unsigned SI/DI mess

[GoogleCodeExporter]

testcase gcc.dg/range-test-1.c, routine test2()

problems occur with -O1, because gcc emits compare:CCUDI to compare and
unsigned SI (0-extended to DI) and immediate DI -4, while it should emit a
compare with immediate SI -4. not really obvious how to prevent gcc from
doing this.

note that there is a (seemingly unrelated) issue with char * compares in
routine test20(): attaching a preprocessed minimal testcase for just this
issue.

Original issue reported on code.google.com by jmoc...@gmail.com on 12 Feb 2009 at 1:37

Attachments:

• range-test-1-min.c

[GoogleCodeExporter]

the rtl causing the mess:

(insn 8 6 9 (set (reg:SI 134)
        (plus:SI (reg/v:SI 133 [ x ])
            (const_int -1 [0xffffffffffffffff]))) -1 (nil)
    (nil))

(insn 9 8 10 (set (reg:DI 135)
        (zero_extend:DI (reg:SI 134))) -1 (nil)
    (nil))

(insn 10 9 11 (set (reg:CCUDI 136)
        (compare:CCUDI (reg:DI 135)
            (const_int -4 [0xfffffffffffffffc]))) -1 (nil)
    (nil))

note the bit pattern of -4 in the last insn. it should really be 
0x00000000fffffffc.
however, in that case, the semantics of the compare is not the same anymore: so 
a
single compare won't do anymore, and the code should also test if 135 is >
0x00000000ffffffff, which makes such an optimization ... well ... a bit 
suboptimal.
now, I wonder whether this will require another CC mode for unsigned SI 
compares? ;)

Original comment by jmoc...@gmail.com on 12 Feb 2009 at 1:41

[GoogleCodeExporter]

the problem was that the mask (M)[1|0] notation sign-extends the negative value 
from
SI to DI. unsigned compares need zero-extension, though ...

fixed in r231 by force_reg()ing negative const_ints in case of unsigned SI 
compares
and explicitly zero-extending them.

might be that a similar solution would work for issue #89 ...

Original comment by jmoc...@gmail.com on 12 Feb 2009 at 2:35

• Changed state: Fixed