commit d11c5c90c2371c2d260a161f8f7d61fc30824615
parent 809a8e5e87c71ee3f3213e7e6f1f31eb972b70da
Author: Ryan Sepassi <rsepassi@gmail.com>
Date: Fri, 5 Jun 2026 21:01:51 -0700
Teach ld -lc to use hosted libc
Diffstat:
6 files changed, 423 insertions(+), 101 deletions(-)
diff --git a/doc/plan/TODO.md b/doc/plan/TODO.md
@@ -193,65 +193,3 @@ program loads at an ASLR base, so captured code addresses don't match link-time
addresses without computing the load slide. `-no-pie` should clear `o->pie` (and
ideally `-static` without `-pie` should default to non-PIE). Found while making a
backtrace demo's addresses line up with `kit addr2line`.
-
-## `kit cc --sysroot=<musl>` one-shot: vestigial interp on a static image, and no way to force a dynamic libc link
-
-Two related problems with one-shot `kit cc` against a musl sysroot (a sysroot
-that ships both `libc.a` and `libc.so`, the normal case):
-
-1. **Vestigial `PT_INTERP` + empty `PT_DYNAMIC` on an effectively-static
- binary.** `kit cc -fPIE --sysroot=<musl> foo.c -o exe` links `libc.a`
- statically — `printf`/`__libc_start_main` come out `T` (defined in the image),
- the Dynamic Section is empty (no `DT_NEEDED`) — yet because the output is
- ET_DYN (PIE) kit ld stamps its default musl `PT_INTERP`
- (`/lib/ld-musl-<arch>.so.1`) and an empty `PT_DYNAMIC`. The result is a static
- image wearing a dynamic costume: it references a loader it doesn't need and
- carries a pointless dynamic segment. It happens to run where that loader path
- exists, but it's malformed and would fail to start on a system without that
- exact interp. kit ld should not emit a `PT_INTERP`/`PT_DYNAMIC` when the link
- produced no dynamic dependencies.
-
-2. **No way to get a genuine dynamic libc link through `kit cc`.** The hosted
- profile selector `hosted_resolve_linux` (`driver/lib/hosted.c:392-397`) routes
- to `hosted_resolve_linux_musl_static` whenever `libc.a` is present — the rule
- at line 394 (`has_libc_a && !(has_libc_so6 && has_glibc_nonshared)`) does not
- consult `req->static_link`, and it precedes the `musl_dynamic` branch (line
- 396), so `hosted_resolve_linux_musl_dynamic` is unreachable for any musl
- sysroot that ships `libc.a`. No `kit cc` flag (`-Bdynamic`, absence of
- `-static`, etc.) flips it. The only way to get a real dynamic exe today
- (`printf` as a `U` import + `DT_NEEDED libc.musl-<arch>.so.1`) is to drop to an
- explicit two-step `kit cc -c` + `kit ld -pie … Scrt1.o crti.o obj libc.so
- libkit_rt.a crtn.o` (what test/libc/musl/run.sh's dynamic lane does). The
- static rule should respect a default-dynamic link mode (or at least let
- `-Bdynamic`/non-`-static` reach the dynamic branch). Found verifying
- `__kit_backtrace` on dynamically-linked binaries (doc/plan/BACKTRACE.md).
-
-## `kit ld` can't link against a sysroot with just `-lc` (no `--sysroot`, no crt auto-add, `-l` is `.a`-only)
-
-The ergonomic expectation — point `kit ld` at a sysroot and pass `-lc`, without
-hand-listing crt objects and the raw `libc.so` path — does not work. Three gaps,
-all reproduced against `build/musl-sysroot`:
-
-1. **No `--sysroot`.** `kit ld --sysroot=<dir> …` → `ld: unknown flag:
- --sysroot=…`. GNU ld supports `--sysroot` (and the `=`-prefix path rewrite);
- kit ld only has `-L`.
-2. **No crt auto-provision.** `kit ld -L <sysroot>/lib -lc -pie -o exe obj
- libkit_rt.a` → `fatal: link: entry symbol '_start' not defined`. kit ld does
- not pull the sysroot's start files, so the caller must pass
- `crt1.o`/`Scrt1.o` + `crti.o` + `crtn.o` explicitly. (crt selection is
- traditionally the cc driver's job, but kit's own one-step `kit cc` can't
- produce a dynamic libc link either — see the item above — so there is no
- convenient route at all.)
-3. **`-l NAME` resolves only `lib<NAME>.a`.** Per `kit ld --help` and observed
- behavior, `-lc` finds `libc.a` (static) only; it never considers
- `libc.so`/`libc.so.N`. So even once crts are sorted, `-lc` yields a static
- libc — a dynamic link still requires handing `libc.so` to kit ld as a
- positional input. GNU ld searches `.so` then `.a` (honoring `-Bstatic`/
- `-Bdynamic`); kit ld should do the same so `-lc` can produce a dynamic
- dependency.
