ARM : code less efficient with gcc-trunk ?

["Vincent R." <forumer at smartmobili dot com>]

Hi,

I am comparing the assembly generated by compilers targeting arm-wince
platform and it seems
that cross-compiler from gcc-trunk is less optimized than an old one based
on gcc 4.1.x
Here is the comparison obtained from objdump:


cegcc-4.1.x : 

00011000 <WinMainCRTStartup>:
   11000:	e92d40f0 	push	{r4, r5, r6, r7, lr}
   11004:	e1a04000 	mov	r4, r0
   11008:	e1a05001 	mov	r5, r1
   1100c:	e1a06002 	mov	r6, r2
   11010:	e1a07003 	mov	r7, r3
   11014:	eb0000de 	bl	11394 <_fpreset>
   11018:	eb00002a 	bl	110c8 <_pei386_runtime_relocator>
   1101c:	eb000099 	bl	11288 <__atexit_init>
   11020:	eb0000d3 	bl	11374 <__gccmain>
   11024:	e1a01005 	mov	r1, r5
   11028:	e1a00004 	mov	r0, r4
   1102c:	e1a02006 	mov	r2, r6
   11030:	e1a03007 	mov	r3, r7
   11034:	eb000005 	bl	11050 <WinMain>
   11038:	e1a04000 	mov	r4, r0
   1103c:	eb000087 	bl	11260 <_cexit>
   11040:	e1a01004 	mov	r1, r4
   11044:	e3a00042 	mov	r0, #66	; 0x42
   11048:	eb0000d4 	bl	113a0 <TerminateProcess>
   1104c:	eafffffe 	b	1104c <WinMainCRTStartup+0x4c>

cegcc-4.4.x

00011000 <WinMainCRTStartup>:
   11000:	e92d4010 	push	{r4, lr}
   11004:	e1a04000 	mov	r4, r0
   11008:	e24dd00c 	sub	sp, sp, #12	; 0xc
   1100c:	e58d1008 	str	r1, [sp, #8]
   11010:	e58d2004 	str	r2, [sp, #4]
   11014:	e58d3000 	str	r3, [sp]
   11018:	eb000120 	bl	114a0 <_fpreset>
   1101c:	eb000043 	bl	11130 <_pei386_runtime_relocator>
   11020:	eb0000ce 	bl	11360 <__atexit_init>
   11024:	eb000111 	bl	11470 <__gccmain>
   11028:	e59d1008 	ldr	r1, [sp, #8]
   1102c:	e1a00004 	mov	r0, r4
   11030:	e59d2004 	ldr	r2, [sp, #4]
   11034:	e59d3000 	ldr	r3, [sp]
   11038:	eb000028 	bl	110e0 <WinMain>
   1103c:	e1a04000 	mov	r4, r0
   11040:	eb0000ba 	bl	11330 <_cexit>
   11044:	e1a01004 	mov	r1, r4
   11048:	e3a00042 	mov	r0, #66	; 0x42
   1104c:	eb000116 	bl	114ac <TerminateProcess>
   11050:	eafffffe 	b	11050 <WinMainCRTStartup+0x50>
   11054:	e1a00000 	nop			(mov r0,r0)
   11058:	e1a00000 	nop			(mov r0,r0)
   1105c:	e1a00000 	nop			(mov r0,r0)

If you have a look at address 11008-1100c you can see that old gcc is using
registers 
but upcoming gcc-4.4 is using memory.

I tried to put some optim flags -O2 but it doesn't modify the situation.
Is there anything to do to improve this situation ? Is it a normal behavior
?
Maybe my remark is not relevant because I didn't try to do some benchmark
and I agree
this is not because gcc-trunk do not optimize this specific part that it
will be slower.
I have also noticed that now I get some nop instructions and when I ask gcc
to generate
assembly I can see that alignement directive is different.
I used to have .align 0 with gcc-4.1 and now I get a .align 4, how can I
change that ?
And finally maybe those nop insn prevents compiler from optimizing ...



Thanks