#3203 3 regtests fail in z80-related ports for --max-allocs-per-node 10

[Philipp Klaus Krause]

Currently, the following tests fail for some z80-related ports when --max-allocs-per-node 10 is added to SDCCFLAGS in spec.mk:

banked:
fails for ez80-z80

gcc-torture-execute-961213-1:
fails for ucr3ka, ez80-z80 and tlcs90.

rotate2:
fails for ucz80-resiy and ucgbz80.

[Philipp Klaus Krause]

The failure of banked for ez80-z80 is actually a bug in peephole rule 33, which apparently gets confused about 8-bit vs. 16-bit loads:

;   genLeftShift
;   AOP_STK for _f_ab_sloc0_1_0
;   AOP_STK for _f_ab_sloc1_1_0
    ld  a, -2 (ix)
    ld  -1 (ix), a
;   genCast
;   AOP_STK for _f_ab_sloc1_1_0
;   AOP_STK for _f_ab_sloc2_1_0
;   genMove_o
; common peephole 33 loaded hl from #0x00 instead of going through -2 (ix).
    ld  -2 (ix), #0x00
    ld  hl, #0x00
    ld  -4 (ix), hl

Looks like a zero byte was loaded into one of the bytes of a variable om the stack, then that varaible was loaded into hl. The rule replaced that by a load of a zero word into hl.

[Sergey Belyashov]

banked fixed for eZ80 in [r12174]

[Sergey Belyashov]

genSwap generate invalid swap32 when left is AOP_STK with offset -4. Currently I do not understand how to compare aopu.aop_stk.

[Sergey Belyashov]

Fixed in [r12178].

But test gcc-torture-execute-961213-1 for tlcs90 and ez80 is failing yet.

[Sergey Belyashov]

I have added new fix for gcc-torture-execute-961213-1 in [r12181], but it cause new regressions. For example, z80n:ast_constant_folding, genPlus generates code for bc = hl + fix:

;ic:3:  iTemp1 [k4 lr4:5 so:0]{ ia0 a2p0 re0 rm0 nos0 ru0 dp0}{volatile-unsigned-int fixed}[c b ] = iTemp0 [k3 lr3:4 so:0]{ ia
0 a2p0 re0 rm0 nos0 ru0 dp0}{unsigned-int fixed}[l h ] + _i [k2 lr0:0 so:0]{ ia0 a2p0 re0 rm0 nos0 ru0 dp0}{volatile-unsigned-
int fixed}
;       genPlus
        ld      a, l ;WTF?!?
;       Shift into pair hl
;fetchLitPair
        ld      hl, #_i
        add     a, (hl)
        ld      c, a
        ld      a, h
        inc     hl
        adc     a, (hl)
        ld      b, a

[Philipp Klaus Krause]

As of [r12258], I do not see any remaining code generation issues when using --max-allocs-per-node 10.
There is one bug triggered for r3ka when using --max-allocs-per-node 10, but it looks like a peephole issue to me, so I'll open a new bug report.