-
-Net: a dynamic libc link today requires the explicit
-`kit ld -pie … Scrt1.o crti.o obj libc.so libkit_rt.a crtn.o` form. Wiring
-`--sysroot`, `.so`-aware `-l` resolution, and (optionally) crt auto-add would let
-`kit ld --sysroot=<dir> -pie -lc -o exe obj` work as expected. Found verifying
-`__kit_backtrace` on dynamically-linked binaries (doc/plan/BACKTRACE.md).
diff --git a/driver/cmd/ld.c b/driver/cmd/ld.c
@@ -5,6 +5,7 @@
#include <string.h>
#include "driver.h"
+#include "hosted.h"
#include "lib_resolve.h"
#include "runtime.h"
@@ -20,8 +21,11 @@
* -e symbol entry symbol
* -T script.ld linker script (parsed, not raw)
* --support-dir DIR kit support root for compiler rt
+ * --sysroot DIR hosted C runtime/sysroot root
* -L dir library search path (-l targets)
- * -l name resolves to lib<name>.a via -L
+ * -l c enable hosted CRT/libc expansion
+ * -l name resolves via -L (.so preferred unless
+ * -Bstatic/-static)
* -static / -pie / -no-pie target.pic
* -shared emit a shared library / dylib
* -r / --relocatable emit a relocatable partial-link object
@@ -82,6 +86,7 @@ typedef struct LdOptions {
int no_default_libs; /* -nostdlib / --no-default-libs */
int pic_explicit; /* -static / -pie / -no-pie / -shared seen */
const char* support_dir; /* --support-dir */
+ const char* sysroot; /* --sysroot / KIT_SYSROOT */
uint16_t pe_subsystem; /* KitPeSubsystem */
/* PT_INTERP path. NULL means "let libkit pick the target default
* (e.g. /lib/ld-musl-aarch64.so.1)". Set by -dynamic-linker /
@@ -105,6 +110,9 @@ typedef struct LdOptions {
const char** lib_dirs; /* -L */
uint32_t nlib_dirs;
+ char** owned_paths; /* sysroot-expanded argv/search paths */
+ size_t* owned_path_sizes;
+ uint32_t nowned_paths;
/* Shared-library output state. */
int shared; /* -shared */
@@ -119,6 +127,9 @@ typedef struct LdOptions {
int gc_sections; /* --gc-sections / --no-gc-sections */
int strip_debug; /* -S / --strip-debug */
int allow_undefined; /* shared output undefined-symbol policy */
+ int static_link; /* -static: hosted libc should pick static profile */
+ int wants_hosted_libc; /* -lc: expand crt + libc through hosted resolver */
+ DriverHostedPlan hosted;
/* --build-id state */
uint8_t build_id_mode; /* KitBuildIdMode */
@@ -182,11 +193,14 @@ void driver_help_ld(void) {
" -no-pie Disable PIE\n"
" --support-dir DIR kit support root for compiler "
"runtime\n"
+ " --sysroot DIR hosted C runtime/sysroot root\n"
" (target is otherwise auto-detected from the first object input)\n"
"\n"
"LIBRARY RESOLUTION\n"
" -L DIR Add library search path\n"
- " -l NAME Resolve to lib<NAME>.a via -L\n"
+ " -l c Add hosted CRT objects and libc\n"
+ " -l NAME Resolve via -L (.so preferred unless "
+ "static)\n"
"\n"
"POSITIONAL ARCHIVE STATE (apply to following .a inputs)\n"
" --whole-archive Pull every member of following "
@@ -237,7 +251,7 @@ void driver_help_ld(void) {
/* ---------- argv-sized scratch arrays ---------- */
static int ld_alloc_arrays(LdOptions* o, int argc) {
- size_t bound = (size_t)argc + 16u;
+ size_t bound = (size_t)argc + 32u;
o->argv_bound = bound;
o->object_files =
driver_alloc_zeroed(o->env, bound * sizeof(*o->object_files));
@@ -245,10 +259,14 @@ static int ld_alloc_arrays(LdOptions* o, int argc) {
o->dsos = driver_alloc_zeroed(o->env, bound * sizeof(*o->dsos));
o->order = driver_alloc_zeroed(o->env, bound * sizeof(*o->order));
o->lib_dirs = driver_alloc_zeroed(o->env, bound * sizeof(*o->lib_dirs));
+ o->owned_paths = driver_alloc_zeroed(o->env, bound * sizeof(*o->owned_paths));
+ o->owned_path_sizes =
+ driver_alloc_zeroed(o->env, bound * sizeof(*o->owned_path_sizes));
o->rpaths = driver_alloc_zeroed(o->env, bound * sizeof(*o->rpaths));
o->rpath_links = driver_alloc_zeroed(o->env, bound * sizeof(*o->rpath_links));
if (!o->object_files || !o->archives || !o->dsos || !o->order ||
- !o->lib_dirs || !o->rpaths || !o->rpath_links) {
+ !o->lib_dirs || !o->owned_paths || !o->owned_path_sizes || !o->rpaths ||
+ !o->rpath_links) {
driver_errf(LD_TOOL, "out of memory");
return 1;
}
@@ -265,6 +283,26 @@ static void ld_push_order(LdOptions* o, uint8_t kind, uint32_t index) {
