[llvm-bugs] [Bug 40265] autovectorization of repeated calls to vectorizable functions fails

[via llvm-bugs]

https://bugs.llvm.org/show_bug.cgi?id=40265

Kay F. Jahnke <kfjahnke at gmail.com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
         Resolution|FIXED                       |---
             Status|RESOLVED                    |REOPENED

--- Comment #8 from Kay F. Jahnke <kfjahnke at gmail.com> ---
I'd like to raise the topic again. Using the compiler flags which resulted in
autovectorization of my example loop running sqrt, I tried a few other standard
functions, notably the trigonometric ones, log and exp. Contrary to the
documentation, I could see no indication of these functions being
autovectorized on my AVX2 machine. Instead the maths is done by issuing a callq
to the respective standard function (like, callq sinf). I assume that these
functions are not autovectorized for my system because there are no
corresponding assembler instructions - the equivalent of what vsqrtps and
vsqrtpd are in the sqrt case - and vectorizing these functions for intel ISA
requires additional code. So I'd like to find out whether I am again missing
something or whether autovectorization of these functions simply does not
happen for SSE/AVX/AVX2, in which case I'd propose to give yet another hint in
the documentation. Here's the example code I used calling sin repeatedly:

#include <cmath>

extern float data [ 32768 ] ;

extern void vf1()
{
#pragma clang loop vectorize(enable)
  for ( int i = 0 ; i < 32768 ; i++ )
    data [ i ] = std::sin ( data [ i ] ) ;
}

I compiled like this:

$ clang++ -fvectorize  -Rpass=loop-vectorize -Rpass-analysis=loop-vectorize
-fno-math-errno -std=c++11 -O3 -mavx2 -S -o xx.s xx.cc
xx.cc:8:3: remark: the cost-model indicates that interleaving is not beneficial
[-Rpass-analysis=loop-vectorize]                                                
  for ( int i = 0 ; i < 32768 ; i++ )                                           
  ^                                                                             
xx.cc:8:3: remark: vectorized loop (vectorization width: 8, interleaved count:
1) [-Rpass=loop-vectorize]

using this clang++ version:

clang version 6.0.0-1ubuntu2 (tags/RELEASE_600/final)                           
Target: x86_64-pc-linux-gnu

Kay

-- 
You are receiving this mail because:
You are on the CC list for the bug.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-bugs/attachments/20190208/77a4fb70/attachment-0001.html>