Omitted conversion float -> double in x86 back-end

[Richard Guenther]

On 11/21/06, Roberto COSTA <roberto.costa@st.com> wrote:
> Hello,
> I've just noticed that, in a sequence of conversions signed long long ->
> float -> double, x86 back-end eats the intermediate conversion away.
> In this case, the conversion to float isn't redundant because it lowers
> the precision of the computation (which is something a programmer may
> want to do).
> It looks to me this this is a bug; I'm not completely sure because of my
> ignorance of what C standards state about it.
> Before filing a report, I'd like to hear other opinions.
>
> Here is a simple example:
>
> ------------- float_conv.c ------------
> double foo(signed long long l)
> {
>      return (float)l;
> }
> ---------------------------------------
>
> Files float_conv.c.026t.fixupcfg and float_conv.s are the outcome of the
> following compilation:
>      gcc -O0 float_conv.c -S -fdump-tree-all
>
> ------ float_conv.c.026t.fixupcfg -----
> ;; Function foo (foo)
>
> foo (l)
> {
>    float D.1524;
>    double D.1523;
>
> <bb 2>:
>    D.1524 = (float) l;
>    D.1523 = (double) D.1524;
>    return D.1523;
> }
> ---------------------------------------
>
> ------------- float_conv.s ------------
>          .file   "float_conv.c"
>          .text
> .globl foo
>          .type   foo, @function
> foo:
>          pushl   %ebp
>          movl    %esp, %ebp
>          subl    $8, %esp
>          movl    8(%ebp), %eax
>          movl    %eax, -8(%ebp)
>          movl    12(%ebp), %eax
>          movl    %eax, -4(%ebp)
>          fildll  -8(%ebp)
>          leave
>          ret
>          .size   foo, .-foo
>          .ident  "GCC: (GNU) 4.2.0 20060826 (experimental)"
>          .section        .note.GNU-stack,"",@progbits
> ---------------------------------------
>
> The conversion is still there at the end of GIMPLE passes, which leads
> me to the conclusion that it is removed in RTL compilation.

Yes, with -O2 after expand we have

;; return (double) (float) l
(insn 10 9 11 (set (reg:SF 62)
        (float:SF (reg/v:DI 60 [ l ]))) -1 (nil)
    (nil))

(insn 11 10 12 (set (reg:DF 61)
        (float_extend:DF (reg:SF 62))) -1 (nil)
    (nil))

(insn 12 11 13 (set (reg:DF 59 [ <result> ])
        (reg:DF 61)) -1 (nil)
    (nil))

which is ok.  Even after combine it's ok:

(insn 10 7 15 2 (set (reg:SF 62)
        (float:SF (mem/c/i:DI (reg/f:SI 16 argp) [3 l+0 S8 A32]))) 172
{*floatdisf2_i387} (nil)
    (nil))

(note 15 10 18 2 NOTE_INSN_FUNCTION_END)

(insn 18 15 24 2 (set (reg/i:DF 8 st [ <result> ])
        (float_extend:DF (reg:SF 62))) 125 {*extendsfdf2_i387}
(insn_list:REG_DEP_TRUE 10 (nil))
    (expr_list:REG_DEAD (reg:SF 62)
        (nil)))

(insn 24 18 0 2 (use (reg/i:DF 8 st [ <result> ])) -1
(insn_list:REG_DEP_TRUE 18 (nil))
    (nil))

but then regstack comes along and

(insn:TI 10 32 33 2 (set (reg:SF 8 st)
        (float:SF (mem/c/i:DI (plus:SI (reg/f:SI 6 bp)
                    (const_int 8 [0x8])) [3 l+0 S8 A32]))) 172
{*floatdisf2_i387} (nil)
    (nil))

(note 33 10 34 2 NOTE_INSN_EPILOGUE_BEG)

(insn:TI 34 33 24 2 (parallel [
            (set (reg/f:SI 6 bp)
                (mem:SI (reg/f:SI 7 sp) [0 S4 A8]))
            (set (reg/f:SI 7 sp)
                (plus:SI (reg/f:SI 7 sp)
                    (const_int 4 [0x4])))
        ]) 37 {popsi1} (nil)
    (nil))

(insn 24 34 35 2 (use (reg/i:DF 8 st [ <result> ])) -1 (nil)
    (nil))

where the *floatdisf2_i387 pattern misses the truncation to SFmode.

(you should file a bugzilla about this)

Richard.