* Combine on IA-64
@ 2002-02-26 6:18 Jakub Jelinek
2002-02-26 10:12 ` Richard Henderson
0 siblings, 1 reply; 4+ messages in thread
From: Jakub Jelinek @ 2002-02-26 6:18 UTC (permalink / raw)
To: gcc
Hi!
With the visibility patch I've noticed IA-64 generates
pretty inefficient code like:
addl r15 = @gprel(l#), gp
addl r16 = @gprel(m#), gp
;;
.mii
// cycle 1
adds r15 = 10, r15
adds r16 = 10, r16
instead of:
addl r15 = @gprel(l#+10), gp
addl r16 = @gprel(m#+10), gp
I've tried to fix this using peephole2, but that doesn't work too well,
since the two instructions aren't usually adjacent.
Now, as I understand, symbolic_operand accepts
(const (plus:DI (symbol_ref:DI ("@foo")) (const_int )))
just fine, so this looks like combiner inefficiency.
try_combine for:
i2:
(insn 9 18 12 (parallel[
(set (reg/f:DI 340)
(symbol_ref:DI ("@l")))
(clobber (reg:DI 341))
(use (reg:DI 1 r1))
] ) 5 {movdi_symbolic} (nil)
(expr_list:REG_UNUSED (reg:DI 341)
(expr_list:REG_EQUAL (symbol_ref:DI ("@l"))
(nil))))
i3:
(insn 12 9 14 (set (reg/f:DI 343)
(plus:DI (reg/f:DI 340)
(const_int 10 [0xa]))) 90 {adddi3} (insn_list 9 (nil))
(expr_list:REG_DEAD (reg/f:DI 340)
(expr_list:REG_EQUAL (const:DI (plus:DI (symbol_ref:DI ("@l"))
(const_int 10 [0xa])))
(nil))))
only attempts to recognize:
(set (reg/f:DI 343)
(const (plus:DI (symbol_ref:DI ("@l"))
(const_int 10 [0xa]))))
and not
(parallel[
(set (reg/f:DI 343)
(const (plus:DI (symbol_ref:DI ("@l"))
(const_int 10 [0xa]))))
(clobber (reg:DI 341))
(use (reg:DI 1 r1))
] )
Cannot this be special cased in try_combine (there is already some other
similar special case which fails because i3's SET_SRC is not
register but PLUS) provided the clobbers aren't mentioned between i2 and i3
and uses are not modified in between?
Jakub
^ permalink raw reply [flat|nested] 4+ messages in thread
* Re: Combine on IA-64
2002-02-26 6:18 Combine on IA-64 Jakub Jelinek
@ 2002-02-26 10:12 ` Richard Henderson
2002-02-26 10:26 ` Jakub Jelinek
0 siblings, 1 reply; 4+ messages in thread
From: Richard Henderson @ 2002-02-26 10:12 UTC (permalink / raw)
To: Jakub Jelinek; +Cc: gcc
On Tue, Feb 26, 2002 at 02:50:31PM +0100, Jakub Jelinek wrote:
> Cannot this be special cased in try_combine (there is already some other
> similar special case which fails because i3's SET_SRC is not
> register but PLUS) provided the clobbers aren't mentioned between i2 and i3
> and uses are not modified in between?
I think better is to do
(define_split
[(set (match_operand:DI 0 "register_operand" "")
(match_operand:DI 1 "symbolic_operand" ""))]
""
[(const_int 0)]
"ia64_expand_load_address (operands[0], operands[1], operands[2]); DONE;")
r~
^ permalink raw reply [flat|nested] 4+ messages in thread
* Re: Combine on IA-64
2002-02-26 10:12 ` Richard Henderson
@ 2002-02-26 10:26 ` Jakub Jelinek
2002-02-26 10:58 ` Richard Henderson
0 siblings, 1 reply; 4+ messages in thread
From: Jakub Jelinek @ 2002-02-26 10:26 UTC (permalink / raw)
To: Richard Henderson, gcc
On Tue, Feb 26, 2002 at 10:08:46AM -0800, Richard Henderson wrote:
> On Tue, Feb 26, 2002 at 02:50:31PM +0100, Jakub Jelinek wrote:
> > Cannot this be special cased in try_combine (there is already some other
> > similar special case which fails because i3's SET_SRC is not
> > register but PLUS) provided the clobbers aren't mentioned between i2 and i3
> > and uses are not modified in between?
>
> I think better is to do
>
> (define_split
> [(set (match_operand:DI 0 "register_operand" "")
> (match_operand:DI 1 "symbolic_operand" ""))]
> ""
> [(const_int 0)]
> "ia64_expand_load_address (operands[0], operands[1], operands[2]); DONE;")
It is already there (well, define_insn_and_split, but I thought it
is the same).
But it is split during first schedule, not before, so nobody
combines into load_gprel.
Jakub
^ permalink raw reply [flat|nested] 4+ messages in thread
* Re: Combine on IA-64
2002-02-26 10:26 ` Jakub Jelinek
@ 2002-02-26 10:58 ` Richard Henderson
0 siblings, 0 replies; 4+ messages in thread
From: Richard Henderson @ 2002-02-26 10:58 UTC (permalink / raw)
To: Jakub Jelinek; +Cc: gcc
On Tue, Feb 26, 2002 at 01:14:15PM -0500, Jakub Jelinek wrote:
> > (define_split
> > [(set (match_operand:DI 0 "register_operand" "")
> > (match_operand:DI 1 "symbolic_operand" ""))]
> > ""
> > [(const_int 0)]
> > "ia64_expand_load_address (operands[0], operands[1], operands[2]); DONE;")
>
> It is already there (well, define_insn_and_split, but I thought it
> is the same).
No, the input pattern is different.
r~
^ permalink raw reply [flat|nested] 4+ messages in thread
end of thread, other threads:[~2002-02-26 18:26 UTC | newest]
Thread overview: 4+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2002-02-26 6:18 Combine on IA-64 Jakub Jelinek
2002-02-26 10:12 ` Richard Henderson
2002-02-26 10:26 ` Jakub Jelinek
2002-02-26 10:58 ` 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).