[Bug middle-end/32708] _mm_cvtsi64x_si128() and _mm_cvtsi128_si64x() inefficient

["ubizjak at gmail dot com" <gcc-bugzilla at gcc dot gnu dot org>]

------- Comment #4 from ubizjak at gmail dot com  2007-07-10 10:36 -------
(In reply to comment #0)

> long2vector() should use a simple MOVQ instruction the way int2vector() uses
> MOVD. It appears that the reason for the stack access is that the original code
> used a reg64->mem->mm->xmm path, which the optimizer partly noticed;
> gcc-4.3-20070617 leaves the full path in place.
> 
> Also, do the vector2<X>() functions really need to access the stack?

Stack access is the remain of partially implemented TARGET_INTER_UNIT_MOVES,
and this was fully fixed in 4.3. The fact that gcc-4.3 leaves full path in
place is due to missing pattern for vec_concat:V2DI (implemented in the patch
below) where 64bit registers can be concatenated with zero (for 64bit targets)
using movq instruction.

With attached patch, 4.3 generates:

long2vector:
.LFB4:
        movq    %rdi, %xmm0
        ret
.LFE4:

for -march=core2 (TARGET_INTER_UNIT_MOVES allowed)

and

long2vector:
.LFB4:
        movq    %rdi, -8(%rsp)
        movq    -8(%rsp), %xmm0
        ret
.LFE4:

for -march=k8 (no TARGET_INTER_UNIT_MOVES allowed).

> Finally, I've noticed several places where instructions involving 64-bit values
> use the "d/l" suffix (e.g. "long i = 0" ==> "xorl %eax, %eax"), or 32-bit
> operations that use 64-bit registers (e.g. "movd %xmm0, %rax" above). Are those
> generally features, bugs, or a "who cares?"

These are "who cares", produced by reload to satisfy register constraints (i.e.
certain alternatives missing as an attempt to solve INTER_UNIT_MOVES
requirements). They are harmles.

Index: sse.md
===================================================================
--- sse.md      (revision 126510)
+++ sse.md      (working copy)
@@ -4717,7 +4717,7 @@
        (vec_concat:V2DI
          (match_operand:DI 1 "nonimmediate_operand" "  m,*y ,0 ,0,0,m")
          (match_operand:DI 2 "vector_move_operand"  "  C,  C,Yt,x,m,0")))]
-  "TARGET_SSE"
+  "!TARGET_64BIT && TARGET_SSE"
   "@
    movq\t{%1, %0|%0, %1}
    movq2dq\t{%1, %0|%0, %1}
@@ -4728,6 +4728,23 @@
   [(set_attr "type" "ssemov,ssemov,sselog,ssemov,ssemov,ssemov")
    (set_attr "mode" "TI,TI,TI,V4SF,V2SF,V2SF")])

+(define_insn "*vec_concatv2di_rex"
+  [(set (match_operand:V2DI 0 "register_operand"     "=Yt,Yi,!Yt,Yt,x,x,x")
+       (vec_concat:V2DI
+         (match_operand:DI 1 "nonimmediate_operand" "  m,r ,*y ,0 ,0,0,m")
+         (match_operand:DI 2 "vector_move_operand"  "  C,C ,C  ,Yt,x,m,0")))]
+  "TARGET_64BIT"
+  "@
+   movq\t{%1, %0|%0, %1}
+   movq\t{%1, %0|%0, %1}
+   movq2dq\t{%1, %0|%0, %1}
+   punpcklqdq\t{%2, %0|%0, %2}
+   movlhps\t{%2, %0|%0, %2}
+   movhps\t{%2, %0|%0, %2}
+   movlps\t{%1, %0|%0, %1}"
+  [(set_attr "type" "ssemov,ssemov,ssemov,sselog,ssemov,ssemov,ssemov")
+   (set_attr "mode" "TI,TI,TI,TI,V4SF,V2SF,V2SF")])
+
 (define_expand "vec_setv2di"
   [(match_operand:V2DI 0 "register_operand" "")
    (match_operand:DI 1 "register_operand" "")


-- 

ubizjak at gmail dot com changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
         AssignedTo|unassigned at gcc dot gnu   |ubizjak at gmail dot com
                   |dot org                     |
             Status|NEW                         |ASSIGNED
   Last reconfirmed|2007-07-10 09:14:14         |2007-07-10 10:36:00
               date|                            |


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=32708