Claude/upgrade ape c library mm z gd

CLAUDE.md

@@ -1559,6 +1559,12 @@ sweep rather than argued, and each named its own error.
   blast radius is every package in the tree, and no instrument here
   measures that* -- the cheap guard is to grep `external/` for the
   bare word before adding any name to `sys/include/ape`.
+  **CONFIRMED ON THE VM: the revert builds.** `mk distclean` plus
+  `mk install` completes, so the tree gets past `lex` and the whole
+  order after it was reached for the first time since the header
+  changed. *That is the weakest useful result and is worth saying so*
+  -- it confirms the collision is gone, and the `net_*` removal
+  beside it costs no link, which is all a completed build can say.
 - **`getopt`: TWO INSTALLED HEADERS DECLARED IT INCOMPATIBLY, and my
   note said there was only one.** `bsd.h:47` had
   `(int, char**, char*)` while **`getopt.h:8` has had POSIX's
@@ -1754,13 +1760,102 @@ live either way.
 ## What is open, as a list
 
 **Ports not started**, in the order their risk was last measured:
-- **EDG** -- investigated and measured on the host (it builds, the
-  default build IS the C-generating one, `eccp -S` translates a toy
-  with templates and virtual bases into C that gcc compiles). Two
-  named risks remain and both are one host afternoon: `targ_def.h`
-  describing kencc (`int` 4, `long` **4**, pointer 8 -- LLP64, which
-  EDG supports for MSVC), and whether its C back end handles the C++
-  subset EDG itself is written in. No mkfile written.
+- **EDG** -- **BOTH NAMED RISKS ARE CLOSED AND THE SELF-TRANSLATION IS
+  IN THE TREE.** `sys/src/external/edg` holds **133 generated C files,
+  ~2.90M lines**: all 75 `src/`, 7 `util/` and 51 `lib_src/`, produced
+  by a `cpfe` built from the same commit with a kencc target, plus
+  EDG's own 17 C++ headers verbatim. `gcc -fsyntax-only` over all 133
+  is **0 errors, 0 warnings** -- which is the host's answer and says
+  nothing about kencc. Full detail in `sys/src/external/edg/NOTE`;
+  only the consequences are here.
+  - **Risk 1, the target description: answered by EDG's own `win64`.**
+    kencc on amd64 is LLP64 and win64 is EDG's LLP64 configuration, so
+    the values were taken from the build's generated
+    `cmake_defines.h` rather than invented. Every scalar width
+    matches, **including `long double` 8** -- win64 says 8 because
+    MSVC has no extended precision and kencc says 8 because
+    `cc/sub.c`'s `simplet()` maps `BDOUBLE|BLONG` to `types[TDOUBLE]`,
+    which was not predicted. `size_t`/`ssize_t`/`ptrdiff_t` are
+    win64's kinds too, read out of `stddef_arch.h`; **`wchar_t` is the
+    one place the two differ** (`unsigned int` here, `unsigned short`
+    there). `kencc_targ.h`, committed beside the NOTE, is the whole
+    configuration -- a **pre-included header**, so nothing in the EDG
+    tree is modified.
+  - **Risk 2, self-translation: it translates.** One file does not and
+    it is the right one -- `util/cfe_daemon_client.c`, the only file
+    in the tree needing a real C++ standard library, and a
+    unix-domain-socket client is out of scope here twice over.
+  - **The target was inferred from the compiler that BUILT cpfe, which
+    is the finding of the round.** `targ_def.h:3505` derives
+    `GCC_IS_GENERATED_CODE_TARGET` from `defined(__GNUC__)`, and that
+    one variable is the sole gate on both `__weak__` emission sites --
+    so cpfe built by g++ decided its OUTPUT was for gcc and emitted
+    **26,712 `__attribute__((__weak__))`**, which kencc has not. A
+    first round also baked in LP64 and glibc's headers. Setting it to
+    0 and staging APExp's own headers takes the applied-attribute
+    count **38,245 -> 0**; the 11 textual occurrences left are all
+    inside string literals, EDG's own emitter text.
+  - **THE LINK MODEL IS THE OPEN PROBLEM, and removing `__weak__`
+    exposed it rather than solving it.** Those attributes were COMDAT
+    emulation for vague linkage. Measured: two trivial TUs sharing one
