From 5f746c8e2704d311cc01c8edb5460a0cb19c74e2 Mon Sep 17 00:00:00 2001 From: Florian Hahn Date: Mon, 19 Jun 2017 12:53:31 +0000 Subject: Recommit rL305677: [CodeGen] Add generic MacroFusion pass Use llvm::make_unique to avoid ambiguity with MSVC. This patch adds a generic MacroFusion pass, that is used on X86 and AArch64, which both define target-specific shouldScheduleAdjacent functions. This generic pass should make it easier for other targets to implement macro fusion and I intend to add macro fusion for ARM shortly. Differential Revision: https://reviews.llvm.org/D34144 llvm-svn: 305690 --- llvm/lib/Target/X86/X86MacroFusion.cpp | 101 ++++++--------------------------- 1 file changed, 16 insertions(+), 85 deletions(-) (limited to 'llvm/lib/Target/X86/X86MacroFusion.cpp') diff --git a/llvm/lib/Target/X86/X86MacroFusion.cpp b/llvm/lib/Target/X86/X86MacroFusion.cpp index dd21e2b7c4a..8fdf1061705 100644 --- a/llvm/lib/Target/X86/X86MacroFusion.cpp +++ b/llvm/lib/Target/X86/X86MacroFusion.cpp @@ -2,39 +2,31 @@ // // The LLVM Compiler Infrastructure // -// \file This file is distributed under the University of Illinois Open Source +// This file is distributed under the University of Illinois Open Source // License. See LICENSE.TXT for details. // //===----------------------------------------------------------------------===// // -// This file contains the X86 implementation of the DAG scheduling mutation to -// pair instructions back to back. +/// \file This file contains the X86 implementation of the DAG scheduling +/// mutation to pair instructions back to back. // //===----------------------------------------------------------------------===// #include "X86MacroFusion.h" #include "X86Subtarget.h" -#include "llvm/ADT/Statistic.h" -#include "llvm/Support/CommandLine.h" #include "llvm/Target/TargetInstrInfo.h" - -#define DEBUG_TYPE "misched" - -STATISTIC(NumFused, "Number of instr pairs fused"); +#include "llvm/CodeGen/MacroFusion.h" using namespace llvm; -static cl::opt EnableMacroFusion("x86-misched-fusion", cl::Hidden, - cl::desc("Enable scheduling for macro fusion."), cl::init(true)); - -namespace { - -/// \brief Verify that the instruction pair, First and Second, -/// should be scheduled back to back. If either instruction is unspecified, -/// then verify that the other instruction may be part of a pair at all. -static bool shouldScheduleAdjacent(const X86Subtarget &ST, - const MachineInstr *First, - const MachineInstr *Second) { +/// \brief Check if the instr pair, FirstMI and SecondMI, should be fused +/// together. Given SecondMI, when FirstMI is unspecified, then check if +/// SecondMI may be part of a fused pair at all. +static bool shouldScheduleAdjacent(const TargetInstrInfo &TII, + const TargetSubtargetInfo &TSI, + const MachineInstr *FirstMI, + const MachineInstr &SecondMI) { + const X86Subtarget &ST = static_cast(TSI); // Check if this processor supports macro-fusion. Since this is a minor // heuristic, we haven't specifically reserved a feature. hasAVX is a decent // proxy for SandyBridge+. @@ -47,13 +39,10 @@ static bool shouldScheduleAdjacent(const X86Subtarget &ST, FuseInc } FuseKind; - assert((First || Second) && "At least one instr must be specified"); - unsigned FirstOpcode = First - ? First->getOpcode() + unsigned FirstOpcode = FirstMI + ? FirstMI->getOpcode() : static_cast(X86::INSTRUCTION_LIST_END); - unsigned SecondOpcode = Second - ? Second->getOpcode() - : static_cast(X86::INSTRUCTION_LIST_END); + unsigned SecondOpcode = SecondMI.getOpcode(); switch (SecondOpcode) { default: @@ -203,69 +192,11 @@ static bool shouldScheduleAdjacent(const X86Subtarget &ST, } } -/// \brief Post-process the DAG to create cluster edges between instructions -/// that may be fused by the processor into a single operation. -class X86MacroFusion : public ScheduleDAGMutation { -public: - X86MacroFusion() {} - - void apply(ScheduleDAGInstrs *DAGInstrs) override; -}; - -void X86MacroFusion::apply(ScheduleDAGInstrs *DAGInstrs) { - ScheduleDAGMI *DAG = static_cast(DAGInstrs); - const X86Subtarget &ST = DAG->MF.getSubtarget(); - - // For now, assume targets can only fuse with the branch. - SUnit &ExitSU = DAG->ExitSU; - MachineInstr *Branch = ExitSU.getInstr(); - if (!Branch || !shouldScheduleAdjacent(ST, nullptr, Branch)) - return; - - for (SDep &PredDep : ExitSU.Preds) { - if (PredDep.isWeak()) - continue; - SUnit &SU = *PredDep.getSUnit(); - MachineInstr &Pred = *SU.getInstr(); - if (!shouldScheduleAdjacent(ST, &Pred, Branch)) - continue; - - // Create a single weak edge from SU to ExitSU. The only effect is to cause - // bottom-up scheduling to heavily prioritize the clustered SU. There is no - // need to copy predecessor edges from ExitSU to SU, since top-down - // scheduling cannot prioritize ExitSU anyway. To defer top-down scheduling - // of SU, we could create an artificial edge from the deepest root, but it - // hasn't been needed yet. - bool Success = DAG->addEdge(&ExitSU, SDep(&SU, SDep::Cluster)); - (void)Success; - assert(Success && "No DAG nodes should be reachable from ExitSU"); - - // Adjust latency of data deps between the nodes. - for (SDep &PredDep : ExitSU.Preds) - if (PredDep.getSUnit() == &SU) - PredDep.setLatency(0); - for (SDep &SuccDep : SU.Succs) - if (SuccDep.getSUnit() == &ExitSU) - SuccDep.setLatency(0); - - ++NumFused; - DEBUG(dbgs() << DAG->MF.getName() << "(): Macro fuse "; - SU.print(dbgs(), DAG); - dbgs() << " - ExitSU" - << " / " << DAG->TII->getName(Pred.getOpcode()) << " - " - << DAG->TII->getName(Branch->getOpcode()) << '\n';); - - break; - } -} - -} // end namespace - namespace llvm { std::unique_ptr createX86MacroFusionDAGMutation () { - return EnableMacroFusion ? make_unique() : nullptr; + return createBranchMacroFusionDAGMutation(shouldScheduleAdjacent); } } // end namespace llvm -- cgit v1.2.3