I tested the global-removal variant based on trunk commit 5453f858146 on x86_64-pc-linux-gnu.
The bootstrap completed successfully, including the stage2/stage3 comparison, and the new PR126355 testcase passed. However, the full regression test exposed ICEs in the following existing tests with -O3 -g: g++.dg/torture/covariant-1.C g++.dg/torture/pr45699.C g++.dg/torture/pr46287.C g++.dg/torture/pr48661.C g++.dg/torture/pr58585.C g++.dg/torture/pr59226.C g++.dg/torture/pr94202.C They all fail at the same assertion: old_die = lookup_decl_die (decl); gcc_assert (old_die != NULL); in dwarf2out_abstract_function. A representative backtrace is: dwarf2out_abstract_function tree_function_versioning save_inline_function_body inline_transform These cases involve C++ thunks expanded through the inline path. Without the DECL_IGNORED_P setting, final debug processing reaches declarations which do not have an early DIE, and lookup_decl_die returns NULL. Thus, although g++.dg/other/pr87574.C itself no longer reproduces its original ICE on current trunk, it is not a sufficient guard for the underlying condition. At least some forced GIMPLE thunks expanded through the inline path still rely on DECL_IGNORED_P. I therefore will not post the global-removal variant as v2. The caller-specific save/restore in cgraph_node::create_wrapper remains the narrower fix: it bypasses the protection only for an ICF wrapper which reuses the original source declaration and its early DIE, while leaving late-created thunks in the other callers unchanged. Unless there is a better way to distinguish forced thunks which already have a usable early DIE, I suggest keeping the ICF-local approach. Richard Biener <[email protected]> 于2026年7月24日周五 15:56写道： > On Thu, 23 Jul 2026, Longjun Luo wrote: > > > Thanks. I traced the setting back to PR middle-end/87574. > > > > r264943 removed the previously unconditional DECL_IGNORED_P setting from > > expand_thunk, which exposed the DWARF ICE reported as PR87574. r265028 > > (Eric Botcazou) then restored it conditionally for forced GIMPLE thunks. > > > > For the record, the PR87574 reproducer itself is not LTO-related: it is a > > plain "-O2 -g" ICE in add_data_member_location_attribute > > (dwarf2out.c:19226), and its testcase g++.dg/other/pr87574.C is a > > { dg-do compile }, so compiling it exercises the original failure mode > > directly. > > > > There are in fact three callers of > > expand_thunk (..., /*force_gimple_thunk=*/true): inline_call > > (ipa-inline-transform.cc), the ICF wrapper > > (cgraph_node::create_wrapper), and coverage instrumentation > > (tree-profile.cc). > > > > I built an x86_64-pc-linux-gnu compiler from trunk commit 5453f858146 > > with only the conditional DECL_IGNORED_P setting removed, and ran these > > targeted compile checks: > > > > - g++.dg/other/pr87574.C compiled with -O2 -g and no ICE; an additional > > IPA inline dump showed the thunk expanded through inline_call; > > - the new PR126355 test source compiled with the test options; the ICF > > dump showed that the wrapper was created, and the assembly showed an > > address-bearing DIE for pr_icf_wrapper_b; > > - g++.dg/tree-prof/devirt.C compiled to assembly with > > -O3 -g -fprofile-generate -fno-profile-values > > -fdump-ipa-cgraph-details -S -dA, and the generated thunks contained > > their __gcov counters, exercising the tree-profile.cc path. > > > > Thus, the original PR87574 failure no longer reproduces on this trunk > > checkout with the setting removed, and g++.dg/other/pr87574.C remains as > > a guard should it ever return. > > > > Given that, removing the setting globally looks worth testing as the > > simpler fix. It addresses the common setting in expand_thunk rather than > > undoing it only in the ICF caller. It also changes debug emission for the > > inline_call and coverage paths, however, so I will do a full bootstrap > and > > regtest, with particular attention to the LTO, debug, guality, dwarf2, > and > > tree-prof suites, before posting such a revision. > > > > Would you prefer that global removal, or the more conservative > > caller-specific patch as posted? I can go either way; the ICF-only > version > > remains a fallback if the removal turns up any fallout. > > I prefer the global removal, but I CCed Eric in case he remembers > details and can argue the actual issue still exists. > > Thanks, > Richard. > > > Richard Biener <[email protected]> 于2026年7月23日周四 16:38写道： > > > > > On Thu, 23 Jul 2026, Longjun Luo wrote: > > > > > > > When IPA ICF keeps a distinct symbol for an address-taken function by > > > > replacing its body with a wrapper, cgraph_node::create_wrapper > reuses the > > > > original function declaration and its early debug DIE. > > > > > > > > create_wrapper then calls expand_thunk with force_gimple_thunk set. > > > Forced > > > > GIMPLE thunks