fix(pkg): compat.gmp 在 MSVC ABI 上给 gmp.h 的 _MSC_VER 分支补守卫

首轮 windows CI:516 个 TU 全部编译通过,链接死在十个重复符号 ——
__gmpz_get_ui / __gmpn_sub / … 各自在 gmp.h 的内联体和带 -D__GMP_FORCE_<fn>
的那个 TU 里各定义一次。

gmp.h 用一串厂商分支决定 inline 怎么写。GCC 分支给
`extern __inline__ __attribute__ ((__gnu_inline__))`,即 GNU89 语义:头里的函数体
只用于内联、永不作为独立符号发射,独立那份由 __GMP_FORCE_<fn> 的 TU 提供。而
`_MSC_VER` 那条是整串里唯一没有 `! defined (__GMP_EXTERN_INLINE)` 守卫的
(SunPro/SCO 都有),clang 打 MSVC ABI 时两个宏都定义,于是它把 GCC 的答案改写成
裸 `__inline` —— MS 语义下每个包含 gmp.h 的 TU 都发射一份外部定义。

上游没碰到,因为它没有 MSVC ABI 构建。补上邻居们已有的守卫即可,别处严格 no-op:
非 windows 不定义 _MSC_VER,真 MSVC 不定义 __GNUC__。生成器对补丁点做了"必须恰好
命中一次"的断言,上游改了这段会报错而不是静默漏打。

顺带修正一处注释:单个 gmpxx TU 的理由是"feature 表只能门控 sources、带不了
__GMP_WITHIN_GMPXX",不是 basename 撞名 —— 2026.8.3.4 起对象路径无条件嵌套,
本包自己就有三个 add.c 各自落在 obj/compat_gmp/gmp-6.3.0/{mpf,mpn/generic,mpz}/add.o。

本地(linux gcc@16.1.0 与 llvm@22.1.8)重跑 gmp / gmp-gmpxx 仍全绿 —— 该补丁在
非 windows 上不改变任何一个字节的行为。

Author: Sunrisepeak

.agents/docs/2026-08-10-gmp-full-platform-no-build-system.md

@@ -131,9 +131,12 @@ GMP 的 C++ 绑定(`mpz_class`/`mpq_class`/`mpf_class`,RAII + 运算符重载 +
 
 **必须由消费端工具链编译**:这些符号的名字里带 `std::ostream`/`std::string`,即 C++ 标准库的 ABI。
 在钩子里用宿主 g++ 编一份 `libgmpxx.a`,在 libstdc++ 腿上能链、在每一条 libc++ 腿上都是 undefined
-reference。所以做成 feature:11 个 `cxx/*.cc` 合并成**一个**唯一命名的 TU
-(`mcpp_gmpxx.cc`)—— 一是 mcpp 按 basename 命名对象,`limits.cc → limits.o` 是个等着撞的名字
-(mcpp#233/#240);二是 feature 表只能门控 sources、带不了 `__GMP_WITHIN_GMPXX`。
+reference。所以做成 feature:11 个 `cxx/*.cc` 合并成**一个** TU(`mcpp_gmpxx.cc`)。
+
+合并的理由是 **feature 表只能门控 sources、带不了 define** —— `__GMP_WITHIN_GMPXX` 必须写在文件里,
+所以无论如何都要一个 wrapper,一个比十一个简单。**不是**因为 basename 撞名:那是 2026.8.3.4 之前的
+情况,现版本对象路径无条件嵌套,本包自己就有三个 `add.c` 各自落在
+`obj/compat_gmp/gmp-6.3.0/{mpf,mpn/generic,mpz}/add.o`(实测)。
 
 合并踩到一个真坑并已修:`gmp.h` 用"`<stdio.h>`/`<stdarg.h>` 是否已被看见"来决定要不要声明
 `FILE*`/`va_list` 入口(`_GMP_H_HAVE_FILE` / `_GMP_H_HAVE_VA_LIST`),`gmp-impl.h` 又拿同一个答案
@@ -142,7 +145,35 @@ reference。所以做成 feature:11 个 `cxx/*.cc` 合并成**一个**唯一命
 `aggregate gmp_asprintf_t has incomplete type`,而在 `<iostream>` 恰好拖进 `<cstdarg>` 的标准库上
 一切正常。解法是 wrapper TU 顶部先包含那两个头。
 
