OpenE2K/gcc
2
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

pa-linux.h (TARGET_HAS_STUBS_AND_ELF_SECTIONS): Delete define.

Browse Source 
* pa-linux.h (TARGET_HAS_STUBS_AND_ELF_SECTIONS): Delete define.
	* pa32-linux.h (FUNCTION_OK_FOR_SIBCALL): Delete define.
	* pa.c (pa_function_ok_for_sibcall): Allow non indirect sibcalls on
	TARGET_ELF32.  Add comment on sibcall issues for TARGET_64BIT.

From-SVN: r60377
This commit is contained in:
John David Anglin 2002-12-20 23:24:39 +00:00 committed by John David Anglin
parent f9d2de4d0d
commit 11f4312758
4 changed files with 26 additions and 20 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

7
gcc/ChangeLog
View File

@ -1,3 +1,10 @@
2002-12-20 John David Anglin <dave.anglin@nrc.gc.ca>
* pa-linux.h (TARGET_HAS_STUBS_AND_ELF_SECTIONS): Delete define.
* pa32-linux.h (FUNCTION_OK_FOR_SIBCALL): Delete define.
* pa.c (pa_function_ok_for_sibcall): Allow non indirect sibcalls on
TARGET_ELF32. Add comment on sibcall issues for TARGET_64BIT.
2002-12-20 Kazu Hirata <kazu@cs.umass.edu>
* config/h8300/h8300-protos.h: Add prototypes for

5
gcc/config/pa/pa-linux.h
View File

@ -167,11 +167,6 @@ Boston, MA 02111-1307, USA. */
/* Globalizing directive for a label. */
#define GLOBAL_ASM_OP ".globl "
/* This definition is used inside pa.c to disable all
sibcall optimization, because sibcalls, stubs and
elf sections don't play well. */
#define TARGET_HAS_STUBS_AND_ELF_SECTIONS 1
/* FIXME: Hacked from the <elfos.h> one so that we avoid multiple
labels in a function declaration (since pa.c seems determined to do
it differently) */

26
gcc/config/pa/pa.c
View File

@ -6878,14 +6878,26 @@ pa_function_ok_for_sibcall (decl, exp)
tree decl;
tree exp ATTRIBUTE_UNUSED;
{
#ifdef TARGET_HAS_STUBS_AND_ELF_SECTIONS
/* Sibcalls, stubs, and elf sections don't play well. */
return false;
#endif
/* Sibcalls are ok for TARGET_ELF32 as along as the linker is used in
single subspace mode and the call is not indirect. As far as I know,
there is no operating system support for the multiple subspace mode.
It might be possible to support indirect calls if we didn't use
$$dyncall (see the indirect sequence generated in output_call). */
if (TARGET_ELF32)
return (decl != NULL_TREE);
/* Sibcalls are not ok because the arg pointer register is not a fixed
register. This prevents the sibcall optimization from occuring. In
addition, there are problems with stub placement using GNU ld. This
is because a normal sibcall branch uses a 17-bit relocation while
a regular call branch uses a 22-bit relocation. As a result, more
care needs to be taken in the placement of long-branch stubs. */
if (TARGET_64BIT)
return false;
return (decl
&& ! TARGET_PORTABLE_RUNTIME
&& ! TARGET_64BIT
&& ! TREE_PUBLIC (decl));
&& !TARGET_PORTABLE_RUNTIME
&& !TREE_PUBLIC (decl));
}
/* Returns 1 if the 6 operands specified in OPERANDS are suitable for

8
gcc/config/pa/pa32-linux.h
View File

@ -22,14 +22,6 @@ Boston, MA 02111-1307, USA. */
#undef TARGET_ELF32
#define TARGET_ELF32 1
/* Sibcalls are ok when ld is used in single subspace mode. The
multiple subspace mode is not compatible with sibcalls to external
functions because the linker generated stubs store the return
pointer into the frame. This target does not need multiple
subspace stubs, so we allow sibcalls to all functions. */
#undef FUNCTION_OK_FOR_SIBCALL
#define FUNCTION_OK_FOR_SIBCALL(DECL) 1
/* The libcall __canonicalize_funcptr_for_compare is referenced in
crtend.o and the reference isn't resolved in objects that don't
compare function pointers. Thus, we need to play games to provide