llvm/llvm/lib/Target/SystemZ/SystemZTDC.cpp

//===-- SystemZTDC.cpp - Utilize Test Data Class instruction --------------===//
//
// 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 pass looks for instructions that can be replaced by a Test Data Class
// instruction, and replaces them when profitable.
//
// Roughly, the following rules are recognized:
//
// 1: fcmp pred X, 0 -> tdc X, mask
// 2: fcmp pred X, +-inf -> tdc X, mask
// 3: fcmp pred X, +-minnorm -> tdc X, mask
// 4: tdc (fabs X), mask -> tdc X, newmask
// 5: icmp slt (bitcast float X to int), 0 -> tdc X, mask [ie. signbit]
// 6: icmp sgt (bitcast float X to int), -1 -> tdc X, mask
// 7: icmp ne/eq (call @llvm.s390.tdc.*(X, mask)) -> tdc X, mask/~mask
// 8: and i1 (tdc X, M1), (tdc X, M2) -> tdc X, (M1 & M2)
// 9: or i1 (tdc X, M1), (tdc X, M2) -> tdc X, (M1 | M2)
// 10: xor i1 (tdc X, M1), (tdc X, M2) -> tdc X, (M1 ^ M2)
//
// The pass works in 4 steps:
//
// 1. All fcmp and icmp instructions in a function are checked for a match
//    with rules 1-3 and 5-7.  Their TDC equivalents are stored in
//    the ConvertedInsts mapping.  If the operand of a fcmp instruction is
//    a fabs, it's also folded according to rule 4.
// 2. All and/or/xor i1 instructions whose both operands have been already
//    mapped are mapped according to rules 8-10.  LogicOpsWorklist is used
//    as a queue of instructions to check.
// 3. All mapped instructions that are considered worthy of conversion (ie.
//    replacing them will actually simplify the final code) are replaced
//    with a call to the s390.tdc intrinsic.
// 4. All intermediate results of replaced instructions are removed if unused.
//
// Instructions that match rules 1-3 are considered unworthy of conversion
// on their own (since a comparison instruction is superior), but are mapped
// in the hopes of folding the result using rules 4 and 8-10 (likely removing
// the original comparison in the process).
//
//===----------------------------------------------------------------------===//

#include "SystemZ.h"
#include "SystemZSubtarget.h"
#include "llvm/ADT/MapVector.h"
#include "llvm/CodeGen/TargetPassConfig.h"
#include "llvm/IR/Constants.h"
#include "llvm/IR/IRBuilder.h"
#include "llvm/IR/InstIterator.h"
#include "llvm/IR/Instructions.h"
#include "llvm/IR/IntrinsicInst.h"
#include "llvm/IR/IntrinsicsS390.h"
#include "llvm/IR/LegacyPassManager.h"
#include "llvm/IR/Module.h"
#include "llvm/Target/TargetMachine.h"
#include <deque>
#include <set>

usingnamespacellvm;

namespace {

class SystemZTDCPass : public FunctionPass { … };

} // end anonymous namespace

char SystemZTDCPass::ID = …;
INITIALIZE_PASS(…)

FunctionPass *llvm::createSystemZTDCPass() { … }

void SystemZTDCPass::convertFCmp(CmpInst &I) { … }

void SystemZTDCPass::convertICmp(CmpInst &I) { … }

void SystemZTDCPass::convertLogicOp(BinaryOperator &I) { … }

bool SystemZTDCPass::runOnFunction(Function &F) { … }