This C code: long set_work_pending_p; _Bool set_work_pending() { _Bool __trans_tmp_1; long mask = 1, old = __atomic_fetch_or(&set_work_pending_p, mask, 0); __trans_tmp_1 = old & mask; return !__trans_tmp_1; } void __queue_work() { _Bool ret = set_work_pending(); if (ret) __queue_work(); } does this with recent gcc: cvise $ ~/gcc/results/bin/gcc -c -g -O3 bug1040.c bug1040.c: In function ‘__queue_work’: bug1040.c:8:6: error: missing definition 8 | void __queue_work() { | ^~~~~~~~~~~~ for SSA_NAME: _6 in statement: # DEBUG ret => (_Bool) _6 during GIMPLE pass: fab bug1040.c:8:6: internal compiler error: verify_ssa failed The bug first seems to appear sometime before g:66d6b1861ec57ba2, dated 20240611.
- _5 = __atomic_fetch_or_8 (&set_work_pending_p, 1, 0); - # DEBUG old => (long int) _5 + _6 = .ATOMIC_BIT_TEST_AND_SET (&set_work_pending_p, 0, 1, 0, __atomic_fetch_or_8); + # DEBUG old => NULL # DEBUG BEGIN_STMT - # DEBUG D#2 => _5 & 1 + # DEBUG D#2 => NULL ... - _10 = ~_5; - _8 = (_Bool) _10; - # DEBUG ret => _8 + _8 = _6 == 0; + # DEBUG ret => (_Bool) _10 confirmed. convert_atomic_bit_not does this, it checks for single_use and removes the def, failing to release the name (which would fix this up IIRC). Note the function removes stmts in "wrong" order (before uses of LHS are removed), so it requires larger surgery. And it leaks SSA names.
Mine.
The releases/gcc-12 branch has been updated by hongtao Liu <liuhongt@gcc.gnu.org>: https://gcc.gnu.org/g:e1427b39d28f382d21e7a0ea1714b3250e0a6e5d commit r12-10617-ge1427b39d28f382d21e7a0ea1714b3250e0a6e5d Author: liuhongt <hongtao.liu@intel.com> Date: Fri Jul 12 09:39:23 2024 +0800 Fix SSA_NAME leak due to def_stmt is removed before use_stmt. - _5 = __atomic_fetch_or_8 (&set_work_pending_p, 1, 0); - # DEBUG old => (long int) _5 + _6 = .ATOMIC_BIT_TEST_AND_SET (&set_work_pending_p, 0, 1, 0, __atomic_fetch_or_8); + # DEBUG old => NULL # DEBUG BEGIN_STMT - # DEBUG D#2 => _5 & 1 + # DEBUG D#2 => NULL ... - _10 = ~_5; - _8 = (_Bool) _10; - # DEBUG ret => _8 + _8 = _6 == 0; + # DEBUG ret => (_Bool) _10 confirmed. convert_atomic_bit_not does this, it checks for single_use and removes the def, failing to release the name (which would fix this up IIRC). Note the function removes stmts in "wrong" order (before uses of LHS are removed), so it requires larger surgery. And it leaks SSA names. gcc/ChangeLog: PR target/115872 * tree-ssa-ccp.cc (convert_atomic_bit_not): Remove use_stmt after use_nop_stmt is removed. (optimize_atomic_bit_test_and): Ditto. gcc/testsuite/ChangeLog: * gcc.target/i386/pr115872.c: New test. (cherry picked from commit a8209237dc46dc4db7d9d8e3807e6c93734c64b5)
The releases/gcc-13 branch has been updated by hongtao Liu <liuhongt@gcc.gnu.org>: https://gcc.gnu.org/g:9a1cdaa5e8441394d613f5f3401e7aab21efe8f0 commit r13-8913-g9a1cdaa5e8441394d613f5f3401e7aab21efe8f0 Author: liuhongt <hongtao.liu@intel.com> Date: Fri Jul 12 09:39:23 2024 +0800 Fix SSA_NAME leak due to def_stmt is removed before use_stmt. - _5 = __atomic_fetch_or_8 (&set_work_pending_p, 1, 0); - # DEBUG old => (long int) _5 + _6 = .ATOMIC_BIT_TEST_AND_SET (&set_work_pending_p, 0, 1, 0, __atomic_fetch_or_8); + # DEBUG old => NULL # DEBUG BEGIN_STMT - # DEBUG D#2 => _5 & 1 + # DEBUG D#2 => NULL ... - _10 = ~_5; - _8 = (_Bool) _10; - # DEBUG ret => _8 + _8 = _6 == 0; + # DEBUG ret => (_Bool) _10 confirmed. convert_atomic_bit_not does this, it checks for single_use and removes the def, failing to release the name (which would fix this up IIRC). Note the function removes stmts in "wrong" order (before uses of LHS are removed), so it requires larger surgery. And it leaks SSA names. gcc/ChangeLog: PR target/115872 * tree-ssa-ccp.cc (convert_atomic_bit_not): Remove use_stmt after use_nop_stmt is removed. (optimize_atomic_bit_test_and): Ditto. gcc/testsuite/ChangeLog: * gcc.target/i386/pr115872.c: New test. (cherry picked from commit a8209237dc46dc4db7d9d8e3807e6c93734c64b5)
The releases/gcc-14 branch has been updated by hongtao Liu <liuhongt@gcc.gnu.org>: https://gcc.gnu.org/g:13bfc385b0baebd22aeabb0d90915f2e9b18febe commit r14-10422-g13bfc385b0baebd22aeabb0d90915f2e9b18febe Author: liuhongt <hongtao.liu@intel.com> Date: Fri Jul 12 09:39:23 2024 +0800 Fix SSA_NAME leak due to def_stmt is removed before use_stmt. - _5 = __atomic_fetch_or_8 (&set_work_pending_p, 1, 0); - # DEBUG old => (long int) _5 + _6 = .ATOMIC_BIT_TEST_AND_SET (&set_work_pending_p, 0, 1, 0, __atomic_fetch_or_8); + # DEBUG old => NULL # DEBUG BEGIN_STMT - # DEBUG D#2 => _5 & 1 + # DEBUG D#2 => NULL ... - _10 = ~_5; - _8 = (_Bool) _10; - # DEBUG ret => _8 + _8 = _6 == 0; + # DEBUG ret => (_Bool) _10 confirmed. convert_atomic_bit_not does this, it checks for single_use and removes the def, failing to release the name (which would fix this up IIRC). Note the function removes stmts in "wrong" order (before uses of LHS are removed), so it requires larger surgery. And it leaks SSA names. gcc/ChangeLog: PR target/115872 * tree-ssa-ccp.cc (convert_atomic_bit_not): Remove use_stmt after use_nop_stmt is removed. (optimize_atomic_bit_test_and): Ditto. gcc/testsuite/ChangeLog: * gcc.target/i386/pr115872.c: New test. (cherry picked from commit a8209237dc46dc4db7d9d8e3807e6c93734c64b5)
Fixed in GCC12.5, GCC13.4, GCC14.2 and main trunk.
The master branch has been updated by hongtao Liu <liuhongt@gcc.gnu.org>: https://gcc.gnu.org/g:f27bf48e0204524ead795fe618cd8b1224f72fd4 commit r15-2038-gf27bf48e0204524ead795fe618cd8b1224f72fd4 Author: liuhongt <hongtao.liu@intel.com> Date: Fri Jul 12 09:39:23 2024 +0800 Fix SSA_NAME leak due to def_stmt is removed before use_stmt. - _5 = __atomic_fetch_or_8 (&set_work_pending_p, 1, 0); - # DEBUG old => (long int) _5 + _6 = .ATOMIC_BIT_TEST_AND_SET (&set_work_pending_p, 0, 1, 0, __atomic_fetch_or_8); + # DEBUG old => NULL # DEBUG BEGIN_STMT - # DEBUG D#2 => _5 & 1 + # DEBUG D#2 => NULL ... - _10 = ~_5; - _8 = (_Bool) _10; - # DEBUG ret => _8 + _8 = _6 == 0; + # DEBUG ret => (_Bool) _10 confirmed. convert_atomic_bit_not does this, it checks for single_use and removes the def, failing to release the name (which would fix this up IIRC). Note the function removes stmts in "wrong" order (before uses of LHS are removed), so it requires larger surgery. And it leaks SSA names. gcc/ChangeLog: PR target/115872 * tree-ssa-ccp.cc (convert_atomic_bit_not): Remove use_stmt after use_nop_stmt is removed. (optimize_atomic_bit_test_and): Ditto. gcc/testsuite/ChangeLog: * gcc.target/i386/pr115872.c: New test.