#3120 Regression test failed bug2942247.c

[Sergey Belyashov]

source:

_Bool bar2(void)
{
        _Bool foo[2] = {1, 2};
        return foo[1];  // Crashed sdcc.
}

Result assembler (no-peep, i-code-in-asm):

; Stack space usage: 2 bytes.
_bar2::
        ld      a, #3
        rst     0x08
        add     sp, #-2
;gen/ucgbz80/bug2942247/bug2942247.c:19: _Bool foo[2] = {1, 2};
;ic:4:  arrayInit iTemp1 [k4 lr3:3 so:0]{ ia0 a2p0 re0 rm0 nos0 ru0 dp0}{_Bool near* fixed}
;       genArrayInit
;fetchPairLong
        ld      e, a
        ld      d, a
        call    __initrleblock
        .db     #2
        .db     #0x01, #0x01
        .db     #0

Here DE is loaded from A register which has gabarge.

[Philipp Klaus Krause]

Fo me that tests does not fail reliably (though I've seen teh failure before). The attached test should be a reliable reproducer.

[Philipp Klaus Krause]

This problem happens in an optimization, as can be seen from --dump-i-code:
The dumplood code still looks correct, but the dumplospre one doesn't.

However, it is not lospre that is at fault here, as --nolospre is not a workround for this bug.

That leaves the four minor optimizations done just before lospre:
optimizeOpWidth, optimizeCastCast, iCodeLabelOptimize, shortenLiveRanges.

[Philipp Klaus Krause]

Apparently the problem is related to offsetFoldGet. If I comment out the call to offsetFoldGet in eBBlockFromiCode, the bug disappears.

P.S.: Looks like this is due to offsetFoldGet is getting wrong information from an earlier analysis, as at offsetFoldGet OP_USES for the ADDRESS_OF does not include the ARRAYINIT iCode.

[Philipp Klaus Krause]

I'm not familiar with the parts of SDCC that set OP_USES, so I hope someone else can have a look at this issue in the arrayinit branch.

[Philipp Klaus Krause]

Apart from the bug, using an __initrleblock support call to initialize an array of total size 16 bits also seems quite inefficient.

[Sergey Belyashov]

agree. there are much of improvements to this part of code, but I have not sufficient experience to do it.

[Tony Pavlov]

i think that __initrleblock simply can be switched for objects bigger than 4 bytes.

[Philipp Klaus Krause]

In [r11936], I made the use of RLE depend on array size and array element size.

[Tony Pavlov]

isn't that function optimized to ld e, #1 when __initrleblock is disabled?

[Philipp Klaus Krause]

Fixed by reverting compressed array initalization support in trunk.

[Sergey Belyashov]

Bug is not fixed. It just hidden, because bug in another place not in AI. AI support just activates it.

[Philipp Klaus Krause]

As far as I know, the bug is specific to the ARRAYINIT iCode. The ARRAYINIT iCode is used only if compressed array initalization is enabled. So disabling the compressed array initalization in trunk fixes the bug (as no trunk user would be exposed to it).

[Philipp Klaus Krause]

I set the state to pending-fixed, since the bug is fixed (by disabling compressed array initialization) in trunk, but we can still use the bug report to track the bug in the arrayinit branch.

[Tony Pavlov]

that RLE initializer correctly inits arrays when using {0}, while current one initializes only the first element.