Initial commit

Fix riscv64 and x86_64_v2 builds of gcc-toolset-16-gcc, and stop the v2
build from configuring the compiler for x86-64-v3.

--with-arch_64 is guarded by %ifarch x86_64 in the spec, and AlmaLinux's
rpm treats x86_64_v2 as matching x86_64, so the v2 toolset would otherwise
default to generating x86-64-v3 code. Same rule as gcc-toolset-15-gcc. No
bugurl rule is needed here: this spec already uses
--with-bugurl=%dist_bug_report_url, which expands to bugs.almalinux.org.

The other two failures are in the libstdc++ "nonshared" compat sources that
Patch1000 (gcc16-libstdc++-compat.patch) creates, so neither has an
upstream fix -- Red Hat builds neither arch.

riscv64: nonshared98/extfloat.S has no __riscv branch in its arch chain and
no #else fallback, so ALIGN1 reaches gas as a bare token:

    extfloat.S:279: Error: unrecognized opcode `align1'

riscv64 was also missing from the 64-bit POINTER/SIZE group, which would
have emitted a 32-bit typeinfo layout on LP64. Patch taken as is from
autopatch/gcc-toolset-15-gcc, where it is already in use.

x86_64_v2: nonshared20/format-inst.cc hardcodes .hidden for
basic_string<wchar_t>::__resize_and_overwrite<lambda in
__format::__formatter_fp<wchar_t>::_M_localize>. That list was derived at
the el10 baseline -march=x86-64-v3; at v2 the compiler inlines the
instantiation away and emits nothing, so the bare .hidden directive leaves
an undefined hidden symbol in libstdc++_nonshared140.a and the compat test
link at the end of %build fails with

    hidden symbol `_ZNSt7__cxx1112basic_string...' isn't defined
    final link failed: bad value

This is driven by %optflags, not by --with-arch_64, so it needs its own
fix. Emission tracks the AVX2 boundary exactly (v1/v2 emit nothing, v3/v4
emit a 670-byte FUNC WEAK HIDDEN), so guard that one directive on __AVX2__.
This is a no-op at v3 -- the symbol table is unchanged -- and it removes the
dangling reference at v2. All 41 nonshared sources carrying .hidden were
checked at both ISA levels; this is the only symbol that differs.

config.yaml

@@ -0,0 +1,42 @@
+actions:
+  - replace:
+# The spec guards --with-arch_64 with %ifarch x86_64, which also matches the
+# x86_64_v2 arch, so without this the v2 toolset would default to generating
+# x86-64-v3 code.
+      - target: "spec"
+        find: |
+          %if 0%{?rhel} > 9
+          	--with-arch_64=x86-64-v3 \
+        replace: |
+          %if 0%{?rhel} > 9
+          %ifarch x86_64_v2
+          	--with-arch_64=x86-64-v2 \
+          %else
+          	--with-arch_64=x86-64-v3 \
+          %endif
+        count: 1
+
+# No bugurl rule here: unlike gcc-toolset-15-gcc, this spec already uses
+# --with-bugurl=%dist_bug_report_url, which almalinux-release expands to
+# https://bugs.almalinux.org/ .
+
+  - modify_release:
+    - suffix: ".alma.1"
+      enabled: true
+      auto_increment: true
+
+  - changelog_entry:
+      - name: "Eduard Abdullin"
+        email: "eabdullin@almalinux.org"
+        line:
+          - "Fix compilation on RISC-V"
+          - "Do not assert .hidden on a libstdc++ instantiation that is not emitted below x86-64-v3 (fixes x86_64_v2)"
+          - "Set --with-arch_64=x86-64-v2 on v2 arch"
+
+  - add_files:
+      - type: "patch"
+        name: "gcc16-riscv64-extfloat.patch"
+        number: 9001
+      - type: "patch"
+        name: "gcc16-x86_64_v2-nonshared-hidden.patch"
+        number: 9002

files/gcc16-riscv64-extfloat.patch

@@ -0,0 +1,29 @@
+From 8afa08292030c4da33ebf80ce39353678d9d9955 Mon Sep 17 00:00:00 2001
+From: Andrew Lukoshko <alukoshko@almalinux.org>
+Date: Wed, 18 Feb 2026 14:37:39 +0000
+Subject: [PATCH] libstdc++-v3: add __riscv constants to extfloat.S
+
+---
+ libstdc++-v3/src/nonshared98/extfloat.S | 5 ++++-
+ 1 file changed, 4 insertions(+), 1 deletion(-)
+
+diff --git a/libstdc++-v3/src/nonshared98/extfloat.S b/libstdc++-v3/src/nonshared98/extfloat.S
+index c72980884..9fad014c4 100644
+--- a/libstdc++-v3/src/nonshared98/extfloat.S
++++ b/libstdc++-v3/src/nonshared98/extfloat.S
+@@ -53,8 +53,11 @@
+ #elif defined __s390__
+ #define ALIGN1	.align 4
+ #define ALIGN3	.align 2
++#elif defined __riscv
++#define ALIGN1	.align 3
++#define ALIGN3	.align 3
+ #endif
+-#if defined __x86_64__ || defined __powerpc64__ || defined __s390x__ || defined __ia64__ || defined __aarch64__
++#if defined __x86_64__ || defined __powerpc64__ || defined __s390x__ || defined __ia64__ || defined __aarch64__ || defined __riscv
+ #define SIZE1	32
+ #define SIZE2	16
+ #define OFF	16
+-- 
+2.43.7
+

files/gcc16-x86_64_v2-nonshared-hidden.patch

@@ -0,0 +1,19 @@
+--- a/libstdc++-v3/src/nonshared20/format-inst.cc
++++ b/libstdc++-v3/src/nonshared20/format-inst.cc
+@@ -30,7 +30,16 @@
+ asm (".hidden _ZNSt8__detail18__to_chars_10_implImEEvPcjT_");
+ #endif
+ #if defined(__x86_64__)
++/* The out-of-line copy of this instantiation only exists when the compiler
++   declines to inline it into __formatter_fp<wchar_t>::_M_localize.  On
++   x86-64 that happens from the x86-64-v3 ISA level upwards; with a lower
++   baseline (e.g. -march=x86-64-v2) the symbol is never emitted, and a bare
++   .hidden directive would then leave an undefined hidden symbol behind in
++   libstdc++_nonshared140.a, which cannot be satisfied from libstdc++.so.6
++   and breaks the link with "hidden symbol ... isn't defined".  */
++#if defined(__AVX2__)
+ asm (".hidden _ZNSt7__cxx1112basic_stringIwSt11char_traitsIwESaIwEE22__resize_and_overwriteIZNKSt8__format14__formatter_fpIwE11_M_localizeESt17basic_string_viewIwS2_EcRKSt6localeEUlPwmE_EEvmT_");
++#endif
+ asm (".hidden _ZNSt8__detail13__to_chars_16IjEESt15to_chars_resultPcS2_T_");
+ #endif
+ asm (".hidden _ZNSt7__cxx1112basic_stringIcSt11char_traitsIcESaIcEE10_M_disposeEv");