| Message ID | 20230228111917.82910-1-jiaxun.yang@flygoat.com (mailing list archive) |
|---|---|
| State | Superseded |
| Headers | show |
| Series | [v2] MIPS: Workaround clang inline compat branch issue | expand |
On Tue, Feb 28, 2023 at 11:19:17AM +0000, Jiaxun Yang wrote: > Clang is unable to handle the situation that a chunk of inline > assembly ends with a compat branch instruction and then compiler > generates another control transfer instruction immediately after > this compat branch. The later instruction will end up in forbidden > slot and cause exception. > > Workaround by add a option to control the use of compact branch. > Currently it's selected by CC_IS_CLANG and hopefully we can change > it to a version check in future if clang manages to fix it. > > Fix boot on boston board. > > Link: https://github.com/llvm/llvm-project/issues/61045 > Signed-off-by: Jiaxun Yang <jiaxun.yang@flygoat.com> Thanks for filing that bug and debugging this. I came across a post from the (now former) MIPS code owner and I am not sure anyone else has stepped up, the vast majority of commits to llvm/lib/Targets/Mips have been either NFC commits or treewide refactorings, so it will be interesting to see if we can get that fixed. https://discourse.llvm.org/t/mips-backend-code-owner/60737 https://github.com/llvm/llvm-project/commit/7daed359111f6d151fef447f520f85ef1dabedf6 For now: Acked-by: Nathan Chancellor <nathan@kernel.org> If a v3 is necessary for some reason, a link to that issue in the source itself would be useful, as we can easily grep for 'llvm/llvm-project' to audit issues; I am not sure it is worth sending a v3 just for that alone though. > --- > v2: Add Link tag to LLVM bug > --- > arch/mips/Kconfig | 3 +++ > arch/mips/include/asm/asm.h | 2 +- > 2 files changed, 4 insertions(+), 1 deletion(-) > > diff --git a/arch/mips/Kconfig b/arch/mips/Kconfig > index 37072e15b263..adf2c5a0bdba 100644 > --- a/arch/mips/Kconfig > +++ b/arch/mips/Kconfig > @@ -3206,6 +3206,9 @@ config CC_HAS_MNO_BRANCH_LIKELY > def_bool y > depends on $(cc-option,-mno-branch-likely) > > +config CC_HAS_BROKEN_INLINE_COMPAT_BRANCH > + def_bool y if CC_IS_CLANG > + > menu "Power management options" > > config ARCH_HIBERNATION_POSSIBLE > diff --git a/arch/mips/include/asm/asm.h b/arch/mips/include/asm/asm.h > index 336ac9b65235..2e99450f4228 100644 > --- a/arch/mips/include/asm/asm.h > +++ b/arch/mips/include/asm/asm.h > @@ -336,7 +336,7 @@ symbol = value > */ > #ifdef CONFIG_WAR_R10000_LLSC > # define SC_BEQZ beqzl > -#elif MIPS_ISA_REV >= 6 > +#elif !defined(CONFIG_CC_HAS_BROKEN_INLINE_COMPAT_BRANCH) && MIPS_ISA_REV >= 6 > # define SC_BEQZ beqzc > #else > # define SC_BEQZ beqz > -- > 2.37.1 (Apple Git-137.1) >
> 2023年2月28日 19:21，Nathan Chancellor <nathan@kernel.org> 写道： > > On Tue, Feb 28, 2023 at 11:19:17AM +0000, Jiaxun Yang wrote: >> Clang is unable to handle the situation that a chunk of inline >> assembly ends with a compat branch instruction and then compiler >> generates another control transfer instruction immediately after >> this compat branch. The later instruction will end up in forbidden >> slot and cause exception. >> >> Workaround by add a option to control the use of compact branch. >> Currently it's selected by CC_IS_CLANG and hopefully we can change >> it to a version check in future if clang manages to fix it. >> >> Fix boot on boston board. >> >> Link: https://github.com/llvm/llvm-project/issues/61045 >> Signed-off-by: Jiaxun Yang <jiaxun.yang@flygoat.com> > > Thanks for filing that bug and debugging this. I came across a post from > the (now former) MIPS code owner and I am not sure anyone else has > stepped up, the vast majority of commits to llvm/lib/Targets/Mips have > been either NFC commits or treewide refactorings, so it will be > interesting to see if we can get that fixed. Yunqiang Su <syq@debian.org>[1] said he will take a look on that issue. > > https://discourse.llvm.org/t/mips-backend-code-owner/60737 > https://github.com/llvm/llvm-project/commit/7daed359111f6d151fef447f520f85ef1dabedf6 > > For now: > > Acked-by: Nathan Chancellor <nathan@kernel.org> > > If a v3 is necessary for some reason, a link to that issue in the source > itself would be useful, as we can easily grep for 'llvm/llvm-project' to > audit issues; I am not sure it is worth sending a v3 just for that alone > though. Will do a v3, it also helps me document the actual issue. Thanks [1]: https://reviews.llvm.org/p/wzssyqa/ - Jiaxun >
diff --git a/arch/mips/Kconfig b/arch/mips/Kconfig index 37072e15b263..adf2c5a0bdba 100644 --- a/arch/mips/Kconfig +++ b/arch/mips/Kconfig @@ -3206,6 +3206,9 @@ config CC_HAS_MNO_BRANCH_LIKELY def_bool y depends on $(cc-option,-mno-branch-likely) +config CC_HAS_BROKEN_INLINE_COMPAT_BRANCH + def_bool y if CC_IS_CLANG + menu "Power management options" config ARCH_HIBERNATION_POSSIBLE diff --git a/arch/mips/include/asm/asm.h b/arch/mips/include/asm/asm.h index 336ac9b65235..2e99450f4228 100644 --- a/arch/mips/include/asm/asm.h +++ b/arch/mips/include/asm/asm.h @@ -336,7 +336,7 @@ symbol = value */ #ifdef CONFIG_WAR_R10000_LLSC # define SC_BEQZ beqzl -#elif MIPS_ISA_REV >= 6 +#elif !defined(CONFIG_CC_HAS_BROKEN_INLINE_COMPAT_BRANCH) && MIPS_ISA_REV >= 6 # define SC_BEQZ beqzc #else # define SC_BEQZ beqz
Clang is unable to handle the situation that a chunk of inline assembly ends with a compat branch instruction and then compiler generates another control transfer instruction immediately after this compat branch. The later instruction will end up in forbidden slot and cause exception. Workaround by add a option to control the use of compact branch. Currently it's selected by CC_IS_CLANG and hopefully we can change it to a version check in future if clang manages to fix it. Fix boot on boston board. Link: https://github.com/llvm/llvm-project/issues/61045 Signed-off-by: Jiaxun Yang <jiaxun.yang@flygoat.com> --- v2: Add Link tag to LLVM bug --- arch/mips/Kconfig | 3 +++ arch/mips/include/asm/asm.h | 2 +- 2 files changed, 4 insertions(+), 1 deletion(-)