[Bug target/40499] [missed optimization] branch to return not threaded on thumb

[Bug target/40499] [missed optimization] branch to return not threaded on thumb

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=40499

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |cptarse-luke at yahoo dot com

--- Comment #7 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
*** Bug 111580 has been marked as a duplicate of this bug. ***

[Bug target/40499] [missed optimization] branch to return not threaded on thumb

[ramana at gcc dot gnu dot org]

-- 

ramana at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Severity|normal                      |enhancement
           Keywords|                            |missed-optimization
      Known to fail|                            |4.5.0


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=40499

[Bug target/40499] [missed optimization] branch to return not threaded on thumb

[rearnsha at gcc dot gnu dot org]

------- Comment #6 from rearnsha at gcc dot gnu dot org  2009-06-22 12:37 -------
Support for single-instruction return insns has been around a lot longer than
rtl-based epilogues, so there's no need to convert the thumb target to RTL
epilogues as a pre-requisite for fixing this.  Someone simply needs to sit down
and work out what the code needs to do in order to support this feature; and
then implement it...


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=40499

[Bug target/40499] [missed optimization] branch to return not threaded on thumb

[steven at gcc dot gnu dot org]

------- Comment #5 from steven at gcc dot gnu dot org  2009-06-22 08:06 -------
The only way to really fix this, is to make an RTL epilogue for thumb. Turning
to Richard Earnshaw again for advice...


-- 

steven at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |rearnsha at gcc dot gnu dot
                   |                            |org


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=40499

[Bug target/40499] [missed optimization] branch to return not threaded on thumb

[carrot at google dot com]

------- Comment #4 from carrot at google dot com  2009-06-22 08:00 -------
Sorry I didn't make it clear. It is a performance bug, not a code size issue.
If the epilogue is a simple return instruction, the branch to return can be
replaced by the return instruction. So we can execute one less instruction at
run time without any code size penalty.

It looks the code at function.c:5078 can't be applied to thumb.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=40499

[Bug target/40499] [missed optimization] branch to return not threaded on thumb

[steven at gcc dot gnu dot org]

------- Comment #3 from steven at gcc dot gnu dot org  2009-06-20 07:23 -------
Output from ./cc1 -march=armv5te -mthumb -Os PR40499.c -dAP:

