patch that FIXED regression PR opt/7687 (still broken in 3.2 branch)

patch that FIXED regression PR opt/7687 (still broken in 3.2 branch)

[Janis Johnson]

The regression reported in PR optimization/7687, which is still broken
on the 3.2 branch, was fixed on the mainline with this patch.  Given
the size of the patch, it's probably unlikely that this information
will be very useful.

2002-10-22  Richard Henderson  <rth@redhat.com>

	* target.h (gcc_target.asm_out): Merge output_mi_thunk and
	output_mi_vcall_thunk into a single hook.  Add can_output_mi_thunk.
	* target-def.h (TARGET_ASM_OUTPUT_MI_THUNK): Don't conditionalize.
	(TARGET_ASM_OUTPUT_MI_VCALL_THUNK): Remove.
	(TARGET_ASM_CAN_OUTPUT_MI_THUNK): New.
	(TARGET_ASM_OUT): Update.
	* hooks.c (hook_bool_tree_hwi_hwi_tree_false): New.
	(hook_bool_tree_hwi_hwi_tree_true): New.
	(default_can_output_mi_thunk_no_vcall): New.
	* hooks.h: Declare them.
	* system.h (ASM_OUTPUT_MI_THUNK): Poison.

	* config/alpha/alpha.c (TARGET_ASM_CAN_OUTPUT_MI_THUNK): New.
	(alpha_output_mi_thunk_osf): Add VCALL_OFFSET parameter.
	* config/arm/arm.c, config/cris/cris.c, config/frv/frv.c,
	config/i960/i960.c, config/ia64/ia64.c, config/m68k/m68k.c,
	config/mmix/mmix.c, config/pa/pa.c, config/sparc/sparc.c,
	config/stormy16/stormy16.c: Similarly.

	* config/i386/i386.c (x86_output_mi_thunk): Merge vcall_offset code.
	Handle 64-bit properly.  Streamline.
	(x86_output_mi_vcall_thunk): Remove.
	(x86_this_parameter): Rename from ia32_this_parameter; handle 64-bit.
	(x86_can_output_mi_thunk): New.
	(TARGET_ASM_OUTPUT_MI_VCALL_THUNK): Remove.
	(TARGET_ASM_CAN_OUTPUT_MI_THUNK): New.
	(override_options): Don't zap targetm.asm_out.output_mi_vcall_thunk.

	* config/rs6000/rs6000.c (rs6000_output_mi_thunk): Rename from
	output_mi_thunk; make static; always use function_section.
	(TARGET_ASM_OUTPUT_MI_THUNK): New.
	(TARGET_ASM_CAN_OUTPUT_MI_THUNK): New.
	(rs6000_ra_ever_killed): Test no_new_pseudos not
	targetm.asm_out.output_mi_thunk in conjunction with thunks.
	* config/rs6000/rs6000-protos.h: Update.
	* config/rs6000/sysv4.h (TARGET_ASM_OUTPUT_MI_THUNK): Remove.
	* config/rs6000/xcoff.h (ASM_DECLARE_FUNCTION_NAME): Don't call
	xcoffout_declare_function when using rs6000_output_mi_thunk.

	* config/s390/s390.c (s390_output_mi_thunk): Rename from
	s390_output_mi_vcall_thunk.
	(TARGET_ASM_OUTPUT_MI_THUNK): Remove.
	(TARGET_ASM_CAN_OUTPUT_MI_THUNK): New.

	* config/vax/vax.c (vax_output_mi_thunk): Static; add vcall_offset.
	(TARGET_ASM_OUTPUT_MI_THUNK, TARGET_ASM_CAN_OUTPUT_MI_THUNK): New.
	* config/vax/vax-protos.h: Update.
	* config/vax/vax.h (ASM_OUTPUT_MI_THUNK): Remove.

2002-10-22  Richard Henderson  <rth@redhat.com>

	* method.c (use_thunk): Always compute vcall_value; assert that
	it is not zero.  Use can_output_mi_thunk; use output_mi_thunk
	for vcall thunks as well.

Here's a small test case that causes the 3.2 branch compiler to ICE
on i686-linux:

-------------------
class A {
public:
  virtual void f1(); 
}; 
class B {
public:
  virtual void f2(); 
}; 
class C : public A, public B 
{
public:
  virtual void f1();
  virtual void f2(); 
};

void C::f2() {}
-------------------

Output from the the 3.2 branch compiler with -fPIC -mregparm=3:

7687.C: In member function `void C::_ZThn4_N1C2f2Ev()':
7687.C:16: Internal compiler error in x86_output_mi_thunk, at
   config/i386/i386.c:12655
Please submit a full bug report,
with preprocessed source if appropriate.
See <URL:http://www.gnu.org/software/gcc/bugs.html> for instructions.

I'll add this information to the PR.

Janis

Re: patch that FIXED regression PR opt/7687 (still broken in 3.2 branch)

[Richard Henderson]

On Fri, Dec 20, 2002 at 03:07:16PM -0800, Janis Johnson wrote:
> The regression reported in PR optimization/7687, which is still broken
> on the 3.2 branch, was fixed on the mainline with this patch.  Given
> the size of the patch, it's probably unlikely that this information
> will be very useful.

The old code simply didn't handle the case in question.

          if (ix86_regparm > 2)
            abort ();

There's really no point in keeping this PR open.  Most
of the infrastructure in the quoted patch is necessary
to avoid this abort.


r~