match.pd: Remove bit set/bit clear branch mispredict
GCC now avoids branch mispredictions for certain bitwise operations by replacing conditional bit setting/clearing with unconditional operations.
GCC now avoids branch mispredictions in specific bitwise operation scenarios. The compiler now uses unconditional bit setting/clearing instead of conditional operations when checking if a single bit is not set (and setting it) or if a bitmask is set (and clearing it). This change eliminates potential branch mispredictions, improving performance by avoiding unnecessary flushing of the instruction pipeline. The update adds two new patterns to match.pd to accomplish this.
In Details
This commit modifies match.pd, adding patterns to handle specific bitwise operations that previously led to branch mispredictions. The patterns address scenarios where a bit is conditionally set/cleared based on whether it's already set/clear. By replacing the conditional operations with unconditional ones, the compiler avoids mispredictions and improves code execution speed. The relevant patterns are (cond (bit_and A IMM) (bit_or A IMM) A) and (cond (bit_and A IMM) (bit_and A ~IMM) A).
For Context
Modern processors use branch prediction to optimize performance by guessing which path a program will take at conditional branches. When the prediction is wrong, it's called a branch misprediction, which can stall the processor. This commit avoids branch mispredictions for certain bitwise operations, common operations for manipulating individual bits within a larger value. By using unconditional operations instead of conditional ones, the compiler ensures the processor always executes the correct path, improving performance.