Hello,
I'm using gcc-4.1.1, built into an EABI tool chain with glibc-2.4 and
binutils-2.16.92.
I have a Fixed-to-Float transition program shown as below. But it can only work
with -O1 option. -O2/-O3/-Os are all failed.
% cat CTL_Math.c
typedef float CTLFLOAT;
typedef int CTLFIXED;
typedef int CTLI32;
typedef unsigned int CTLU32;
inline CTLFLOAT _CTL_Fixed2Float( CTLFIXED In )
{
CTLI32 lm = In;
CTLI32 tm = lm;
CTLI32 le;
CTLU32 bits;
CTLI32 man;
if (0 == In) return 0;
if (tm < 0)
tm = -tm;
// determine shift necessary to move radix to the 22-bit position
le = CTL_FASTFLOAT_MANTISSA_LEADINGZEROS - (CTLI32)(CLZ(tm));
if (le >= 0)
lm >>= le;
else
lm <<= -le;
// le = le + (CTL_FASTFLOAT_MANTISSA_BITS-17);
le += 5;
if (0==lm) le = 0;
bits = ((CTLI32)(le + 127) & 0x00FF) << 23;
man = lm;
if (lm < 0)
{
man = -man;
bits |= 0x80000000;
}
man = man << CTL_FASTFLOAT_IEEE_BITDIF_MANTISSA;
bits |= man & 0x007FFFFF;
return *(CTLFLOAT*)&bits;
}
I compiled it with following command.
% arm-iwmmxt-linux-gnueabi-gcc -O1 -Wall -mcpu=xscale -mtune=xscale
-mabi=aapcs-linux -c CTL_Math.c -o CTL_Math.o1
% arm-iwmmxt-linux-gnueabi-gcc -O2 -Wall -mcpu=xscale -mtune=xscale
-mabi=aapcs-linux -c CTL_Math.c -o CTL_Math.o2
I disassembled the object file and found the instruction order was different
between -O1 and -O2. With -O2 option, it seems GCC pull in 'ldr r0, [sp,
#4]', which will only get an obsolete r0. So the return value is incorrect. The
disassembly result is as below.
% arm-iwmmxt-linux-gnueabi-objdump -d CTL_Math.o1 > o1
% arm-iwmmxt-linux-gnueabi-objdump -d CTL_Math.o2 > o2
% vimdiff o1 o2
00000564 <_CTL_Fixed2Float>: |00000560
<_CTL_Fixed2Float>:
564: e92d4010 stmdb sp!, {r4, lr} | 560:
e92d4010 stmdb sp!, {r4, lr}
568: e24dd008 sub sp, sp, #8 ; 0x8 | 564:
e2504000 subs r4, r0, #0 ; 0x0
56c: e2504000 subs r4, r0, #0 ; 0x0 | 568:
e24dd008 sub sp, sp, #8 ; 0x8
570: 03a00000 moveq r0, #0 ; 0x0 | 56c:
03a02000 moveq r2, #0 ; 0x0
574: 0a000019 beq 5e0 <_CTL_Fixed2Float+0x7c> | 570:
1a000002 bne 580 <_CTL_Fixed2Float+0x20>
578: e0240fc4 eor r0, r4, r4, asr #31 | 574:
e1a00002 mov r0, r2
57c: e0400fc4 sub r0, r0, r4, asr #31 | 578:
e28dd008 add sp, sp, #8 ; 0x8
580: ebfffffe bl 0 <E3D_CLZ> | 57c:
e8bd8010 ldmia sp!, {r4, pc}
584: e270200a rsbs r2, r0, #10 ; 0xa | 580:
e0240fc4 eor r0, r4, r4, asr #31
588: 51a00254 movpl r0, r4, asr r2 | 584:
e0400fc4 sub r0, r0, r4, asr #31
58c: 42623000 rsbmi r3, r2, #0 ; 0x0 | 588:
ebfffffe bl 0 <E3D_CLZ>
