Extend can_open_code_p to handle bswap operations for store merging
GCC now uses `can_open_code_p` to determine if `bswap` operations can be open-coded during store merging, improving scalar optimization and fixing a RISC-V tes…
This commit refactors GCC’s store merging pass to rely on can_open_code_p for determining if bswap (byte swap) operations can be generated as inline code. Previously, this logic was duplicated and less efficient within gimple-ssa-store-merging.cc. By centralizing this check in can_open_code_p and extending it to handle various bswap cases, the compiler can more effectively optimize scalar code and resolve a bug in gcc.dg/optimize-bswapsi-6.c on RISC-V architectures.
In Details
GCC's gimple-ssa-store-merging pass attempts to coalesce multiple small stores into larger, more efficient ones, particularly when dealing with vector constructors or byte-swapping operations. The can_open_code_p function in optabs-query.cc serves as a centralized mechanism for determining if an operation can be expressed using direct machine instructions rather than a library call. This change integrates bswap operation checks into can_open_code_p, moving logic previously scattered across maybe_optimize_vector_constructor, pass_optimize_bswap::execute, and `imm_store_chain_info…
For Context
When a computer program stores data in memory, it often does so in small pieces. Compilers like GCC try to optimize this by combining several small storage operations into one larger, more efficient operation, a process called "store merging." One specific operation that can benefit from this is "byte swapping" (bswap), which reorders the bytes within a word of data. This commit improves how GCC handles byte swapping during store merging. Previously, the compiler had separate, somewhat redundant checks to see if it could generate direct machine instructions for byte swaps. This change centralizes that decision-making into a single function, can_open_code_p, which is responsible for determining if an operation can be directly translated into machine code. This makes the compiler more efficient and fixes a specific problem that caused incorrect code generation for byte-swapping operations on RISC-V processors.