[Bug rtl-optimization/81538] Optimization problem compiling op.c (Perl_custom_op_get_field) in perl 5.26

[danglin at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81538

--- Comment #8 from John David Anglin  ---
Created attachment 41897
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=41897=edit
miniperl debug session

The flow and optimization of the Perl_custom_op_get_field function is quite
complex.  Two switch statements are merged and it appears that ultimately the
wrong case is selected to load the return value.

On entry, register %r3 contains the value 0x5aedb0.  %r3 is not modified
in the routine but it is eventually used to load the return value.

   0x0001ab70 :   ldw 10(r3),ret0

It seems there is some confusion as to where the variable xop resides.  When
we hit the following statement, we have

14879   if(field == XOPe_xop_ptr) {(gdb) disass $pc-16,$pc+16
Dump of assembler code from 0x1aa28 to 0x1aa48:
   0x0001aa28 :   b,l 0x3b410,rp
   0x0001aa2c :   stw r0,0(r8)
   0x0001aa30 :   movb,<> ret0,r13,0x1aa74

   0x0001aa34 :   ldi 14,r25
=> 0x0001aa38 :   cmpib,= 0,r5,0x1ab94

   0x0001aa3c :   ldil L%186800,ret0
   0x0001aa40 :   ldo -1(r5),r5
   0x0001aa44 :   ldil L%1a800,ret0
End of assembler dump.
(gdb) info address xop
Symbol "xop" is multi-location:
  Range 0x1a988-0x1a9a4: a variable in $r3
  Range 0x1a9ac-0x1a9b8: a variable in $r3
  Range 0x1aa38-0x1aa40: a complex DWARF expression:
 0: DW_OP_addr 0x186a88
 5: DW_OP_stack_value

  Range 0x1aa44-0x1aa74: a variable in $r3
  Range 0x1aa84-0x1ab94: a variable in $r3
  Range 0x1ab94-0x1ab9c: a complex DWARF expression:
 0: DW_OP_addr 0x186a88
 5: DW_OP_stack_value

  Range 0x1ab9c-0x1abb4: a variable in $r3
.
(gdb) p/x $r3
$2 = 0x5aedb0
(gdb) p xop
$3 = (XOP *) 0x186a88 

The variable he is always 0 in this call.  So, nominally xop should point
at xop_null.

After we execute the switch statement, gdb thinks we are at line 14896:

(gdb) stepi
Perl_custom_op_get_field (my_perl=0x1e3008, o=0x5aee58, field=XOPe_xop_peep)
at op.c:14896
14896   break;
(gdb) disass $pc-16,$pc+16
Dump of assembler code from 0x1ab5c to 0x1ab7c:
   0x0001ab5c :   b,l 0x1a9b8
,r0
   0x0001ab60 :   ldw 8(r3),ret0
   0x0001ab64 :   b,l 0x1a9b8
,r0
   0x0001ab68 :   ldw 4(r3),ret0
=> 0x0001ab6c :   b,l 0x1a9b8
,r0
   0x0001ab70 :   ldw 10(r3),ret0
   0x0001ab74 :   b,l 0x1a9b8
,r0
   0x0001ab78 :   ldw c(r3),ret0

Line 14896 is in the first switch expression.  This should result in
xop->xop_peep being returned.  However, here xop is supposed in register
%r3.

However, the flags value from xop_null is 0.  So, the code should be
returning XOPd_xop_peep (ie., ((Perl_cpeep_t)0)).

[Bug rtl-optimization/81538] Optimization problem compiling op.c (Perl_custom_op_get_field) in perl 5.26

[danglin at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81538

--- Comment #7 from John David Anglin  ---
Created attachment 41895
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=41895=edit
Function Perl_custom_op_get_field

[Bug rtl-optimization/81538] Optimization problem compiling op.c (Perl_custom_op_get_field) in perl 5.26

[dave.anglin at bell dot net]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81538

--- Comment #6 from dave.anglin at bell dot net ---
On 2017-07-25, at 2:46 PM, dave.anglin at bell dot net wrote:

> Both gcc-5 and gcc-6 are broken.

Take that back; only gcc-6 is broken.  I incorrectly set the package up for
gcc-5
in my first attempt and it used gcc-6.

--
John David Anglin   dave.ang...@bell.net

[Bug rtl-optimization/81538] Optimization problem compiling op.c (Perl_custom_op_get_field) in perl 5.26

[dave.anglin at bell dot net]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81538

--- Comment #5 from dave.anglin at bell dot net ---
On 2017-07-25, at 10:02 AM, dave.anglin at bell dot net wrote:

> Will check gcc-5 when current build
> completes.

Both gcc-5 and gcc-6 are broken.

--
John David Anglin   dave.ang...@bell.net

[Bug rtl-optimization/81538] Optimization problem compiling op.c (Perl_custom_op_get_field) in perl 5.26

[dave.anglin at bell dot net]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81538

--- Comment #4 from dave.anglin at bell dot net ---
On 2017-07-25, at 3:46 AM, rguenth at gcc dot gnu.org wrote:

> Any other compiler versions to put in known-to-work/fail?

The fault doesn't occur with gcc-7:
gcc version 7.1.0 (Debian 7.1.0-9).  Will check gcc-5 when current build
completes.

--
John David Anglin   dave.ang...@bell.net

[Bug rtl-optimization/81538] Optimization problem compiling op.c (Perl_custom_op_get_field) in perl 5.26

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81538

--- Comment #3 from Richard Biener  ---
Any other compiler versions to put in known-to-work/fail?

[Bug rtl-optimization/81538] Optimization problem compiling op.c (Perl_custom_op_get_field) in perl 5.26

[glaubitz at physik dot fu-berlin.de]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81538

--- Comment #2 from John Paul Adrian Glaubitz  ---
Adding Kaz and Oleg to CC since it affects sh*-*-* as well [1].

> [1] https://bugs.debian.org/cgi-bin/bugreport.cgi?bug=869373

[Bug rtl-optimization/81538] Optimization problem compiling op.c (Perl_custom_op_get_field) in perl 5.26

[danglin at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81538

--- Comment #1 from John David Anglin  ---
Created attachment 41820
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=41820=edit
.s file

Assembler output for Perl_custom_op_get_field.