Created attachment 24649 [details] Output of "ajo-gcc -std=c99 -O2 -g -w -c test.c -v" This failure reproduces for me with svn revision 175547 (2011-06-27). I've seen it twice in random testing, and never saw it back in May, so it must be a fairly new regression. I'm on Ubuntu 10.10, x86-64. cat >test.c <<EOF static void use(int *p_29) { } void func_25(int *p_29) { for (short p_26 = 0; p_26 == 1; ++p_26) p_29 = 0; use(p_29); } EOF gcc -std=c99 -O2 -g -w -c test.c test.c: In function ‘func_25’: test.c:6:1: error: definition in block 3 does not dominate use in block 5 for SSA_NAME: p_29_3 in statement: # DEBUG D#1 => p_29_3 test.c:6:1: internal compiler error: verify_ssa failed Bug 40711 (from 2009) has the same symptom, but doesn't require "-g", so it probably has a different root cause. This test case is reduced from the output of Csmith 2.1.0 (git hash 01aa8b04, https://github.com/Quuxplusone/csmith/), using the following command line: csmith --no-paranoid --longlong --pointers --no-arrays --jumps --no-consts --volatiles --checksum --no-divs --muls --no-bitfields --no-packed-struct -s 651222746
Huh, this is from into-SSA. Confirmed.
Caused by my http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=175288 so I'll have a look...
The problem is that for the debug uses before going into SSA we obviously don't want the uses to affect code generation and thus we don't call set_livein_block. Unfortunately that means get_current_def is sometimes incorrect during rewrite_ --- tree-into-ssa.c.jj 22011-06-23 10:13:58.000000000 +0200 +++ tree-into-ssa.c 2011-07-01 16:23:04.000000000 +0200 @@ -1343,7 +1343,15 @@ rewrite_debug_stmt_uses (gimple stmt) } } else - def = get_current_def (var); + { + def = get_current_def (var); + if (def + && !SSA_NAME_IS_DEFAULT_DEF (def) + && gimple_bb (SSA_NAME_DEF_STMT (def)) != gimple_bb (stmt) + && !dominated_by_p (CDI_DOMINATORS, gimple_bb (stmt), + gimple_bb (SSA_NAME_DEF_STMT (def)))) + def = NULL; + } if (def == NULL) { gimple_debug_bind_reset_value (stmt); seems to fix the ICE, the question is if get_current_def can be trusted to be the right SSA_NAME even after this check. I guess if the definition bb is the same as stmt's bb, it can, similarly if get_phi_state (var) == NEED_PHI_STATE_NO (plus the dominated_by_p check above), or if bitmap_bit_p (get_def_blocks_for (var)->livein_blocks, gimple_bb (stmt)->index). Any other cases?
Author: jakub Date: Mon Jul 4 17:19:52 2011 New Revision: 175818 URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=175818 Log: PR debug/49602 * tree-into-ssa.c (rewrite_debug_stmt_uses): Disregard get_current_def return value if it can't be trusted to be the current value of the variable in the current bb. * gcc.dg/pr49602.c: New test. Added: trunk/gcc/testsuite/gcc.dg/pr49602.c Modified: trunk/gcc/ChangeLog trunk/gcc/testsuite/ChangeLog trunk/gcc/tree-into-ssa.c
Fixed.