+    header collide on **10 symbols**, and across `src/` there are
+    **3,094 COMDAT symbols named in more than one of the 75 files**.
+    `--one_instantiation_per_object` is deliberately NOT used -- it
+    covers only the template instantiations (3 of the 10) and makes
+    the output depend on an `edg_prelink` phase mk cannot naturally
+    express. Upstream's answer for a target without COMDAT is the
+    **Cfront-like ABI**, and that was attempted and **backed out**:
+    `IA64_ABI=0` makes cpfe fail to *compile* on three
+    `TARGET_CONFIGURATION`-suffixed macros the Cfront ABI needs, so it
+    is reachable by regenerating the macro configuration rather than
+    by setting switches, and it changes mangling, the ctor/dtor model
+    and vtable layout. Its own round.
+  - **NO MKFILE, AND NOTHING HERE HAS BEEN THROUGH pcc.** That is the
+    next step and the link model is what it will run into.
+  - **Its staging list doubled as a to-do list for our own headers.**
+    Eight APE headers could not be included from C++; **four were
+    fixed on their own evidence** later in the same session
+    (`signal()`'s prototype, the `SIG_*` casts, `features.h`'s
+    `hidden`, `bsd.h`'s `getopt`) and a re-translation need not repeat
+    them. **The other four are fixed now too, so the staging list is
+    closed**: `stdlib.h`'s unguarded `_Noreturn`, its parameter named
+    `template`, `unistd.h`'s parameter named `new`, and `signal.h`'s
+    `restrict` read as a duplicate parameter name -- every one a
+    parameter name or a keyword, so invisible to a C caller.
+    *EDG is a stricter compiler than gcc, and that is what it bought.*
+  - **AND THE OBVIOUS FIX FOR TWO OF THEM WAS THE ONE THAT HAD JUST
+    BROKEN THE BUILD.** `#define _Noreturn` or `#define restrict`
+    under `__cplusplus` would put an UNRESERVED name in a public APE
+    header that external packages define themselves -- gnulib ships a
+    whole `_Noreturn.h`, and a dozen `config.h` here define
+    `restrict` -- which is `reject` again, one round later. **The
+    bare-word sweep was run FIRST this time**, and it is what chose
+    the spellings: everything used is in the reserved
+    double-underscore space, which nothing outside an implementation
+    may define.
+    **Three of the four need no macro at all.** `restrict` becomes
+    `__restrict`, which costs nothing because *kencc lexes all three
+    spellings to the same `LRESTRICT`* (`cc/lex.c:1652,1678-1679`) --
+    and where a package's own `cdefs.h` defines `__restrict` (twelve
+    do) it expands to the qualifier or to nothing, correct either way.
+    The two parameters are renamed, which no caller can see.
+    `__ape_template` was not invented here: **EDG's generated C
+    already carries it**, so the staged edit had used this convention
+    and the tree now matches its own output.
+  - **The control is a before-and-after pair on IDENTICAL staging, and
+    it found a FIFTH item.** A C++ translation unit including
+    `<stdlib.h>`, `<unistd.h>` and `<signal.h>` gives **12 errors
+    before and 1 after** -- and the survivor is `stddef.h:50`
+    typedef'ing **`wchar_t`, which is a built-in type in C++**,
+    present in BOTH runs and named nowhere in the NOTE's list of
+    eight. *Recorded, not fixed.* The half that could have cost the
+    tree is the other one: the same unit compiled as **C is 0
+    errors**. `apehdr-sweep` 148 headers / 0 findings with the
+    together-case holding at its recorded 8; `apdecl-sweep` 0.
+    *The first staging recipe differed slightly between the two runs
+    and the after-count was re-taken with the before's exact
+    commands* -- two compiles differing in anything but the change
+    are not a control, which this file records from the `-J` round.
 - **muon** -- in `_OPTIONAL_APPS` commented out. Never built here.
 - **go** -- `go1.4` is in the tree, commented out of `_CORE_APPS`,
   and is the only thing that mentions `Ureg` outside libap.

