Running ../../../gdb/testsuite/gdb.base/store.exp ... FAIL: gdb.base/store.exp: upvar charest l; print old l, expecting -1 .* FAIL: gdb.base/store.exp: upvar short l; print old l, expecting -1 FAIL: gdb.base/store.exp: upvar int l; print old l, expecting -1 FAIL: gdb.base/store.exp: upvar long l; print old l, expecting -1 FAIL: gdb.base/store.exp: upvar longest l; print old l, expecting -1 FAIL: gdb.base/store.exp: upvar doublest l; print old l, expecting -1 are new failures in gdb testsuite when compiled with 4.4 compared to store.c compiled with 4.3.x. The difference seems to be introduced by IRA, when compiled with -fno-ira (when trunk still had that option) the test worked. With the old RA, -g -O0 -dA -fverbose-asm compiled: unsigned int foo (unsigned int, unsigned int); unsigned int bar (register unsigned int a, register unsigned int b) { register unsigned int c = a, d = b; c = foo (c, d); return c + d; } d is allocated in a call-saved register (%ebx), which is fine and c is assigned a stack slot. With IRA d is still allocated in %ebx, but c is in %eax, call-clobbered register. So when gdb in the foo call does up and checks the value in c and d, it finds correct value of d, but garbage in c. Could we perhaps at -O0 avoid allocating user variables that are live accross function calls in call-clobbered registers? For -O1 it is obviously a fine decision.
Jakub, how is about the following patch. Is it ok for you? I mean correct user variable identification. 2009-03-11 Vladimir Makarov <vmakarov@redhat.com> PR debug/39432 * ira-int.h (struct allocno): Fix comment for calls_crossed_num. * ira-conflicts.c (ira_build_conflicts): Prohibit call used registers for allocnos created from user-defined variables. Index: ira-int.h =================================================================== --- ira-int.h (revision 144543) +++ ira-int.h (working copy) @@ -333,7 +333,7 @@ struct ira_allocno /* Accumulated frequency of calls which given allocno intersects. */ int call_freq; - /* Length of the previous array (number of the intersected calls). */ + /* Accumulated number of the intersected calls. */ int calls_crossed_num; /* Non NULL if we remove restoring value from given allocno to MEM_OPTIMIZED_DEST at loop exit (see ira-emit.c) because the Index: ira-conflicts.c =================================================================== --- ira-conflicts.c (revision 144543) +++ ira-conflicts.c (working copy) @@ -800,29 +800,33 @@ ira_build_conflicts (void) } FOR_EACH_ALLOCNO (a, ai) { - if (ALLOCNO_CALLS_CROSSED_NUM (a) == 0) - continue; - if (! flag_caller_saves) + reg_attrs *attrs; + tree decl; + + if ((! flag_caller_saves && ALLOCNO_CALLS_CROSSED_NUM (a) != 0) + /* For debugging purposes don't put user defined variables in + callee-clobbered registers. */ + || (optimize <= 1 + && (attrs = REG_ATTRS (regno_reg_rtx [ALLOCNO_REGNO (a)])) != NULL + && (decl = attrs->decl) != NULL + && VAR_OR_FUNCTION_DECL_P (decl) + && DECL_NAME (decl) != NULL)) { IOR_HARD_REG_SET (ALLOCNO_TOTAL_CONFLICT_HARD_REGS (a), call_used_reg_set); - if (ALLOCNO_CALLS_CROSSED_NUM (a) != 0) - IOR_HARD_REG_SET (ALLOCNO_CONFLICT_HARD_REGS (a), - call_used_reg_set); + IOR_HARD_REG_SET (ALLOCNO_CONFLICT_HARD_REGS (a), + call_used_reg_set); } - else + else if (ALLOCNO_CALLS_CROSSED_NUM (a) != 0) { IOR_HARD_REG_SET (ALLOCNO_TOTAL_CONFLICT_HARD_REGS (a), no_caller_save_reg_set); IOR_HARD_REG_SET (ALLOCNO_TOTAL_CONFLICT_HARD_REGS (a), temp_hard_reg_set); - if (ALLOCNO_CALLS_CROSSED_NUM (a) != 0) - { - IOR_HARD_REG_SET (ALLOCNO_CONFLICT_HARD_REGS (a), - no_caller_save_reg_set); - IOR_HARD_REG_SET (ALLOCNO_CONFLICT_HARD_REGS (a), - temp_hard_reg_set); - } + IOR_HARD_REG_SET (ALLOCNO_CONFLICT_HARD_REGS (a), + no_caller_save_reg_set); + IOR_HARD_REG_SET (ALLOCNO_CONFLICT_HARD_REGS (a), + temp_hard_reg_set); } } if (optimize && ira_conflicts_p
You should use DECL_ARTIFICIAL I think.
Thanks, Richard. So instead of "DECL_NAME (decl) != NULL" I should use "! DECL_ARTIFICIAL (decl)". Right? Ok, I'll test the new patch then and send it for approval after testing.
Also perhaps should test DECL_HARD_REGISTER, for DECL_HARD_REGISTER we shouldn't limit them in any way to allow the user to shoot himself. In any case, I'll test your patch momentarily.
As for DECL_HARD_REGISTER, such decl regs are never considered by IRA for allocation. So I think there is no necessity to check them here.
I can confirm that trunk with the #c1 patch modified as mentioned in #c3 cures all the gdb.base/store.exp failures.
Subject: Bug 39432 Author: vmakarov Date: Thu Mar 12 14:39:55 2009 New Revision: 144812 URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=144812 Log: 2009-03-12 Vladimir Makarov <vmakarov@redhat.com> PR debug/39432 * ira-int.h (struct allocno): Fix comment for calls_crossed_num. * ira-conflicts.c (ira_build_conflicts): Prohibit call used registers for allocnos created from user-defined variables. Modified: trunk/gcc/ChangeLog trunk/gcc/ira-conflicts.c trunk/gcc/ira-int.h
Fixed, thanks.