slot->index = index;
}
+static void ld_insert_order(LdOptions* o, uint32_t pos, uint8_t kind,
+ uint32_t index) {
+ uint32_t i;
+ if (pos > o->norder) pos = o->norder;
+ for (i = o->norder; i > pos; --i) o->order[i] = o->order[i - 1u];
+ o->order[pos].kind = kind;
+ o->order[pos].index = index;
+ o->norder++;
+}
+
+static void ld_push_object(LdOptions* o, const char* path) {
+ o->object_files[o->nobject_files++] = path;
+ ld_push_order(o, KIT_LINK_INPUT_OBJ_BYTES, o->nobject_files - 1u);
+}
+
+static void ld_insert_object(LdOptions* o, const char* path, uint32_t pos) {
+ o->object_files[o->nobject_files++] = path;
+ ld_insert_order(o, pos, KIT_LINK_INPUT_OBJ_BYTES, o->nobject_files - 1u);
+}
+
static void ld_push_archive(LdOptions* o, const char* path, int owned,
size_t owned_size) {
LdArchive* a = &o->archives[o->narchives++];
@@ -330,6 +368,81 @@ static int driver_is_so_filename(const char* path) {
return 0;
}
+static char* ld_join2(DriverEnv* env, const char* a, const char* b,
+ size_t* out_size) {
+ size_t alen = driver_strlen(a);
+ size_t blen = driver_strlen(b);
+ size_t slash = (alen > 0 && a[alen - 1] != '/') ? 1u : 0u;
+ size_t bytes = alen + slash + blen + 1u;
+ char* out = driver_alloc(env, bytes);
+ size_t off = 0;
+ if (!out) return NULL;
+ if (alen) {
+ driver_memcpy(out + off, a, alen);
+ off += alen;
+ }
+ if (slash) out[off++] = '/';
+ if (blen) {
+ driver_memcpy(out + off, b, blen);
+ off += blen;
+ }
+ out[off] = '\0';
+ if (out_size) *out_size = bytes;
+ return out;
+}
+
+static int ld_own_path(LdOptions* o, char* path, size_t size,
+ const char** out) {
+ if (!path) {
+ driver_errf(LD_TOOL, "out of memory");
+ return 1;
+ }
+ if (o->nowned_paths >= o->argv_bound) {
+ driver_free(o->env, path, size);
+ driver_errf(LD_TOOL, "too many sysroot-expanded paths");
+ return 1;
+ }
+ o->owned_paths[o->nowned_paths] = path;
+ o->owned_path_sizes[o->nowned_paths] = size;
+ o->nowned_paths++;
+ if (out) *out = path;
+ return 0;
+}
+
+static int ld_add_sysroot_libdir(LdOptions* o) {
+ char* path;
+ size_t size;
+ const char* owned;
+ if (!o->sysroot || !o->sysroot[0]) return 0;
+ path = ld_join2(o->env, o->sysroot, "lib", &size);
+ if (ld_own_path(o, path, size, &owned) != 0) return 1;
+ o->lib_dirs[o->nlib_dirs++] = owned;
+ return 0;
+}
+
+static int ld_sysroot_rewrite_path(LdOptions* o, const char* path,
+ const char** out) {
+ const char* tail;
+ char* joined;
+ size_t size;
+ if (!path || path[0] != '=' || !o->sysroot || !o->sysroot[0]) {
+ *out = path;
+ return 0;
+ }
+ tail = path + 1;
+ if (tail[0] == '/') tail++;
+ joined = ld_join2(o->env, o->sysroot, tail, &size);
+ return ld_own_path(o, joined, size, out);
+}
+
+static int ld_note_library_request(LdOptions* o, const char* name) {
+ if (driver_streq(name, "c") && !o->no_default_libs) {
+ o->wants_hosted_libc = 1;
+ return 1;
+ }
+ return 0;
+}
+
/* ---------- --build-id parsing ---------- */
static int hex_nibble(char c) {
@@ -576,14 +689,31 @@ static int ld_parse(int argc, char** argv, LdOptions* o) {
/* First pass: detect --ms-link-driver up front so the option can
* appear anywhere on the command line and still affect earlier
- * `/...` tokens. (Matches how `link.exe` treats option order as
- * non-positional.) */
+ * `/...` tokens. Also capture --sysroot before resolving any -l
+ * entries; GNU ld treats the sysroot as a global search-prefix
+ * setting, not positional state. */
for (i = 1; i < argc; ++i) {
if (driver_streq(argv[i], "--ms-link-driver")) {
o->ms_link_driver = 1;
- break;
+ continue;
}
+ if (driver_streq(argv[i], "--sysroot")) {
+ if (i + 1 < argc) {
+ o->sysroot = argv[i + 1];
+ i++;
+ }
+ continue;
+ }
+ if (driver_strneq(argv[i], "--sysroot=", 10)) {
+ o->sysroot = argv[i] + 10;
+ continue;
+ }
+ }
+ if (!o->sysroot || !o->sysroot[0]) {
+ const char* env_sysroot = driver_getenv("KIT_SYSROOT");
+ if (env_sysroot && env_sysroot[0]) o->sysroot = env_sysroot;
}
+ if (ld_add_sysroot_libdir(o) != 0) return 1;
for (i = 1; i < argc; ++i) {
const char* a = argv[i];
