On Fri, 21 Aug 2026 18:42:07 PDT (-0700), [email protected] wrote:
Hi,
Thanks for the detailed explanation. I understand the concern regarding the breaking change and the lack of ISA-based dynamic multilib resolution.
To address this, I have experimented with alternative approaches and laid out
the multilibs in the sysroot using the following directory structure in my
local environment:
They're: usr/lib<xlen>/<arch>/<abi>
Could you please share your thoughts on this sysroot layout? Do you think this
structure could serve as a practical workaround or a step towards better
multilib organization?
Something like that would be part of the equation, but it doesn't handle
the dynamic detection side of things. It'd basically end up in the same
spot we're in for the -elf toolchains, where we're just doing build-time
multilib detection. That works fine in -elf land, but I think most
people who want to do build-time stuff in the Linux toolchains just
build a non-multilib toolchain because they're going to end up
rebuilding the whole distro anyway (as opposed to in -elf, where people
tend to have externally built toolchains and projects with somewhat
self-contained build systems).
Evan had a patch for glibc-hwcaps a while ago, the general opinion was
that the implementation was fine but we needed to figure out what
targets we were going to support. I think none of that has really
changed, but if you're intrested in this stuff that's probably the best
place to start. I'm pretty confident Evan won't mind someone picking it
up ;)
[email protected]
From: Palmer Dabbelt
Date: 2026-08-22 00:23
To: Kito Cheng
CC: panciyan; gcc-patches; Richard.Earnshaw; jeffreyalaw
Subject: Re: [PATCH] RISC-V: Fix glibc multilib configure
On Fri, 21 Aug 2026 00:22:42 PDT (-0700), Kito Cheng wrote:
This seems more like a patch for riscv-gnu-toolchain rather than GCC?
Ya, I agree this is basically just a distro problem: if you're building
binaries you need to know what's implemented on the systems you're going
to be running them on.
That said: we did change the default behavior here so it's going to
break some users. We also don't have a great way to fix that, as we're
not doing ISA-based dynamic multilib resolution. So if I understand how
this all works, we've basically changed the default multilib to include
V when it didn't on older releases.
We've talked about trying to fix this ISA-dependent multilib resolution
stuff a few times. Last round we sort of got hung up on the "what
should the supported set of multilibs actually be", and I suspect we'd
end up there again (and likely will continue to end up there until
someone makes hardware that's actually interesting in these
binary-compatible use cases).
panciyan <[email protected]> 於 2026年8月19日週三 下午1:48寫道：
Building glibc with -march=rv64gcv causes GCC to auto-vectorize generic
C code (e.g. calc_state_hash in regex), generating RVV instructions. On
hardware without the V extension, this results in SIGILL.
Signed-off-by: Ciyan Pan <[email protected]>
ChangeLog:
* configure: Regenerate.
* configure.ac: Remove rv64gcv-lp64d.
---
configure | 5 -----
configure.ac | 2 +-
2 files changed, 1 insertion(+), 6 deletions(-)
diff --git a/configure b/configure
index d610aa16..6328fcee 100755
--- a/configure
+++ b/configure
@@ -4159,11 +4159,6 @@ fi
if test "x$enable_multilib" = xyes
then :
glibc_multilib_names="rv32imac-ilp32 rv32gc-ilp32d rv64imac-lp64
rv64gc-lp64d"
- # Vector extension is currently not supported on Big-endian systems.
- if test "x$endian" = xlittle
- then :
- glibc_multilib_names+=" rv64gcv-lp64d"
- fi
else $as_nop
glibc_multilib_names="$with_arch-$with_abi"
diff --git a/configure.ac b/configure.ac
index 4e534faa..dd32ec6c 100644
--- a/configure.ac
+++ b/configure.ac
@@ -158,7 +158,7 @@ AS_IF([test "x$enable_multilib" = xyes || test
"x$with_multilib_generator" != xn
[AC_SUBST(multilib_flags,--disable-multilib)])
AS_IF([test "x$enable_multilib" = xyes],
- [AC_SUBST(glibc_multilib_names,"rv32imac-ilp32 rv32gc-ilp32d rv64imac-lp64
rv64gc-lp64d rv64gcv-lp64d")],
+ [AC_SUBST(glibc_multilib_names,"rv32imac-ilp32 rv32gc-ilp32d rv64imac-lp64
rv64gc-lp64d")],
[AC_SUBST(glibc_multilib_names,"$with_arch-$with_abi")])
AS_IF([test "x$enable_multilib" = xyes],
--
2.43.0
