On Jul 29, 2026, at 20:08, Konstantin Belousov <[email protected]> wrote:
> On Wed, Jul 29, 2026 at 03:08:20PM -0700, Mark Millard wrote: >> On 7/29/26 13:44, John Baldwin wrote: >>> On 7/20/26 17:51, Konstantin Belousov wrote: >>>> The branch main has been updated by kib: >>>> >>>> URL: https://cgit.FreeBSD.org/src/commit/? >>>> id=6e93f5e4d6932c423b89dff8fc08d86f8bdeb7b9 >>>> >>>> commit 6e93f5e4d6932c423b89dff8fc08d86f8bdeb7b9 >>>> Author: Konstantin Belousov <[email protected]> >>>> AuthorDate: 2026-02-13 11:24:40 +0000 >>>> Commit: Konstantin Belousov <[email protected]> >>>> CommitDate: 2026-07-20 21:51:02 +0000 >>>> >>>> amd64: FRED support >>>> FRED support as defined starting from the SDM rev. 90, >>>> requires a new >>>> 'events' entry point to receive user and kernel mode exceptions and >>>> interrupts notifications from the hardware. A minimal asm >>>> trampoline is >>>> enough, rest can be implemented in C due to the clean FRED >>>> organization >>>> of the event reporting. >>>> The syscall entry is handled by a microptimized assembly >>>> path, directly >>>> calling into the amd64_syscall() handler, instead of the generic >>>> events >>>> entry point. >>>> Tested by: emaste >>>> Sponsored by: The FreeBSD Foundation >>>> MFC after: 1 week >>>> Differential revision: https://reviews.freebsd.org/D55829 >>> >>> This appears to have broken the cross-build jobs on GitHub for the past >>> week >>> which build with clang 15 and clang 18. Are those versions too old to >>> support >>> the instructions used here? >>> >>> Sample output: >>> >>> -------------------------------------------------------------- >>>>>> stage 3.1: building everything >>> -------------------------------------------------------------- >>> /home/runner/work/freebsd-src/freebsd-src/sys/amd64/amd64/ >>> exception.S:1455:2: error: invalid instruction mnemonic 'eretu' >>> eretu >>> ^~~~~ >>> /home/runner/work/freebsd-src/freebsd-src/sys/amd64/amd64/ >>> exception.S:1521:2: error: invalid instruction mnemonic 'lkgs' >>> lkgs %ax >>> ^~~~ >>> /home/runner/work/freebsd-src/freebsd-src/sys/amd64/amd64/ >>> exception.S:1542:2: error: invalid instruction mnemonic 'eretu' >>> eretu >>> ^~~~~ >>> /home/runner/work/freebsd-src/freebsd-src/sys/amd64/amd64/ >>> exception.S:1662:2: error: invalid instruction mnemonic 'erets' >>> erets >>> ^~~~~ >>> /home/runner/work/freebsd-src/freebsd-src/sys/amd64/amd64/ >>> exception.S:1682:2: error: invalid instruction mnemonic 'erets' >>> erets >>> ^~~~~ >>> --- exception.o --- >>> *** [exception.o] Error code 1 >>> >>> (Or maybe clang needs some -m<foo> or .set <foo> to enable use of these >>> instructions?) >>> >> >>> From what I read (but its is only reading): >> >> clang-15 predates any FRED support. >> clang-17 was more incomplete/preliminary/experimental for FRED support >> >> >> clang-18 has .S and inline assembly FRED support but various (most?) >> target triples do not enable it by default. >> >> Possibly add to appropriate .S file(s): >> >> .arch x86-64 >> .arch .fred >> .arch .lkgs >> >> for defined(__clang__) && (__clang_major__ >= 18) >> >> Or, make use of: >> >> -Wa,-mwave-fred -Wa,-mwave-lkgs >> >> >> clang-19 has the intended fully general handling of FRED. > > I think that the adequate workaround would be to #ifdef out the > instructions for old versions of clang. I doubt that anybody would > run the built kernel, esp. on the new enough hardware. > > But I am open for other suggestions. So far as I can tell, clang documentation does not cover this area, making what I read solidly hear-say. (The -mwave-* notation is a possible hint at that status.) It seems that linux may have gotten its source to go through all the tools via use of (what translates to) .byte use for the 3 instructions. It is not clear to me what versions of clang may require -mattr= or -mcpu= to specify something known to support LKGS/FRED in order to enable the notation for the 3 instructions. [Note: For some reason, I only found your note in my Email, not in dev-commits-src-all@ or dev-commits-src-main@ (via multiple ways of accessing those). So this is an Email based reply, not thunderbird based use of news.gmane.io .] === Mark Millard marklmi at yahoo.com
