[llvm] [X86] Set up the framework for optimization of CCMP/CTEST (PR #84603)

Shengchen Kan via llvm-commits llvm-commits at lists.llvm.org
Wed May 8 12:16:37 PDT 2024


================
@@ -0,0 +1,887 @@
+//==========-- X86ConditionalCompares.cpp --- CCMP formation for X86 -------===//
+//
+// 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
+//
+//===----------------------------------------------------------------------===//
+//
+// This file implements the X86ConditionalCompares pass which reduces
+// branching by using the conditional compare instructions CCMP, CTEST.
+//
+// The CFG transformations for forming conditional compares are very similar to
+// if-conversion, and this pass should run immediately before the early
+// if-conversion pass.
+//
+//===----------------------------------------------------------------------===//
+
+#include "X86.h"
+#include "X86InstrInfo.h"
+#include "X86Subtarget.h"
+#include "llvm/ADT/DepthFirstIterator.h"
+#include "llvm/ADT/SmallPtrSet.h"
+#include "llvm/ADT/Statistic.h"
+#include "llvm/CodeGen/MachineBranchProbabilityInfo.h"
+#include "llvm/CodeGen/MachineDominators.h"
+#include "llvm/CodeGen/MachineFunction.h"
+#include "llvm/CodeGen/MachineFunctionPass.h"
+#include "llvm/CodeGen/MachineInstrBuilder.h"
+#include "llvm/CodeGen/MachineLoopInfo.h"
+#include "llvm/CodeGen/MachineRegisterInfo.h"
+#include "llvm/CodeGen/MachineTraceMetrics.h"
+#include "llvm/CodeGen/Passes.h"
+#include "llvm/CodeGen/TargetInstrInfo.h"
+#include "llvm/CodeGen/TargetRegisterInfo.h"
+#include "llvm/CodeGen/TargetSubtargetInfo.h"
+#include "llvm/InitializePasses.h"
+#include "llvm/Support/CommandLine.h"
+#include "llvm/Support/Debug.h"
+#include "llvm/Support/raw_ostream.h"
+
+using namespace llvm;
+
+#define DEBUG_TYPE "x86-ccmp"
+
+// Absolute maximum number of instructions allowed per speculated block.
+// This bypasses all other heuristics, so it should be set fairly high.
+static cl::opt<unsigned> BlockInstrLimit(
+    "x86-ccmp-limit", cl::init(30), cl::Hidden,
+    cl::desc("Maximum number of instructions per speculated block."));
+
+STATISTIC(NumConsidered, "Number of ccmps considered");
+STATISTIC(NumPhiRejs, "Number of ccmps rejected (PHI)");
+STATISTIC(NumPhysRejs, "Number of ccmps rejected (Physregs)");
+STATISTIC(NumPhi2Rejs, "Number of ccmps rejected (PHI2)");
+STATISTIC(NumHeadBranchRejs, "Number of ccmps rejected (Head branch)");
+STATISTIC(NumCmpBranchRejs, "Number of ccmps rejected (CmpBB branch)");
+STATISTIC(NumCmpTermRejs, "Number of ccmps rejected (CmpBB is cbz...)");
+STATISTIC(NumMultEFLAGSUses, "Number of ccmps rejected (EFLAGS used)");
+STATISTIC(NumUnknEFLAGSDefs, "Number of ccmps rejected (EFLAGS def unknown)");
+STATISTIC(NumSpeculateRejs, "Number of ccmps rejected (Can't speculate)");
+STATISTIC(NumConverted, "Number of ccmp instructions created");
+
+//===----------------------------------------------------------------------===//
+//                                 SSACCmpConv
+//===----------------------------------------------------------------------===//
+//
+// The SSACCmpConv class performs ccmp-conversion on SSA form machine code
+// after determining if it is possible. The class contains no heuristics;
+// external code should be used to determine when ccmp-conversion is a good
+// idea.
