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

Shengchen Kan via llvm-commits llvm-commits at lists.llvm.org
Fri May 10 19:59:10 PDT 2024


================
@@ -0,0 +1,886 @@
+//=========-- 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) {
----------------
KanRobert wrote:

Done

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


More information about the llvm-commits mailing list