invoke.texi (Option Summary): Document -mbranch-predict, -mreg-stack-fill-bug-workaround and their negatives.
* doc/invoke.texi (Option Summary) <MMIX Options>: Document
-mbranch-predict, -mreg-stack-fill-bug-workaround and their
negatives.
(MMIX Options): Ditto. Fix item/itemx typo for -mno-zero-extend.
* config/mmix/mmix.c (mmix_target_asm_function_prologue): Rework
kludge for pre-october-14th mmix versions to handle new-found bug
with PUSHJ/PUSHGO and the register stack.
* config/mmix/mmix.h (struct machine_function): Rename member
has_call_value_without_parameters to has_call_without_parameters.
All referers changed.
(TARGET_MASK_REG_STACK_FILL_BUG, TARGET_DEFAULT
TARGET_MASK_BRANCH_PREDICT): New macros.
(TARGET_SWITCHES): New options -mreg-stack-fill-bug-workaround,
-mno-reg-stack-fill-bug-workaround.
* config/mmix/mmix.md ("call"): Set struct machine member
has_call_without_parameters.
From-SVN: r48558
This commit is contained in:
Hans-Peter Nilsson
Hans-Peter Nilsson
parent
c6c3221d21
commit
3e0f61acb3
@ -1,3 +1,22 @@
|
||||
2002-01-05 Hans-Peter Nilsson <hp@bitrange.com>
|
||||
|
||||
* doc/invoke.texi (Option Summary) <MMIX Options>: Document
|
||||
-mbranch-predict, -mreg-stack-fill-bug-workaround and their
|
||||
negatives.
|
||||
(MMIX Options): Ditto. Fix item/itemx typo for -mno-zero-extend.
|
||||
* config/mmix/mmix.c (mmix_target_asm_function_prologue): Rework
|
||||
kludge for pre-october-14th mmix versions to handle new-found bug
|
||||
with PUSHJ/PUSHGO and the register stack.
|
||||
* config/mmix/mmix.h (struct machine_function): Rename member
|
||||
has_call_value_without_parameters to has_call_without_parameters.
|
||||
All referers changed.
|
||||
(TARGET_MASK_REG_STACK_FILL_BUG, TARGET_DEFAULT
|
||||
TARGET_MASK_BRANCH_PREDICT): New macros.
|
||||
(TARGET_SWITCHES): New options -mreg-stack-fill-bug-workaround,
|
||||
-mno-reg-stack-fill-bug-workaround.
|
||||
* config/mmix/mmix.md ("call"): Set struct machine member
|
||||
has_call_without_parameters.
|
||||
|
||||
Sat Jan 5 02:20:22 CET 2002 Jan Hubicka <jh@suse.cz>
|
||||
|
||||
* cfgcleanup.c (thread_jump): Fix handling of reversed branches.
|
||||
|
||||
@ -967,19 +967,51 @@ mmix_target_asm_function_prologue (stream, locals_size)
|
||||
|
||||
mmix_highest_saved_stack_register = regno;
|
||||
|
||||
/* FIXME: A kludge for the MMIXware ABI. The return value comes back in
|
||||
L of the caller, not just the register number of the X field of
|
||||
PUSH{J,GO}. So we need to make L agree with that number if there's a
|
||||
function call in this function that returns a value but takes no
|
||||
parameters (if there were parameters, L would be set to at least the
|
||||
first parameter register, $16). A real solution includes a pass to
|
||||
test that settings of $15 (MMIX_RETURN_VALUE_REGNUM for the MMIXware
|
||||
ABI) dominate all function calls that return a value. This could be
|
||||
done in the planned machine_dep_reorg pass to rename all registers. */
|
||||
if (! TARGET_ABI_GNU && cfun->machine->has_call_value_without_parameters)
|
||||
/* FIXME: Remove this when a corrected mmix version is released.
|
||||
|
||||
This kludge is a work-around for a presumed bug in the mmix simulator
|
||||
(reported to knuth-bug), all versions up and including "Version of 14
|
||||
October 2001". When the circular register stack fills up, the parts
|
||||
that would be overwritten need to be written to memory. If the
|
||||
"filling" instruction is a PUSHJ or PUSHGO, rL == 0 afterwards. That
|
||||
precise condition (rS == rO && rL == 0) is the same as for an empty
|
||||
register stack, which means no more data is written to memory for
|
||||
that round. A hack is to remove the "&& L!=0" from "@<Increase
|
||||
rL@>=" in mmix-sim.w: the register stack isn't empty under normal
|
||||
circumstances, unless SAVE or UNSAVE is used, interrupts are enabled
|
||||
or cases where rS == rO and rL is explicitly written to 0 as in
|
||||
"PUT rL,0".