+//
+// CCmp-formation works on a CFG representing chained conditions, typically
+// from C's short-circuit || and && operators:
+//
+//   From:         Head            To:         Head
+//                 / |                         CmpBB
+//                /  |                         / |
+//               |  CmpBB                     /  |
+//               |  / |                    Tail  |
+//               | /  |                      |   |
+//              Tail  |                      |   |
+//                |   |                      |   |
+//               ... ...                    ... ...
+//
+// The Head block is terminated by a br.cond instruction, and the CmpBB block
+// contains compare + br.cond. Tail must be a successor of both.
+//
+// The cmp-conversion turns the compare instruction in CmpBB into a conditional
+// compare, and merges CmpBB into Head, speculatively executing its
+// instructions. The X86 conditional compare instructions have an operand that
+// specifies the conditional flags to set values when the condition is false and
+// the compare isn't executed. This makes it possible to chain compares with
+// different condition codes.
+//
+// Example:
+//
+// void f(int a, int b) {
+//   if (a == 5 || b == 17)
+//     foo();
+// }
+//
+//    Head:
+//      cmpl  $5, $edi
+//      je Tail
+//    CmpBB:
+//      cmpl  $17, $esi
+//      je Tail
+//    ...
+//    Tail:
+//      call foo
+//
+//  Becomes:
+//
+//    Head:
+//      cmpl  $5, $edi
+//      ccmpel {dfv=zf} $17, $edi
+//      je Tail
+//    ...
+//    Tail:
+//      call foo
+//
+// The ccmp condition code is the one that would cause the Head terminator to
+// branch to CmpBB.
+
+namespace {
+class SSACCmpConv {
+  MachineFunction *MF;
+  const X86Subtarget *STI;
+  const TargetInstrInfo *TII;
+  const TargetRegisterInfo *TRI;
+  MachineRegisterInfo *MRI;
+  const MachineBranchProbabilityInfo *MBPI;
+
+public:
+  /// The first block containing a conditional branch, dominating everything
+  /// else.
+  MachineBasicBlock *Head;
+
+  /// The block containing cmp+br.cond with a successor shared with Head.
+  MachineBasicBlock *CmpBB;
+
+  /// The common successor for Head and CmpBB.
+  MachineBasicBlock *Tail;
+
+  /// The compare instruction in CmpBB that can be converted to a ccmp.
+  MachineInstr *CmpMI;
+
+private:
+  /// The branch condition in Head as determined by analyzeBranch.
+  SmallVector<MachineOperand, 4> HeadCond;
+
+  /// The condition code that makes Head branch to CmpBB.
+  X86::CondCode HeadCmpBBCC;
+
+  /// The branch condition in CmpBB.
+  SmallVector<MachineOperand, 4> CmpBBCond;
+
+  /// The condition code that makes CmpBB branch to Tail.
+  X86::CondCode CmpBBTailCC;
+
+  /// Check if the Tail PHIs are trivially convertible.
+  bool trivialTailPHIs();
+
+  /// Remove CmpBB from the Tail PHIs.
+  void updateTailPHIs();
+
+  /// Check if an operand defining DstReg is dead.
+  bool isDeadDef(unsigned DstReg);
+
+  /// Find the compare instruction in MBB that controls the conditional branch.
+  /// Return NULL if a convertible instruction can't be found.
+  MachineInstr *findConvertibleCompare(MachineBasicBlock *MBB);
+
+  /// Return true if all non-terminator instructions in MBB can be safely
+  /// speculated.
+  bool canSpeculateInstrs(MachineBasicBlock *MBB, const MachineInstr *CmpMI);
+
+public:
+  /// runOnMachineFunction - Initialize per-function data structures.
+  void runOnMachineFunction(MachineFunction &MF,
+                            const MachineBranchProbabilityInfo *MBPI) {
+    this->MF = &MF;
+    this->MBPI = MBPI;
+    STI = &MF.getSubtarget<X86Subtarget>();
+    TII = MF.getSubtarget().getInstrInfo();
+    TRI = MF.getSubtarget().getRegisterInfo();
+    MRI = &MF.getRegInfo();
+  }
+
+  /// If the sub-CFG headed by MBB can be cmp-converted, initialize the
+  /// internal state, and return true.