-### 3.5 `-fPIC`
+### 3.5 `gmp.h` 的一处补丁:MSVC ABI 上的 `__GMP_EXTERN_INLINE`
+
+除 configure 的 8 处 substitute 外,生成的 `gmp.h` 还打了**一个**补丁,由首轮 windows CI 抓出来。
+
+`gmp.h` 用一串厂商分支决定 "inline" 怎么写。GCC 分支给的是
+`extern __inline__ __attribute__ ((__gnu_inline__))` —— GNU89 语义:**头里的函数体只用于内联,
+永远不作为独立符号发射**,而独立那份由带 `-D__GMP_FORCE_<fn>` 的那个 TU 提供
+(`mpz/get_ui.c`、`mpn/generic/sub.c` …)。整个库靠这个分工成立。
+
+问题在于 `_MSC_VER` 那条分支是整串里**唯一没有** `! defined (__GMP_EXTERN_INLINE)` 守卫的
+(SunPro / SCO 两条都有)。而 clang 打 MSVC ABI 时**同时**定义 `__GNUC__` 与 `_MSC_VER`,于是它
+默默把 GCC 的答案改写成裸 `__inline`。MS 的 inline 语义下,每个包含 `gmp.h` 的 TU 都会发射一份外部
+定义,链接直接死在十个重复符号上:
+
+```
+lld-link: error: duplicate symbol: __gmpz_get_ui
+>>> defined at gmp.h:1793   obj/.../mpz/pprime_p.o
+>>> defined at              obj/.../mpz/get_ui.o
+```
+
+上游从来没碰到,因为它压根没有 MSVC ABI 的构建。补丁就是给这条分支补上邻居们已经有的守卫
+(`#if defined (_MSC_VER) && ! defined (__GNUC__)`),在别处是严格 no-op:非 windows 目标不定义
+`_MSC_VER`,真 MSVC 不定义 `__GNUC__`。生成器里对补丁点做了"必须恰好命中一次"的断言,
+上游改动了这段会直接报错而不是静默漏打。
+
+值得记一笔的是**首轮 windows CI 的 516 个 TU 全部编译通过**,只挂在链接 —— 也就是说
+`config.h`/`gmp.h`/内联表在 LLP64 + MSVC ABI 上本来就是对的。
+
+### 3.6 `-fPIC`
 
 `linux`/`macosx` 加 `-fPIC`(`cflags` + `cxxflags`)。静态库经常被链进消费者的 `.so`;而且 clang++
 默认 PIE 时链接本包就直接

pkgs/c/compat.gmp.lua

@@ -931,7 +931,14 @@ typedef const __gmp_randstate_struct *gmp_randstate_srcptr;
 #endif
 
 /* Microsoft's C compiler accepts __inline */
-#ifdef _MSC_VER
+/* mcpp-index (compat.gmp): `&& ! defined (__GNUC__)`, which upstream's
+   neighbouring vendor branches already carry in spirit.  clang targeting the
+   MSVC ABI defines BOTH __GNUC__ and _MSC_VER; without the guard this line
+   replaces the GCC branch's gnu_inline spelling with MS inline semantics, the
+   header starts emitting an external definition of every __GMP_EXTERN_INLINE
+   function in every TU that includes it, and the link fails with ten
+   duplicate symbols against the -D__GMP_FORCE_<fn> copies. */
+#if defined (_MSC_VER) && ! defined (__GNUC__)
 #define __GMP_EXTERN_INLINE  __inline
 #endif
 
@@ -3449,11 +3456,14 @@ const struct bases mp_bases[257] =
             ["gmp-6.3.0/mcpp/mcpp_gmpxx.cc"] =
 [[
 /* compat.gmp `gmpxx` feature: GMP's C++ bindings.
-   ONE translation unit, not eleven source entries, for two reasons:
-   mcpp names objects by BASENAME in a flat obj/ dir, and `limits.cc`
-   -> `limits.o` collides with anything else in the build that has
-   the same basename (mcpp#233/#240); and the feature table gates
-   sources, not defines, so __GMP_WITHIN_GMPXX has to live in a file.
+   ONE translation unit rather than eleven source entries because the
+   feature table gates SOURCES, not defines: __GMP_WITHIN_GMPXX has to
+   live inside a file, so a wrapper is needed either way and one is
+   simpler than eleven. (Object-name collisions are NOT the reason --
+   `limits.cc` would have been one before mcpp 2026.8.3.4, but object
+   paths are nested unconditionally now: this package alone compiles
+   three different add.c into obj/compat_gmp/gmp-6.3.0/{mpf,mpn/generic,
+   mpz}/add.o.)
 
