cc65: Regression: "0 < i--" is optimized to "0 < --i".

Fixed with https://github.com/cc65/cc65/pull/1053.

Register usage isn’t currently tracked correctly for OptStackOps transformations, which is the cause of this issue as well as #167. I’m working on an improved way to handle this.

The error is introduced in the compiler pass “OptStackOps”. I used a slightly simplified version of the test program posted by @greg-king5 .

static unsigned char failure = 1;
static unsigned char count = 1;

int main(void)
{
    while (0 < count--)
    {
        --failure;
    }
    return failure;
}

Last good pass:

;
; while (0 < count--)
;
	jmp     L0006
;
; --failure;
;
L0004:	ldx     #$00
	dec     _failure
	lda     _failure
;
; while (0 < count--)
;
L0006:	ldx     #$00
	txa
	jsr     push0
	lda     _count
	dec     _count
	jsr     tosicmp0
	bcc     L0004
;
; return failure;
;
	ldx     #$00
	lda     _failure
;
; }
;
	rts

After OptStackOps:

;
; while (0 < count--)
;
	jmp     L0006
;
; --failure;
;
L000A:	dec     _failure
	lda     _failure
;
; while (0 < count--)
;
L0006:	ldx     #$00
	txa
	dec     _count
	cmp     _count	; (A)
	bcc     L000A
;
; return failure;
;
	lda     _failure
;
; }
;
	rts

On line (A), a new cmp instruction is inserted by the optimizer pass, testing count after it has already been decremented, despite the fact that it was post-decremented.

static int failure = 1;
static unsigned char count = 1;

int main(void)
{
    while (0 < count--) {
        --failure;
    }
    return failure;
}

The unoptimized program succeeds. The optimized program fails.

its still close to impossible to fix anything without a minimal testcase (and then the testcase will go into the regression tests, to make sure it wont break again)