+  bool canConvert(MachineBasicBlock *MBB);
+
+  /// Cmo-convert the last block passed to canConvertCmp(), assuming
+  /// it is possible. Add any erased blocks to RemovedBlocks.
+  void convert(SmallVectorImpl<MachineBasicBlock *> &RemovedBlocks);
+};
+} // end anonymous namespace
+
+// Check that all PHIs in Tail are selecting the same value from Head and CmpBB.
+// This means that no if-conversion is required when merging CmpBB into Head.
+bool SSACCmpConv::trivialTailPHIs() {
+  for (auto &I : *Tail) {
+    if (!I.isPHI())
+      break;
+    unsigned HeadReg = 0, CmpBBReg = 0;
+    // PHI operands come in (VReg, MBB) pairs.
+    for (unsigned oi = 1, oe = I.getNumOperands(); oi != oe; oi += 2) {
+      MachineBasicBlock *MBB = I.getOperand(oi + 1).getMBB();
+      Register Reg = I.getOperand(oi).getReg();
+      if (MBB == Head) {
+        assert((!HeadReg || HeadReg == Reg) && "Inconsistent PHI operands");
+        HeadReg = Reg;
+      }
+      if (MBB == CmpBB) {
+        assert((!CmpBBReg || CmpBBReg == Reg) && "Inconsistent PHI operands");
+        CmpBBReg = Reg;
+      }
+    }
+    if (HeadReg != CmpBBReg)
+      return false;
+  }
+  return true;
+}
+
+// Assuming that trivialTailPHIs() is true, update the Tail PHIs by simply
+// removing the CmpBB operands. The Head operands will be identical.
+void SSACCmpConv::updateTailPHIs() {
+  for (auto &I : *Tail) {
+    if (!I.isPHI())
+      break;
+    // I is a PHI. It can have multiple entries for CmpBB.
+    for (unsigned Idx = I.getNumOperands(); Idx > 2; Idx -= 2) {
+      // PHI operands are (Reg, MBB) at (Idx-2, Idx-1).
+      if (I.getOperand(Idx - 1).getMBB() == CmpBB) {
+        I.removeOperand(Idx - 1);
+        I.removeOperand(Idx - 2);
+      }
+    }
+  }
+}
+
+bool SSACCmpConv::isDeadDef(unsigned DstReg) {
+  if (!Register::isVirtualRegister(DstReg))
+    return false;
+  // A virtual register def without any uses will be marked dead later, and
+  // eventually replaced by the zero register.
+  return MRI->use_nodbg_empty(DstReg);
+}
+
+MachineInstr *SSACCmpConv::findConvertibleCompare(MachineBasicBlock *MBB) {
+  MachineBasicBlock::iterator I = MBB->getFirstTerminator();
+  if (I == MBB->end())
+    return nullptr;
+  // The terminator must be controlled by the flags.
+  if (!I->readsRegister(X86::EFLAGS)) {
+    ++NumCmpTermRejs;
+    LLVM_DEBUG(dbgs() << "Flags not used by terminator: " << *I);
+    return nullptr;
+  }
+
+  // Now find the instruction controlling the terminator.
+  for (MachineBasicBlock::iterator B = MBB->begin(); I != B;) {
+    I = prev_nodbg(I, MBB->begin());
+    assert(!I->isTerminator() && "Spurious terminator");
+
+    switch (I->getOpcode()) {
+    // This pass run before peephole optimization, so the SUB has not been
+    // optimized to CMP yet.
