On 11/10/20 2:17 PM, Jakub Jelinek wrote:
> On Tue, Nov 10, 2020 at 02:09:20PM -0700, Jeff Law wrote:
>> On 11/10/20 1:14 PM, Jakub Jelinek via Gcc-patches wrote:
>>> On Tue, Nov 10, 2020 at 08:44:32PM +0100, Stefan Kanthak wrote:
>>>> Eric Botcazou <botca...@adacore.com> wrote:
>>>>
>>>>>> The implementation of the __ashlDI3(), __ashrDI3() and __lshrDI3()
>>>>>> functions
>>>>>> is rather bad, it yields bad machine code at least on i386 and AMD64.
>>>>>> Since
>>>>>> GCC knows how to shift integers twice the register size these functions
>>>>>> can
>>>>>> be written as one-liners.
>>>>> These functions are precisely meant to be used when GCC cannot do that.
>>>> On which processor(s) is GCC unable to generate code for DWtype shifts?
>>> E.g. avr-none, msp430-elf, pdp11-aout.
>>> And I see recursive __cmpdi2 calls on avr-none too.
>> ACK. I'll pull those [u]cmpdi changes. They were iffy at best, this
>> confirms the concerns we both had.
> To be precise, I haven't tried to build libgcc for all of those, just
> checked what code is produced for
> unsigned long long foo (unsigned long long x) { return x << 3; }
> etc. by all trunk cross-compilers I have lying around.
> It might be that their __ashlDI3 etc. are actually __ashlsi3 and they
> have some other implementation of __ashldi3 etc.
Well for avr libgcc's cmpdi compiles into something reasonable, but
ucmpdi calls cmpdi and I can't convince myself it's right. Regardless,
the [u]cmpdi stuff was very much borderline from a correctness
standpoint and I'm going to revert that part.
>
> Anyway, I'm not sure if we really need to optimize too much functions which
> are never used; and the -ftrapv stuff should one day be reimplemented using
> -fsanitize=signed-integer-overflow with an abort behavior on overflows.
>
> libgcc functions which are used heavily of course should be optimized as
> much as possible.
Agreed. The improvements to use the overflow intrinsics are somewhat
useful -- they make the code clearer and that's a win. The potential
self-recursive stuff in [u]cmpdi and the shifts probably aren't useful
(I haven't looked closely at the shifts yet). Stefan has a fundamental
mis-understanding of how those routines are used and more importantly
the constraints they have to work within.
Jeff
Jeff
>
> Jakub
>