|
||||
|
||||
A workaround is to make sure PUSHJ or PUSHGO isn't filling up the
|
||||
register stac. This is accomplished if $16 or higher is written
|
||||
before the function call. This doesn't happen from a leaf functions
|
||||
of course. For the MMIXware ABI, this can't happen if all called
|
||||
functions have parameters, because parameters start at $16.
|
||||
Otherwise, and for the GNU ABI, if any register $16 and up is used,
|
||||
we can see if it's mentioned before any function-call without
|
||||
parameters. This isn't too important; the bug will probably be fixed
|
||||
soon and there's an option to not emit the work-around code. The
|
||||
call-with-parameters kludge wouldn't be there if it hadn't been for
|
||||
it being left-over from a previous mmix version.
|
||||
|
||||
The actual code makes sure any register stack fill happens as early
|
||||
as in the function prologue with a "SET $16,$16" (essentially a nop
|
||||
except for the effects on the register stack). */
|
||||
if (TARGET_REG_STACK_FILL_BUG
|
||||
&& ((TARGET_ABI_GNU && !leaf_function_p ())
|
||||
|| (!TARGET_ABI_GNU
|
||||
&& cfun->machine->has_call_without_parameters)))
|
||||
{
|
||||
/* We don't have a specific macro or derivable expression for the
|
||||
first non-call-saved register. If we need it in other places
|
||||
than here (which is temporary code anyway), such a macro should
|
||||
be added. */
|
||||
int flush_regno
|
||||
= TARGET_ABI_GNU ? mmix_highest_saved_stack_register + 2 : 16;
|
||||
|
||||
fprintf (stream, "\tSET %s,%s\n",
|
||||
reg_names[MMIX_RETURN_VALUE_REGNUM],
|
||||
reg_names[MMIX_RETURN_VALUE_REGNUM]);
|
||||
reg_names[flush_regno], reg_names[flush_regno]);
|
||||
}
|
||||
}
|
||||
|
||||
/* TARGET_ASM_FUNCTION_EPILOGUE. */
|
||||
|
||||
@ -90,7 +90,7 @@ extern struct rtx_def *mmix_compare_op1;
|
||||
mmix.md too. */
|
||||
struct machine_function
|
||||
{
|
||||
int has_call_value_without_parameters;
|
||||
int has_call_without_parameters;
|
||||
int has_landing_pad;
|
||||
};
|
||||
|
||||
@ -157,6 +157,7 @@ extern int target_flags;
|
||||
#define TARGET_MASK_KNUTH_DIVISION 16
|
||||
#define TARGET_MASK_TOPLEVEL_SYMBOLS 32
|
||||
#define TARGET_MASK_BRANCH_PREDICT 64
|
||||
#define TARGET_MASK_REG_STACK_FILL_BUG 128
|
||||
|
||||
/* FIXME: Get rid of this one. */
|
||||
#define TARGET_LIBFUNC (target_flags & TARGET_MASK_LIBFUNCS)
|
||||
@ -166,8 +167,11 @@ extern int target_flags;
|
||||
#define TARGET_KNUTH_DIVISION (target_flags & TARGET_MASK_KNUTH_DIVISION)
|
||||
#define TARGET_TOPLEVEL_SYMBOLS (target_flags & TARGET_MASK_TOPLEVEL_SYMBOLS)
|
||||
#define TARGET_BRANCH_PREDICT (target_flags & TARGET_MASK_BRANCH_PREDICT)
|
||||
#define TARGET_REG_STACK_FILL_BUG \
|
||||
(target_flags & TARGET_MASK_REG_STACK_FILL_BUG)
|
||||
|
||||
#define TARGET_DEFAULT TARGET_MASK_BRANCH_PREDICT
|
||||
#define TARGET_DEFAULT \
|
||||
(TARGET_MASK_BRANCH_PREDICT | TARGET_MASK_REG_STACK_FILL_BUG)
|
||||
|
||||
/* FIXME: Provide a way to *load* the epsilon register. */
|
||||
#define TARGET_SWITCHES \
|
||||
@ -198,6 +202,12 @@ extern int target_flags;
|
||||
N_("Use P-mnemonics for branches statically predicted as taken")}, \
|
||||
{"no-branch-predict", -TARGET_MASK_BRANCH_PREDICT, \
|
||||
N_("Don't use P-mnemonics for branches")}, \
|
||||
{"reg-stack-fill-bug-workaround", TARGET_MASK_REG_STACK_FILL_BUG, \
