Re: [patch] ppc prologue scheduling/ -fPIC and -msched-prolog

[Clinton Popetz <cpopetz at cygnus dot com>]

On Wed, Feb 09, 2000 at 01:41:13AM +0100, Franz Sirl wrote:
> Am Mit, 09 Feb 2000 schrieb Clinton Popetz:

> >2000-02-08  Clinton Popetz  <cpopetz@cygnus.com>
> >
> >        * config/rs6000/rs6000-protos.h: (get_TOC_alias_set, uses_TOC,
> >	constant_pool_expr_p): Declare them.
> >        (constant_pool_expr_p): Declare it.
> >        * config/rs6000/rs6000.c (toc_label_name): Define.
> >	(rs6000_override_options): Set toc_label_name.
> >        (input_operand): Allow any TOC_RELATIVE_EXPR_P.
> >        (get_TOC_alias_set, constant_pool_expr_p, constant_pool_expr_1,
> >	uses_TOC): New functions.
> >        (print_operand): Delete old '*' case.
> >	(print_operand_address): Use LEGITIMATE_CONSTANT_POOL_ADDRESS_P,
> >        and strip off rtl for TOC before calling output_addr_const.
> >	(rs6000_emit_load_toc_table): Use toc_label_name.
> >        * config/rs6000/rs6000.h (TARGET_SWITCHES): Make msched-prolog
> >        the default.
> >        (LEGITIMATE_CONSTANT_POOL_BASE_P): Delete.
> >        (CONSTANT_POOL_EXPR_P, TOC_RELATIVE_EXPR_P): New macros.
> >        (LEGITIMATE_CONSTANT_POOL_ADDRESS_P): Use CONSTANT_POOL_EXPR_P.
> >        (LEGITIMIZE_ADDRESS, LEGITIMIZE_RELOAD_ADDRESS): Turn symbol_refs
> >        into explicit TOC_REGISTER offsets.
> >        (TOC_REGISTER): New macro.
> >        * config/rs6000/rs6000.md (movsi, movdi): Emit rtl under 
> >	TARGET_TOC to reference TOC_REGISTER.
> >        (load_toc_v4_PIC_1b): Add 4 to offset for toc reload.
> >        * config/rs6000/sysv4.h (ASM_DECLARE_FUNCTION_NAME): Call  
> >	uses_TOC before emitting label references.
> 
> Hi Clinton,
> 
> unfortunately, with --enable-checking=tree,gc,misc, bootstrap with your patch
> applied to todays backend with newppc-branch on powerpc-linux-gnu fails with:
> 
> ../../../cvsx/gccm/gcc/real.c:1566: Internal compiler error in `verify_flow_info', at flow.c:5914

This seems to be happening on x86 as well, so it is not my patch.  It looks
to be due to rth's direct_return work.

	http://gcc.gnu.org/ml/gcc-bugs/2000-02/msg00227.html

				-Clint