LLVM 17.0.0git
SpeculativeExecution.h
Go to the documentation of this file.
1//===- SpeculativeExecution.h -----------------------------------*- C++ -*-===//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8//
9// This pass hoists instructions to enable speculative execution on
10// targets where branches are expensive. This is aimed at GPUs. It
11// currently works on simple if-then and if-then-else
12// patterns.
13//
14// Removing branches is not the only motivation for this
15// pass. E.g. consider this code and assume that there is no
16// addressing mode for multiplying by sizeof(*a):
17//
18// if (b > 0)
19// c = a[i + 1]
20// if (d > 0)
21// e = a[i + 2]
22//
23// turns into
24//
25// p = &a[i + 1];
26// if (b > 0)
27// c = *p;
28// q = &a[i + 2];
29// if (d > 0)
30// e = *q;
31//
32// which could later be optimized to
33//
34// r = &a[i];
35// if (b > 0)
36// c = r[1];
37// if (d > 0)
38// e = r[2];
39//
40// Later passes sink back much of the speculated code that did not enable
41// further optimization.
42//
43// This pass is more aggressive than the function SpeculativeyExecuteBB in
44// SimplifyCFG. SimplifyCFG will not speculate if no selects are introduced and
45// it will speculate at most one instruction. It also will not speculate if
46// there is a value defined in the if-block that is only used in the then-block.
47// These restrictions make sense since the speculation in SimplifyCFG seems
48// aimed at introducing cheap selects, while this pass is intended to do more
49// aggressive speculation while counting on later passes to either capitalize on
50// that or clean it up.
51//
52// If the pass was created by calling
53// createSpeculativeExecutionIfHasBranchDivergencePass or the
54// -spec-exec-only-if-divergent-target option is present, this pass only has an
55// effect on targets where TargetTransformInfo::hasBranchDivergence() is true;
56// on other targets, it is a nop.
57//
58// This lets you include this pass unconditionally in the IR pass pipeline, but
59// only enable it for relevant targets.
60//
61//===----------------------------------------------------------------------===//
62#ifndef LLVM_TRANSFORMS_SCALAR_SPECULATIVEEXECUTION_H
63#define LLVM_TRANSFORMS_SCALAR_SPECULATIVEEXECUTION_H
64
65#include "llvm/IR/PassManager.h"
66
67namespace llvm {
68class TargetTransformInfo;
70 : public PassInfoMixin<SpeculativeExecutionPass> {
71public:
72 SpeculativeExecutionPass(bool OnlyIfDivergentTarget = false);
73
75
76 // Glue for old PM
78
79private:
80 bool runOnBasicBlock(BasicBlock &B);
81 bool considerHoistingFromTo(BasicBlock &FromBlock, BasicBlock &ToBlock);
82
83 // If true, this pass is a nop unless the target architecture has branch
84 // divergence.
85 const bool OnlyIfDivergentTarget = false;
86
87 TargetTransformInfo *TTI = nullptr;
88};
89}
90
91#endif // LLVM_TRANSFORMS_SCALAR_SPECULATIVEEXECUTION_H
static GCRegistry::Add< OcamlGC > B("ocaml", "ocaml 3.10-compatible GC")
#define F(x, y, z)
Definition: MD5.cpp:55
This header defines various interfaces for pass management in LLVM.
A container for analyses that lazily runs them and caches their results.
Definition: PassManager.h:620
LLVM Basic Block Representation.
Definition: BasicBlock.h:56
A set of analyses that are preserved following a run of a transformation pass.
Definition: PassManager.h:152
bool runImpl(Function &F, TargetTransformInfo *TTI)
PreservedAnalyses run(Function &F, FunctionAnalysisManager &AM)
This pass provides access to the codegen interfaces that are needed for IR-level transformations.
This is an optimization pass for GlobalISel generic memory operations.
Definition: AddressRanges.h:18
A CRTP mix-in to automatically provide informational APIs needed for passes.
Definition: PassManager.h:371