Merge pull request #22049 from alexrp/musl-gen-stubs

`gen_stubs`: No compiler-rt hacks, more compact output, and more arch support

lib/libc/musl/arch/i386/crt_arch.h

@@ -3,6 +3,7 @@ __asm__(
 ".weak _DYNAMIC \n"
 ".hidden _DYNAMIC \n"
 ".global " START "\n"
+".type " START ",%function \n"
 START ":\n"
 "	xor %ebp,%ebp \n"
 "	mov %esp,%eax \n"

lib/libc/musl/arch/x86_64/crt_arch.h

@@ -1,6 +1,7 @@
 __asm__(
 ".text \n"
 ".global " START " \n"
+".type " START ",%function \n"
 START ": \n"
 "	xor %rbp,%rbp \n"
 "	mov %rsp,%rdi \n"

lib/libc/musl/libc.S

@@ -7,6 +7,13 @@
 #define PTR_SIZE_BYTES 4
 #define PTR2_SIZE_BYTES 8
 #endif
+
+#ifdef TIME32
+#define WEAKTIME64 .globl
+#else
+#define WEAKTIME64 .weak
+#endif
+
 .bss
 .weak ___environ
 .type ___environ, %object;
@@ -168,18 +175,64 @@ _IO_putc:
 .weak _IO_putc_unlocked
 .type _IO_putc_unlocked, %function;
 _IO_putc_unlocked:
-#if !defined(ARCH_riscv64) && !defined(ARCH_mips) && !defined(ARCH_mips64) && !defined(ARCH_x86_64) && !defined(ARCH_powerpc) && !defined(ARCH_powerpc64) && !defined(ARCH_aarch64) && !defined(ARCH_riscv32) && !defined(ARCH_loongarch64)
+#ifdef ARCH_i386
 .globl ___tls_get_addr
 .type ___tls_get_addr, %function;
 ___tls_get_addr:
 #endif
-#ifdef PTR32
+#ifdef TIME32
 .globl __adjtime64
 .type __adjtime64, %function;
 __adjtime64:
 .globl __adjtimex_time64
 .type __adjtimex_time64, %function;
 __adjtimex_time64:
+#endif
+#ifdef ARCH_arm
+.globl __aeabi_atexit
+.type __aeabi_atexit, %function;
+__aeabi_atexit:
+.globl __aeabi_memclr
+.type __aeabi_memclr, %function;
+__aeabi_memclr:
+.globl __aeabi_memclr4
+.type __aeabi_memclr4, %function;
+__aeabi_memclr4:
+.globl __aeabi_memclr8
+.type __aeabi_memclr8, %function;
+__aeabi_memclr8:
+.globl __aeabi_memcpy
+.type __aeabi_memcpy, %function;
+__aeabi_memcpy:
+.globl __aeabi_memcpy4
+.type __aeabi_memcpy4, %function;
+__aeabi_memcpy4:
+.globl __aeabi_memcpy8
+.type __aeabi_memcpy8, %function;
+__aeabi_memcpy8:
+.globl __aeabi_memmove
+.type __aeabi_memmove, %function;
+__aeabi_memmove:
+.globl __aeabi_memmove4
+.type __aeabi_memmove4, %function;
+__aeabi_memmove4:
+.globl __aeabi_memmove8
+.type __aeabi_memmove8, %function;
+__aeabi_memmove8:
+.globl __aeabi_memset
+.type __aeabi_memset, %function;
+__aeabi_memset:
+.globl __aeabi_memset4
+.type __aeabi_memset4, %function;
+__aeabi_memset4:
+.globl __aeabi_memset8
+.type __aeabi_memset8, %function;
+__aeabi_memset8:
+.globl __aeabi_read_tp
+.type __aeabi_read_tp, %function;
+__aeabi_read_tp:
+#endif
+#ifdef TIME32
 .globl __aio_suspend_time64
 .type __aio_suspend_time64, %function;
 __aio_suspend_time64:
@@ -187,12 +240,12 @@ __aio_suspend_time64:
 .globl __assert_fail
 .type __assert_fail, %function;
