Currently polly is run at the beginning of the pass chain as the code has at this point less scalar dependences as later in the pass chain, such that our optimizations are more often triggered. However, running Polly early requires us to schedule additional canonicalization passes which increase compile time (slightly, but consistently) and also cause execution time noise compared to a plain 'clang -O3' compilation.
To avoid any execution time noise, as well as compile time increases for compilations that do not benefit from Polly, we want to fold Polly into the pass pipeline: -mllvm -polly-position=before-vectorizer
(this also enables better optimizations for C++ code, which requires inlining to expose interesting code snippets)
This bug tracks the different issues that currently prevent us from doing so.
Quuxplusone
commented
8 years ago
Currently polly is run at the beginning of the pass chain as the code has at this point less scalar dependences as later in the pass chain, such that our optimizations are more often triggered. However, running Polly early requires us to schedule additional canonicalization passes which increase compile time (slightly, but consistently) and also cause execution time noise compared to a plain 'clang -O3' compilation.
To avoid any execution time noise, as well as compile time increases for compilations that do not benefit from Polly, we want to fold Polly into the pass pipeline: -mllvm -polly-position=before-vectorizer
(this also enables better optimizations for C++ code, which requires inlining to expose interesting code snippets)
This bug tracks the different issues that currently prevent us from doing so.