@@ -732,26 +862,44 @@ static int ld_parse(int argc, char** argv, LdOptions* o) {
o->support_dir = a + 14;
continue;
}
+ if (driver_streq(a, "--sysroot")) {
+ if (++i >= argc) {
+ driver_errf(LD_TOOL, "--sysroot requires an argument");
+ return 1;
+ }
+ o->sysroot = argv[i];
+ continue;
+ }
+ if (driver_strneq(a, "--sysroot=", 10)) {
+ o->sysroot = a + 10;
+ continue;
+ }
if (driver_strneq(a, "-L", 2)) {
const char* dir = a[2] ? a + 2 : (++i < argc ? argv[i] : NULL);
+ const char* rewritten;
if (!dir) {
driver_errf(LD_TOOL, "-L requires an argument");
return 1;
}
- o->lib_dirs[o->nlib_dirs++] = dir;
+ if (ld_sysroot_rewrite_path(o, dir, &rewritten) != 0) return 1;
+ o->lib_dirs[o->nlib_dirs++] = rewritten;
continue;
}
if ((val = arg_eq_value(a, "--library-path")) != NULL) {
- o->lib_dirs[o->nlib_dirs++] = val;
+ const char* rewritten;
+ if (ld_sysroot_rewrite_path(o, val, &rewritten) != 0) return 1;
+ o->lib_dirs[o->nlib_dirs++] = rewritten;
continue;
}
if (driver_streq(a, "--library-path")) {
+ const char* rewritten;
if (++i >= argc) {
driver_errf(LD_TOOL, "--library-path requires an argument");
return 1;
}
- o->lib_dirs[o->nlib_dirs++] = argv[i];
+ if (ld_sysroot_rewrite_path(o, argv[i], &rewritten) != 0) return 1;
+ o->lib_dirs[o->nlib_dirs++] = rewritten;
continue;
}
@@ -766,6 +914,7 @@ static int ld_parse(int argc, char** argv, LdOptions* o) {
driver_errf(LD_TOOL, "-l requires an argument");
return 1;
}
+ if (ld_note_library_request(o, name)) continue;
/* -Bstatic forces .a only; everything else (default,
* -Bdynamic, --as-needed) prefers .so but falls back to .a. */
mode = (o->cur_link_mode == KIT_LM_STATIC) ? LIB_RESOLVE_STATIC_ONLY
@@ -796,6 +945,7 @@ static int ld_parse(int argc, char** argv, LdOptions* o) {
LibResolveOS resolve_os = (o->target.os == KIT_OS_WINDOWS)
? LIB_RESOLVE_OS_WINDOWS
: LIB_RESOLVE_OS_POSIX;
+ if (ld_note_library_request(o, val)) continue;
if (driver_lib_resolve_for_os(o->env, val, mode, resolve_os, o->lib_dirs,
o->nlib_dirs, &resolved, &resolved_size,
&kind) != 0) {
@@ -820,6 +970,7 @@ static int ld_parse(int argc, char** argv, LdOptions* o) {
driver_errf(LD_TOOL, "--library requires an argument");
return 1;
}
+ if (ld_note_library_request(o, argv[i])) continue;
mode = (o->cur_link_mode == KIT_LM_STATIC) ? LIB_RESOLVE_STATIC_ONLY
: LIB_RESOLVE_DYNAMIC_PREFER;
resolve_os = (o->target.os == KIT_OS_WINDOWS) ? LIB_RESOLVE_OS_WINDOWS
@@ -841,6 +992,8 @@ static int ld_parse(int argc, char** argv, LdOptions* o) {
if (driver_streq(a, "-static")) {
o->target.pic = KIT_PIC_NONE;
+ o->static_link = 1;
+ o->cur_link_mode = KIT_LM_STATIC;
o->pic_explicit = 1;
continue;
}
@@ -1025,13 +1178,16 @@ static int ld_parse(int argc, char** argv, LdOptions* o) {
return 1;
}
- if (driver_has_suffix(a, ".a")) {
- ld_push_archive(o, a, 0, 0);
- } else if (driver_is_so_filename(a)) {
- ld_push_dso(o, a, 0, 0);
- } else {
- o->object_files[o->nobject_files++] = a;
- ld_push_order(o, KIT_LINK_INPUT_OBJ_BYTES, o->nobject_files - 1u);
+ {
+ const char* path;
+ if (ld_sysroot_rewrite_path(o, a, &path) != 0) return 1;
+ if (driver_has_suffix(path, ".a")) {
+ ld_push_archive(o, path, 0, 0);
+ } else if (driver_is_so_filename(path)) {
+ ld_push_dso(o, path, 0, 0);
+ } else {
+ ld_push_object(o, path);
+ }
}
}
@@ -1048,11 +1204,21 @@ static int ld_parse(int argc, char** argv, LdOptions* o) {
ld_usage();
return 1;
}
+ if (o->wants_hosted_libc && o->no_default_libs) {
+ driver_errf(LD_TOOL,
+ "-lc hosted expansion is disabled by -nostdlib/"
+ "--no-default-libs");
+ return 1;
+ }
if (o->relocatable) {
if (o->shared) {
driver_errf(LD_TOOL, "-r and -shared are incompatible");
return 1;
}
+ if (o->wants_hosted_libc) {
+ driver_errf(LD_TOOL, "-lc hosted expansion requires executable output");
+ return 1;
+ }
/* Only an explicit -pie conflicts; a default PIE target.pic is simply
* overridden by -r (driver_link_pie suppresses pie when relocatable). */
if (o->pie) {
@@ -1096,14 +1262,22 @@ static void ld_options_release(LdOptions* o) {
driver_free(o->env, (void*)d->path, d->owned_size);
}
}
+ for (i = 0; i < o->nowned_paths; ++i) {
