This repository contains a PIP package which is an OpenAI environment for simulating an environment for OptClang.

gym-OptClang

The gym-OptClang is an OpanAI Gym compatiable interface.
However, do not assume the usage of APIs are as same as the Gym
Read the document and source code may help you a lot.

Installation

Install the OpenAI gym.

Then install this package via

sudo -H pip3 install -e . --upgrade

Usage

import gym
import gym_OptClang

env = gym.make('OptClang-v0')

The Environment

• Observation
  • Type: BOX(4176)

Num	Observation	Min	Max
0	Added by PassRewriter	0	Maximum number of unsigned int(Based on the instrumentation runtime implementation)
…	…	…	…
4175	Added by PassRewriter	0	Maximum number of unsigned int(Based on the instrumentation runtime implementation)

• Actions
  • Type: Discrete(34)
  • The action start from “0”, but the modified Clang start from 1
    • This is already handled inside gym-OptClang

Num	Action (Apply pass)
0	PGOMemOPSizeOptLegacyPass
1	InstructionCombiningPass
2	InstructionSimplifierPass
3	LowerSwitchPass
4	PromoteMemoryToRegisterPass
5	BitTrackingDCEPass
6	LoopDataPrefetchPass
7	ConstantHoistingPass
8	SROAPass
9	GVNSinkPass
10	SCCPPass
11	ScalarizerPass
12	JumpThreadingPass
13	NaryReassociatePass
14	MergedLoadStoreMotionPass
15	DeadStoreEliminationPass
16	SinkingPass
17	EarlyCSEPass
18	FlattenCFGPass
19	DeadCodeEliminationPass
20	DemoteRegisterToMemoryPass
21	PlaceSafepointsPass
22	AlignmentFromAssumptionsPass
23	PartiallyInlineLibCallsPass
24	AggressiveDCEPass
25	StraightLineStrengthReducePass
26	GVNPass
27	TailCallEliminationPass
28	NewGVNPass
29	GVNHoistPass
30	ConstantPropagationPass
31	ReassociatePass
32	MemCpyOptPass
33	LowerExpectIntrinsicPass

• Reward

  • Indictating status
    • Success: 1
    • Fail: -1
  • Real reward calculating based on the returned "Dict" during runtime
    • Please refer to source code of the agent for more details.

• Starting State

  • Random select a program

• Episode Termination

  • The episode ends after you apply N passes
    • Based on the random results, 9 is good choice

• Solved Requirements

  • As same as the “Episode Termination”

License

Refer to LICENSE