[llvm] [LoongArch] Enable callee saved register optimization (PR #226809)
via llvm-commits
llvm-commits at lists.llvm.org
Sat Oct 3 04:08:48 PDT 2026
lrzlin wrote:
@bojle I've tested SPEC CPU2017's `500.perlbench` on Loongson 3B6000 too, it shows a 0.8% improments on `CSRCostScale` set to 30 and 1.6% on 80, which basically on par with the AArch64 results. And on SPEC CPU2006, the regression on 30 shrink to 0.5% this time, but got worse on 80 (-3.01%), So the difference on testsuite makes sense.
The perl which used by 2006 heavily depend on recursive, while the 2017 one moved to iteration, this is the reason why IMO. We still need more work on the heuristic algorthim for CSR optimization.
https://github.com/llvm/llvm-project/pull/226809
More information about the llvm-commits
mailing list