sys/include/ape/signal.h

@@ -158,10 +158,25 @@ extern int sigaction(int, const struct sigaction *, struct sigaction *);
 extern int sigprocmask(int, sigset_t *, sigset_t *);
 extern int sigpending(sigset_t *);
 extern int sigsuspend(const sigset_t *);
-extern int sigwait(const sigset_t *restrict, int *restrict);
-extern int sigwaitinfo(const sigset_t *restrict, siginfo_t *restrict);
+/*
+ * `__restrict' RATHER THAN `restrict', AND IT COSTS NOTHING: kencc
+ * lexes all three spellings to the same LRESTRICT (`cc/lex.c:1652,
+ * 1678-1679'). `restrict' is a C keyword and NOT a C++ one, so in C++
+ * these read as PARAMETER NAMES -- and two per prototype is then a
+ * duplicate parameter name, which is why a C++ compiler could not
+ * read these three lines at all.
+ *
+ * `#define restrict' away for C++ would be the wrong fix twice over:
+ * it is an unreserved name, and gnulib's `config.h' already defines
+ * it in a dozen packages here. `__restrict' is reserved, so nothing
+ * outside an implementation may define it -- and where a package's
+ * own `cdefs.h' does (twelve do), it expands to the qualifier or to
+ * nothing, which is correct either way.
+ */
+extern int sigwait(const sigset_t *__restrict, int *__restrict);
+extern int sigwaitinfo(const sigset_t *__restrict, siginfo_t *__restrict);
 struct timespec;	/* avoid pulling in time.h */
-extern int sigtimedwait(const sigset_t *restrict, siginfo_t *restrict, const struct timespec *restrict);
+extern int sigtimedwait(const sigset_t *__restrict, siginfo_t *__restrict, const struct timespec *__restrict);
 extern int sigqueue(pid_t, int, const union sigval);
 
 /*

sys/include/ape/stdlib.h

@@ -12,6 +12,28 @@
 typedef struct { int quot, rem; } div_t;
 typedef struct { long quot, rem; } ldiv_t;
 
+/*
+ * `_Noreturn' IS A C11 KEYWORD AND C++ HAS NO SUCH SPELLING, so the
+ * three declarations using it below could not be read by a C++
+ * compiler at all. Found by EDG, which is stricter than gcc: its own
+ * sources had to be translated against a STAGED copy of this header,
+ * and the staging list is in `sys/src/external/edg/NOTE'.
+ *
+ * **The obvious fix is the one this tree has just been burned by.**
+ * `#define _Noreturn' under __cplusplus would be a public APE header
+ * defining a name ten external packages define themselves -- gnulib
+ * ships a whole `_Noreturn.h' -- which is the `reject' collision
+ * waiting for the next build. The spelling used instead is in the
+ * RESERVED double-underscore space, which nothing outside an
+ * implementation may define, and `__ape_template' below is the same
+ * convention; EDG's generated C already carries that one.
+ */
+#if defined(__cplusplus)
+#define __ape_Noreturn
+#else
+#define __ape_Noreturn _Noreturn
+#endif
+
 #ifdef __cplusplus
 extern "C" {
 #endif
@@ -44,9 +66,9 @@ extern void *calloc(size_t, size_t);
 extern void free(void *);
 extern void *malloc(size_t);
 extern void *realloc(void *, size_t);
-extern _Noreturn void abort(void);
+extern __ape_Noreturn void abort(void);
 extern int atexit(void (*func)(void));
-extern _Noreturn void exit(int);
+extern __ape_Noreturn void exit(int);
 extern char *getenv(const char *);
 extern int putenv(char *);
 extern int system(const char *);
@@ -66,8 +88,8 @@ extern size_t wcstombs(char *, const wchar_t *, size_t);
 #include <bsd.h>
 
 extern char *mktemp(char *);
-extern int mkstemp(char *template);
-extern int mkostemp(char *template, int);
+extern int mkstemp(char *__ape_template);
+extern int mkostemp(char *__ape_template, int);
 
