks(MachineBasicBlock &MBB,
+                           MachineBasicBlock::iterator I, const DebugLoc &DL,
+                           Register DstReg, Register PrevReg,
+                           Register CurReg) override;
+  void constrainIncomingRegisterTakenAsIs(Incoming &In) override;
+};
+
+DivergenceLoweringHelper::DivergenceLoweringHelper(
+    MachineFunction *MF, MachineDominatorTree *DT,
+    MachinePostDominatorTree *PDT, MachineUniformityInfo *MUI)
+    : PhiLoweringHelper(MF, DT, PDT), MUI(MUI) {}
+
+// _(s1) -> SReg_32/64(s1)
+void DivergenceLoweringHelper::markAsLaneMask(Register DstReg) const {
+  assert(MRI->getType(DstReg) == LLT::scalar(1));
+
+  // Can't blindly set a register class on phi, users could have reg class
+  // constraints (e.g. sreg_32/64..._xexec classes for control flow intrinsics).
+  if (MRI->getRegClassOrNull(DstReg))
+    return;
+
+  MRI->setRegClass(DstReg, ST->getBoolRC());
+  return;
+}
+
+void DivergenceLoweringHelper::getCandidatesForLowering(
+    SmallVectorImpl<MachineInstr *> &Vreg1Phis) const {
+  LLT S1 = LLT::scalar(1);
+
+  // Add divergent i1 phis to the list
+  for (MachineBasicBlock &MBB : *MF) {
+    for (MachineInstr &MI : MBB.phis()) {
+      Register Dst = MI.getOperand(0).getReg();
+      if (MRI->getType(Dst) == S1 && MUI->isDivergent(Dst))
+        Vreg1Phis.push_back(&MI);
+    }
+  }
+
+  return;
+}
+
+void DivergenceLoweringHelper::collectIncomingValuesFromPhi(
+    const MachineInstr *MI, SmallVectorImpl<Incoming> &Incomings) const {
+  for (unsigned i = 1; i < MI->getNumOperands(); i += 2) {
+    Incomings.emplace_back(MI->getOperand(i).getReg(),
+                           MI->getOperand(i + 1).getMBB(), Register());
+  }
+}
+
+void DivergenceLoweringHelper::replaceDstReg(Register NewReg, Register OldReg,
+                                             MachineBasicBlock *MBB) {
+  BuildMI(*MBB, MBB->getFirstNonPHI(), {}, TII->get(AMDGPU::COPY), OldReg)
+      .addReg(NewReg);
+}
+
+// Get pointers to build instruction just after MI (skips phis if needed)
+static std::pair<MachineBasicBlock *, MachineBasicBlock::iterator>
+getInsertAfterPtrs(MachineInstr *MI) {
+  MachineBasicBlock *InsertMBB = MI->getParent();
+  return std::make_pair(
----------------
Pierre-vh wrote:

small nit: i think `make_pair ` isn't needed, you can just use `{A, B}`
Also perhaps put `InsertMBB->SkipPHIsAndLabels(std::next(MI->getIterator())` in a variable, or the `std::next` part at least, to make it less dense

https://github.com/llvm/llvm-project/pull/73337