 __assert_fail:
-#if !defined(ARCH_riscv64) && !defined(ARCH_x86) && !defined(ARCH_x86_64) && !defined(ARCH_powerpc) && !defined(ARCH_powerpc64) && !defined(ARCH_aarch64) && !defined(ARCH_riscv32) && !defined(ARCH_loongarch64)
+#ifdef FAMILY_mips
 .globl __cachectl
 .type __cachectl, %function;
 __cachectl:
 #endif
-#ifdef PTR32
+#ifdef TIME32
 .globl __clock_adjtime64
 .type __clock_adjtime64, %function;
 __clock_adjtime64:
@@ -236,7 +289,7 @@ __cxa_atexit:
 .globl __cxa_finalize
 .type __cxa_finalize, %function;
 __cxa_finalize:
-#ifdef PTR32
+#ifdef TIME32
 .globl __difftime64
 .type __difftime64, %function;
 __difftime64:
@@ -247,7 +300,12 @@ __dls2b:
 .globl __dls3
 .type __dls3, %function;
 __dls3:
-#ifdef PTR32
+#ifdef FAMILY_mips
+.globl __dlstart
+.type __dlstart, %function;
+__dlstart:
+#endif
+#ifdef TIME32
 .globl __dlsym_time64
 .type __dlsym_time64, %function;
 __dlsym_time64:
@@ -312,7 +370,7 @@ __fseterr:
 .globl __fsetlocking
 .type __fsetlocking, %function;
 __fsetlocking:
-#ifdef PTR32
+#ifdef TIME32
 .weak __fstat_time64
 .type __fstat_time64, %function;
 __fstat_time64:
@@ -347,7 +405,7 @@ __fxstatat:
 .weak __getdelim
 .type __getdelim, %function;
 __getdelim:
-#ifdef PTR32
+#ifdef TIME32
 .globl __getitimer_time64
 .type __getitimer_time64, %function;
 __getitimer_time64:
@@ -364,6 +422,11 @@ __gmtime64:
 .type __gmtime64_r, %function;
 __gmtime64_r:
 #endif
+#ifdef ARCH_arm
+.globl __gnu_Unwind_Find_exidx
+.type __gnu_Unwind_Find_exidx, %function;
+__gnu_Unwind_Find_exidx:
+#endif
 .globl __h_errno_location
 .type __h_errno_location, %function;
 __h_errno_location:
@@ -490,7 +553,7 @@ __libc_current_sigrtmin:
 .globl __libc_start_main
 .type __libc_start_main, %function;
 __libc_start_main:
-#ifdef PTR32
+#ifdef TIME32
 .globl __localtime64
 .type __localtime64, %function;
 __localtime64:
@@ -498,12 +561,12 @@ __localtime64:
 .type __localtime64_r, %function;
 __localtime64_r:
 #endif
-#if !defined(ARCH_mips) && !defined(ARCH_mips64) && !defined(ARCH_x86) && !defined(ARCH_x86_64) && !defined(ARCH_powerpc) && !defined(ARCH_powerpc64) && !defined(ARCH_aarch64) && !defined(ARCH_loongarch64)
+#ifdef FAMILY_riscv
 .globl __longjmp
 .type __longjmp, %function;
 __longjmp:
 #endif
-#ifdef PTR32
+#ifdef TIME32
 .globl __lstat_time64
 .type __lstat_time64, %function;
 __lstat_time64:
@@ -514,7 +577,7 @@ __lutimes_time64:
 .globl __lxstat
 .type __lxstat, %function;
 __lxstat:
-#ifdef PTR32
+#ifdef TIME32
 .globl __mktime64
 .type __mktime64, %function;
 __mktime64:
@@ -547,7 +610,7 @@ __overflow:
 .weak __posix_getopt
 .type __posix_getopt, %function;
 __posix_getopt:
-#ifdef PTR32
+#ifdef TIME32
 .globl __ppoll_time64
 .type __ppoll_time64, %function;
 __ppoll_time64:
@@ -576,17 +639,7 @@ __recvmmsg_time64:
 .globl __res_state
