I have written patches for FEAT_CMPBR support in QEMU, and wanted to test them out with gcc. The easiest way, seemed to be bootstrapping gcc with cmpbr enabled. The attempt failed on stage1 libgcc.
My bug report is target/121385. Pinski did some analyis, which was very helpful, and guides most of this patch set. Pinski filed a second PR, target/121388, vs bid_binarydecimal.c. I think that might now be in hiding rather than actually being fixed, since I did nothing that should have affected changing the sense of the branch during cfglayout. I suspect we need to accept a symmetric range of constants for all comparisons, expand those that don't fit to cmp+bcond. While we could have post-reload splitting, I'm not sure how split works with shorten_branches, so we might have to retain cmp+bcond all the way through to output. If that's true, and we assume that performant cpus will fuse cmp+bcond anyway, it might not be worth bothering to split. My bootstrap under emulation is still going, 8 hours in. In hindsight I should have used a bigger machine, but laziness used the one with the chroot already set up. This passes native bootstrap and make check on a neoverse-n1, so I don't seem to have broken anything obvious. r~ Richard Henderson (8): aarch64: Drop label format argument from aarch64_gen_far_branch aarch64: Fix spelling of BRANCH_LEN_N_1KiB aarch64: Drop cbranch<SHORT>4 expander aarch64: Disable TARGET_CMPBR with aarch64_track_speculation aarch64: Use aarch64_gen_compare_zero_and_branch in aarch64_restore_za aarch64: Add cc clobber to compare-and-branch patterns aarch64: Consider TARGET_CMPBR in rtx costs aarch64: Use cc when CB/CBB/CBH is out-of-range gcc/config/aarch64/aarch64-protos.h | 5 +- gcc/config/aarch64/aarch64.h | 5 +- gcc/config/aarch64/aarch64-sme.md | 3 +- gcc/config/aarch64/aarch64.cc | 22 +++++-- gcc/config/aarch64/aarch64.md | 94 ++++++++++++++--------------- 5 files changed, 74 insertions(+), 55 deletions(-) -- 2.43.0