+ if (o->owned_paths[i])
+ driver_free(o->env, o->owned_paths[i], o->owned_path_sizes[i]);
+ }
if (o->build_id_bytes) {
driver_free(o->env, o->build_id_bytes, o->build_id_alloc);
}
+ driver_hosted_plan_fini(o->env, &o->hosted);
driver_free(o->env, o->object_files, bound * sizeof(*o->object_files));
driver_free(o->env, o->archives, bound * sizeof(*o->archives));
driver_free(o->env, o->dsos, bound * sizeof(*o->dsos));
driver_free(o->env, o->order, bound * sizeof(*o->order));
driver_free(o->env, o->lib_dirs, bound * sizeof(*o->lib_dirs));
+ driver_free(o->env, o->owned_paths, bound * sizeof(*o->owned_paths));
+ driver_free(o->env, o->owned_path_sizes,
+ bound * sizeof(*o->owned_path_sizes));
driver_free(o->env, o->rpaths, bound * sizeof(*o->rpaths));
driver_free(o->env, o->rpath_links, bound * sizeof(*o->rpath_links));
}
@@ -1140,6 +1314,82 @@ static void release_all(LoadedFile* arr, uint32_t n) {
for (i = 0; i < n; ++i) release_file(&arr[i]);
}
+static int ld_append_hosted_input(LdOptions* o, const DriverHostedInput* in,
+ uint32_t insert_pos, int insert) {
+ switch ((DriverHostedInputKind)in->kind) {
+ case DRIVER_HOSTED_INPUT_OBJECT:
+ if (insert)
+ ld_insert_object(o, in->path, insert_pos);
+ else
+ ld_push_object(o, in->path);
+ return 0;
+ case DRIVER_HOSTED_INPUT_ARCHIVE: {
+ LdArchive* a = &o->archives[o->narchives++];
+ a->path = in->path;
+ a->owned = 0;
+ a->owned_size = 0;
+ a->whole_archive = 0;
+ a->link_mode = KIT_LM_DEFAULT;
+ a->group_id = 0;
+ if (insert)
+ ld_insert_order(o, insert_pos, KIT_LINK_INPUT_ARCHIVE,
+ o->narchives - 1u);
+ else
+ ld_push_order(o, KIT_LINK_INPUT_ARCHIVE, o->narchives - 1u);
+ return 0;
+ }
+ case DRIVER_HOSTED_INPUT_DSO: {
+ LdDso* d = &o->dsos[o->ndsos++];
+ d->path = in->path;
+ d->owned = 0;
+ d->owned_size = 0;
+ if (insert)
+ ld_insert_order(o, insert_pos, KIT_LINK_INPUT_DSO, o->ndsos - 1u);
+ else
+ ld_push_order(o, KIT_LINK_INPUT_DSO, o->ndsos - 1u);
+ return 0;
+ }
+ default:
+ driver_errf(LD_TOOL, "internal error: unknown hosted input kind");
+ return 1;
+ }
+}
+
+static int ld_apply_hosted_before_after(LdOptions* o) {
+ DriverHostedRequest req;
+ uint32_t i;
+ uint32_t insert_pos = 0;
+ if (!o->wants_hosted_libc || o->shared) return 0;
+ memset(&req, 0, sizeof req);
+ req.env = o->env;
+ req.tool = LD_TOOL;
+ req.target = o->target;
+ req.sysroot = o->sysroot;
+ req.static_link = o->static_link;
+ req.link_inputs = 1;
+ if (driver_hosted_resolve(&req, &o->hosted) != 0) return 1;
+ for (i = 0; i < o->hosted.nbefore; ++i) {
+ if (ld_append_hosted_input(o, &o->hosted.before[i], insert_pos, 1) != 0)
+ return 1;
+ insert_pos++;
+ }
+ for (i = 0; i < o->hosted.nafter; ++i) {
+ if (ld_append_hosted_input(o, &o->hosted.after[i], 0, 0) != 0) return 1;
+ }
+ if (!o->interp_path && o->hosted.interp_path)
+ o->interp_path = o->hosted.interp_path;
+ return 0;
+}
+
+static int ld_apply_hosted_final(LdOptions* o) {
+ uint32_t i;
+ if (!o->hosted.profile_name) return 0;
+ for (i = 0; i < o->hosted.nfinal; ++i) {
+ if (ld_append_hosted_input(o, &o->hosted.final[i], 0, 0) != 0) return 1;
+ }
+ return 0;
+}
+
/* ---------- link execution ---------- */
/* Run the link. Returns 0 on success, nonzero on error. The caller owns
@@ -1163,6 +1413,7 @@ static int ld_run_link(LdOptions* o) {
DriverRuntimeSupport runtime = {0};
DriverRuntimeArchive rt_archive = {0};
uint32_t i;
+ uint32_t initial_nobject_files;
int runtime_resolved = 0;
int rc = 1;
@@ -1171,17 +1422,20 @@ static int ld_run_link(LdOptions* o) {
return 1;
}
- /* Load object files first so the final target is known before deciding
- * which compiler-runtime archive to auto-link. */
- if (o->nobject_files) {
- obj_lf = driver_alloc_zeroed(o->env, o->nobject_files * sizeof(*obj_lf));
- obj_in = driver_alloc_zeroed(o->env, o->nobject_files * sizeof(*obj_in));
+ /* Load the caller's object files first so the final target is known before
+ * deciding which hosted CRT/libc profile and compiler-runtime archive to
+ * add. The arrays are sized to the argv bound because hosted expansion may
+ * append start files after target detection. */
