[Bug rtl-optimization/15422] New: fmod loop exposes non-efficient code generation in reg-stack.c

["uros at kss-loka dot si" <gcc-bugzilla at gcc dot gnu dot org>]

At the bottom of reg-stack.c file, there is a comment in convert_regs() function:

  /* ??? Future: process inner loops first, and give them arbitrary
     initial stacks which emit_swap_insn can modify.  This ought to
     prevent double fxch that often appears at the head of a loop.  */

This anomaly is exposed with fmod() instruction with -O2 -ffast-math flags. This
testcase:
double test1d(double x, double y)
{
  return fmod(x, y);
}

compiles to:

test1d:
        fldl 4(%esp)   <- parameter loads should be inverted in this case
        fldl 12(%esp)
        jmp .L2        <- jump
        .p2align 4,,7
.L6:
        fxch %st(1)    <- double fxch
.L2:
        fxch %st(1)    <- double fxch
        fprem
        fnstsw %ax
        sahf
        jp .L6
        fstp %st(1)
        ret
 
Optimum code would be something like this:
test1d:
        fldl 12(%esp)
        fldl 4(%esp)
        .p2align 4,,7
.L2:
        fprem
        fnstsw %ax
        sahf
        jp .L6
        fstp %st(1)
        ret
 
The problem can be seen in  _.c.35.stack dump in the edge 1->1 in BB 1:

Basic block 0
Input stack: empty
  insn 10 input stack: empty
  insn 11 input stack: [ 8 ]
Expected live registers [ 8 9 ]
Output stack: [ 8 9 ]
Edge 0->1: new block; copying stack position

Basic block 1
Input stack: [ 8 9 ]
  insn 13 input stack: [ 8 9 ]
Expected live registers [ 8 9 ]
Output stack: [ 9 8 ]
Edge 1->1: correcting stack to [ 8 9 ]
Edge 1->2: new block; pops needed

Basic block 2
Input stack: [ 8 ]
  insn 17 input stack: [ 8 ]
  insn 27 input stack: [ 8 ]
Expected live registers [ 8 ]
Output stack: [ 8 ]
Edge 2->-2: no changes needed
Edge 1->1 redirected to 3

-- 
           Summary: fmod loop exposes non-efficient code generation in reg-
                    stack.c
           Product: gcc
           Version: 3.5.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P2
         Component: rtl-optimization
        AssignedTo: unassigned at gcc dot gnu dot org
        ReportedBy: uros at kss-loka dot si
                CC: gcc-bugs at gcc dot gnu dot org
 GCC build triplet: i686-pc-linux-gnu
  GCC host triplet: i686-pc-linux-gnu
GCC target triplet: i686-pc-linux-gnu


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=15422