Prevent reduction simplification cycle in match.pd
Fixes a compiler infinite loop caused by reduction simplification rules.
A cycle in the reduction simplification rules within match.pd could cause the compiler to enter an infinite loop, exhausting memory. This issue occurred when transformation rules for reductions, particularly those involving constants and scalable vectors, repeatedly applied and undid each other. The fix separates reduction rules with inverse pairs from those without, ensuring that constant reductions only apply when they simplify to a scalar, thus breaking the cycle.
In Details
This commit addresses PR tree-optimization/127039 by modifying the match.pd reduction rules. The REDUC (@0 op VECTOR_CST) rule, which transforms REDUC (V op C) into REDUC (V) op REDUC (C), was prone to cycling with its inverse REDUC_PAIR rule, especially when REDUC (C) did not fold for scalable vectors. The fix restricts the constant reduction rule to only succeed when REDUC (C) simplifies to a scalar operand, preventing the inverse rule from matching and breaking the infinite loop during forward propagation.
- match.pd
- A file within GCC that defines pattern-matching rules for various optimization passes, particularly for hardware-specific transformations and general tree optimizations.
- reduction
- An optimization pattern where an operation is applied cumulatively across a sequence of elements, such as summing all elements in a vector.
REDUC(op)is a GCC internal representation for such operations. - simplification cycle
- A situation where compiler optimization rules repeatedly transform code into a different form, and then other rules transform it back, leading to an infinite loop or excessive resource consumption.
- scalable vector
- A type of vector processing where the vector length is determined at runtime or is programmable, rather than being fixed at compile time. This is common in architectures like ARM's SVE or RISC-V's V extension.
- forward propagation
- A compiler optimization phase that involves propagating values, properties, or conditions through the control flow graph of a program to discover new optimization opportunities.