 .type __res_state, %function;
 __res_state:
-#if !defined(ARCH_mips) && !defined(ARCH_mips64) && !defined(ARCH_x86) && !defined(ARCH_x86_64) && !defined(ARCH_powerpc) && !defined(ARCH_powerpc64) && !defined(ARCH_aarch64) && !defined(ARCH_riscv32) && !defined(ARCH_loongarch64)
+#ifdef FAMILY_riscv
-.globl __restore
-.type __restore, %function;
-__restore:
-#endif
-#if !defined(ARCH_mips) && !defined(ARCH_mips64) && !defined(ARCH_x86) && !defined(ARCH_x86_64) && !defined(ARCH_powerpc) && !defined(ARCH_powerpc64) && !defined(ARCH_aarch64) && !defined(ARCH_riscv32) && !defined(ARCH_loongarch64)
-.globl __restore_rt
-.type __restore_rt, %function;
-__restore_rt:
-#endif
-#if !defined(ARCH_mips) && !defined(ARCH_mips64) && !defined(ARCH_x86) && !defined(ARCH_x86_64) && !defined(ARCH_powerpc) && !defined(ARCH_powerpc64) && !defined(ARCH_aarch64) && !defined(ARCH_loongarch64)
 .globl __riscv_flush_icache
 .type __riscv_flush_icache, %function;
 __riscv_flush_icache:
@@ -594,7 +647,7 @@ __riscv_flush_icache:
 .globl __sched_cpucount
 .type __sched_cpucount, %function;
 __sched_cpucount:
-#ifdef PTR32
+#ifdef TIME32
 .globl __sched_rr_get_interval_time64
 .type __sched_rr_get_interval_time64, %function;
 __sched_rr_get_interval_time64:
@@ -614,7 +667,7 @@ __setitimer_time64:
 .globl __setjmp
 .type __setjmp, %function;
 __setjmp:
-#ifdef PTR32
+#ifdef TIME32
 .globl __settimeofday_time64
 .type __settimeofday_time64, %function;
 __settimeofday_time64:
@@ -631,7 +684,7 @@ __signbitl:
 .globl __sigsetjmp
 .type __sigsetjmp, %function;
 __sigsetjmp:
-#ifdef PTR32
+#ifdef TIME32
 .globl __sigtimedwait_time64
 .type __sigtimedwait_time64, %function;
 __sigtimedwait_time64:
@@ -639,7 +692,7 @@ __sigtimedwait_time64:
 .globl __stack_chk_fail
 .type __stack_chk_fail, %function;
 __stack_chk_fail:
-#ifdef PTR32
+#ifdef TIME32
 .globl __stat_time64
 .type __stat_time64, %function;
 __stat_time64:
@@ -692,7 +745,7 @@ __strxfrm_l:
 .weak __sysv_signal
 .type __sysv_signal, %function;
 __sysv_signal:
-#ifdef PTR32
+#ifdef TIME32
 .globl __thrd_sleep_time64
 .type __thrd_sleep_time64, %function;
 __thrd_sleep_time64:
@@ -718,9 +771,16 @@ __timerfd_settime64:
 .type __timespec_get_time64, %function;
 __timespec_get_time64:
 #endif
+#if !defined(ARCH_s390x)
 .globl __tls_get_addr
 .type __tls_get_addr, %function;
 __tls_get_addr:
+#endif
+#ifdef ARCH_s390x
+.globl __tls_get_offset
+.type __tls_get_offset, %function;
+__tls_get_offset:
+#endif
 .globl __tolower_l
 .type __tolower_l, %function;
 __tolower_l:
@@ -743,7 +803,7 @@ __uflow:
 .globl __uselocale
 .type __uselocale, %function;
 __uselocale:
-#ifdef PTR32
+#ifdef TIME32
 .globl __utime64
 .type __utime64, %function;
 __utime64:
@@ -796,7 +856,7 @@ _dl_debug_state:
 .globl _dlstart
