12345678910111213141516171819202122232425262728293031323334 |
- //===- llvm/Transforms/Utils/BypassSlowDivision.h --------------*- C++ -*-===//
- //
- // The LLVM Compiler Infrastructure
- //
- // This file is distributed under the University of Illinois Open Source
- // License. See LICENSE.TXT for details.
- //
- //===----------------------------------------------------------------------===//
- //
- // This file contains an optimization for div and rem on architectures that
- // execute short instructions significantly faster than longer instructions.
- // For example, on Intel Atom 32-bit divides are slow enough that during
- // runtime it is profitable to check the value of the operands, and if they are
- // positive and less than 256 use an unsigned 8-bit divide.
- //
- //===----------------------------------------------------------------------===//
- #ifndef LLVM_TRANSFORMS_UTILS_BYPASSSLOWDIVISION_H
- #define LLVM_TRANSFORMS_UTILS_BYPASSSLOWDIVISION_H
- #include "llvm/ADT/DenseMap.h"
- #include "llvm/IR/Function.h"
- namespace llvm {
- /// This optimization identifies DIV instructions that can be
- /// profitably bypassed and carried out with a shorter, faster divide.
- bool bypassSlowDivision(Function &F,
- Function::iterator &I,
- const DenseMap<unsigned int, unsigned int> &BypassWidth);
- } // End llvm namespace
- #endif
|