[Bug middle-end/70801] New: IRA caller-saves does not support rematerialization

["wdijkstr at arm dot com" <gcc-bugzilla at gcc dot gnu dot org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=70801

            Bug ID: 70801
           Summary: IRA caller-saves does not support rematerialization
           Product: gcc
           Version: 7.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: middle-end
          Assignee: unassigned at gcc dot gnu.org
          Reporter: wdijkstr at arm dot com
  Target Milestone: ---

GCC emits the same code for caller-saves in all cases, even if the caller-save
is an immediate which can be trivially rematerialized. The caller-save code
should support rematerialization directly by omitting saves and emitting the
init value for restores. This allows the cost of rematerializeable caller-saves
to be lowered and avoids a stack slot. 

Alternatively given the spill code already supports rematerialization, it may
be possible to stop immediates from being caller-saved by increasing their
caller-save cost significantly.

void g(void);
float bad_remat(float x)
{
  x += 3.0f;
  g();
  x *= 3.0f;
  x *= (3.0f + x);
  g();
  x *= 3.0f;
  x *= (3.0f + x);
  return x;
}

AArch64 gives with -O2 -fomit-frame-pointer -ffixed-d8 -ffixed-d9 -ffixed-d10
-ffixed-d11 -ffixed-d12 -ffixed-d13 -ffixed-d14 -ffixed-d15:

        fmov    s1, 3.0e+0
        str     x30, [sp, -32]!
        fadd    s0, s0, s1
        stp     s0, s1, [sp, 24]
        bl      g
        ldp     s0, s1, [sp, 24]
        fmul    s0, s0, s1
        fadd    s2, s0, s1
        fmul    s0, s0, s2
        str     s0, [sp, 24]
        bl      g
        ldp     s0, s1, [sp, 24]
        ldr     x30, [sp], 32
        fmul    s0, s0, s1
        fadd    s1, s0, s1
        fmul    s0, s0, s1
        ret


x86_64 with plain -O2:
        subq    $24, %rsp
        movss   .LC0(%rip), %xmm1
        addss   %xmm1, %xmm0
        movss   %xmm1, 12(%rsp)
        movss   %xmm0, 8(%rsp)
        call    g
        movss   8(%rsp), %xmm0
        movaps  %xmm0, %xmm2
        movss   12(%rsp), %xmm1
        mulss   %xmm1, %xmm2
        movaps  %xmm2, %xmm0
        addss   %xmm1, %xmm0
        mulss   %xmm2, %xmm0
        movss   %xmm0, 8(%rsp)
        call    g
        movss   12(%rsp), %xmm1
        movss   8(%rsp), %xmm0
        addq    $24, %rsp
        mulss   %xmm1, %xmm0
        addss   %xmm0, %xmm1
        mulss   %xmm1, %xmm0
        ret