[Bug middle-end/47379] New: fwprop1 generates bad codes for x32

[Bug middle-end/47379] New: fwprop1 generates bad codes for x32

[hjl.tools at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=47379

           Summary: fwprop1 generates bad codes for x32
           Product: gcc
           Version: 4.6.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: middle-end
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: hjl.tools@gmail.com


On x32 branch, for

---
typedef unsigned IPos;
typedef unsigned char uch;
extern uch window[];
unsigned max_chain_length;
unsigned strstart;
int longest_match(IPos cur_match, int len, int best_len)
{
    unsigned chain_length = max_chain_length;
    register uch *scan = window + strstart;
    register uch *match;
    register uch scan_end1 = scan[best_len-1];
    register uch scan_end = scan[best_len];
    do {
        ;
        match = window + cur_match;
        if (match[best_len] != scan_end ||
            match[best_len-1] != scan_end1 ||
            *match != *scan ||
            *++match != scan[1]) continue;
            best_len = len;
    } while ( --chain_length != 0);
    return best_len;
}
---

compiled with -mx32 -O, fwprop1 pass turns:

(insn 14 13 15 2 (parallel [
            (set (reg/f:SI 86)
                (plus:SI (reg:SI 63 [ strstart.0 ])
                    (symbol_ref:SI ("window") [flags 0x40]  <var_decl
0x7f7a44198000 window>)))
            (clobber (reg:CC 17 flags))
        ]) x.i:9 253 {*addsi_1}
     (nil))

(insn 15 14 16 2 (set (reg/v/f:DI 64 [ scan ])
        (zero_extend:DI (reg/f:SI 86))) x.i:9 115 {*zero_extendsidi2_rex64}
     (nil))

(insn 16 15 17 2 (set (reg:SI 65 [ best_len.1 ])
        (reg/v:SI 85 [ best_len ])) x.i:11 64 {*movsi_internal}
     (nil))

(insn 17 16 18 2 (parallel [
            (set (reg:SI 89)
                (plus:SI (reg/v:SI 85 [ best_len ])
                    (subreg/s/u:SI (reg/v/f:DI 64 [ scan ]) 0)))
            (clobber (reg:CC 17 flags))
        ]) x.i:11 253 {*addsi_1}
     (nil))

into

(insn 14 13 15 2 (parallel [
            (set (reg/f:SI 86) 
                (plus:SI (reg:SI 63 [ strstart.0 ])
                    (symbol_ref:SI ("window") [flags 0x40]  <var_decl
0x7f7a44198000 window>)))
            (clobber (reg:CC 17 flags))
        ]) x.i:9 253 {*addsi_1}
     (expr_list:REG_UNUSED (reg:CC 17 flags)
        (nil)))

(insn 15 14 17 2 (set (reg/v/f:DI 64 [ scan ])
        (zero_extend:DI (reg/f:SI 86))) x.i:9 115 {*zero_extendsidi2_rex64}
     (expr_list:REG_DEAD (reg/f:SI 86) 
        (nil)))

(insn 17 15 18 2 (parallel [
            (set (reg:SI 89) 
                (plus:SI (reg/v:SI 85 [ best_len ])
                    (reg/f:SI 86)))
            (clobber (reg:CC 17 flags))
        ]) x.i:11 253 {*addsi_1}
     (expr_list:REG_UNUSED (reg:CC 17 flags)
        (nil)))

with

In insn 17, replacing
 (subreg/s/u:SI (reg/v/f:DI 64 [ scan ]) 0)
 with (reg/f:SI 86)
Changed insn 17
deferring rescan insn with uid = 17.
deferring rescan insn with uid = 17.

For x32, (subreg/s/u:SI (reg/v/f:DI 64 [ scan ]) 0) is
free while (reg/f:SI 86) is an extra register.

[Bug middle-end/47379] fwprop1 generates bad codes for x32

[pinskia at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=47379

Andrew Pinski <pinskia at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |pinskia at gcc dot gnu.org

--- Comment #1 from Andrew Pinski <pinskia at gcc dot gnu.org> 2011-01-20 19:36:25 UTC ---
I saw something like this on MIPS64-linux recently too where fwprop1 would prop
into the subreg when it should only do it if it was used once rather than one
than once.

[Bug middle-end/47379] fwprop1 generates bad codes for x32

[hjl.tools at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=47379

--- Comment #2 from H.J. Lu <hjl.tools at gmail dot com> 2011-01-20 23:28:32 UTC ---
When fwprop props subreg, it should check

1. If inner register is dead after use.
2. If subreg is used in multi-word operations.
3. If subreg is a paradoxical subreg.

[Bug middle-end/47379] fwprop1 generates bad codes for x32

[hjl.tools at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=47379

H.J. Lu <hjl.tools at gmail dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |amodra at gmail dot com,
                   |                            |ubizjak at gmail dot com

--- Comment #3 from H.J. Lu <hjl.tools at gmail dot com> 2011-01-26 15:39:05 UTC ---
This regression may be introduced by revision 151022:

http://gcc.gnu.org/ml/gcc-cvs/2009-08/msg00704.html

Propagate zero extended register is bad for x86-64.

[Bug rtl-optimization/47379] fwprop1 generates bad codes for x86-64

[pinskia at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=47379

Andrew Pinski <pinskia at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |WAITING
   Last reconfirmed|                            |2012-02-02
     Ever Confirmed|0                           |1

--- Comment #4 from Andrew Pinski <pinskia at gcc dot gnu.org> 2012-02-02 08:49:21 UTC ---
Can you try after:
2011-12-19  Richard Sandiford  <rdsandiford@googlemail.com>

        PR rtl-optimization/42839
        * fwprop.c (forward_propagate_subreg): Skip the SIGN/ZERO_EXTEND
        optimization if the source register is already extended.

[Bug rtl-optimization/47379] fwprop1 generates bad codes for x86-64

[hjl.tools at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=47379

--- Comment #5 from H.J. Lu <hjl.tools at gmail dot com> 2012-02-02 18:11:24 UTC ---
(In reply to comment #4)
> Can you try after:
> 2011-12-19  Richard Sandiford  <rdsandiford@googlemail.com>
> 
>         PR rtl-optimization/42839
>         * fwprop.c (forward_propagate_subreg): Skip the SIGN/ZERO_EXTEND
>         optimization if the source register is already extended.

GCC 4.7.0 20120131 still does it.