 .type _dlstart, %function;
 _dlstart:
-#if !defined(ARCH_riscv64) && !defined(ARCH_mips) && !defined(ARCH_x86) && !defined(ARCH_x86_64) && !defined(ARCH_powerpc) && !defined(ARCH_powerpc64) && !defined(ARCH_aarch64) && !defined(ARCH_riscv32) && !defined(ARCH_loongarch64)
+#ifdef FAMILY_mips
 .globl _dlstart_data
 .type _dlstart_data, %function;
 _dlstart_data:
@@ -807,7 +867,7 @@ _exit:
 .weak _fini
 .type _fini, %function;
 _fini:
-#if !defined(ARCH_riscv64) && !defined(ARCH_x86) && !defined(ARCH_x86_64) && !defined(ARCH_powerpc) && !defined(ARCH_powerpc64) && !defined(ARCH_aarch64) && !defined(ARCH_riscv32) && !defined(ARCH_loongarch64)
+#ifdef FAMILY_mips
 .globl _flush_cache
 .type _flush_cache, %function;
 _flush_cache:
@@ -908,7 +968,7 @@ aligned_alloc:
 .globl alphasort
 .type alphasort, %function;
 alphasort:
-#if !defined(ARCH_riscv64) && !defined(ARCH_mips) && !defined(ARCH_mips64) && !defined(ARCH_powerpc) && !defined(ARCH_powerpc64) && !defined(ARCH_aarch64) && !defined(ARCH_riscv32) && !defined(ARCH_loongarch64)
+#ifdef FAMILY_x86
 .globl arch_prctl
 .type arch_prctl, %function;
 arch_prctl:
@@ -1033,12 +1093,10 @@ cabsf:
 .globl cabsl
 .type cabsl, %function;
 cabsl:
-#if !defined(ARCH_riscv64) && !defined(ARCH_x86) && !defined(ARCH_x86_64) && !defined(ARCH_powerpc) && !defined(ARCH_powerpc64) && !defined(ARCH_aarch64) && !defined(ARCH_riscv32) && !defined(ARCH_loongarch64)
+#ifdef FAMILY_mips
 .weak cachectl
 .type cachectl, %function;
 cachectl:
-#endif
-#if !defined(ARCH_riscv64) && !defined(ARCH_x86) && !defined(ARCH_x86_64) && !defined(ARCH_powerpc) && !defined(ARCH_powerpc64) && !defined(ARCH_aarch64) && !defined(ARCH_riscv32) && !defined(ARCH_loongarch64)
 .weak cacheflush
 .type cacheflush, %function;
 cacheflush:
@@ -1232,10 +1290,10 @@ clock_getcpuclockid:
 .globl clock_getres
 .type clock_getres, %function;
 clock_getres:
-WEAK64 clock_gettime
+WEAKTIME64 clock_gettime
 .type clock_gettime, %function;
 clock_gettime:
-WEAK64 clock_nanosleep
+WEAKTIME64 clock_nanosleep
 .type clock_nanosleep, %function;
 clock_nanosleep:
 .globl clock_settime
@@ -2018,10 +2076,10 @@ fsetpos:
 .globl fsetxattr
 .type fsetxattr, %function;
 fsetxattr:
-WEAK64 fstat
+WEAKTIME64 fstat
 .type fstat, %function;
 fstat:
-WEAK64 fstatat
+WEAKTIME64 fstatat
 .type fstatat, %function;
 fstatat:
 .weak fstatfs
@@ -2063,7 +2121,7 @@ futimens:
 .globl futimes
 .type futimes, %function;
 futimes:
-WEAK64 futimesat
+WEAKTIME64 futimesat
 .type futimesat, %function;
 futimesat:
 .globl fwide
@@ -2408,7 +2466,7 @@ globfree:
 .globl gmtime
 .type gmtime, %function;
 gmtime:
-WEAK64 gmtime_r
+WEAKTIME64 gmtime_r
 .type gmtime_r, %function;
 gmtime_r:
 .globl grantpt
@@ -2549,12 +2607,12 @@ insque:
 .globl ioctl