        .file   "PR40499.c"
        .text
        .align  1
        .global dual_feasible
        .code   16
        .thumb_func
        .type   dual_feasible, %function
dual_feasible:
        @ basic block 2
@(insn 42 37 43 PR40499.c:11 (set (reg/v/f:SI 2 r2 [orig:169 stop ] [169])
@        (mem/s/f:SI (plus:SI (reg/v/f:SI 0 r0 [orig:172 net ] [172])
@                (const_int 4 [0x4])) [3 <variable>.stop_arcs+0 S4 A32])) 168
{*thumb1_movsi_insn} (nil))
@ 0x0000
        ldr     r2, [r0, #4]    @ 42    *thumb1_movsi_insn/7    [length = 2]
@(insn 43 42 108 PR40499.c:13 (set (reg/v/f:SI 3 r3 [orig:167 arc ] [167])
@        (mem/s/f:SI (reg/v/f:SI 0 r0 [orig:172 net ] [172]) [3
<variable>.arcs+0 S4 A32])) 168 {*thumb1_movsi_insn} (expr_list:REG_DEAD
(reg/v/f:SI 0 r0 [orig:172 net ] [172])
@        (nil)))
@ 0x0002
        ldr     r3, [r0]        @ 43    *thumb1_movsi_insn/7    [length = 2]
@(jump_insn 108 43 109 PR40499.c:13 (set (pc)
@        (label_ref 51)) 243 {*thumb_jump} (nil))
@ 0x0004
        b       .L2     @ 108   *thumb_jump     [length = 2]
.L4:
        @ basic block 3
@(insn 47 46 48 PR40499.c:15 (set (reg:SI 1 r1 [173])
@        (mem:SI (reg/v/f:SI 3 r3 [orig:167 arc ] [167]) [4 S4 A32])) 168
{*thumb1_movsi_insn} (expr_list:REG_EQUIV (mem:SI (reg/v/f:SI 3 r3 [orig:167
arc ] [167]) [4 S4 A32])
@        (nil)))
@ 0x0006
        ldr     r1, [r3]        @ 47    *thumb1_movsi_insn/7    [length = 2]
@(jump_insn 48 47 49 PR40499.c:15 (set (pc)
@        (if_then_else (le (reg:SI 1 r1 [173])
@                (const_int 7 [0x7]))
@            (label_ref:SI 70)
@            (pc))) 200 {*cbranchsi4_insn} (expr_list:REG_DEAD (reg:SI 1 r1
[173])
@        (expr_list:REG_BR_PROB (const_int 450 [0x1c2])
@            (nil))))
@ 0x0008
        cmp     r1, #7  @ 48    *cbranchsi4_insn/1      [length = 4]
        ble     .L5
        @ basic block 4
@(insn 50 49 51 PR40499.c:13 (set (reg/v/f:SI 3 r3 [orig:167 arc ] [167])
@        (plus:SI (reg/v/f:SI 3 r3 [orig:167 arc ] [167])
@            (const_int 4 [0x4]))) 5 {*thumb1_addsi3} (nil))
@ 0x000c
        add     r3, r3, #4      @ 50    *thumb1_addsi3/1        [length = 2]
.L2:
        @ basic block 5
@(jump_insn 54 52 66 PR40499.c:13 (set (pc)
@        (if_then_else (ltu (reg/v/f:SI 3 r3 [orig:167 arc ] [167])
@                (reg/v/f:SI 2 r2 [orig:169 stop ] [169]))
@            (label_ref 53)
@            (pc))) 200 {*cbranchsi4_insn} (expr_list:REG_BR_PROB (const_int
9550 [0x254e])
@        (nil)))
@ 0x000e
        cmp     r3, r2  @ 54    *cbranchsi4_insn/1      [length = 4]
        bcc     .L4
        @ basic block 6
@(insn 38 66 110 PR40499.c:13 (set (reg:SI 0 r0 [orig:168 D.1218 ] [168])
@        (const_int 0 [0x0])) 168 {*thumb1_movsi_insn} (expr_list:REG_EQUAL
(const_int 0 [0x0])
@        (nil)))
@ 0x0012
        mov     r0, #0  @ 38    *thumb1_movsi_insn/2    [length = 2]
@(jump_insn 110 38 111 (set (pc)
@        (label_ref 55)) 243 {*thumb_jump} (nil))
@ 0x0014
        b       .L3     @ 110   *thumb_jump     [length = 2]
.L5:
        @ basic block 7
@(insn 39 69 55 PR40499.c:15 (set (reg:SI 0 r0 [orig:168 D.1218 ] [168])
@        (const_int 1 [0x1])) 168 {*thumb1_movsi_insn} (expr_list:REG_EQUAL
(const_int 1 [0x1])
@        (nil)))
@ 0x0016
        mov     r0, #1  @ 39    *thumb1_movsi_insn/2    [length = 2]
.L3:
        @ basic block 8
@(insn 98 97 99 PR40499.c:22 (unspec:SI [
@            (reg/f:SI 13 sp)
@        ] 6) 341 {prologue_use} (nil))
@ 0x0018
        @ sp needed for prologue        @ 98    prologue_use    [length = 4]
@(jump_insn 100 99 101 PR40499.c:22 (unspec_volatile [
@            (return)
@        ] 1) 323 {*epilogue_insns} (nil))
@ 0x001c
        bx      lr
        .size   dual_feasible, .-dual_feasible
        .ident  "GCC: (GNU) 4.5.0 20090614 (experimental) [trunk revision
148472]"



There are several interesting things going on here:

1. Note how "prologue_use" has length 4, that's probably wrong (should be
length 0?).

2. Note "epilogue_insns".  For thumb this pattern outputs instructions directly
to the cc1 output (i.e. the .s file), so these insns are never seen by
thread_prologue_and_epilogue_insns. Basically the thumb1 epilogue_insns is an
old-style text epilogue_insns instead of an RTL epilogue.


-- 

steven at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|WAITING                     |NEW
     Ever Confirmed|0                           |1
   GCC host triplet|i686-linux                  |
   Last reconfirmed|0000-00-00 00:00:00         |2009-06-20 07:23:38
               date|                            |
            Summary|[missed optimization] branch|[missed optimization] branch
                   |to return                   |to return not threaded on
                   |                            |thumb


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=40499