Re: How to do scan-tree-dump for test.o

[Richard Biener <richard dot guenther at gmail dot com>]

On Tue, Jun 6, 2017 at 3:39 PM, Tom de Vries <Tom_deVries@mentor.com> wrote:
> [ was: Re: [nvptx, PATCH, 3/3] Add v2di support ]
>
> On 06/06/2017 03:12 PM, Tom de Vries wrote:
>>
>> diff --git a/libgomp/testsuite/libgomp.oacc-c/vec.c
>> b/libgomp/testsuite/libgomp.oacc-c/vec.c
>> new file mode 100644
>> index 0000000..79c1c17
>> --- /dev/null
>> +++ b/libgomp/testsuite/libgomp.oacc-c/vec.c
>> @@ -0,0 +1,48 @@
>> +/* { dg-skip-if "" { *-*-* } { "-O0" } { "" } } */
>> +/* { dg-skip-if "" { *-*-* } { "*" } { "-DACC_MEM_SHARED=0" } } */
>> +/* { dg-additional-options "-std=c99 -ftree-slp-vectorize
>> -foffload=-ftree-slp-vectorize -foffload=-fdump-tree-slp1
>> -foffload=-save-temps -save-temps" } */
>> +
>> +#include <stdio.h>
>> +#include <sys/time.h>
>> +
>> +long long int p[32 *1000] __attribute__((aligned(16)));
>> +long long int p2[32 *1000] __attribute__((aligned(16)));
>> +
>> +int
>> +main (void)
>> +{
>> +#pragma acc parallel num_gangs(1) num_workers(1) vector_length(32)
>> +  {
>> +    if (((unsigned long int)p & (0xfULL)) != 0)
>> +      __builtin_abort ();
>> +    if (((unsigned long int)p2 & (0xfULL)) != 0)
>> +      __builtin_abort ();
>> +
>> +    for (unsigned int k = 0; k < 10000; k += 1)
>> +      {
>> +#pragma acc loop vector
>> +       for (unsigned long long int j = 0; j < 32; j += 1)
>> +         {
>> +           unsigned long long a, b;
>> +           unsigned long long *p3, *p4;
>> +           p3 = (unsigned long long *)((unsigned long long int)p &
>> (~0xfULL));
>> +           p4 = (unsigned long long *)((unsigned long long int)p2 &
>> (~0xfULL));
>> +
>> +           for (unsigned int i = 0; i < 1000; i += 2)
>> +             {
>> +               a = p3[j * 1000 + i];
>> +               b = p3[j * 1000 + i + 1];
>> +
>> +               p4[j * 1000 + i] = a;
>> +               p4[j * 1000 + i + 1] = b;
>> +             }
>> +         }
>> +      }
>> +  }
>> +
>> +  return 0;
>> +}
>> +
>> +/* Todo: make a scan-tree-dump variant that scans vec.o instead.  */
>> +/* { dg-final { file copy -force [glob vec.o.*] [regsub \.o\. [glob
>> vec.o.*] \.c\.] } } */
>> +/* { dg-final { scan-tree-dump "vector\\(2\\) long long unsigned int"
>> "slp1" } } */
>
>
> Hi,
>
> we have scan-tree-dump that scans in test.c.* files.  But when we run lto1
> for the offloaded region, we produce test.o.* files instead.  In the
> test-case above, I work around that by using 'dg-final { file copy }'. What
> is a good way to get rid of this workaround ?
>
> Add scan-o-tree-dump ?
>
> Or make the "slp1" field smarter, and allow f.i. "o.slp1" ?

There is the same issue with regular LTO tests using scan-tree-dump which
end up scanning the "fat" compilation dumpfile.  Maybe add
scan-ltrans-tree-dump and scan-wpa-ipa-dump that look at appropriate
files plus passing appropriate flags to generate dumpfiles in known locations
(I think part of them end up in /tmp).

Richard.

> Thanks,
> - Tom