Let's make sure we're looking at the same example. The x86 instruction
that
I think we're discussing is cmpps/cmppd with constant predicate TRUE_UQ
(0xF), FALSE_OQ (0xB), etc.
The expected output for cases including NaN inputs is shown by the program
in PR28110:
https://bugs.llvm.org/show_bug.cgi?id=28110
Based on the Intel documentation and the experimental data, the output
value (ignoring exceptions) is never affected by the input value. It
doesn't matter if we have NaN input, we should still produce a constant
with -1 elements, so I don't think we need any relaxed-mode FP flags for
this transform.
Here's a simpler test program:
#include <immintrin.h>
#include <math.h>
#include <stdio.h>
#include <string.h>
__m128 get_all_ones(__m128 a, __m128 b) {
return _mm_cmp_ps(a, b, 15);
}
int main() {
int elts[4];
__m128 nan = (__m128) _mm_set1_ps(nanf(""));
memcpy(elts, &nan, 16);
printf("4 nans as hex: %x %x %x %x\n", elts[0], elts[1], elts[2],
elts[3]);
// The params to cmpps with pred #15 should not matter - it always
returns all ones.
// In this case one param is a nan.
__m128 ones = get_all_ones(nan, _mm_set1_ps(42.0));
memcpy(elts, &ones, 16);
printf("expecting 4 all ones: %x %x %x %x\n", elts[0], elts[1],
elts[2],
elts[3]);
// The params to cmpps with pred #15 should not matter - it always
returns all ones.
// In this case both params are nan.
__m128 more_ones = get_all_ones(nan, nan);
memcpy(elts, &more_ones, 16);
printf("expecting 4 all ones: %x %x %x %x\n", elts[0], elts[1],
elts[2],
elts[3]);
return 0;
}
$ clang cmptrue.c -mavx -O1
$ ./a.out
4 nans as hex: 7fc00000 7fc00000 7fc00000 7fc00000
expecting 4 all ones: ffffffff ffffffff ffffffff ffffffff
expecting 4 all ones: ffffffff ffffffff ffffffff ffffffff
On Wed, Jun 14, 2017 at 7:34 AM, Hal Finkel <hfinkel at anl.gov> wrote:
> Hi, Dinar,
>
> You should assume that FPU exceptions are not enabled. LLVM does not
> (currently) support them. That having been said, in this particular case,
> the output is also different if a NaN is produced (i.e. it produces a NaN
> and not a -1), and so I'd think you can only do this transformations
when
> the call has the nnan flag (e.g. because the code was compiled with
> -ffinite-math-only).
>
> -Hal
>
>
> On 06/14/2017 05:06 AM, Dinar Temirbulatov wrote:
>
>> Hi,
>> We are interesting in expanding some vector operations directly in the
>> IR form as constants https://reviews.llvm.org/D33406,
>> for example: _mm256_cmp_ps("any input", "any
input", _CMP_TRUE_UQ)
>> should produce -1, -1, -1, ... vector, but for some values for example
>> "1.00 -nan" if FPU exceptions were enabled this operation
triggers the
>> exception. Here is the question: Should we assume that FPENV was
>> initialized with FE_ALL_EXCEPT by default or we could rely for example
>> on "-fno-trapping-math" flag or we could completely ignore
the FPU
>> exception issue(see https://bugs.llvm.org/show_bug.cgi?id=6050)?
>>
>> Thanks, Dinar.
>>
>
> --
> Hal Finkel
> Lead, Compiler Technology and Programming Languages
> Leadership Computing Facility
> Argonne National Laboratory
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20170614/b75f6b26/attachment.html>