AMD has posted a new patch for the GNU Compiler Collection (GCC) aimed at improving optimization for x86_64 code. The optimization, termed Extended Load Optimization, is designed to benefit code where a narrower data type is loaded and then used in an arithmetic operation with a wider data type.
The patch, submitted by Sridevi from AMD’s compiler team, addresses a situation where GCC currently uses `vextract` instructions. Sridevi explained that the patch will allow GCC to instead utilize extended loads with ZMM registers, potentially achieving better performance. The patch identifies a “load + convert” pattern within the SLP tree and narrows the load vector type if the target supports fused load-and-extend functionality. The function `vect_optimize_slp_widening_loads` performs this optimization within the `tree-vect-loop.cc` file.
This optimization is not specific to AMD’s Zen processors but applies to the broader x86_64 architecture. AMD has recently shifted towards upstreaming new compiler support, including Zen target support for both GCC and LLVM/Clang, earlier in the development process. Previously, much of this work was handled by partners like SUSE post-launch. The patch is currently under review for inclusion in GCC 17. No performance metrics were included with the patch submission.
Read the original coverage
💬 Comments
📜 Comment Policy