[Bug target/40537] wrong instr. dependency with some SSE intrinsics

[ubizjak at gmail dot com]

--- Comment #9 from ubizjak at gmail dot com  2009-06-25 09:03 ---


*** This bug has been marked as a duplicate of 21920 ***


-- 

ubizjak at gmail dot com changed:

   What|Removed |Added

 Status|NEW |RESOLVED
 Resolution||DUPLICATE


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=40537

[Bug target/40537] wrong instr. dependency with some SSE intrinsics

[gael dot guennebaud at gmail dot com]

--- Comment #1 from gael dot guennebaud at gmail dot com  2009-06-24 08:07 
---
Created an attachment (id=18055)
 -- (http://gcc.gnu.org/bugzilla/attachment.cgi?id=18055action=view)
a complete example showing the problem

usage:

Works:
  g++ -O3 instr_dependency.cpp -o instr_dependency  ./instr_dependency

Fails:
  g++ -O3 instr_dependency.cpp -o instr_dependency -DFAST  ./instr_dependency


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=40537

[Bug target/40537] wrong instr. dependency with some SSE intrinsics

[rguenth at gcc dot gnu dot org]

--- Comment #2 from rguenth at gcc dot gnu dot org  2009-06-24 09:48 ---
I can't verify it (you do not provide complete source that can be compiled),
but I think this is a duplicate of PR40141 which is fixed in GCC 4.4.


-- 

rguenth at gcc dot gnu dot org changed:

   What|Removed |Added

 CC||rguenth at gcc dot gnu dot
   ||org


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=40537

[Bug target/40537] wrong instr. dependency with some SSE intrinsics

[gael dot guennebaud at gmail dot com]

--- Comment #3 from gael dot guennebaud at gmail dot com  2009-06-24 10:53 
---
There is a compilable example attached to comment #1.

Furthermore, I can reproduce the problem with gcc 4.1.3, 4.2.4, 4.3.2, and
4.4.0, so I don't think it is a duplicate of PR40141.


FYI, in the meantime I workaround the issue using inline assembly:

inline __m128 ploadu(const float* from)
{
  __m128 res; 
  asm(movsd  %[from0], %[r]
 : [r] =x (res) : [from0] m (*from), [dummy] m (*(from+1)) );
  asm(movhps %[from2], %[r]
 : [r] +x (res) : [from2] m (*(from+2)), [dummy] m (*(from+3)) );
  return res;
}

but that's not as portable as intrinsics.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=40537

[Bug target/40537] wrong instr. dependency with some SSE intrinsics

[gael dot guennebaud at gmail dot com]

--- Comment #4 from gael dot guennebaud at gmail dot com  2009-06-24 11:12 
---
some additional info:

- compiling with -fno-strict-aliasing fix the issue, so perhaps this is not a
real bug but a feature ?

- on the other hand using the may_alias type attribute for casting does not
help:

inline __m128 ploadu(const float*   from) { 
  typedef double __attribute__((may_alias)) doubleA;
  typedef __m128 __attribute__((may_alias)) __m128A;

  __m128 r;
  r = (__m128A)(_mm_load_sd( (const doubleA*)(from) ));
  r = _mm_loadh_pi(r, (const __m64*)(from+2));
  return r;
}

I guess this is because the const doubleA* is then casted to a const
double* when calling the _mm_load_sd() function.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=40537

[Bug target/40537] wrong instr. dependency with some SSE intrinsics

[rguenth at gcc dot gnu dot org]

--- Comment #5 from rguenth at gcc dot gnu dot org  2009-06-24 11:23 ---
There are aliasing issues with your code / the intrinsics implementation:

  float data[4] = {1, 2, 3, 4};
...
  r = _mm_castpd_ps(_mm_load_sd((double*)(from)));

ends up loading from float data via a pointer to double.  That is invalid
unless the intrinsic specifies it should work ok in which case it better
had implemented counter-measures here:

  data[0] ={v} 1.0e+0;
  data[1] ={v} 2.0e+0;
  data[2] ={v} 3.0e+0;
  data[3] ={v} 4.0e+0;
  from.92_14 = (const double *) data[0];
  D.25303_15 = *from.92_14;
  D.25304_16 = {D.25303_15, 0.0};
  r_17 = VIEW_CONVERT_EXPR__m128(D.25304_16);

I would suggest using

  typedef double __attribute__((may_alias)) aliased_double;
  r = _mm_castpd_ps(_mm_set_sd (*(aliased_double *)from));

instead.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=40537

[Bug target/40537] wrong instr. dependency with some SSE intrinsics

[ubizjak at gmail dot com]

--- Comment #7 from ubizjak at gmail dot com  2009-06-24 11:57 ---
Adding x86 intrinsic expert...


-- 

ubizjak at gmail dot com changed:

   What|Removed |Added

 CC||hjl dot tools at gmail dot
   ||com
 Status|UNCONFIRMED |NEW
 Ever Confirmed|0   |1
   Last reconfirmed|-00-00 00:00:00 |2009-06-24 11:57:04
   date||


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=40537

[Bug target/40537] wrong instr. dependency with some SSE intrinsics

[hjl dot tools at gmail dot com]

--- Comment #8 from hjl dot tools at gmail dot com  2009-06-24 13:09 ---
(In reply to comment #6)
 _mm_load_sd( (const doubleA*)(from) )
 
 does not work because the prototype of _mm_load_sd does not have a type
 with the may-alias attribute, so it gets stripped again.
 

I think movsd/movhpd intrinsics with union of double/float or
movlps/movhps intrinsics should be used.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=40537