Re: PING^2 [PATCH] rs6000: Fix an assertion in update_target_cost_per_stmt [PR103702]

["Kewen.Lin" <linkw@linux.ibm.com>]

on 2022/1/26 下午3:28, Richard Biener wrote:
> On Wed, Jan 26, 2022 at 3:15 AM Kewen.Lin via Gcc-patches
> <gcc-patches@gcc.gnu.org> wrote:
>>
>> Gentle ping:
>>
>> https://gcc.gnu.org/pipermail/gcc-patches/2021-December/587309.html
> 
> OK.
> 

Thanks Richi!  Rebased, re-tested and committed as r12-6891.

BR,
Kewen

> Thanks,
> Richard.
> 
>> BR,
>> Kewen
>>
>>> on 2021/12/23 上午10:06, Kewen.Lin via Gcc-patches wrote:
>>>> Hi,
>>>>
>>>> This patch is to fix one wrong assertion which is too aggressive.
>>>> Vectorizer can do vec_construct costing for the vector type which
>>>> only has one unit.  For the failed case, the passed-in vector type
>>>> is "vector(1) int", though it doesn't end up with any construction
>>>> eventually.  We have to handle this kind of input in function
>>>> rs6000_cost_data::update_target_cost_per_stmt.
>>>>
>>>> Bootstrapped and regtested on powerpc64le-linux-gnu P9 and
>>>> powerpc64-linux-gnu P8.
>>>>
>>>> Is it ok for trunk?
>>>>
>>>> BR,
>>>> Kewen
>>>> -----
>>>> gcc/ChangeLog:
>>>>
>>>>      PR target/103702
>>>>      * config/rs6000/rs6000.c
>>>>      (rs6000_cost_data::update_target_cost_per_stmt): Fix one wrong
>>>>      assertion with early return.
>>>>
>>>> gcc/testsuite/ChangeLog:
>>>>
>>>>      PR target/103702
>>>>      * gcc.target/powerpc/pr103702.c: New test.
>>>> ---
>>>>  gcc/config/rs6000/rs6000.c                  |  7 ++++--
>>>>  gcc/testsuite/gcc.target/powerpc/pr103702.c | 24 +++++++++++++++++++++
>>>>  2 files changed, 29 insertions(+), 2 deletions(-)
>>>>  create mode 100644 gcc/testsuite/gcc.target/powerpc/pr103702.c
>>>>
>>>> diff --git a/gcc/config/rs6000/rs6000.c b/gcc/config/rs6000/rs6000.c
>>>> index 0b09713b2f5..37f07fe5358 100644
>>>> --- a/gcc/config/rs6000/rs6000.c
>>>> +++ b/gcc/config/rs6000/rs6000.c
>>>> @@ -5461,8 +5461,11 @@ rs6000_cost_data::update_target_cost_per_stmt (vect_cost_for_stmt kind,
>>>>      {
>>>>        tree vectype = STMT_VINFO_VECTYPE (stmt_info);
>>>>        unsigned int nunits = vect_nunits_for_cost (vectype);
>>>> -      /* We don't expect strided/elementwise loads for just 1 nunit.  */
>>>> -      gcc_assert (nunits > 1);
>>>> +      /* As PR103702 shows, it's possible that vectorizer wants to do
>>>> +         costings for only one unit here, it's no need to do any
>>>> +         penalization for it, so simply early return here.  */
>>>> +      if (nunits == 1)
>>>> +        return;
>>>>        /* i386 port adopts nunits * stmt_cost as the penalized cost
>>>>           for this kind of penalization, we used to follow it but
>>>>           found it could result in an unreliable body cost especially
>>>> diff --git a/gcc/testsuite/gcc.target/powerpc/pr103702.c b/gcc/testsuite/gcc.target/powerpc/pr103702.c
>>>> new file mode 100644
>>>> index 00000000000..585946fd64b
>>>> --- /dev/null
>>>> +++ b/gcc/testsuite/gcc.target/powerpc/pr103702.c
>>>> @@ -0,0 +1,24 @@
>>>> +/* We don't have one powerpc.*_ok for Power6, use altivec_ok conservatively.  */
>>>> +/* { dg-require-effective-target powerpc_altivec_ok } */
>>>> +/* { dg-options "-mdejagnu-cpu=power6 -O2 -ftree-loop-vectorize -fno-tree-scev-cprop" } */
>>>> +
>>>> +/* Verify there is no ICE.  */
>>>> +
>>>> +unsigned short a, e;
>>>> +int *b, *d;
>>>> +int c;
>>>> +extern int fn2 ();
>>>> +void
>>>> +fn1 ()
>>>> +{
>>>> +  void *f;
>>>> +  for (;;)
>>>> +    {
>>>> +      fn2 ();
>>>> +      b = f;
>>>> +      e = 0;
>>>> +      for (; e < a; ++e)
>>>> +    b[e] = d[e * c];
>>>> +    }
>>>> +}
>>>> +
>>>> --
>>>> 2.27.0
>>>>
>>