 .type ioctl, %function;
 ioctl:
-#if !defined(ARCH_riscv64) && !defined(ARCH_mips64) && !defined(ARCH_aarch64) && !defined(ARCH_riscv32) && !defined(ARCH_loongarch64)
+#if !defined(ARCH_aarch64) && !defined(ARCH_arm) && !defined(ARCH_loongarch64) && !defined(ARCH_mips64) && !defined(ARCH_mipsn32) && !defined(ARCH_riscv32) && !defined(ARCH_riscv64) && !defined(ARCH_s390x)
 .globl ioperm
 .type ioperm, %function;
 ioperm:
 #endif
-#if !defined(ARCH_riscv64) && !defined(ARCH_mips64) && !defined(ARCH_aarch64) && !defined(ARCH_riscv32) && !defined(ARCH_loongarch64)
+#if !defined(ARCH_aarch64) && !defined(ARCH_arm) && !defined(ARCH_loongarch64) && !defined(ARCH_mips64) && !defined(ARCH_mipsn32) && !defined(ARCH_riscv32) && !defined(ARCH_riscv64) && !defined(ARCH_s390x)
 .globl iopl
 .type iopl, %function;
 iopl:
@@ -2853,7 +2911,7 @@ localeconv:
 .globl localtime
 .type localtime, %function;
 localtime:
-WEAK64 localtime_r
+WEAKTIME64 localtime_r
 .type localtime_r, %function;
 localtime_r:
 .globl lockf
@@ -3552,7 +3610,7 @@ pthread_cond_init:
 .globl pthread_cond_signal
 .type pthread_cond_signal, %function;
 pthread_cond_signal:
-WEAK64 pthread_cond_timedwait
+WEAKTIME64 pthread_cond_timedwait
 .type pthread_cond_timedwait, %function;
 pthread_cond_timedwait:
 .globl pthread_cond_wait
@@ -3642,7 +3700,7 @@ pthread_mutex_lock:
 .globl pthread_mutex_setprioceiling
 .type pthread_mutex_setprioceiling, %function;
 pthread_mutex_setprioceiling:
-WEAK64 pthread_mutex_timedlock
+WEAKTIME64 pthread_mutex_timedlock
 .type pthread_mutex_timedlock, %function;
 pthread_mutex_timedlock:
 .weak pthread_mutex_trylock
@@ -3693,10 +3751,10 @@ pthread_rwlock_init:
 .weak pthread_rwlock_rdlock
 .type pthread_rwlock_rdlock, %function;
 pthread_rwlock_rdlock:
-WEAK64 pthread_rwlock_timedrdlock
+WEAKTIME64 pthread_rwlock_timedrdlock
 .type pthread_rwlock_timedrdlock, %function;
 pthread_rwlock_timedrdlock:
-WEAK64 pthread_rwlock_timedwrlock
+WEAKTIME64 pthread_rwlock_timedwrlock
 .type pthread_rwlock_timedwrlock, %function;
 pthread_rwlock_timedwrlock:
 .weak pthread_rwlock_tryrdlock
@@ -3774,7 +3832,7 @@ pthread_spin_unlock:
 .weak pthread_testcancel
 .type pthread_testcancel, %function;
 pthread_testcancel:
-WEAK64 pthread_timedjoin_np
+WEAKTIME64 pthread_timedjoin_np
 .type pthread_timedjoin_np, %function;
 pthread_timedjoin_np:
 .weak pthread_tryjoin_np
@@ -3999,7 +4057,7 @@ rintf:
 .globl rintl
 .type rintl, %function;
 rintl:
-#if !defined(ARCH_mips) && !defined(ARCH_mips64) && !defined(ARCH_x86) && !defined(ARCH_x86_64) && !defined(ARCH_powerpc) && !defined(ARCH_powerpc64) && !defined(ARCH_aarch64) && !defined(ARCH_loongarch64)
+#ifdef FAMILY_riscv
 .weak riscv_flush_icache
 .type riscv_flush_icache, %function;
 riscv_flush_icache:

lib/libc/musl/src/signal/riscv32/restore.s

@@ -1,7 +1,9 @@
 .global __restore
+.hidden __restore
 .type __restore, %function
 __restore:
 .global __restore_rt
+.hidden __restore_rt
 .type __restore_rt, %function
 __restore_rt:
 	li a7, 139 # SYS_rt_sigreturn

lib/libc/musl/src/signal/riscv64/restore.s

@@ -1,7 +1,9 @@
 .global __restore
+.hidden __restore
 .type __restore, %function
 __restore:
 .global __restore_rt
+.hidden __restore_rt
 .type __restore_rt, %function
 __restore_rt:
 	li a7, 139 # SYS_rt_sigreturn

lib/libc/musl/src/thread/s390x/__tls_get_offset.s

@@ -5,6 +5,7 @@ __tls_get_offset:
 	aghi  %r15, -160
 
 	la    %r2, 0(%r2, %r12)
+.hidden __tls_get_addr
 	brasl %r14, __tls_get_addr
 
 	ear   %r1, %a0

src/Compilation.zig

@@ -5884,7 +5884,9 @@ pub const FileExt = enum {
 };
 
 pub fn hasObjectExt(filename: []const u8) bool {
-    return mem.endsWith(u8, filename, ".o") or mem.endsWith(u8, filename, ".obj");
+    return mem.endsWith(u8, filename, ".o") or
+        mem.endsWith(u8, filename, ".lo") or
+        mem.endsWith(u8, filename, ".obj");
 }
 
 pub fn hasStaticLibraryExt(filename: []const u8) bool {

src/musl.zig

@@ -138,17 +138,6 @@ pub fn buildCrtFile(comp: *Compilation, in_crt_file: CrtFile, prog_node: std.Pro
                 try addSrcFile(arena, &source_table, src_file);
             }
 
-            const time32_compat_arch_list = [_][]const u8{
-                "arm",
-                "i386",
-                "m68k",
-                "microblaze",
-                "mips",
-                "mipsn32",
-                "or1k",
-                "powerpc",
-                "sh",
-            };
             for (time32_compat_arch_list) |time32_compat_arch| {
                 if (mem.eql(u8, arch_name, time32_compat_arch)) {
                     for (compat_time32_files) |compat_time32_file| {
@@ -239,13 +228,29 @@ pub fn buildCrtFile(comp: *Compilation, in_crt_file: CrtFile, prog_node: std.Pro
             });
 
             const target = comp.root_mod.resolved_target.result;
-            const arch_define = try std.fmt.allocPrint(arena, "-DARCH_{s}", .{
+            const arch_name = std.zig.target.muslArchName(target.cpu.arch, target.abi);
-                @tagName(target.cpu.arch),
+            const time32 = for (time32_compat_arch_list) |time32_compat_arch| {
-            });
+                if (mem.eql(u8, arch_name, time32_compat_arch)) break true;
+            } else false;
+            const arch_define = try std.fmt.allocPrint(arena, "-DARCH_{s}", .{arch_name});
+            const family_define = switch (target.cpu.arch) {
+                .arm, .armeb, .thumb, .thumbeb => "-DFAMILY_arm",
+                .aarch64, .aarch64_be => "-DFAMILY_aarch64",
+                .loongarch64 => "-DFAMILY_loongarch",
+                .m68k => "-DFAMILY_m68k",
+                .mips, .mipsel, .mips64, .mips64el => "-DFAMILY_mips",
+                .powerpc, .powerpc64, .powerpc64le => "-DFAMILY_powerpc",
+                .riscv32, .riscv64 => "-DFAMILY_riscv",
+                .s390x => "-DFAMILY_s390x",
+                .x86, .x86_64 => "-DFAMILY_x86",
+                else => unreachable,
+            };
             const cc_argv: []const []const u8 = if (target.ptrBitWidth() == 64)
-                &.{ "-DPTR64", arch_define }
+                &.{ "-DPTR64", arch_define, family_define }
+            else if (time32)
+                &.{ "-DTIME32", arch_define, family_define }
             else
-                &.{arch_define};
+                &.{ arch_define, family_define };
 
             const root_mod = try Module.create(arena, .{
                 .global_cache_directory = comp.global_cache_directory,
@@ -347,6 +352,18 @@ pub fn needsCrt0(output_mode: std.builtin.OutputMode, link_mode: std.builtin.Lin
     };
 }
 
+const time32_compat_arch_list = [_][]const u8{
+    "arm",
+    "i386",
+    "m68k",
+    "microblaze",
+    "mips",
+    "mipsn32",
+    "or1k",
+    "powerpc",
+    "sh",
+};
+
 fn isArchName(name: []const u8) bool {
     const musl_arch_names = [_][]const u8{
         "aarch64",