590: 41a00314 movmi r0, r4, lsl r3 | 58c:
e270200a rsbs r2, r0, #10 ; 0xa
594: e3500000 cmp r0, #0 ; 0x0 | 590:
42623000 rsbmi r3, r2, #0 ; 0x0
598: 03a035fe moveq r3, #1065353216 ; 0x3f800000| 594:
41a00314 movmi r0, r4, lsl r3
59c: 058d3004 streq r3, [sp, #4] | 598:
51a00254 movpl r0, r4, asr r2
5a0: 03a00000 moveq r0, #0 ; 0x0 | 59c:
e3500000 cmp r0, #0 ; 0x0
5a4: 0a000006 beq 5c4 <_CTL_Fixed2Float+0x60> | 5a0:
03a035fe moveq r3, #1065353216 ; 0x3f800000
5a8: e2823084 add r3, r2, #132 ; 0x84 | 5a4:
058d3004 streq r3, [sp, #4]
5ac: e20330ff and r3, r3, #255 ; 0xff | 5a8:
0a00000c beq 5e0 <_CTL_Fixed2Float+0x80>
5b0: e1a03b83 mov r3, r3, lsl #23 | 5ac:
e2823084 add r3, r2, #132 ; 0x84
5b4: e58d3004 str r3, [sp, #4] | 5b0:
e20330ff and r3, r3, #255 ; 0xff
5b8: b2600000 rsblt r0, r0, #0 ; 0x0 | 5b4:
e1a02b83 mov r2, r3, lsl #23
5bc: b3833102 orrlt r3, r3, #-2147483648 ; 0x| 5b8:
b2603000 rsblt r3, r0, #0 ; 0x0
5c0: b58d3004 strlt r3, [sp, #4] | 5bc:
a1a03100 movge r3, r0, lsl #2
5c4: e3a02502 mov r2, #8388608 ; 0x800000 | 5c0:
b1a03103 movlt r3, r3, lsl #2
5c8: e2422001 sub r2, r2, #1 ; 0x1 | 5c4:
e58d2004 str r2, [sp, #4]
5cc: e0022100 and r2, r2, r0, lsl #2 | 5c8:
a3c304ff bicge r0, r3, #-16777216 ; 0x
5d0: e59d3004 ldr r3, [sp, #4] | 5cc:
b3822102 orrlt r2, r2, #-2147483648 ; 0x
5d4: e1833002 orr r3, r3, r2 | 5d0:
b3c304ff biclt r0, r3, #-16777216 ; 0x
5d8: e58d3004 str r3, [sp, #4] | 5d4:
a3c00502 bicge r0, r0, #8388608 ; 0x
5dc: e59d0004 ldr r0, [sp, #4] | 5d8:
b58d2004 strlt r2, [sp, #4]
5e0: e28dd008 add sp, sp, #8 ; 0x8 | 5dc:
b3c00502 biclt r0, r0, #8388608 ; 0x
5e4: e8bd8010 ldmia sp!, {r4, pc} | 5e0:
e59d3004 ldr r3, [sp, #4]
| 5e4:
e59d2004 ldr r2, [sp, #4]
| 5e8:
e1833000 orr r3, r3, r0
| 5ec:
e58d3004 str r3, [sp, #4]
| 5f0:
eaffffdf b 574 <_CTL_Fixed2Float+0x14>
The 'ldr r0, [sp, #4]' at 5dc line (left, O1) is put at 5e4 line (right,
O2).
I tried to narrow the test case to small enough. But I found the -O2 is fine to
this function in simplized test case while failed to other functions.
--
Summary: optimization -O2/-O3/-Os generate wrong instruction
order
Product: gcc
Version: 4.1.1
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c
AssignedTo: unassigned at gcc dot gnu dot org
ReportedBy: mingqiao dot wu at gmail dot com
GCC build triplet: i686-pc-linux-gnu
GCC host triplet: i686-pc-linux-gnu
GCC target triplet: arm-iwmmxt-linux-gnueabi
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=29140