|
||||
N_("Work around inconsistent behavior when a PUSHJ or PUSHGO\
|
||||
instruction fills the register stack")}, \
|
||||
{"no-reg-stack-fill-bug-workaround", -TARGET_MASK_REG_STACK_FILL_BUG,\
|
||||
N_("Don't work around inconsistent behavior when a PUSHJ or PUSHGO\
|
||||
instruction fills the register stack")}, \
|
||||
{"", TARGET_DEFAULT, ""}}
|
||||
|
||||
/* Unfortunately, this must not reference anything in "mmix.c". */
|
||||
|
||||
@ -995,6 +995,12 @@ DIVU %1,%1,%2\;GET %0,:rR\;NEGU %2,0,%0\;CSNN %0,$255,%2")
|
||||
error too. */
|
||||
if (operands[2] == NULL_RTX)
|
||||
operands[2] = const0_rtx;
|
||||
|
||||
/* FIXME: Documentation bug: operands[3] (operands[2] for 'call') is the
|
||||
*next* argument register, not the number of arguments in registers. */
|
||||
cfun->machine->has_call_without_parameters
|
||||
|= REG_P (operands[2]) && REGNO (operands[2]) == MMIX_FIRST_ARG_REGNUM;
|
||||
|
||||
operands[4] = gen_rtx_REG (DImode, MMIX_INCOMING_RETURN_ADDRESS_REGNUM);
|
||||
}")
|
||||
|
||||
@ -1020,7 +1026,7 @@ DIVU %1,%1,%2\;GET %0,:rR\;NEGU %2,0,%0\;CSNN %0,$255,%2")
|
||||
|
||||
/* FIXME: Documentation bug: operands[3] (operands[2] for 'call') is the
|
||||
*next* argument register, not the number of arguments in registers. */
|
||||
cfun->machine->has_call_value_without_parameters
|
||||
cfun->machine->has_call_without_parameters
|
||||
|= REG_P (operands[3]) && REGNO (operands[3]) == MMIX_FIRST_ARG_REGNUM;
|
||||
|
||||
operands[5] = gen_rtx_REG (DImode, MMIX_INCOMING_RETURN_ADDRESS_REGNUM);
|
||||
|
||||
@ -589,7 +589,8 @@ in the following sections.
|
||||
@gccoptlist{
|
||||
-mlibfuncs -mno-libfuncs -mepsilon -mno-epsilon -mabi=gnu @gol
|
||||
-mabi=mmixware -mzero-extend -mknuthdiv -mtoplevel-symbols @gol
|
||||
-melf}
|
||||
-melf -mbranch-predict -mreg-stack-fill-bug-workaround @gol
|
||||
-mno-branch-predict -mno-reg-stack-fill-bug-workaround}
|
||||
|
||||
@emph{IA-64 Options}
|
||||
@gccoptlist{
|
||||
@ -9716,7 +9717,7 @@ the called function) are seen as registers @code{$0} and up, as opposed to
|
||||
the GNU ABI which uses global registers @code{$231} and up.
|
||||
|
||||
@item -mzero-extend
|
||||
@item -mno-zero-extend
|
||||
@itemx -mno-zero-extend
|
||||
When reading data from memory in sizes shorter than 64 bits, use (do not
|
||||
use) zero-extending load instructions by default, rather than
|
||||
sign-extending ones.
|
||||
@ -9736,6 +9737,18 @@ code can be used with the @code{PREFIX} assembly directive.
|
||||
@item -melf
|
||||
Generate an executable in the ELF format, rather than the default
|
||||
@samp{mmo} format used by the @command{mmix} simulator.
|
||||
|
||||
@item -mbranch-predict
|
||||
@itemx -mno-branch-predict
|
||||
Use (do not use) the probable-branch instructions, when static branch
|
||||
prediction indicates a probable branch.
|
||||
|
||||
@item -mreg-stack-fill-bug-workaround
|
||||
@itemx -mno-reg-stack-fill-bug-workaround
|
||||
Work around (do not work around) an inconsistency in the circular
|
||||
register stack mechanism in the @command{mmix} simulator, which
|
||||
causes entries in the register stack to not be flushed to memory if
|
||||
the instruction causing the fill-up is @code{PUSHJ} or @code{PUSHGO}.
|
||||
@end table
|
||||
|
||||
@node Code Gen Options
|
||||
|
||||
Loading…
Reference in New Issue
Block a user