* [IA64] Hookize MEMORY_MOVE_COST
@ 2010-06-28 21:52 Anatoly Sokolov
2010-06-29 7:14 ` Richard Henderson
2010-06-29 17:26 ` Richard Henderson
0 siblings, 2 replies; 6+ messages in thread
From: Anatoly Sokolov @ 2010-06-28 21:52 UTC (permalink / raw)
To: gcc-patches; +Cc: Steve Ellcey
Hello.
This patch removes obsolete MEMORY_MOVE_COST macro from IA64 back end in
the GCC and introduces equivalent TARGET_MEMORY_MOVE_COST target hooks.
Bootstrapped and regression tested on ia64-unknown-linux-gnu.
OK to install?
* config/ia64/ia64.h (MEMORY_MOVE_COST): Remove macro.
* config/ia64/t-ia64 (ia64.o): Depend on reload.h.
* config/ia64/ia64.c Include reload.h.
(ia64_memory_move_cost): New function.
(TARGET_MEMORY_MOVE_COST): Define.
(ia64_register_move_cost): Replace MEMORY_MOVE_COST with
memory_move_cost.
Index: gcc/config/ia64/ia64.c
===================================================================
--- gcc/config/ia64/ia64.c (revision 161470)
+++ gcc/config/ia64/ia64.c (working copy)
@@ -59,6 +59,7 @@
#include "dbgcnt.h"
#include "tm-constrs.h"
#include "sel-sched.h"
+#include "reload.h"
/* This is used for communication between ASM_OUTPUT_LABEL and
ASM_OUTPUT_LABELREF. */
@@ -212,6 +213,8 @@
static bool ia64_function_value_regno_p (const unsigned int);
static int ia64_register_move_cost (enum machine_mode, enum reg_class,
enum reg_class);
+static int ia64_memory_move_cost (enum machine_mode mode, enum reg_class,
+ bool);
static bool ia64_rtx_costs (rtx, int, int, int *, bool);
static int ia64_unspec_may_trap_p (const_rtx, unsigned);
static void fix_range (const char *);
@@ -458,6 +461,8 @@
#undef TARGET_REGISTER_MOVE_COST
#define TARGET_REGISTER_MOVE_COST ia64_register_move_cost
+#undef TARGET_MEMORY_MOVE_COST
+#define TARGET_MEMORY_MOVE_COST ia64_memory_move_cost
#undef TARGET_RTX_COSTS
#define TARGET_RTX_COSTS ia64_rtx_costs
#undef TARGET_ADDRESS_COST
@@ -5226,12 +5231,12 @@
/* Moving from FR<->GR in XFmode must be more expensive than 2,
so that we get secondary memory reloads. Between FR_REGS,
- we have to make this at least as expensive as MEMORY_MOVE_COST
+ we have to make this at least as expensive as memory_move_cost
to avoid spectacularly poor register class preferencing. */
if (mode == XFmode || mode == RFmode)
{
if (to != GR_REGS || from != GR_REGS)
- return MEMORY_MOVE_COST (mode, to, 0);
+ return memory_move_cost (mode, to, false);
else
return 3;
}
@@ -5244,20 +5249,20 @@
return 3;
/* Moving between PR and anything but GR is impossible. */
if (from != GR_REGS)
- return MEMORY_MOVE_COST (mode, to, 0);
+ return memory_move_cost (mode, to, false);
break;
case BR_REGS:
/* Moving between BR and anything but GR is impossible. */
if (from != GR_REGS && from != GR_AND_BR_REGS)
- return MEMORY_MOVE_COST (mode, to, 0);
+ return memory_move_cost (mode, to, false);
break;
case AR_I_REGS:
case AR_M_REGS:
/* Moving between AR and anything but GR is impossible. */
if (from != GR_REGS)
- return MEMORY_MOVE_COST (mode, to, 0);
+ return memory_move_cost (mode, to, false);
break;
case GR_REGS:
@@ -5275,6 +5280,23 @@
return 2;
}
+/* Calculate the cost of moving data of MODE from a register to or from
+ memory. */
+
+static int
+ia64_memory_move_cost (enum machine_mode mode ATTRIBUTE_UNUSED,
+ enum reg_class rclass,
+ bool in ATTRIBUTE_UNUSED)
+{
+ if (rclass == GENERAL_REGS
+ || rclass == FR_REGS
+ || rclass == FP_REGS
+ || rclass == GR_AND_FR_REGS)
+ return 4;
+ else
+ return 10;
+}
+
/* Implement PREFERRED_RELOAD_CLASS. Place additional restrictions on RCLASS
to use when copying X into that class. */
Index: gcc/config/ia64/ia64.h
===================================================================
--- gcc/config/ia64/ia64.h (revision 161470)
+++ gcc/config/ia64/ia64.h (working copy)
@@ -1310,12 +1310,6 @@
\f
/* Describing Relative Costs of Operations */
-/* A C expression for the cost of moving data of mode M between a
- register and memory. */
-#define MEMORY_MOVE_COST(MODE,CLASS,IN) \
- ((CLASS) == GENERAL_REGS || (CLASS) == FR_REGS || (CLASS) == FP_REGS \
- || (CLASS) == GR_AND_FR_REGS ? 4 : 10)
-
/* A C expression for the cost of a branch instruction. A value of 1 is the
default; other values are interpreted relative to that. Used by the
if-conversion code as max instruction count. */
Index: gcc/config/ia64/t-ia64
===================================================================
--- gcc/config/ia64/t-ia64 (revision 161470)
+++ gcc/config/ia64/t-ia64 (working copy)
