Proposes factoring common pure operations across PHIs
Guoce Feng proposes an optimization to factor out common pure operations like `sqrt` across PHIs and branches to reduce redundant calls, especially in vectoriz…
Guoce Feng proposes an optimization to factor out common pure operations, such as sqrt, from control-flow joins in GCC. This aims to transform patterns like phi(sqrt(a), sqrt(b), sqrt(c)) into sqrt(phi(a, b, c)) when safe and profitable, potentially reducing redundant calculations in vectorized code. The author seeks guidance on the best implementation approach, considering extending the existing PHI factoring logic or adding a more general common instruction sinking transformation.
- proposer
Proposes factoring common pure operations out of control-flow joins to reduce duplicated calls, using `sqrt` as an example.
“That is, transform a conceptual pattern like phi(sqrt(a), sqrt(b), sqrt(c)) into sqrt(phi(a, b, c)) when the operation is safe to factor out and doing so is profitable.”
- other
Suggests looking at PR123113, which relaxes restrictions in phiopt/cselim, and at the cselib pass, which may already perform this optimization.
- other
Clarifies that PR123113 is about merge blocks with more than two predecessors and suggests the cselib pass might handle the requested optimization.
Technical Tradeoffs
- Factoring out operations may introduce additional overhead if the merged expression is more complex.
- The transformation needs to ensure the operation is truly side-effect-free and safe to move.
- Profitability analysis is required to determine whether the transformation results in a net performance gain.
In Details
This RFC proposes an optimization to sink pure operations out of PHI nodes in GCC's SSA representation. The goal is to transform phi(op(a), op(b), op(c)) into op(phi(a, b, c)) to avoid redundant computation, especially in vectorized code. This touches on the SSA form, PHI nodes, and optimization passes like phiopt, cselim, and cselib.
For Context
In compiler optimization, a PHI node merges values from different control flow paths. This proposal aims to reduce redundant calculations by moving common, side-effect-free operations (like square root) out of these merge points. By performing the operation once on a merged value, instead of multiple times on individual values, the code can become more efficient, particularly when dealing with vector operations.