Re: gcc 2.95.2-13 for arm

[Nicholas Clark <nick@ccl4.org>]

On Sun, Sep 10, 2000 at 10:38:53PM +0100, Philip Blundell wrote:
> >/tmp/ccqBnxST.s: Assembler messages:
> >/tmp/ccqBnxST.s:8196: Warning: destination register same as write-back base
> >
> >This would be a bad thing. This would appear to be the bad thing I had
> >before with the previous (debian supplied) compiler. How (where) do I go
> >about working out if it's a known compiler bug?
> 
> What's the offending line of the file?  I just sent Scott a patch this morning 
> for some missing constraints on post-inc `str' peepholes; it sounds like you 
> might have run into an analogous situation with `ldr', though I thought they 
> were all covered already.

	ldr	r3, [r3], #-4

in

Perl_pp_entersub:
	@ args = 0, pretend = 0, frame = 40
	@ frame_needed = 1, current_function_anonymous_args = 0
	mov	ip, sp
	stmfd	sp!, {r4, r5, r6, r7, r8, r9, sl, fp, ip, lr, pc}
	sub	fp, ip, #4
	sub	sp, sp, #40
	ldr	r3, .L1502
	ldr	r3, [r3, #0]
	ldr	r3, [r3], #-4
	str	r3, [fp, #-60]
	str	r3, [fp, #-56]
	ldr	r3, .L1502+4
	ldr	r3, [r3, #0]
	ldrb	r3, [r3, #20]	@ zero_extendqisi2
	mov	r3, r3, lsr #6
	and	r3, r3, #1
	str	r3, [fp, #-68]
	ldr	r3, [fp, #-56]
	cmp	r3, #0
	beq	.L1497
	ldr	r3, [fp, #-56]
	ldrb	r3, [r3, #8]	@ zero_extendqisi2

[snip]

.L1502:
	.word	PL_stack_sp
	.word	PL_op
	.word	PL_sv_yes
	.word	PL_stack_base
	.word	PL_markstack_ptr
	.word	PL_no_usym
	.word	.LC41
	.word	PL_no_symref
	.word	PL_amagic_generation
	.word	PL_sv_undef
	.word	.LC4
	.word	PL_tmps_floor
	.word	PL_tmps_ix
	.word	.LC42
	.word	.LC43
	.word	.LC40



OP * Perl_pp_entersub (void ) 
{
    register SV **sp = PL_stack_sp ; SV *sv = (*sp--)  ;
    GV *gv;
    HV *stash;
    register CV *cv;
    register PERL_CONTEXT *cx;
    I32 gimme;
    char  hasargs = (PL_op->op_flags & 64 ) != 0;

    if (!sv)
	return Perl_die (  "Not a CODE reference");



it looks like it's that *sp--;


-O2 assembler looks like this:

Perl_pp_entersub:
	@ args = 0, pretend = 0, frame = 44
	@ frame_needed = 1, current_function_anonymous_args = 0
	mov	ip, sp
	stmfd	sp!, {r4, r5, r6, r7, r8, r9, sl, fp, ip, lr, pc}
	ldr	r0, .L1488
	sub	sp, sp, #44
	ldr	r1, [r0, #0]
	sub	fp, ip, #4
	ldr	ip, [r1], #-4
	str	r1, [fp, #-60]
	str	ip, [fp, #-56]
	ldr	r1, .L1488+4
	ldr	r2, [r1, #0]
	ldrb	r3, [r2, #20]	@ zero_extendqisi2
	cmp	ip, #0
	mov	r3, r3, lsr #6
	and	r3, r3, #1
	str	r3, [fp, #-68]
	beq	.L1483
	ldrb	r3, [ip, #8]	@ zero_extendqisi2

.L1488:
	.word	PL_stack_sp
	.word	PL_op
	.word	PL_sv_yes
	.word	PL_markstack_ptr
	.word	PL_stack_base
	.word	PL_no_usym


which I think is correct (r1 is sp, ip is sv)
and all my SEGVs seem to be caused by PL_stack_sp[1] containing something
bogus (ie odd number < 2048) rather than a deferenceable pointer.


Having spent a long time staring at the output of the optimiser I'm generally
very impressed - a job well done, thanks.

That was going to be unconditional, but I've just spotted this. Can't:

	str	ip, [fp, #-56]
	ldr	r1, .L1488+4
	ldr	r2, [r1, #0]
	ldrb	r3, [r2, #20]	@ zero_extendqisi2
	cmp	ip, #0

be re-ordered to:

	ldr	r1, .L1488+4
	str	ip, [fp, #-56]
	ldr	r2, [r1, #0]
	cmp	ip, #0
	ldrb	r3, [r2, #20]	@ zero_extendqisi2

?

Nicholas Clark


-- 
To UNSUBSCRIBE, email to debian-arm-request@lists.debian.org
with a subject of "unsubscribe". Trouble? Contact listmaster@lists.debian.org