 /* from musl */
 typedef struct { long long quot, rem; } lldiv_t;
@@ -84,7 +106,7 @@ extern long long llabs(long long);
 extern lldiv_t lldiv(long long, long long);
 
 extern int at_quick_exit(void (*)(void));
-extern _Noreturn void quick_exit(int);
+extern __ape_Noreturn void quick_exit(int);
 
 /* musl and other ports */
 

sys/include/ape/unistd.h

@@ -183,13 +183,20 @@ int openat(int dirfd, const char *path, int flags, ...); /* unistd/at_functions.
 int fstatat(int dirfd, const char *path, struct stat *buf, int flags);
 int unlinkat(int dirfd, const char *path, int flags);
 int mkdirat(int dirfd, const char *path, mode_t mode);
-int renameat(int olddirfd, const char *old, int newdirfd, const char *new);
+/*
+ * `__ape_new' because `new' is a C++ KEYWORD, so these two lines
+ * could not be read from C++ at all. A parameter name in a prototype
+ * is documentation -- no caller can see it -- so this changes nothing
+ * for C. Same convention as `<stdlib.h>'s `__ape_template'; the
+ * reason both are in the reserved space is recorded there.
+ */
+int renameat(int olddirfd, const char *old, int newdirfd, const char *__ape_new);
 ssize_t readlinkat(int dirfd, const char *path, char *buf, size_t n);
 int symlinkat(const char *target, int dirfd, const char *linkpath);
 int faccessat(int dirfd, const char *path, int mode, int flags);
 int fchownat(int dirfd, const char *path, uid_t uid, gid_t gid, int flags);
 int fchmodat(int dirfd, const char *path, mode_t mode, int flags);
-int linkat(int old_dfd, const char *old, int new_dfd, const char *new, int flags);
+int linkat(int old_dfd, const char *old, int new_dfd, const char *__ape_new, int flags);
 
 /* GNU extensions */
 extern int pipe2(int [2], int);

sys/src/external/edg/NOTE

@@ -26,7 +26,11 @@ Converted
            cmake --preset linux-gcc-release -B build/kencc \
                  -DCMAKE_CXX_FLAGS="-include <path>/kencc_targ.h"
            ninja cpfe
-         kencc_targ.h is the target configuration; see below.  It is a
+         kencc_targ.h IS COMMITTED BESIDE THIS FILE -- it is the one
+         input to this directory that is neither upstream's nor
+         generated, so a scratch copy of it would have made the whole
+         translation unreproducible once the container went.
+         It is the target configuration; see below.  It is a
          pre-included header rather than a patch: NOTHING IN THE EDG
          TREE IS MODIFIED.
   Driver: dev_tools/bin/eccp, with build/kencc/environment.sh sourced.
@@ -270,31 +274,74 @@ it makes the output DEPEND on a prelink step.
 
 APExp HEADER CHANGES THIS TRANSLATION DEPENDS ON
 ------------------------------------------------
-The headers were STAGED (copied) and edited in the staging copy only.
-NOTHING IN APExp's HEADER TREE WAS MODIFIED.  To reproduce this
-directory those edits have to be repeated, and they are all cases of
-an APE header not being usable from C++:
+At translation time the headers were STAGED (copied) and edited in the
+staging copy only, and NOTHING IN APExp's HEADER TREE WAS MODIFIED.
+Eight edits were needed, every one a case of an APE header not being
+usable from C++.
+
+**ALL EIGHT ARE FIXED IN THE TREE NOW, so a re-translation needs NONE
+of them** -- stage the headers as they stand.  The list is kept
+because it is the record of what a stricter compiler found, and
+because each was fixed on its own evidence rather than to suit EDG.
+
+The last four, each a parameter name or a keyword, so none of them
+changes what a C caller sees:
 
