Re: [PATCH][i386][3/3] PR target/84164: Make *cmpqi_ext_<n> patterns accept more zero_extract modes

[Uros Bizjak <ubizjak at gmail dot com>]

On Wed, Feb 14, 2018 at 7:04 PM, Kyrill  Tkachov
<kyrylo.tkachov@foss.arm.com> wrote:
>
> On 13/02/18 16:45, Jeff Law wrote:
>>
>> On 02/09/2018 07:50 AM, Kyrill Tkachov wrote:
>>>
>>> Hi Uros,
>>>
>>> On 08/02/18 22:54, Uros Bizjak wrote:
>>>>
>>>> On Thu, Feb 8, 2018 at 6:11 PM, Kyrill  Tkachov
>>>> <kyrylo.tkachov@foss.arm.com> wrote:
>>>>>
>>>>> Hi all,
>>>>>
>>>>> This patch fixes some fallout in the i386 testsuite that occurs after
>>>>> the
>>>>> simplification in patch [1/3] [1].
>>>>> The gcc.target/i386/extract-2.c FAILs because it expects to match:
>>>>> (set (reg:CC 17 flags)
>>>>>       (compare:CC (subreg:QI (zero_extract:SI (reg:HI 98)
>>>>>                   (const_int 8 [0x8])
>>>>>                   (const_int 8 [0x8])) 0)
>>>>>           (const_int 4 [0x4])))
>>>>>
>>>>> which is the *cmpqi_ext_2 pattern in i386.md but with the new
>>>>> simplification
>>>>> the combine/simplify-rtx
>>>>> machinery produces:
>>>>> (set (reg:CC 17 flags)
>>>>>       (compare:CC (subreg:QI (zero_extract:HI (reg:HI 98)
>>>>>                   (const_int 8 [0x8])
>>>>>                   (const_int 8 [0x8])) 0)
>>>>>           (const_int 4 [0x4])))
>>>>>
>>>>> Notice that the zero_extract now has HImode like the register source
>>>>> rather
>>>>> than SImode.
>>>>> The existing *cmpqi_ext_<n> patterns however explicitly demand an
>>>>> SImode on
>>>>> the zero_extract.
>>>>> I'm not overly familiar with the i386 port but I think that's too
>>>>> restrictive.
>>>>> The RTL documentation says:
>>>>> For (zero_extract:m loc size pos) "The mode m is the same as the mode
>>>>> that
>>>>> would be used for loc if it were a register."
>>>>> I'm not sure if that means that the mode of the zero_extract and the
>>>>> source
>>>>> register must always match (as is the
>>>>> case after patch [1/3]) but in any case it shouldn't matter
>>>>> semantically
>>>>> since we're taking a QImode subreg of the whole
>>>>> thing anyway.
>>>>>
>>>>> So the proposed solution in this patch is to allow HI, SI and DImode
>>>>> zero_extracts in these patterns as these are the
>>>>> modes that the ext_register_operand predicate accepts, so that the
>>>>> patterns
>>>>> can match the new form above.
>>>>>
>>>>> With this patch the aforementioned test passes again and bootstrap and
>>>>> testing on x86_64-unknown-linux-gnu shows
>>>>> no regressions.
>>>>>
>>>>> Is this ok for trunk if the first patch is accepted?
>>>>
>>>> Huh, there are many other zero-extract patterns besides cmpqi_ext_*
>>>> with QImode subreg of SImode zero_extract in i386.md, used to access
>>>> high QImode register of HImode pair. A quick grep shows these that
>>>> have _ext_ in their name:
>>>>
>>>> (define_insn "*cmpqi_ext_1"
>>>> (define_insn "*cmpqi_ext_2"
>>>> (define_expand "cmpqi_ext_3"
>>>> (define_insn "*cmpqi_ext_3"
>>>> (define_insn "*cmpqi_ext_4"
>>>> (define_insn "addqi_ext_1"
>>>> (define_insn "*addqi_ext_2"
>>>> (define_expand "testqi_ext_1_ccno"
>>>> (define_insn "*testqi_ext_1"
>>>> (define_insn "*testqi_ext_2"
>>>> (define_insn_and_split "*testqi_ext_3"
>>>> (define_insn "andqi_ext_1"
>>>> (define_insn "*andqi_ext_1_cc"
>>>> (define_insn "*andqi_ext_2"
>>>> (define_insn "*<code>qi_ext_1"
>>>> (define_insn "*<code>qi_ext_2"
>>>> (define_expand "xorqi_ext_1_cc"
>>>> (define_insn "*xorqi_ext_1_cc"
>>>>
>>>> There are also relevant splitters and peephole2 patterns.
>>>
>>> I see. Another approach I've looked at is removing the mode specifier
>>> from
>>> the zero_extract in these patterns. This means that they can be of any
>>> mode
>>> so they will match all of these modes without creating new patterns
>>> through
>>> iterators. That also works for the testcase and passes bootstrap and
>>> testing
>>> however there is the snag that the define_insns that don't start with a
>>> "*"
>>> are used to generate RTL through the gen_* mechanism and in that context
>>> the absence of a mode on the zero_extract would mean a VOIDmode
>>> zero_extract
>>> would be created, which I'm fairly sure is not good. So in my
>>> experiments I left
>>> those patterns alone (with an explicit SI on the zero_extract).
>>>
>>>> IIRC, SImode zero_extract was enough to catch all high-register uses.
>>>> There will be a pattern explosion if we want to handle all other
>>>> integer modes here. However, I'm not a RTL expert, so someone will
>>>> have to say what is the correct RTX form here.
>>>
>>> Jeff, Richard, could you please give us some guidance on this issue?
>>> Sorry for the trouble.
>>>
>> I don't think any of the patterns above are known to the generic code.
>> So you just have to check the x86 backend to see their precise uses in a
>> generator (ie gen_cmpqi_ext_3) and verify those do not allow a VOIDmode
>> (or any other undesirable mode) to slip through.
>>
>> Jeff
>
>
> Thanks Jeff, I did have a look. I think we want to maintain the SImode on
> the
> RTL that gets created through these named expanders, as generating a
> VOIDmode
> zero_extract is not valid. So my patch leaves those intact.
> The patch removes the mode from the zero_extract RTXes in patterns that are
> only ever going to get matched (as opposed to generated). That is the ones
> that
> start with "*" in their name.
> This should allow them to match any of the zero_extract modes that
> might get generated by the midend.
>
> Bootstrapped and tested on x86_64-unknown-linux-gnu.
> Is this a preferable approach?

We are trying to avoid VOIDmode RTXes as much as possible (to tighten
pattern matching to avoid various surpsrises). Looking at these
instructions, I think that using SWI248 in insn and peephole2 patterns
should be OK.

So, if it survives regression tests, the patch is OK with SWI248 mode
instead of void mode.

Thanks,
Uros.

>
> Thanks,
> Kyrill
>
> 2018-02-14  Kyrylo Tkachov  <kyrylo.tkachov@arm.com>
>
>     PR target/84164
>     * config/i386/i386.md (*cmpqi_ext_1, *cmpqi_ext_2, *cmpqi_ext_3,
>     *cmpqi_ext_4, *extzvqi_mem_rex64, *extzvqi, QImode zero_extract
>     peephole, *addqi_ext_2, *testqi_ext_1, *testqi_ext_2, *andqi_ext_1_cc,
>     *andqi_ext_2, *<code>qi_ext_1, *<code>qi_ext_2, *xorqi_ext_1_cc
>     AND QImode subreg of zero_extract peephole): Remove mode from
> zero_extract.
>