+ initial_nobject_files = o->nobject_files;
+ if (o->argv_bound) {
+ obj_lf = driver_alloc_zeroed(o->env, o->argv_bound * sizeof(*obj_lf));
+ obj_in = driver_alloc_zeroed(o->env, o->argv_bound * sizeof(*obj_in));
if (!obj_lf || !obj_in) {
driver_errf(LD_TOOL, "out of memory");
goto out;
}
}
- for (i = 0; i < o->nobject_files; ++i) {
+ for (i = 0; i < initial_nobject_files; ++i) {
const char* path = o->object_files[i];
if (load_file(io, path, &obj_lf[i]) != 0) {
driver_errf(LD_TOOL, "failed to read: %.*s",
@@ -1202,13 +1456,13 @@ static int ld_run_link(LdOptions* o) {
* a freestanding (`*-none-elf`, EI_OSABI=STANDALONE) object means a freestanding
* link even if a foreign object (e.g. a clang-assembled startup stub, which
* stamps EI_OSABI=SysV and so decodes as Linux) appears first. */
- if (o->nobject_files > 0) {
+ if (initial_nobject_files > 0) {
KitTargetSpec detected;
if (kit_detect_target(obj_lf[0].data.data, obj_lf[0].data.size,
&detected) == KIT_OK) {
uint8_t pic = o->target.pic;
uint32_t oi;
- for (oi = 1; oi < o->nobject_files; ++oi) {
+ for (oi = 1; oi < initial_nobject_files; ++oi) {
KitTargetSpec t;
if (kit_detect_target(obj_lf[oi].data.data, obj_lf[oi].data.size, &t) ==
KIT_OK) {
@@ -1234,6 +1488,8 @@ static int ld_run_link(LdOptions* o) {
}
}
+ if (ld_apply_hosted_before_after(o) != 0) goto out;
+
/* Auto-link kit's compiler runtime for any target that has a variant —
* including the freestanding riscv32-none-elf / riscv64-none-elf targets,
* whose runtime (and, for rv32, the float-ABI it was detected with) is
@@ -1254,6 +1510,19 @@ static int ld_run_link(LdOptions* o) {
ld_push_runtime_archive(o, &rt_archive);
}
+ if (ld_apply_hosted_final(o) != 0) goto out;
+
+ for (i = initial_nobject_files; i < o->nobject_files; ++i) {
+ const char* path = o->object_files[i];
+ if (load_file(io, path, &obj_lf[i]) != 0) {
+ driver_errf(LD_TOOL, "failed to read: %.*s",
+ KIT_SLICE_ARG(kit_slice_cstr(path)));
+ goto out;
+ }
+ obj_in[i].data = obj_lf[i].data.data;
+ obj_in[i].len = obj_lf[i].data.size;
+ }
+
if (o->narchives) {
arch_lf = driver_alloc_zeroed(o->env, o->narchives * sizeof(*arch_lf));
arch_in = driver_alloc_zeroed(o->env, o->narchives * sizeof(*arch_in));
@@ -1461,8 +1730,8 @@ out:
release_all(dso_lf, o->ndsos);
if (arch_in) driver_free(o->env, arch_in, o->narchives * sizeof(*arch_in));
if (arch_lf) driver_free(o->env, arch_lf, o->narchives * sizeof(*arch_lf));
- if (obj_in) driver_free(o->env, obj_in, o->nobject_files * sizeof(*obj_in));
- if (obj_lf) driver_free(o->env, obj_lf, o->nobject_files * sizeof(*obj_lf));
+ if (obj_in) driver_free(o->env, obj_in, o->argv_bound * sizeof(*obj_in));
+ if (obj_lf) driver_free(o->env, obj_lf, o->argv_bound * sizeof(*obj_lf));
if (dso_in) driver_free(o->env, dso_in, o->ndsos * sizeof(*dso_in));
if (dso_lf) driver_free(o->env, dso_lf, o->ndsos * sizeof(*dso_lf));
return rc;
diff --git a/driver/lib/hosted.c b/driver/lib/hosted.c
@@ -391,10 +391,10 @@ static int hosted_resolve_linux(const DriverHostedRequest* req,
has_glibc_nonshared = hosted_libdir_has(req->env, dirs, "libc_nonshared.a");
if (!req->static_link && has_libc_so6 && has_glibc_nonshared)
return hosted_resolve_linux_glibc_dynamic(req, dirs, plan);
- if (has_libc_a && !(has_libc_so6 && has_glibc_nonshared))
- return hosted_resolve_linux_musl_static(req, dirs, plan);
if (!req->static_link && has_libc_so)
return hosted_resolve_linux_musl_dynamic(req, dirs, plan);
+ if (has_libc_a && !(has_libc_so6 && has_glibc_nonshared))
+ return hosted_resolve_linux_musl_static(req, dirs, plan);
driver_errf(req->tool,
"no supported Linux hosted libc found (searched %u library "
"dir(s))",
diff --git a/src/obj/elf/link.c b/src/obj/elf/link.c
@@ -922,7 +922,8 @@ void link_emit_elf(LinkImage* img, Writer* w) {
*
* 1 headers PT_LOAD + nsegments PT_LOAD + 1 PT_NOTE (build-id)
* + 1 PT_TLS when this image carries any TLS sections.
- * + 4 dyn phdrs (PT_PHDR / PT_INTERP / PT_DYNAMIC / PT_GNU_STACK) on PIE.
+ * + 4 dyn phdrs (PT_PHDR / PT_INTERP / PT_DYNAMIC / PT_GNU_STACK) when PIE
+ * also has real dynamic-link state.