@@ -53,4 +53,4 @@
# genattrtab generates very long string literals.
insn-attrtab.o-warn = -Wno-error
-ia64.o: debug.h $(PARAMS_H) sel-sched.h
+ia64.o: debug.h $(PARAMS_H) sel-sched.h reload.h
Anatoly.
^ permalink raw reply [flat|nested] 6+ messages in thread
* Re: [IA64] Hookize MEMORY_MOVE_COST
2010-06-28 21:52 [IA64] Hookize MEMORY_MOVE_COST Anatoly Sokolov
@ 2010-06-29 7:14 ` Richard Henderson
2010-06-29 9:26 ` Anatoly Sokolov
2010-06-29 17:26 ` Richard Henderson
1 sibling, 1 reply; 6+ messages in thread
From: Richard Henderson @ 2010-06-29 7:14 UTC (permalink / raw)
To: Anatoly Sokolov; +Cc: gcc-patches, Steve Ellcey
On 06/28/2010 01:42 PM, Anatoly Sokolov wrote:
> Hello.
>
> This patch removes obsolete MEMORY_MOVE_COST macro from IA64 back end in
> the GCC and introduces equivalent TARGET_MEMORY_MOVE_COST target hooks.
>
> Bootstrapped and regression tested on ia64-unknown-linux-gnu.
>
> OK to install?
>
> * config/ia64/ia64.h (MEMORY_MOVE_COST): Remove macro.
> * config/ia64/t-ia64 (ia64.o): Depend on reload.h.
> * config/ia64/ia64.c Include reload.h.
> (ia64_memory_move_cost): New function.
> (TARGET_MEMORY_MOVE_COST): Define.
> (ia64_register_move_cost): Replace MEMORY_MOVE_COST with
> memory_move_cost.
>
>
> Index: gcc/config/ia64/ia64.c
> ===================================================================
> --- gcc/config/ia64/ia64.c (revision 161470)
> +++ gcc/config/ia64/ia64.c (working copy)
> @@ -59,6 +59,7 @@
> #include "dbgcnt.h"
> #include "tm-constrs.h"
> #include "sel-sched.h"
> +#include "reload.h"
What bit of reload.h is being used here?
r~
^ permalink raw reply [flat|nested] 6+ messages in thread
* Re: [IA64] Hookize MEMORY_MOVE_COST
2010-06-29 7:14 ` Richard Henderson
@ 2010-06-29 9:26 ` Anatoly Sokolov
2010-06-29 17:24 ` Richard Henderson
0 siblings, 1 reply; 6+ messages in thread
From: Anatoly Sokolov @ 2010-06-29 9:26 UTC (permalink / raw)
To: Richard Henderson; +Cc: gcc-patches, Steve Ellcey
Hi.
From: "Richard Henderson" <rth@redhat.com>
To: "Anatoly Sokolov" <aesok@post.ru>
Sent: Tuesday, June 29, 2010 4:14 AM
> On 06/28/2010 01:42 PM, Anatoly Sokolov wrote:
>>
>> Index: gcc/config/ia64/ia64.c
>> ===================================================================
>> --- gcc/config/ia64/ia64.c (revision 161470)
>> +++ gcc/config/ia64/ia64.c (working copy)
>> @@ -59,6 +59,7 @@
>> #include "dbgcnt.h"
>> #include "tm-constrs.h"
>> #include "sel-sched.h"
>> +#include "reload.h"
>
> What bit of reload.h is being used here?
