blob: 0d081c0194bfbdfe12229875251ed92ee88d7476 (
plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
|
//===- 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
|
