Alive-Infer: data-driven precondition inference for peephole optimizations in LLVM

Abstract

Peephole optimizations are a common source of compiler bugs. Compiler developers typically transform an incorrect peephole optimization into a valid one by strengthening the precondition. This process is challenging and tedious. This paper proposes ALIVE-INFER, a data-driven approach that infers preconditions for peephole optimizations expressed in Alive. ALIVE-INFER generates positive and negative examples for an optimization, enumerates predicates on-demand, and learns a set of predicates that separate the positive and negative examples. ALIVE-INFER repeats this process until it finds a precondition that ensures the validity of the optimization. ALIVE-INFER reports both a weakest precondition and a set of succinct partial preconditions to the developer. Our prototype generates preconditions that are weaker than LLVM's preconditions for 73 optimizations in the Alive suite. We also demonstrate the applicability of this technique to generalize 54 optimization patterns generated by Souper, an LLVM IR-based superoptimizer.

Rewrite Rule Inference Using Equality Saturation

“…Furthermore, some rewrites are total but still depend on some condition being met: for example |𝑥 | ↭ 𝑥 only when 𝑥 is non-negative. An extension to Ruler could possibly to infer such side conditions based on "near cvec matches" where only a few entries differ between two e-classes's cvecs, building on prior work [Menendez and Nagarakatte 2017] that infers preconditions for peephole optimizations.…”

“…These rulesets are typically written by a programmer and therefore can have errors or may not be complete. Several tools have automated peephole optimization generation [Bansal and Aiken 2006;Davidson and Fraser 2004;Menendez and Nagarakatte 2017]. Bansal and Aiken [2006] presented a tool for automatically inferring peephole optimizations using superoptimization, using exhaustive enumeration to generate terms up to a certain depth, and leveraging canonicalization to reduce the search space.…”

“…Several other papers [Schkufza et al 2014;Sharma et al 2015] have extended and/or use STOKE for synthesizing superoptimizations. Alive-Infer [Menendez and Nagarakatte 2017] is a tool for automatically generating pre-conditions for peephole optimizations for LLVM. Alive-Infer works in three stages: first, it generates positive and negative examples whose validity is checked using an SMT solver.…”

“…Several noteworthy projects have developed tool-specific techniques for checking or inferring rules [Bansal and Aiken 2006;Joshi et al 2002;Menendez and Nagarakatte 2017;Singh and Solar-Lezama 2016], but implementing a rewrite system still generally requires domain experts to first manually develop rulesets by trial and error. Such slow, ad hoc, and error-prone approaches hinder design space exploration for new domains and discourage updating existing systems.…”

Synthesizing Safe and Efficient Kernel Extensions for Packet Processing

“…Synthesis and optimization typically scale to large programs by operating over a sequence of smaller windows, where a window is a contiguous sequence of instructions as they appear in the text of the program [88,94,101,111]. Small windows (rather than the full programs) lead to smaller verification conditions, which are much faster to solve.…”