THURSDAY, JULY 30, 2026
match.pd Performance Win
match.pd: Add rule for masked OR comparisons.
Adds a new pattern matching rule for optimizing masked OR comparisons.
This commit introduces a new optimization rule to GCC’s pattern matching system (match.pd) for simplifying masked OR comparisons. It addresses the dual of an existing rule: (X | C) == (Y | C) now folds to ((X ^ Y) & C) == 0. This optimization is demonstrated with an example of case-insensitive ASCII comparison, allowing for more efficient code generation on architectures like aarch64.
In Details
Adds a new simplification rule to match.pd: (X | C) ==/!= (Y | C) now simplifies to ((X ^ Y) & C) == 0. This is the dual of the existing (X & C) == (Y & C) rule. The rule is placed alongside its bit_and dual within the eq and ne iterators. Tested on aarch64-none-linux-gnu.
For Context
- match.pd
- A file in GCC containing pattern-matching rules used for instruction selection and optimization. It defines how to recognize and transform specific code patterns.
- masked OR comparison
- A comparison operation where two values are first bitwise ORed with a mask constant, and then the results are compared for equality or inequality.
- dual
- In this context, the dual refers to a complementary mathematical or logical operation. If a rule applies to AND, its dual might apply to OR, XOR, etc.
- aarch64
- The 64-bit ARM architecture. Compilers generate code specifically optimized for this processor family.
- XOR
- Exclusive OR. A bitwise logical operation that outputs true only when inputs differ.
- TST
- Test bits. An instruction that performs a bitwise AND between a register and an immediate value, and sets the condition flags based on the result.