*
* Scripted images skip the headers PT_LOAD and PT_NOTE: phdrs are
* just the per-segment PT_LOADs. */
@@ -937,7 +938,7 @@ void link_emit_elf(LinkImage* img, Writer* w) {
}
}
}
- u32 nphdr_extra_dyn = pie ? 4u : 0u;
+ u32 nphdr_extra_dyn = (pie && img->dyn) ? 4u : 0u;
u32 nphdr_headers = scripted ? 0u : 1u;
u32 nphdr_buildid = scripted ? 0u : 1u;
u32 nphdr_total = nphdr_headers + img->nsegments + nphdr_buildid +
@@ -1359,7 +1360,7 @@ void link_emit_elf(LinkImage* img, Writer* w) {
* PT_LOAD. Required by the runtime loader for ET_DYN to know
* where its own program headers live. Must appear before the
* first PT_LOAD on dynamic exes (musl checks). */
- if (pie) {
+ if (pie && img->dyn) {
phdrs[pi].p_type = PT_PHDR;
phdrs[pi].p_flags = PF_R;
phdrs[pi].p_offset = sizeof(Ehdr64);
@@ -1488,8 +1489,7 @@ void link_emit_elf(LinkImage* img, Writer* w) {
* PT_LOAD that's never made writable. */
} else if (pie) {
/* dyn was nominally requested but layout_dyn early-out — no
- * imports and no DSO inputs. The image still needs a PT_GNU_STACK
- * for kernels that demand it; INTERP/DYNAMIC are skipped. */
+ * imports and no DSO inputs. INTERP/DYNAMIC loader headers are skipped. */
(void)0;
}
(void)pi;
diff --git a/src/obj/elf/link_dyn.c b/src/obj/elf/link_dyn.c
@@ -525,6 +525,7 @@ static u32 count_dynamic_entries(const LinkDynState* dyn) {
void layout_dyn(Linker* l, LinkImage* img) {
Heap* h = img->heap;
LinkDynState* dyn;
+ LinkDynState dyn_probe;
ImportLists imports;
ByteBuf dynstr;
u64 page;
@@ -555,9 +556,22 @@ void layout_dyn(Linker* l, LinkImage* img) {
"link: layout_dyn: no ELF arch descriptor");
}
+ /* Step 1: enumerate imports + DT_NEEDED. A PIE with no imports and no
+ * DSO inputs is effectively static; keep ET_DYN output but do not stamp
+ * PT_INTERP/PT_DYNAMIC or an empty .dynamic section. */
+ memset(&dyn_probe, 0, sizeof dyn_probe);
+ collect_imports(l, img, h, &imports);
+ collect_needed(l, img, &dyn_probe);
+ if (l->emit_static_exe && imports.nfuncs == 0 && imports.ndatas == 0 &&
+ dyn_probe.nneeded == 0) {
+ img->pie = 1;
+ free_imports(h, &imports);
+ return;
+ }
+
dyn = (LinkDynState*)h->alloc(h, sizeof(*dyn), _Alignof(LinkDynState));
if (!dyn) compiler_panic(img->c, SRCLOC_NONE, "link: oom on dyn state");
- memset(dyn, 0, sizeof(*dyn));
+ *dyn = dyn_probe;
img->dyn = dyn;
img->pie = 1;
@@ -573,10 +587,6 @@ void layout_dyn(Linker* l, LinkImage* img) {
: pool_intern_slice(l->c->global,
slice_from_cstr(elf_arch->default_musl_interp));
- /* Step 1: enumerate imports + DT_NEEDED. */
- collect_imports(l, img, h, &imports);
- collect_needed(l, img, dyn);
-
/* Step 2: build .dynstr + .dynsym. .dynstr must also carry the
* DT_NEEDED soname strings the .dynamic body references; intern
* them after the import names so build_dynsym's de-dup also covers
diff --git a/test/driver/run.sh b/test/driver/run.sh
@@ -360,6 +360,111 @@ else
not_ok "ld-no-undefined" "$work/shared-undef-cc.err"
fi
+# ---- ld -lc expands hosted CRT/libc from --sysroot ----
+mkdir -p "$work/ld-hosted-sr/lib" "$work/ld-hosted-sr/include"
+cat > "$work/ld-hosted-main.c" <<'SRC'
+int main(void) { return 0; }
+SRC
+cat > "$work/ld-hosted-crt.c" <<'SRC'
+extern int main(void);
+void _start(void) { (void)main(); for (;;) {} }
+SRC
+cat > "$work/ld-hosted-crti.c" <<'SRC'
+void __kit_fake_crti(void) {}
+SRC
+cat > "$work/ld-hosted-crtn.c" <<'SRC'
+void __kit_fake_crtn(void) {}
+SRC
+cat > "$work/ld-hosted-libc.c" <<'SRC'
+int libc_marker(void) { return 7; }
+SRC
+if "$KIT" cc -target x86_64-linux -fPIE -c "$work/ld-hosted-main.c" \
+ -o "$work/ld-hosted-main.o" > "$work/ld-hosted-main.out" \
+ 2> "$work/ld-hosted-main.err" &&
+ "$KIT" cc -target x86_64-linux -fPIE -c "$work/ld-hosted-crt.c" \
+ -o "$work/ld-hosted-sr/lib/Scrt1.o" > "$work/ld-hosted-scrt.out" \
+ 2> "$work/ld-hosted-scrt.err" &&
