Potentially codegen issue with ZP variables about llvm-mos

[mysterymath]

Yes, this definitely appears to be in error. I'll have to spend some time diagnosing this; the .zp.bss prefix should be sufficient to cause the assembler to emit zero page addressing and 8-bit relocations, but for some reason it isn't.

[mysterymath]

Closing this with a fix for the original issue, since it's more bug-like than the other tack-on issues.

Please do feel free to file separate issues for those, though; some of these may also be relatively easy to improve, but the work would likely be altogether unlike what was done here.

[johnwbyrd]

Just for the hell of it, have you tried:

__attribute__((section(".zp"))) static char color = 0;

[nolemor]

@johnwbyrd Interesting. That actually fixes it 🤔

00008034 <main>:
    8034: a2 00        	ldx	#$0
    8036: ac 02 20     	ldy	$2002
    8039: 10 fb        	bpl	$8036
    803b: a0 08        	ldy	#$8
    803d: 8c 01 20     	sty	$2001
    8040: a9 00        	lda	#$0
    8042: 4c 6f 80     	jmp	$806f
    8045: ac 02 20     	ldy	$2002
    8048: 10 fb        	bpl	$8045
    804a: e6 03        	inc	$3
    804c: d0 02        	bne	$8050
    804e: e6 02        	inc	$2
    8050: a4 03        	ldy	$3
    8052: e6 02        	inc	$2
    8054: c6 02        	dec	$2
    8056: d0 ed        	bne	$8045
    8058: c0 1e        	cpy	#$1e
    805a: d0 e9        	bne	$8045
    805c: ac 02 20     	ldy	$2002
    805f: a0 3f        	ldy	#$3f
    8061: 8c 06 20     	sty	$2006
    8064: 8d 06 20     	sta	$2006

    8067: a4 20        	ldy	$20
    8069: c8           	iny
    806a: e6 20        	inc	$20

    806c: 8c 07 20     	sty	$2007
    806f: 86 02        	stx	$2
    8071: 86 03        	stx	$3
    8073: 4c 45 80     	jmp	$8045

[nolemor]

Also another thing, while perhaps not related to the issue above, if I change the loop counter from int to char I get this generated code

00008034 <main>:
    8034: a9 00        	lda	#$0
    8036: ae 02 20     	ldx	$2002
    8039: 10 fb        	bpl	$8036
    803b: a2 08        	ldx	#$8
    803d: 8e 01 20     	stx	$2001
    8040: a0 00        	ldy	#$0
    8042: ae 02 20     	ldx	$2002
    8045: 10 fb        	bpl	$8042
    8047: 18           	clc
    8048: 69 01        	adc	#$1
    804a: c9 1e        	cmp	#$1e
    804c: d0 f4        	bne	$8042
    804e: ad 02 20     	lda	$2002
    8051: a2 3f        	ldx	#$3f
    8053: 8e 06 20     	stx	$2006
    8056: 8c 06 20     	sty	$2006

    8059: a5 20        	lda	$20
    805b: 18           	clc
    805c: 69 01        	adc	#$1
    805e: e6 20        	inc	$20

    8060: 8d 07 20     	sta	$2007
    8063: a9 00        	lda	#$0
    8065: 4c 42 80     	jmp	$8042

Two things to observe there is that the loop counter is now smaller (which makes sense as it's a 8-bit value) but it decides to change the inc into clc / adc which is a bit unexpected.

[johnwbyrd]

This bug can be addressed by rewriting

llvm-mos/llvm/lib/Target/MOS/MCTargetDesc/MOSAsmBackend.cpp

Line 278 in 86fbb3b

bool MOSAsmBackend::isZeroPageSectionName(StringRef Name) {

. Currently, it only checks to see if the name is contained in the ZeroPageSectionTable, when what we really want to do is check the Name's prefix against the table entries.

[mysterymath]

Thanks, that's exactly the deal.