This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
[PATCH] Improve length estimate for direct calls on PA
- From: "John David Anglin" <dave at hiauly1 dot hia dot nrc dot ca>
- To: gcc-patches at gcc dot gnu dot org
- Date: Thu, 14 Aug 2003 14:20:21 -0400 (EDT)
- Subject: [PATCH] Improve length estimate for direct calls on PA
This patch improves the length estimate for direct calls on the PA.
When C++ changed to using a local alias for methods jumped and called
from thunks, we had to use different call sequences for local calls.
These are a little shorter than the sequences we had been using before.
This patch mainly is an update to attr_length_call to allow it to
provide better length estimates for local calls.
There is also a small improvement in the length estimate for call
sequences that always fill the delay slot.
Tested on hppa-unknown-linux-gnu, hppa2.0w-hp-hpux11.11 and
hppa64-hp-hpux11.11. Applied to the trunk.
Dave
--
J. David Anglin dave.anglin@nrc-cnrc.gc.ca
National Research Council of Canada (613) 990-0752 (FAX: 952-6602)
2003-08-14 John David Anglin <dave.anglin@nrc-cnrc.gc.ca>
* pa.c (pa_adjust_insn_length): Delete adjustment for delay slot in
direct calls.
(attr_length_call): Include it here. Improve length estimate for
local calls.
(output_call): Use targetm.binds_local_p.
Index: config/pa/pa.c
===================================================================
RCS file: /cvs/gcc/gcc/gcc/config/pa/pa.c,v
retrieving revision 1.220
diff -u -3 -p -r1.220 pa.c
--- config/pa/pa.c 9 Aug 2003 16:40:33 -0000 1.220
+++ config/pa/pa.c 14 Aug 2003 17:33:58 -0000
@@ -4417,23 +4417,9 @@ pa_adjust_insn_length (insn, length)
{
rtx pat = PATTERN (insn);
- /* Call insns which are *not* indirect and have unfilled delay slots. */
- if (GET_CODE (insn) == CALL_INSN)
- {
-
- if (GET_CODE (XVECEXP (pat, 0, 0)) == CALL
- && GET_CODE (XEXP (XEXP (XVECEXP (pat, 0, 0), 0), 0)) == SYMBOL_REF)
- return 4;
- else if (GET_CODE (XVECEXP (pat, 0, 0)) == SET
- && GET_CODE (XEXP (XEXP (XEXP (XVECEXP (pat, 0, 0), 1), 0), 0))
- == SYMBOL_REF)
- return 4;
- else
- return 0;
- }
- /* Jumps inside switch tables which have unfilled delay slots
- also need adjustment. */
- else if (GET_CODE (insn) == JUMP_INSN
+ /* Jumps inside switch tables which have unfilled delay slots need
+ adjustment. */
+ if (GET_CODE (insn) == JUMP_INSN
&& simplejump_p (insn)
&& GET_MODE (insn) == SImode)
return 4;
@@ -6758,77 +6744,95 @@ output_millicode_call (insn, call_dest)
/* Return the attribute length of the call instruction INSN. The SIBCALL
flag indicates whether INSN is a regular call or a sibling call. The
- length returned must be longer than the code generated by output_call.
