[Bug c/58405] New: Unoptimal code generated for computed goto

[Bug c/58405] New: Unoptimal code generated for computed goto

[semicontinuity at yandex dot ru]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=58405

            Bug ID: 58405
           Summary: Unoptimal code generated for computed goto
           Product: gcc
           Version: 4.7.2
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c
          Assignee: unassigned at gcc dot gnu.org
          Reporter: semicontinuity at yandex dot ru

The following code
//
-----------------------------------------------------------------------------
static void test(void) {
    char flag = 1;
    static void* address = &&L0;
    goto *address;

    for(;;) {

L0:
    asm volatile("nop\t\n");


    flag = 0;
L1: (void)&&L1;
    if (flag==0) { address = &&L1; return; }

    asm volatile("nop\t\n");

    flag = 0;
L2: (void)&&L2;
    if (flag==0) { address = &&L2; return; }

    }
}


int main(void) {

    for(;;) {
        test();
    }

    return 0;
}
//
-----------------------------------------------------------------------------

Compiles to:

//
-----------------------------------------------------------------------------
00000052 <test>:
static void test(void) {
  52:    cf 93           push    r28
  54:    df 93           push    r29
  56:    00 d0           rcall    .+0          ; 0x58 <test+0x6>
  58:    cd b7           in    r28, 0x3d    ; 61
  5a:    de b7           in    r29, 0x3e    ; 62
  5c:    80 91 60 00     lds    r24, 0x0060
  60:    90 91 61 00     lds    r25, 0x0061
  64:    8f 93           push    r24
  66:    9f 93           push    r25
  68:    08 95           ret
    goto *address;

    for(;;) {

L0:
    asm volatile("nop\t\n");
  6a:    00 00           nop


    flag = 0;
L1: (void)&&L1;
    if (flag==0) { address = &&L1; return; }
  6c:    81 e4           ldi    r24, 0x41    ; 65
  6e:    90 e0           ldi    r25, 0x00    ; 0
  70:    90 93 61 00     sts    0x0061, r25
  74:    80 93 60 00     sts    0x0060, r24
    flag = 0;
L2: (void)&&L2;
    if (flag==0) { address = &&L2; return; }

    }
}
  78:    0f 90           pop    r0
  7a:    0f 90           pop    r0
  7c:    df 91           pop    r29
  7e:    cf 91           pop    r28
  80:    08 95           ret
  ...
//
-----------------------------------------------------------------------------


This code works but is non optimal:
- It saves stack pointer to r28:r29, but later r28:r29 is never used
- Because it uses r28:r29, it saves r28:r29 to stack.
- It makes rcall .+0 and later removes 2 bytes from stack (pop r0) - this is
not necessary; rcall .+0 and pop r0's can be removed.

- Computed goto is implemented with 5 instructions, 7 words, with stack usage:
  5c:    80 91 60 00     lds    r24, 0x0060
  60:    90 91 61 00     lds    r25, 0x0061
  64:    8f 93           push    r24
  66:    9f 93           push    r25
  68:    08 95           ret

Instead, it could be implemented with

lds    r30, 0x0060
lds    r31, 0x0061
ijmp

with 3 instructions, 5 words, no additional memory access, but with clobbering
of r30:r31 instead of r24:r25 which is OK.

// ------------------------------------

Compiling C: main.c
avr-gcc -c -mmcu=atmega8535 -I. -gdwarf-2  -I../../bsp -I../../../.. -O3
-save-temps -v -funsigned-char -funsigned-bitfields -fpack-struct -fshort-enums
-Wall -Wstrict-prototypes -Wundef -Wa,-adhlns=.obj/main.lst  -std=gnu99 -Wundef
-MD -MP -MF .dep/main.o.d main.c -o .obj/main.o
Using built-in specs.
COLLECT_GCC=avr-gcc
Target: avr
Configured with: ../../gcc.gnu.org/gcc-4_7-branch/configure --target=avr
--prefix=/local/gnu/install/gcc-4.7-mingw32 --host=i386-mingw32
--build=i686-linux-gnu --enable-languages=c,c++ --disable-nls --disable-shared
--with-dwarf2 --with-avrlibc=yes
Thread model: single
gcc version 4.7.2 (GCC)

[Bug target/58405] Unoptimal code generated for computed goto

[gjl at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=58405

Georg-Johann Lay <gjl at gcc dot gnu.org> changed:

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

--- Comment #1 from Georg-Johann Lay <gjl at gcc dot gnu.org> ---
Created attachment 30822
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=30822&action=edit
Generated Assembly

I cannot reproduce this, see the attached assembly code generated with the same
4.7.2 build, same applies to 4.8.

[Bug target/58405] Unoptimal code generated for computed goto

[gjl at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=58405

Georg-Johann Lay <gjl at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |WAITING
   Last reconfirmed|                            |2013-09-14
     Ever confirmed|0                           |1

[Bug target/58405] Unoptimal code generated for computed goto

[gjl at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=58405

Georg-Johann Lay <gjl at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|WAITING                     |RESOLVED
         Resolution|---                         |WORKSFORME

--- Comment #3 from Georg-Johann Lay <gjl at gcc dot gnu.org> ---
(In reply to semicontinuity from comment #2)
> Now I cannot reproduce that as well..
> It seems that I've compiled it with -ffixed-r30 -ffixed-r31 (different from
> original intention) - in this case it produces this kind of assembly .

I can reproduce it with -ffixed-r30 -ffixed-r31.  But as I already wrote above,
without Z (R30/R31) you cannot no more write sensible programs -- except
completely trivial ones. I don't think it makes sense to put effort into
optimizing useless programs or for programs that very likely will break by
occupying Z.

> Still, the code could be more optimal, without rjmp, push/pop - just with
> ijmp replaced by 2 pushes + ret.

Without -ffixed*, the code /does/ use IJMP.  If you take away Z, it's not
possible to use IJMP, of course.