http://gcc.gnu.org/bugzilla/show_bug.cgi?id=53433
--- Comment #21 from Uros Bizjak <ubizjak at gmail dot com> 2012-07-04 10:10:13
UTC ---
It looks that expand_builtin_strncmp gets miscompiled.
Try to compile the testcase from Comment #5 with -O2 on x86_64-pc-linux-gnu.
Put breakpoint in gen_cmpstrnsi.
When debugging "normal", non-LTO, non-profiled build, we get:
Breakpoint 1, gen_cmpstrnsi (operand0=0x7ffff1ace8a0, operand1=0x7ffff1acb8b8,
operand2=0x7ffff1acb918, operand3=0x7ffff19a84f0, operand4=0x7ffff19a8480)
at ../../gcc-svn/trunk/gcc/config/i386/i386.md:15989
15989 (set (match_operand:P 0 "register_operand" "=D")
(gdb) bt
#0 gen_cmpstrnsi (operand0=0x7ffff1ace8a0, operand1=0x7ffff1acb8b8,
operand2=0x7ffff1acb918,
operand3=0x7ffff19a84f0, operand4=0x7ffff19a8480)
at ../../gcc-svn/trunk/gcc/config/i386/i386.md:15989
#1 0x00000000005bb0c0 in expand_builtin_strncmp (exp=0x7ffff1acdd20,
target=0x7ffff1ace8a0,
mode=<optimized out>) at ../../gcc-svn/trunk/gcc/builtins.c:4000
(gdb) up
#1 0x00000000005bb0c0 in expand_builtin_strncmp (exp=0x7ffff1acdd20,
target=0x7ffff1ace8a0,
mode=<optimized out>) at ../../gcc-svn/trunk/gcc/builtins.c:4000
4000 GEN_INT (MIN (arg1_align, arg2_align)));
(gdb) li
3995
3996 arg1_rtx = get_memory_rtx (arg1, len);
3997 arg2_rtx = get_memory_rtx (arg2, len);
3998 arg3_rtx = expand_normal (len);
3999 insn = gen_cmpstrnsi (result, arg1_rtx, arg2_rtx, arg3_rtx,
4000 GEN_INT (MIN (arg1_align, arg2_align)));
4001 if (insn)
4002 {
4003 emit_insn (insn);
4004
(gdb) p target
$11 = (rtx_def *) 0x7ffff1ace8a0
(gdb) p result
$12 = (rtx_def *) 0x7ffff1ace8a0
(gdb)
debugging profiledbootstrap (--with-build-config=bootstrap-lto):
Breakpoint 3, gen_cmpstrnsi (operand0=0x7ffff1ace9e0, operand1=0x7ffff1acb8b8,
operand2=0x7ffff1acb918, operand3=0x7ffff19a84f0, operand4=0x7ffff19a8480)
at ../../gcc-svn/trunk/gcc/config/i386/i386.md:15989
15989 (set (match_operand:P 0 "register_operand" "=D")
(gdb) bt
#0 gen_cmpstrnsi (operand0=0x7ffff1ace9e0, operand1=0x7ffff1acb8b8,
operand2=0x7ffff1acb918,
operand3=0x7ffff19a84f0, operand4=0x7ffff19a8480)
at ../../gcc-svn/trunk/gcc/config/i386/i386.md:15989
#1 0x0000000000b3310f in
_ZL22expand_builtin_strncmpP9tree_nodeP7rtx_def12machine_mode.isra.13 (
target=0x7ffff1ace8a0, exp=0x7ffff1acdd20) at
../../gcc-svn/trunk/gcc/builtins.c:4000
(gdb) p target
$12 = (struct rtx_def *) 0x7ffff1ace8a0
(gdb) p result
$13 = '\000' <repeats 113 times>
(gdb)
The trick is that a couple of lines above, we have:
/* Make a place to write the result of the instruction. */
result = target;
if (! (result != 0
&& REG_P (result) && GET_MODE (result) == insn_mode
&& REGNO (result) >= FIRST_PSEUDO_REGISTER))
result = gen_reg_rtx (insn_mode);
Probably "result" doesn't get propagated correctly to gen_cmpstrnsi, we have:
(gdb) p *target
$15 = {code = REG, mode = SImode, jump = 0, call = 0, unchanging = 0, volatil =
0, in_struct = 0,
(gdb) down
#0 gen_cmpstrnsi (operand0=0x7ffff1ace9e0, operand1=0x7ffff1acb8b8,
operand2=0x7ffff1acb918,
operand3=0x7ffff19a84f0, operand4=0x7ffff19a8480)
at ../../gcc-svn/trunk/gcc/config/i386/i386.md:15989
15989 (set (match_operand:P 0 "register_operand" "=D")
(gdb) p *operand0
$14 = {code = REG, mode = VOIDmode, jump = 0, call = 0, unchanging = 0, volatil
= 0, in_struct = 0,
Everything goes downhill there.
(sorry for not using debug_rtx here, calling it gdb says "Cannot resolve
function debug_rtx to any overloaded instance".)
