Hi all,
On 15/01/18 11:23, Kyrill Tkachov wrote:
Hi all,
In this wrong-code bug we combine a VSUB.I8 and a VABS.S8
into a VABD.S8 instruction . This combination is not valid
for integer operands because in the VABD instruction the semantics
are that the difference is computed in notionally infinite precision
and the absolute difference is computed on that, whereas for a
VSUB.I8 + VABS.S8 sequence the VSUB operation will perform any
wrapping that's needed for the 8-bit signed type before the VABS
gets its hands on it.
This leads to the wrong-code in the PR where the expected
sequence from the intrinsics:
VSUB + VABS of two vectors {-100, -100, -100...}, {100, 100, 100...}
gives a result of {56, 56, 56...} (-100 - 100)
but GCC optimises it into a single
VABD of {-100, -100, -100...}, {100, 100, 100...}
which produces a result of {200, 200, 200...}
The transformation is still valid for floating-point operands,
which is why it was added in the first place I believe (r178817)
but this patch disables it for integer operands.
The HFmode variants though only exist for TARGET_NEON_FP16INST, so
this patch adds the appropriate guards to the new mode iterator
Bootstrapped and tested on arm-none-linux-gnueabihf.
Committing to trunk.
I've backported this patch to the GCC 7 branch after
bootstrapping and testing on arm-none-linux-gnueabihf.
Thanks,
Kyrill
Thanks,
Kyrill
2018-01-15 Kyrylo Tkachov <kyrylo.tkac...@arm.com>
PR target/83687
* config/arm/iterators.md (VF): New mode iterator.
* config/arm/neon.md (neon_vabd<mode>_2): Use the above.
Remove integer-related logic from pattern.
(neon_vabd<mode>_3): Likewise.
2018-01-15 Kyrylo Tkachov <kyrylo.tkac...@arm.com>
PR target/83687
* gcc.target/arm/neon-combine-sub-abs-into-vabd.c: Delete integer
tests.
* gcc.target/arm/pr83687.c: New test.
