Message ID | 20231003222700.909771-1-patrick@rivosinc.com |
---|---|
Headers | show |
Series | Add Ztso atomic mappings | expand |
On 10/3/23 16:26, Patrick O'Neill wrote: > I vaugely recall some discussion about backporting the Ztso mappings > along with the RVWMO mappings. Now that the RVWMO mappings have been > backported for 13.3, is there interest in also backporting the Ztso > mappings? > > Tested using for regressions using rv32gc/rv64gc glibc. > > Jeff Law (1): > [RISCV][committed] Remove spurious newline in ztso sequence > > Patrick O'Neill (2): > RISC-V: Add Ztso atomic mappings > RISC-V: Specify -mabi for ztso testcases I recall discussing Ztso mappings, but not the final conclusion. I think the final decision comes down to the motivation behind the changes. If they're primarily optimized sequences utilizing the stronger ordering guarantees from Ztso, then it's probably not a good candidate for gcc-13. If the primary motivation is to make it easier to port code from targets with stronger memory models (ie x86), then the Ztso work is a reasonable candidate for backporting. Jeff
On 10/4/23 08:53, Jeff Law wrote: > > > On 10/3/23 16:26, Patrick O'Neill wrote: >> I vaugely recall some discussion about backporting the Ztso mappings >> along with the RVWMO mappings. Now that the RVWMO mappings have been >> backported for 13.3, is there interest in also backporting the Ztso >> mappings? >> >> Tested using for regressions using rv32gc/rv64gc glibc. >> >> Jeff Law (1): >> [RISCV][committed] Remove spurious newline in ztso sequence >> >> Patrick O'Neill (2): >> RISC-V: Add Ztso atomic mappings >> RISC-V: Specify -mabi for ztso testcases > I recall discussing Ztso mappings, but not the final conclusion. I > think the final decision comes down to the motivation behind the changes. > > If they're primarily optimized sequences utilizing the stronger > ordering guarantees from Ztso, then it's probably not a good candidate > for gcc-13. If the primary motivation is to make it easier to port > code from targets with stronger memory models (ie x86), then the Ztso > work is a reasonable candidate for backporting. > > Jeff Discussed during the GCC patchworks meeting. The Ztso mappings are a subset of RVWMO fences/ordering constraints (it's an optimization change) [1]. This backport also makes tip-of-tree sync.md files match gcc 13.3+ which simplifies future atomic-related backports. If in the future someone needs Ztso/there is a need to backport we can re-evaluate this backport then. Thanks Patrick [1] https://github.com/riscv-non-isa/riscv-elf-psabi-doc/pull/391