[Bug middle-end/33989] New: Extra load/store for float with union

[pinskia at gcc dot gnu dot org]

Testcase:
union a
{
  int i;
  float f;
};

void f(float *a, int *b, float e)
{
  union a c;
  c.f = *a + e;
  *b = c.i;
}

--- CUT ---
Currently we get (on x86):
        subl    $28, %esp
        movl    32(%esp), %eax
        flds    40(%esp)
        fadds   (%eax)
        movl    36(%esp), %eax
        fstps   12(%esp)  <--- extra store
        movl    12(%esp), %edx  <--- extra load
        movl    %edx, (%eax)  <--- store result to *b
        addl    $28, %esp
        ret
Or with -mfpmath=sse:
_f:
        subl    $28, %esp
        movl    32(%esp), %eax
        movss   40(%esp), %xmm0
        addss   (%eax), %xmm0
        movl    36(%esp), %eax
        movss   %xmm0, 12(%esp)  <--- extra store
        movl    12(%esp), %edx <--- extra load
        movl    %edx, (%eax)  <---store result to *b
        addl    $28, %esp
        ret

Or on PPC:
f:
        lfs 0,0(3)
        stwu 1,-16(1)
        fadds 0,1,0
        stfs 0,8(1)  <--- extra store
        lwz 0,8(1)  <--- extra load
        addi 1,1,16
        stw 0,0(4)  <--- store result to *b
        blr

The issue is that SFmode cannot be in integer registers.
The rtl looks like:
(insn 8 7 9 3 t1.c:10 (set (reg:SF 124)
        (mem:SF (reg/v/f:SI 120 [ a ]) [2 S4 A32])) -1 (nil))

(insn 9 8 10 3 t1.c:10 (set (reg:SF 123)
        (plus:SF (reg/v:SF 122 [ e ])
            (reg:SF 124))) -1 (nil))

(insn 10 9 11 3 t1.c:10 (set (subreg:SF (reg/v:SI 119 [ c ]) 0)
        (reg:SF 123)) -1 (nil))

(insn 11 10 16 3 t1.c:11 (set (mem:SI (reg/v/f:SI 121 [ b ]) [3 S4 A32])
        (reg/v:SI 119 [ c ])) -1 (nil))

See how we could translate register 119 (SImode) into a register that is in
SFmode and get away with it.


-- 
           Summary: Extra load/store for float with union
           Product: gcc
           Version: 4.3.0
            Status: UNCONFIRMED
          Keywords: missed-optimization
          Severity: enhancement
          Priority: P3
         Component: middle-end
        AssignedTo: unassigned at gcc dot gnu dot org
        ReportedBy: pinskia at gcc dot gnu dot org


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33989