    Compiled by the CONSUMER's toolchain on purpose. libgmpxx's
    interface is std::ostream / std::istream / std::string, so its

tools/gmp/generate_descriptor.py

@@ -298,6 +298,51 @@ def prepare(workdir):
     "@CFLAGS@":                       "",
 }
 
+# --------------------------------------------------------------------------
+# One patch to gmp.h beyond configure's substitutions.
+#
+# gmp.h picks how to spell "inline" through a vendor cascade.  The GCC branch
+# gives `extern __inline__ __attribute__ ((__gnu_inline__))` -- GNU89 inline
+# semantics: the body in the header is for inlining only and is NEVER emitted
+# as a standalone symbol, because the out-of-line copy comes from the one TU
+# that compiles with -D__GMP_FORCE_<fn> (mpz/get_ui.c, mpn/generic/sub.c, ...).
+# The whole library depends on that division.
+#
+# The `_MSC_VER` branch further down is the only one in the cascade WITHOUT a
+# `! defined (__GMP_EXTERN_INLINE)` guard (the SunPro and SCO branches have
+# one), so on a compiler that is BOTH -- clang targeting the MSVC ABI, which
+# is what this index uses on Windows -- it silently overwrites the GCC answer
+# with plain `__inline`.  Under MS inline semantics that emits an external
+# definition in every TU that includes gmp.h, and the link dies on ten
+# duplicate symbols against the forced copies:
+#
+#   lld-link: error: duplicate symbol: __gmpz_get_ui
+#   >>> defined at gmp.h:1793  obj/.../mpz/pprime_p.o
+#   >>> defined at            obj/.../mpz/get_ui.o
+#
+# Upstream never hit this because it has no MSVC-ABI build at all.  Adding the
+# same guard the neighbouring branches already carry is the minimal fix, and
+# it is a no-op everywhere else: non-Windows targets never define _MSC_VER,
+# and real MSVC never defines __GNUC__.
+GMP_H_PATCHES = [(
+    """/* Microsoft's C compiler accepts __inline */
+#ifdef _MSC_VER
+#define __GMP_EXTERN_INLINE  __inline
+#endif
+""",
+    """/* Microsoft's C compiler accepts __inline */
+/* mcpp-index (compat.gmp): `&& ! defined (__GNUC__)`, which upstream's
+   neighbouring vendor branches already carry in spirit.  clang targeting the
+   MSVC ABI defines BOTH __GNUC__ and _MSC_VER; without the guard this line
+   replaces the GCC branch's gnu_inline spelling with MS inline semantics, the
+   header starts emitting an external definition of every __GMP_EXTERN_INLINE
+   function in every TU that includes it, and the link fails with ten
+   duplicate symbols against the -D__GMP_FORCE_<fn> copies. */
+#if defined (_MSC_VER) && ! defined (__GNUC__)
+#define __GMP_EXTERN_INLINE  __inline
+#endif
+""")]
+
 # --------------------------------------------------------------------------
 # The mulfunc sources: one file compiled once per OPERATION_* macro, each
 # pass defining a different mpn entry point.  Upstream spells this with a
@@ -345,9 +390,13 @@ def build_files():
     # 1. gmp.h, from the upstream template.
     gmp_h = read(os.path.join(SRC, "gmp-h.in"))
     for k, v in GMP_H_SUBS.items():
-        assert k in gmp_h, k
+        if k not in gmp_h:
+            raise SystemExit("gmp-h.in no longer contains %s" % k)
         gmp_h = gmp_h.replace(k, v)
-    assert "@" not in gmp_h.split("__GMP_CC")[0][-4000:] or True
+    for old, new in GMP_H_PATCHES:
+        if gmp_h.count(old) != 1:
+            raise SystemExit("gmp-h.in patch target not found exactly once:\n" + old)
+        gmp_h = gmp_h.replace(old, new)
     f[f"{WRAP}/gmp.h"] = gmp_h
 
     # 2. config.h.
@@ -437,11 +486,14 @@ def fwd(where, header, target):
     # 10. The gmpxx feature's single TU.
     gmpxx = [
         "/* compat.gmp `gmpxx` feature: GMP's C++ bindings.\n"
-        "   ONE translation unit, not eleven source entries, for two reasons:\n"
-        "   mcpp names objects by BASENAME in a flat obj/ dir, and `limits.cc`\n"
-        "   -> `limits.o` collides with anything else in the build that has\n"
-        "   the same basename (mcpp#233/#240); and the feature table gates\n"
-        "   sources, not defines, so __GMP_WITHIN_GMPXX has to live in a file.\n"
+        "   ONE translation unit rather than eleven source entries because the\n"
+        "   feature table gates SOURCES, not defines: __GMP_WITHIN_GMPXX has to\n"
+        "   live inside a file, so a wrapper is needed either way and one is\n"
+        "   simpler than eleven. (Object-name collisions are NOT the reason --\n"
+        "   `limits.cc` would have been one before mcpp 2026.8.3.4, but object\n"
+        "   paths are nested unconditionally now: this package alone compiles\n"
+        "   three different add.c into obj/compat_gmp/gmp-6.3.0/{mpf,mpn/generic,\n"
+        "   mpz}/add.o.)\n"
         "\n"
         "   Compiled by the CONSUMER's toolchain on purpose. libgmpxx's\n"
         "   interface is std::ostream / std::istream / std::string, so its\n"