   sys/include/ape/stdlib.h:47,49,87  `_Noreturn` (a C11 keyword that
       does not exist in C++) used unguarded
   sys/include/ape/stdlib.h:69,70     parameter named `template`
   sys/include/ape/unistd.h:186,192   parameter named `new`
-  sys/include/ape/signal.h:132-135   `restrict` as a parameter name;
+  sys/include/ape/signal.h:161-164   `restrict` as a parameter name;
       two per prototype reads as a duplicate parameter name in C++
-  sys/include/ape/signal.h:57        `signal()`'s handler parameter
-      declared `void (*)()` -- unprototyped -- where POSIX says
-      `void (*)(int)`
-  sys/include/ape/signal.h:16-18     SIG_DFL/SIG_ERR/SIG_IGN cast to
-      `void (*)()`, same defect as the declaration beside them
-  sys/include/ape/features.h:55      `#define hidden` -- musl's
-      internal visibility macro, in a PUBLIC header
-  sys/include/ape/bsd.h:47           `getopt(int, char**, char*)`,
-      non-const, where POSIX says `(int, char *const *, const char *)`
+
+And the first four, fixed earlier in the same session for their own
+reasons.  Verified against the tree rather than assumed:
+
+  signal.h   `signal()` is prototyped now --
+             `extern void (*signal(int, void (*)(int)))(int)` at :86
+  signal.h   SIG_DFL/SIG_ERR/SIG_IGN are `(void (*)(int))` at :45-47;
+             they had to move with the declaration, since callers
+             compare the return value against them
+  features.h `#define hidden` is GONE from the public header -- it had
+             a measured victim in `sqlite3.h:10957`'s
+             `unsigned char hidden[48]`
+  bsd.h:53   `getopt(int, char * const [], const char *)`, POSIX's
+             spelling, agreeing with `<getopt.h>` at last
+
+*That is the translation's own staging list working as a to-do list
+for the HEADER TREE*: it was written as eight things EDG could not
+use, and every one turned out to be a defect worth fixing on its own
+evidence, found by a compiler stricter than gcc.
+
+The control is a before-and-after pair on IDENTICAL staging: a C++
+translation unit including <stdlib.h>, <unistd.h> and <signal.h>
+gives 12 errors before and 1 after.  The same unit as C is 0 errors,
+which is the half that could have cost the tree.
+
+THE SURVIVOR IS A NINTH ITEM THIS LIST NEVER NAMED: stddef.h:50
+typedefs `wchar_t', which is a BUILT-IN TYPE in C++.  It is present
+in both runs, so it is pre-existing rather than anything done here,
+and it is recorded and not fixed.  cpfe does not trip on it --
+`--no_standard_includes' plus EDG's own include_c99 is why -- so it
+blocks nothing in this directory.
+
+Two of the four could not be fixed the obvious way.  `#define
+_Noreturn' or `#define restrict' under __cplusplus would put an
+UNRESERVED name in a public APE header that external packages define
+themselves: gnulib ships a whole `_Noreturn.h' and a dozen `config.h'
+here define `restrict'.  That is exactly the collision that broke
+flex over `reject' one round earlier, so the spellings used are all
+in the reserved double-underscore space.  Three of the four need no
+macro at all -- `restrict' becomes `__restrict', which kencc lexes to
+the same LRESTRICT (cc/lex.c:1652,1678-1679), and the two parameters
+are simply renamed.  `__ape_template' is this directory's own
+convention arriving back in the tree: the generated C already carries
+it.
 
 One absolute include also had to be rewritten for a host compile:
 amd64/include/ape/stddef_arch.h's `#include "/sys/include/ape/..."`.
 Four others of the form `#include "/sys/include/..."` name native
 Plan 9 headers and are unreachable from EDG, so they were left.
 
-These are reported as findings in their own right; see the round-2
-report.  They are not EDG's problem and they are not fixed here.
+The four that remain are reported as findings in their own right and
+are not fixed here; they are not EDG's problem either.