- When the target supports jumps in the delay slot, we need an extra
- four bytes to handle the situation where the jump can't reach its
- destination. */
+ length returned must be longer than the code actually generated by
+ output_call. Since branch shortening is done before delay branch
+ sequencing, there is no way to determine whether or not the delay
+ slot will be filled during branch shortening. Even when the delay
+ slot is filled, we may have to add a nop if the delay slot contains
+ a branch that can't reach its target. Thus, we always have to include
+ the delay slot in the length estimate. This used to be done in
+ pa_adjust_insn_length but we do it here now as some sequences always
+ fill the delay slot and we can save four bytes in the estimate for
+ these sequences. */
int
attr_length_call (insn, sibcall)
rtx insn;
int sibcall;
{
+ int local_call;
+ rtx call_dest;
+ tree call_decl;
+ int length = 0;
+ rtx pat = PATTERN (insn);
unsigned long distance = -1;
- unsigned long total = IN_NAMED_SECTION_P (cfun->decl) ? 0 : total_code_bytes;
if (INSN_ADDRESSES_SET_P ())
{
+ unsigned long total;
+
+ total = IN_NAMED_SECTION_P (cfun->decl) ? 0 : total_code_bytes;
distance = (total + insn_current_reference_address (insn));
if (distance < total)
distance = -1;
}
- if (TARGET_64BIT)
- {
- if (!TARGET_LONG_CALLS
- && ((!sibcall && distance < 7600000) || distance < 240000))
- return 8;
-
- return (sibcall ? 28 : 24);
- }
+ /* Determine if this is a local call. */
+ if (GET_CODE (XVECEXP (pat, 0, 0)) == CALL)
+ call_dest = XEXP (XEXP (XVECEXP (pat, 0, 0), 0), 0);
else
- {
- if (!TARGET_LONG_CALLS
- && ((TARGET_PA_20 && !sibcall && distance < 7600000)
- || distance < 240000))
- return 8;
+ call_dest = XEXP (XEXP (XEXP (XVECEXP (pat, 0, 0), 1), 0), 0);
- if (TARGET_LONG_ABS_CALL && !flag_pic)
- return 16;
+ call_decl = SYMBOL_REF_DECL (call_dest);
+ local_call = call_decl && (*targetm.binds_local_p) (call_decl);
- if ((TARGET_SOM && TARGET_LONG_PIC_SDIFF_CALL)
- || (TARGET_GAS && TARGET_LONG_PIC_PCREL_CALL))
- {
- if (TARGET_PA_20)
- return 20;
+ /* pc-relative branch. */
+ if (!TARGET_LONG_CALLS
+ && ((TARGET_PA_20 && !sibcall && distance < 7600000)
+ || distance < 240000))
+ length += 8;
- return 28;
- }
- else
- {
- int length = 28;
+ /* 64-bit plabel sequence. */
+ else if (TARGET_64BIT && !local_call)
+ length += sibcall ? 28 : 24;
- if (TARGET_SOM)
- length += length_fp_args (insn);
+ /* non-pic long absolute branch sequence. */
+ else if ((TARGET_LONG_ABS_CALL || local_call) && !flag_pic)
+ length += 12;
- if (flag_pic)
- length += 4;
+ /* long pc-relative branch sequence. */
+ else if ((TARGET_SOM && TARGET_LONG_PIC_SDIFF_CALL)
+ || (TARGET_64BIT && !TARGET_GAS)
+ || (TARGET_GAS && (TARGET_LONG_PIC_PCREL_CALL || local_call)))
+ {
+ length += 20;
- if (!sibcall)
- length += 4;
+ if (!TARGET_PA_20 && !TARGET_NO_SPACE_REGS)
+ length += 8;
+ }
- if (TARGET_PA_20)
- return length;
+ /* 32-bit plabel sequence. */
+ else
+ {
+ length += 32;
- if (!TARGET_NO_SPACE_REGS)
- length += 8;
+ if (TARGET_SOM)
+ length += length_fp_args (insn);
+ if (flag_pic)
+ length += 4;
+
+ if (!TARGET_PA_20)
+ {
if (!sibcall)
length += 8;
- return length;
+ if (!TARGET_NO_SPACE_REGS)
+ length += 8;
}
}
+
+ return length;
}
/* INSN is a function call. It may have an unconditional jump
@@ -6846,7 +6850,7 @@ output_call (insn, call_dest, sibcall)
int delay_slot_filled = 0;
int seq_length = dbr_sequence_length ();
tree call_decl = SYMBOL_REF_DECL (call_dest);
- int local_call = call_decl && !TREE_PUBLIC (call_decl);
+ int local_call = call_decl && (*targetm.binds_local_p) (call_decl);
rtx xoperands[2];
xoperands[0] = call_dest;