Re: lto: from insert_attributes to file_end?
Georg-Johann Lay via Gcc <[email protected]>
| Newsgroups | gmane.comp.gcc.devel |
|---|---|
| Message-ID | <[email protected]> |
Am 06.05.26 um 09:51 schrieb Richard Biener: > On Tue, May 5, 2026 at 8:16 PM Georg-Johann Lay <[email protected]> wrote: >> Am 05.05.26 um 19:05 schrieb Richard Biener: >>> On Tue, May 5, 2026 at 5:41 PM Georg-Johann Lay via Gcc <[email protected]> wrote: >>>> >>>> In the avr backend there is a static variable avr_no_call_main_p that >>>> is set in TARGET_INSERT_ATTRIBUTES and used in TARGET_ASM_FILE_END. >>>> >>>> This works as expected in non-LTO compilations, with LTO however, >>>> TARGET_INSERT_ATTRIBUTES runs in cc1[plus] but TARGET_ASM_FILE_END >>>> is run by lto1, hence the variable is not set as expected. >>>> >>>> What's the recommended way to handle such a situation? >>> >>> Without looking too closely I assume there's an actual attribute somewhere >>> (on the main function decl?). Instead of using a global variable you'd >>> check for the presence of the attribute. Now - the question is whether >>> the TARGET_ASM_FILE_END behavior is only required in the TU with >>> the main() definition? In that case this should work. >> >> What works is to attach a new attribute to main when the conditions are >> right, then check for that attribute in, say, >> TARGET_DECLARE_FUNCTION_NAME and set a variable accordingly, and then >> check for that var in TARGET_ASM_FILE_END. > > It's not technically required to be handled from TARGET_ASM_FILE_END, > correct? Exactly. Background: main is called by __call_main which is provided by some default lib*.a. It is referenced by crt*.o to pull it in. With -mno-call-main, main is effectively a part of the startup-code. The application emits .global __call_main __call_main = 0 so __call_main it is no more pulled in since it is no more needed. > It should work to emit the symbol from ASM_OUTPUT_FUNCTION_LABEL?> (not sure if there's sth more appropriate for 'before function' or > 'after function') You'd at least have access to the functions decl there. Using a new attribute and emit .global __call_main + __call_main=0 in ASM_OUTPUT_FUNCTION_LABEL (or TARGET_ASM_FILE END) generates code as expected, however I am getting an error from the linker because __call_main is defined in lib*.a and also in some app module. What works fine is adding %{mno-call-main: --defsym __call_main=0} to the link specs, and skip messing with __call_main in avr.cc altogether. See attached delta. When the user sets -mcall-main as well as -mno-call-main in the same app, then it's fine for me it that doesn't work. Johann >> What's a bit ugly is that it requires a new target attribute that's only >> for internal usage and should not to be set by the user. Maybe not >> document it is okay then. > > As Sam says, an attribute with a space cannot be used by users. Like > "avr no_call_main". > > But also see the PR iff the only way for the user to get the behavior > with -mno-call-main
pr125194-v2.diff
(text/x-patch, 2.9 KB)
diff --git a/gcc/config/avr/avr.cc b/gcc/config/avr/avr.cc
index b0e9d9000e..25cb518b0f 100644
--- a/gcc/config/avr/avr.cc
+++ b/gcc/config/avr/avr.cc
@@ -249,9 +249,6 @@ bool avr_need_clear_bss_p = false;
bool avr_need_copy_data_p = false;
bool avr_has_rodata_p = false;
-/* To track if we satisfy __call_main from AVR-LibC. */
-bool avr_no_call_main_p = false;
-
/* Counts how often pass avr-fuse-add has been executed. Is is kept in
sync with cfun->machine->n_avr_fuse_add_executed and serves as an
insn condition for shift insn splitters. */
@@ -11906,7 +11903,9 @@ avr_insert_attributes (tree node, tree *attributes)
*attributes = tree_cons (get_identifier ("section"),
arg, *attributes);
}
- avr_no_call_main_p = true;
+ if (!lookup_attribute ("used", *attributes))
+ *attributes = tree_cons (get_identifier ("used"),
+ NULL_TREE, *attributes);
}
} // -mno-call-main
#endif // AVR-LibC
@@ -12467,15 +12466,6 @@ avr_file_end (void)
if (avr_need_clear_bss_p)
fputs (".global __do_clear_bss\n", asm_out_file);
-
- /* Don't let __call_main call main() and exit().
- Defining this symbol will keep the code from being pulled
- in from lib<mcu>.a as requested by AVR-LibC's gcrt1.S.
- We invoke main() by other means: putting it in .init9. */
-
- if (avr_no_call_main_p)
- fputs (".global __call_main\n"
- "__call_main = 0\n", asm_out_file);
}
diff --git a/gcc/config/avr/avr.opt b/gcc/config/avr/avr.opt
index 8bc2f5f173..56264ee6d1 100644
--- a/gcc/config/avr/avr.opt
+++ b/gcc/config/avr/avr.opt
@@ -39,7 +39,7 @@ Target RejectNegative Joined Var(avropt_n_flash) UInteger Init(-1)
This option is used internally. Set the number of 64 KiB flash segments.
mcall-main
-Target Var(avropt_call_main) UInteger Init(1)
+Target Var(avropt_call_main) UInteger Init(1) Save
Call main and exit (default).
mskip-bug
diff --git a/gcc/config/avr/gen-avr-mmcu-specs.cc b/gcc/config/avr/gen-avr-mmcu-specs.cc
index 42623d7e87..470d40538e 100644
--- a/gcc/config/avr/gen-avr-mmcu-specs.cc
+++ b/gcc/config/avr/gen-avr-mmcu-specs.cc
@@ -355,6 +355,8 @@ print_mcu (const avr_mcu_t *mcu, const McuInfo &mi)
fprintf (f, "*link_relax:\n\t%s\n\n", LINK_RELAX_SPEC);
+ fprintf (f, "*link_no_call_main:\n\t%s\n\n", "%{mno-call-main: --defsym __call_main=0}");
+
fprintf (f, "*link_arch:\n\t%s", link_arch_spec);
if (mi.is_device
&& rodata_pm_offset)
diff --git a/gcc/config/avr/specs.h b/gcc/config/avr/specs.h
index 4f71f8a657..1fca4a1d6e 100644
--- a/gcc/config/avr/specs.h
+++ b/gcc/config/avr/specs.h
@@ -67,6 +67,7 @@ along with GCC; see the file COPYING3. If not see
"%(link_relax) " \
"%(link_pmem_wrap) " \
"%(link_rodata_in_ram) " \
+ "%(link_no_call_main) " \
"%{shared:%eshared is not supported} "
#undef LIB_SPEC