+ "$KIT" cc -target x86_64-linux -fno-PIC -c "$work/ld-hosted-crt.c" \
+ -o "$work/ld-hosted-sr/lib/crt1.o" > "$work/ld-hosted-crt1.out" \
+ 2> "$work/ld-hosted-crt1.err" &&
+ "$KIT" cc -target x86_64-linux -c "$work/ld-hosted-crti.c" \
+ -o "$work/ld-hosted-sr/lib/crti.o" > "$work/ld-hosted-crti.out" \
+ 2> "$work/ld-hosted-crti.err" &&
+ "$KIT" cc -target x86_64-linux -c "$work/ld-hosted-crtn.c" \
+ -o "$work/ld-hosted-sr/lib/crtn.o" > "$work/ld-hosted-crtn.out" \
+ 2> "$work/ld-hosted-crtn.err" &&
+ "$KIT" cc -target x86_64-linux -fPIC -c "$work/ld-hosted-libc.c" \
+ -o "$work/ld-hosted-libc-pic.o" > "$work/ld-hosted-libc-pic.out" \
+ 2> "$work/ld-hosted-libc-pic.err" &&
+ "$KIT" ld -shared -nostdlib -e libc_marker "$work/ld-hosted-libc-pic.o" \
+ -o "$work/ld-hosted-sr/lib/libc.so" > "$work/ld-hosted-so.out" \
+ 2> "$work/ld-hosted-so.err" &&
+ "$KIT" cc -target x86_64-linux -fno-PIC -c "$work/ld-hosted-libc.c" \
+ -o "$work/ld-hosted-libc-static.o" > "$work/ld-hosted-libc-static.out" \
+ 2> "$work/ld-hosted-libc-static.err" &&
+ "$KIT" ar rc "$work/ld-hosted-sr/lib/libc.a" \
+ "$work/ld-hosted-libc-static.o" > "$work/ld-hosted-ar.out" \
+ 2> "$work/ld-hosted-ar.err"; then
+ if "$KIT" ld --support-dir "$work/ld-rt-support" \
+ --sysroot "$work/ld-hosted-sr" -pie -lc \
+ "$work/ld-hosted-main.o" -o "$work/ld-hosted" \
+ > "$work/ld-hosted.out" 2> "$work/ld-hosted.err" &&
+ "$KIT" objdump -p "$work/ld-hosted" > "$work/ld-hosted-p.out" \
+ 2> "$work/ld-hosted-p.err" &&
+ grep -q "interpreter /lib/ld-musl-x86_64.so.1" \
+ "$work/ld-hosted-p.out" &&
+ grep -q "NEEDED libc.so" "$work/ld-hosted-p.out"; then
+ ok "ld-hosted-lc-sysroot"
+ else
+ { sed 's/^/ld: /' "$work/ld-hosted.err" 2>/dev/null
+ sed 's/^/dump: /' "$work/ld-hosted-p.err" 2>/dev/null
+ sed 's/^/ | /' "$work/ld-hosted-p.out" 2>/dev/null; } \
+ > "$work/ld-hosted.diag"
+ not_ok "ld-hosted-lc-sysroot" "$work/ld-hosted.diag"
+ fi
+else
+ { sed 's/^/main: /' "$work/ld-hosted-main.err" 2>/dev/null
+ sed 's/^/scrt: /' "$work/ld-hosted-scrt.err" 2>/dev/null
+ sed 's/^/crt1: /' "$work/ld-hosted-crt1.err" 2>/dev/null
+ sed 's/^/crti: /' "$work/ld-hosted-crti.err" 2>/dev/null
+ sed 's/^/crtn: /' "$work/ld-hosted-crtn.err" 2>/dev/null
+ sed 's/^/so: /' "$work/ld-hosted-so.err" 2>/dev/null
+ sed 's/^/ar: /' "$work/ld-hosted-ar.err" 2>/dev/null; } \
+ > "$work/ld-hosted-setup.diag"
+ not_ok "ld-hosted-lc-sysroot" "$work/ld-hosted-setup.diag"
+fi
+
+# ---- ld PIE without dynamic deps has no INTERP/DYNAMIC program headers ----
+cat > "$work/ld-static-pie.c" <<'SRC'
+void _start(void) { for (;;) {} }
+SRC
+if "$KIT" cc -target x86_64-linux -fPIE -c "$work/ld-static-pie.c" \
+ -o "$work/ld-static-pie.o" > "$work/ld-static-pie-cc.out" \
+ 2> "$work/ld-static-pie-cc.err" &&
+ "$KIT" ld -pie -nostdlib "$work/ld-static-pie.o" \
+ -o "$work/ld-static-pie" > "$work/ld-static-pie.out" \
+ 2> "$work/ld-static-pie.err" &&
+ "$KIT" objdump -p "$work/ld-static-pie" \
+ > "$work/ld-static-pie-p.out" 2> "$work/ld-static-pie-p.err"; then
+ if ! grep -q "interpreter " "$work/ld-static-pie-p.out" &&
+ ! grep -q "^ DYNAMIC" "$work/ld-static-pie-p.out" &&
+ ! grep -q "NEEDED" "$work/ld-static-pie-p.out"; then
+ ok "ld-static-pie-no-dynamic-headers"
+ else
+ sed 's/^/ | /' "$work/ld-static-pie-p.out" \
+ > "$work/ld-static-pie.diag"
+ not_ok "ld-static-pie-no-dynamic-headers" \
+ "$work/ld-static-pie.diag"
+ fi
+else
+ { sed 's/^/cc: /' "$work/ld-static-pie-cc.err" 2>/dev/null
+ sed 's/^/ld: /' "$work/ld-static-pie.err" 2>/dev/null
+ sed 's/^/dump: /' "$work/ld-static-pie-p.err" 2>/dev/null; } \
+ > "$work/ld-static-pie-setup.diag"
+ not_ok "ld-static-pie-no-dynamic-headers" \
+ "$work/ld-static-pie-setup.diag"
+fi
+
# ---- objdump -x aggregate (sections + symbol table) ----
if "$KIT" objdump -x "$work/main.o" \
> "$work/objdump-x.out" 2> "$work/objdump-x.err" &&