+    case X86::SUB8rr:
+    case X86::SUB16rr:
+    case X86::SUB32rr:
+    case X86::SUB64rr:
+    case X86::SUB8ri:
+    case X86::SUB16ri:
+    case X86::SUB32ri:
+    case X86::SUB64ri32:
+    case X86::SUB8rr_ND:
+    case X86::SUB16rr_ND:
+    case X86::SUB32rr_ND:
+    case X86::SUB64rr_ND:
+    case X86::SUB8ri_ND:
+    case X86::SUB16ri_ND:
+    case X86::SUB32ri_ND:
+    case X86::SUB64ri32_ND: {
+      if (!isDeadDef(I->getOperand(0).getReg()))
+        return nullptr;
+      return STI->hasCCMP() ? &*I : nullptr;
+    }
+    case X86::CMP8rr:
+    case X86::CMP16rr:
+    case X86::CMP32rr:
+    case X86::CMP64rr:
+    case X86::CMP8ri:
+    case X86::CMP16ri:
+    case X86::CMP32ri:
+    case X86::CMP64ri32:
+    case X86::TEST8rr:
+    case X86::TEST16rr:
+    case X86::TEST32rr:
+    case X86::TEST64rr:
+    case X86::TEST8ri:
+    case X86::TEST16ri:
+    case X86::TEST32ri:
+    case X86::TEST64ri32:
+      return STI->hasCCMP() ? &*I : nullptr;
+    default:
+      break;
+    }
+
+    // Check for flag reads and clobbers.
+    PhysRegInfo PRI = AnalyzePhysRegInBundle(*I, X86::EFLAGS, TRI);
+
+    if (PRI.Read) {
+      // The ccmp doesn't produce exactly the same flags as the original
+      // compare, so reject the transform if there are uses of the flags
+      // besides the terminators.
+      LLVM_DEBUG(dbgs() << "Can't create ccmp with multiple uses: " << *I);
+      ++NumMultEFLAGSUses;
+      return nullptr;
+    }
+
+    if (PRI.Defined || PRI.Clobbered) {
+      LLVM_DEBUG(dbgs() << "Not convertible compare: " << *I);
+      ++NumUnknEFLAGSDefs;
+      return nullptr;
+    }
+  }
+  LLVM_DEBUG(dbgs() << "Flags not defined in " << printMBBReference(*MBB)
+                    << '\n');
+  return nullptr;
+}
+
+/// Determine if all the instructions in MBB can safely
+/// be speculated. The terminators are not considered.
+///
+/// Only CmpMI is allowed to clobber the flags.
+///
+bool SSACCmpConv::canSpeculateInstrs(MachineBasicBlock *MBB,
+                                     const MachineInstr *CmpMI) {
+  // Reject any live-in physregs. It's very hard to get right.
+  if (!MBB->livein_empty()) {
+    LLVM_DEBUG(dbgs() << printMBBReference(*MBB) << " has live-ins.\n");
+    return false;
+  }
+
+  unsigned InstrCount = 0;
+
+  // Check all instructions, except the terminators. It is assumed that
+  // terminators never have side effects or define any used register values.
+  for (auto &I : make_range(MBB->begin(), MBB->getFirstTerminator())) {
+    if (I.isDebugInstr())
+      continue;
+
+    if (++InstrCount > BlockInstrLimit) {
+      LLVM_DEBUG(dbgs() << printMBBReference(*MBB) << " has more than "
+                        << BlockInstrLimit << " instructions.\n");
+      return false;
+    }
+
+    // There shouldn't normally be any phis in a single-predecessor block.
+    if (I.isPHI()) {
+      LLVM_DEBUG(dbgs() << "Can't hoist: " << I);
+      return false;
+    }
+
+    // Don't speculate loads. Note that it may be possible and desirable to
+    // speculate GOT or constant pool loads that are guaranteed not to trap,
+    // but we don't support that for now.
+    if (I.mayLoad()) {
+      LLVM_DEBUG(dbgs() << "Won't speculate load: " << I);
+      return false;
+    }
+
+    // We never speculate stores, so an AA pointer isn't necessary.
+    bool DontMoveAcrossStore = true;
+    if (!I.isSafeToMove(nullptr, DontMoveAcrossStore)) {
+      LLVM_DEBUG(dbgs() << "Can't speculate: " << I);
+      return false;
+    }
+
+    // Only CmpMI is allowed to clobber the flags.
+    if (&I != CmpMI && I.modifiesRegister(X86::EFLAGS, TRI)) {
----------------
KanRobert wrote:

It's already considered in `MachineInstr::isSafeToMove`

https://github.com/llvm/llvm-project/pull/84603


More information about the llvm-commits mailing list