>
The memory_move_cost function declared in reload.h.
Anatoly.
^ permalink raw reply [flat|nested] 6+ messages in thread
* Re: [IA64] Hookize MEMORY_MOVE_COST
2010-06-29 9:26 ` Anatoly Sokolov
@ 2010-06-29 17:24 ` Richard Henderson
2010-06-29 20:47 ` Anatoly Sokolov
0 siblings, 1 reply; 6+ messages in thread
From: Richard Henderson @ 2010-06-29 17:24 UTC (permalink / raw)
To: Anatoly Sokolov; +Cc: gcc-patches, Steve Ellcey
On 06/29/2010 12:00 AM, Anatoly Sokolov wrote:
>> What bit of reload.h is being used here?
>>
>
> The memory_move_cost function declared in reload.h.
... which is not called by ia64_memory_move_cost,
so I ask again.
r~
^ permalink raw reply [flat|nested] 6+ messages in thread
* Re: [IA64] Hookize MEMORY_MOVE_COST
2010-06-28 21:52 [IA64] Hookize MEMORY_MOVE_COST Anatoly Sokolov
2010-06-29 7:14 ` Richard Henderson
@ 2010-06-29 17:26 ` Richard Henderson
1 sibling, 0 replies; 6+ messages in thread
From: Richard Henderson @ 2010-06-29 17:26 UTC (permalink / raw)
To: Anatoly Sokolov; +Cc: gcc-patches, Steve Ellcey
On 06/28/2010 01:42 PM, Anatoly Sokolov wrote:
> Hello.
>
> This patch removes obsolete MEMORY_MOVE_COST macro from IA64 back end in
> the GCC and introduces equivalent TARGET_MEMORY_MOVE_COST target hooks.
>
> Bootstrapped and regression tested on ia64-unknown-linux-gnu.
>
> OK to install?
>
> * config/ia64/ia64.h (MEMORY_MOVE_COST): Remove macro.
> * config/ia64/t-ia64 (ia64.o): Depend on reload.h.
> * config/ia64/ia64.c Include reload.h.
> (ia64_memory_move_cost): New function.
> (TARGET_MEMORY_MOVE_COST): Define.
> (ia64_register_move_cost): Replace MEMORY_MOVE_COST with
> memory_move_cost.
Patch is ok.
r~
^ permalink raw reply [flat|nested] 6+ messages in thread
* Re: [IA64] Hookize MEMORY_MOVE_COST
2010-06-29 17:24 ` Richard Henderson
@ 2010-06-29 20:47 ` Anatoly Sokolov
0 siblings, 0 replies; 6+ messages in thread
From: Anatoly Sokolov @ 2010-06-29 20:47 UTC (permalink / raw)
To: Richard Henderson; +Cc: gcc-patches, Steve Ellcey
Hi.
From: "Richard Henderson" <rth@redhat.com>
To: "Anatoly Sokolov" <aesok@post.ru>
Sent: Tuesday, June 29, 2010 7:22 PM
> On 06/29/2010 12:00 AM, Anatoly Sokolov wrote:
>>> What bit of reload.h is being used here?
>>>
>>
>> The memory_move_cost function declared in reload.h.
>
> ... which is not called by ia64_memory_move_cost,
> so I ask again.
>
I wish split the back end interface from middle end interface. If the back
end
interface is once again will be changed, for example be implemented as
C++ class and TARGET_MEMORY_MOVE_COST target hooks changed to virtual
function,
only memory_move_cost function should be changed, all code which uses it
will
stay unchanged.
Anatoly.
^ permalink raw reply [flat|nested] 6+ messages in thread
end of thread, other threads:[~2010-06-29 19:18 UTC | newest]
Thread overview: 6+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2010-06-28 21:52 [IA64] Hookize MEMORY_MOVE_COST Anatoly Sokolov
2010-06-29 7:14 ` Richard Henderson
2010-06-29 9:26 ` Anatoly Sokolov
2010-06-29 17:24 ` Richard Henderson
2010-06-29 20:47 ` Anatoly Sokolov
2010-06-29 17:26 ` Richard Henderson
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).