https://gcc.gnu.org/bugzilla/show_bug.cgi?id=60947
--- Comment #13 from amker at gcc dot gnu.org ---
OK, I compared generated assembly before/after revision 206552.
BEFORE)
@ frame_needed = 1, uses_anonymous_args = 0
mov ip, sp
stmfd sp!, {r4, r5, r6, r7, r8, r9, r10, fp, ip, lr, pc}
sub fp, ip, #4
sub sp, sp, #20
mov r4, #0
ldr r8, [r0, #112]
.L1064:
add r0, r4, #232
add r0, r8, r0
mov r1, #255
mov r2, #8
add r4, r4, #32
bl memset
cmp r4, #96
bne .L1064
AFTER)
@ frame_needed = 1, uses_anonymous_args = 0
mov ip, sp
stmfd sp!, {r4, r5, r6, r7, r8, r9, r10, fp, ip, lr, pc}
sub fp, ip, #4
sub sp, sp, #20
ldr r8, [r0, #112]
add r3, r8, #232
add r4, r8, #328
.L1064:
mov r0, r3
mov r1, #255
mov r2, #8
bl memset
add r3, r0, #32 <---X
cmp r3, r4
bne .L1064
The only possibility I can see is GCC now uses r3 as the induction variable,
which is a scratch register according to arm acps. Problem is according to
posix standard:
#include <string.h>
void *memset(void *s, int c, size_t n);
RETURN VALUE:
The memset() function shall return s; no return value is reserved to indicate
an error.
So r0 returned in INSN X should be set r3 back to right value.
This can be confirmed by rtl dumps before/after reload:
BEFORE reload)
24: NOTE_INSN_BASIC_BLOCK 2
2: NOTE_INSN_DELETED
3: NOTE_INSN_FUNCTION_BEG
26: r114:SI=[r0:SI+0x70]
REG_DEAD r0:SI
27: r165:SI=r114:SI+0xe8
28: r162:SI=r114:SI+0x148
32: L32:
33: NOTE_INSN_BASIC_BLOCK 3
34: r0:SI=r165:SI
684: r1:SI=0xff
685: r2:SI=0x8
37: r0:SI=call [`memset'] argc:0
REG_RETURNED r165:SI
REG_DEAD r2:SI
REG_DEAD r1:SI
REG_UNUSED r0:SI
REG_EH_REGION 0
38: r165:SI=r165:SI+0x20
40: cc:CC=cmp(r165:SI,r162:SI)
41: pc={(cc:CC!=0)?L32:pc}
REG_DEAD cc:CC
REG_BR_PROB 6667
AFTER reload)
24: NOTE_INSN_BASIC_BLOCK 2
2: NOTE_INSN_DELETED
3: NOTE_INSN_FUNCTION_BEG
26: r8:SI=[r0:SI+0x70]
27: r3:SI=r8:SI+0xe8
28: r4:SI=r8:SI+0x148
32: L32:
33: NOTE_INSN_BASIC_BLOCK 3
34: r0:SI=r3:SI
684: r1:SI=0xff
685: r2:SI=0x8
37: r0:SI=call [`memset'] argc:0
REG_RETURNED r3:SI
REG_EH_REGION 0
755: r3:SI=r0:SI <--------restore r3 from r0
38: r3:SI=r3:SI+0x20
40: cc:CC=cmp(r3:SI,r4:SI)
41: pc={(cc:CC!=0)?L32:pc}
REG_BR_PROB 6667
So here I really don't know why this program can fail.
As far as ivopt is concerned, the dumps of ivopt are like:
BEFORE patch)
sizetype ivtmp.551;
<bb 2>:
runtime_9 = substream_8(D)->runtime;
goto <bb 4>;
<bb 3>:
<bb 4>:
# ivtmp.551_154 = PHI <ivtmp.551_150(3), 0(2)>
_13 = ivtmp.551_154 + 232;
_1 = runtime_9 + _13;
_59 = _1;
memset (_59, 255, 8);
ivtmp.551_150 = ivtmp.551_154 + 32;
if (ivtmp.551_150 != 96)
goto <bb 3>;
else
goto <bb 48>;
AFTER patch)
unsigned int ivtmp.550;
<bb 2>:
runtime_9 = substream_8(D)->runtime;
_1 = runtime_9 + 232;
ivtmp.550_13 = (unsigned int) _1;
_143 = runtime_9 + 328;
_142 = (unsigned int) _143;
goto <bb 4>;
<bb 3>:
<bb 4>:
# ivtmp.550_154 = PHI <ivtmp.550_150(3), ivtmp.550_13(2)>
_59 = (struct snd_mask *) ivtmp.550_154;
memset (_59, 255, 8);
ivtmp.550_150 = ivtmp.550_154 + 32;
if (ivtmp.550_150 != _142)
goto <bb 3>;
else
goto <bb 48>;
The transformation looks fine to me, right?