are normally created after early debug, so expand_thunk > > > marks > > > > their declarations ignored. For an ICF wrapper this also prevents > final > > > > debug emission from attaching the wrapper address to the existing > DIE. > > > > The wrapper consequently has code and an STT_FUNC symbol but no > > > > address-bearing DW_TAG_subprogram DIE. > > > > > > > > Preserve the declaration's original DECL_IGNORED_P value across > forced > > > > GIMPLE thunk expansion in create_wrapper. Declarations that were > already > > > > ignored remain ignored, while an original source declaration can > receive > > > > its final address information. > > > > > > > > Tested on x86_64-pc-linux-gnu with the new gcc.dg/debug/dwarf2 > test. The > > > > test fails before the change and passes afterwards. > > > > > > There's two calls to expand_thunk (..., true), one from ICF as you > > > figured and one from inline_call (). I wonder if it's better to > > > remove > > > > > > /* We need to force DECL_IGNORED_P when the thunk is created > > > after early debug was run. */ > > > if (force_gimple_thunk) > > > DECL_IGNORED_P (thunk_fndecl) = 1; > > > > > > or figure why it was put there and see if the reason still holds > > > (sounds LTO-ish). > > > > > > Richard. > > > > > > > PR debug/126355 > > > > > > > > gcc/ChangeLog: > > > > > > > > * cgraphunit.cc (cgraph_node::create_wrapper): Preserve > > > > DECL_IGNORED_P across forced GIMPLE thunk expansion. > > > > > > > > gcc/testsuite/ChangeLog: > > > > > > > > * gcc.dg/debug/dwarf2/pr126355.c: New test. > > > > > > > > Signed-off-by: Longjun Luo <[email protected]> > > > > --- > > > > gcc/cgraphunit.cc | 5 +++ > > > > gcc/testsuite/gcc.dg/debug/dwarf2/pr126355.c | 35 > ++++++++++++++++++++ > > > > 2 files changed, 40 insertions(+) > > > > create mode 100644 gcc/testsuite/gcc.dg/debug/dwarf2/pr126355.c > > > > > > > > diff --git a/gcc/cgraphunit.cc b/gcc/cgraphunit.cc > > > > index 67aaebf108b..d397c9ae4cc 100644 > > > > --- a/gcc/cgraphunit.cc > > > > +++ b/gcc/cgraphunit.cc > > > > @@ -2683,7 +2683,12 @@ cgraph_node::create_wrapper (cgraph_node > *target) > > > > arguments = TREE_CHAIN (arguments); > > > > } > > > > > > > > + /* Forced GIMPLE thunks are normally ignored because they are > created > > > > + after early debug. ICF wrappers retain the original function > decl > > > and > > > > + its early DIE, so preserve its original debug state. */ > > > > + bool ignored_p = DECL_IGNORED_P (decl); > > > > expand_thunk (this, false, true); > > > > + DECL_IGNORED_P (decl) = ignored_p; > > > > thunk_info::remove (this); > > > > > > > > /* Inline summary set-up. */ > > > > diff --git a/gcc/testsuite/gcc.dg/debug/dwarf2/pr126355.c > > > b/gcc/testsuite/gcc.dg/debug/dwarf2/pr126355.c > > > > new file mode 100644 > > > > index 00000000000..35e321008aa > > > > --- /dev/null > > > > +++ b/gcc/testsuite/gcc.dg/debug/dwarf2/pr126355.c > > > > @@ -0,0 +1,35 @@ > > > > +/* PR debug/126355 */ > > > > +/* Verify that an IPA ICF wrapper for an address-taken function > keeps an > > > > + address-bearing subprogram DIE. */ > > > > +/* { dg-do compile } */ > > > > +/* { dg-options "-O2 -g -gdwarf -dA -fdump-ipa-icf-details" } */ > > > > + > > > > +int > > > > +pr_icf_wrapper_a (const char *host) > > > > +{ > > > > + (void) host; > > > > + return 0; > > > > +} > > > > + > > > > +int > > > > +pr_icf_wrapper_b (const char *host) > > > > +{ > > > > + (void) host; > > > > + return 0; > > > > +} > > > > + > > > > +int (*keep_a) (const char *) = pr_icf_wrapper_a; > > > > +int (*keep_b) (const char *) = pr_icf_wrapper_b; > > > > + > > > > +int > > > > +main (int argc, char **argv) > > > > +{ > > > > + const char *arg = argc > 1 ? argv[1] : "x"; > > > > + return keep_a (arg) + keep_b (arg); > > > > +} > > > > + > > > > +/* Check the ICF direction explicitly because the DWARF scan below > > > inspects > > > > + pr_icf_wrapper_b, the wrapper. */ > > > > +/* { dg-final { scan-ipa-dump "Semantic equality > > > hit:pr_icf_wrapper_a/\[0-9+\]+->pr_icf_wrapper_b/\[0-9+\]+" "icf" } } > */ > > > > +/* { dg-final { scan-ipa-dump "Wrapper has been created" "icf" } } > */ > > > > +/* { dg-final { scan-assembler "\\(DIE \\(0x\[0-9a-f\]+\\) > > > > DW_TAG_subprogram\\)\[\r\n\]+(\[^\r\n\]*\[\r\n\]+){1,12}\[^\r\n\]*DW_AT_name: > > > > \"pr_icf_wrapper_b\"\[\r\n\]+(\[^\r\n\]*\[\r\n\]+){1,12}\[^\r\n\]*DW_AT_low_pc" > > > } } */ > > > > > > > > > > -- > > > Richard Biener <[email protected]> > > > SUSE Software Solutions Germany GmbH, > > > Frankenstrasse 146, 90461 Nuernberg, Germany; > > > GF: Jochen Jaser, Andrew McDonald, Abhinav Puri; (HRB 36809, AG > Nuernberg) > > > > > > > -- > Richard Biener <[email protected]> > SUSE Software Solutions Germany GmbH, > Frankenstrasse 146, 90461 Nuernberg, Germany; > GF: Jochen Jaser, Andrew McDonald, Abhinav Puri; (HRB 36809, AG Nuernberg)
