* [SH] PR 51244 - Improve T bit to reg store
@ 2012-07-29 22:53 Oleg Endo
2012-07-30 3:15 ` Kaz Kojima
0 siblings, 1 reply; 2+ messages in thread
From: Oleg Endo @ 2012-07-29 22:53 UTC (permalink / raw)
To: gcc-patches
[-- Attachment #1: Type: text/plain, Size: 671 bytes --]
Hello,
This improves some cases where the T bit is stored as all ones or all
zeros to a GP reg.
Tested on rev 189916 with
make -k check RUNTESTFLAGS="--target_board=sh-sim
\{-m2/-ml,-m2/-mb,-m2a/-mb,-m2a-single/-mb,-m4/-ml,-m4/-mb,
-m4-single/-ml,-m4-single/-mb,-m4a-single/-ml,-m4a-single/-mb}"
and no new failures.
OK?
Cheers,
Oleg
ChangeLog:
PR target/51244
* config/sh/sh.md (mov_neg_si_t): Move to Scc instructions
section.
Use t_reg_operand predicate. Add split for negated case.
(ashrsi2_31): Pass get_t_reg_rtx to gen_mov_neg_si_t.
* config/sh/sh.c (expand_ashiftrt): Likewise.
testsuite/ChangeLog:
PR target/51244
* gcc.target/sh/pr51244-4.c: New.
[-- Attachment #2: sh_pr51244-8.patch --]
[-- Type: text/x-patch, Size: 3121 bytes --]
Index: gcc/config/sh/sh.c
===================================================================
--- gcc/config/sh/sh.c (revision 189917)
+++ gcc/config/sh/sh.c (working copy)
@@ -3481,7 +3481,7 @@
{
emit_insn (gen_cmpgtsi_t (force_reg (SImode, CONST0_RTX (SImode)),
operands[1]));
- emit_insn (gen_mov_neg_si_t (operands[0]));
+ emit_insn (gen_mov_neg_si_t (operands[0], get_t_reg_rtx ()));
return true;
}
emit_insn (gen_ashrsi2_31 (operands[0], operands[1]));
Index: gcc/config/sh/sh.md
===================================================================
--- gcc/config/sh/sh.md (revision 189916)
+++ gcc/config/sh/sh.md (working copy)
@@ -1584,15 +1584,6 @@
"subc %2,%0"
[(set_attr "type" "arith")])
-;; life_analysis thinks rn is live before subc rn,rn, so make a special
-;; pattern for this case. This helps multimedia applications that compute
-;; the sum of absolute differences.
-(define_insn "mov_neg_si_t"
- [(set (match_operand:SI 0 "arith_reg_dest" "=r") (neg:SI (reg:SI T_REG)))]
- "TARGET_SH1"
- "subc %0,%0"
- [(set_attr "type" "arith")])
-
(define_insn "*subsi3_internal"
[(set (match_operand:SI 0 "arith_reg_dest" "=r")
(minus:SI (match_operand:SI 1 "arith_reg_operand" "0")
@@ -3797,7 +3788,7 @@
[(const_int 0)]
{
emit_insn (gen_ashlsi_c (operands[0], operands[1]));
- emit_insn (gen_mov_neg_si_t (copy_rtx (operands[0])));
+ emit_insn (gen_mov_neg_si_t (operands[0], get_t_reg_rtx ()));
DONE;
})
@@ -9704,6 +9695,25 @@
""
[(const_int 0)])
+;; Store T bit as all zeros or ones in a reg.
+(define_insn "mov_neg_si_t"
+ [(set (match_operand:SI 0 "arith_reg_dest" "=r")
+ (neg:SI (match_operand 1 "t_reg_operand" "")))]
+ "TARGET_SH1"
+ "subc %0,%0"
+ [(set_attr "type" "arith")])
+
+;; Store negated T bit as all zeros or ones in a reg.
+;; Use the following sequence:
+;; subc Rn,Rn ! Rn = Rn - Rn - T; T = T
+;; not Rn,Rn ! Rn = 0 - Rn
+(define_split
+ [(set (match_operand:SI 0 "arith_reg_dest" "")
+ (neg:SI (match_operand 1 "negt_reg_operand" "")))]
+ "TARGET_SH1"
+ [(set (match_dup 0) (neg:SI (reg:SI T_REG)))
+ (set (match_dup 0) (not:SI (match_dup 0)))])
+
;; The *movtt pattern eliminates redundant T bit to T bit moves / tests.
(define_insn_and_split "*movtt"
[(set (reg:SI T_REG)
Index: gcc/testsuite/gcc.target/sh/pr51244-4.c
===================================================================
--- gcc/testsuite/gcc.target/sh/pr51244-4.c (revision 0)
+++ gcc/testsuite/gcc.target/sh/pr51244-4.c (revision 0)
@@ -0,0 +1,19 @@
+/* Check that storing the (negated) T bit as all ones or zeros in a reg
+ uses the subc instruction. On SH2A a sequence with the movrt instruction
+ is also OK instead of subc. */
+/* { dg-do compile { target "sh*-*-*" } } */
+/* { dg-options "-O1 -mbranch-cost=2" } */
+/* { dg-skip-if "" { "sh*-*-*" } { "-m5*"} { "" } } */
+/* { dg-final { scan-assembler-not "movt|tst|negc" } } */
+/* { dg-final { scan-assembler "subc|movrt|neg|not" } } */
+
+int test_00 (int x, int y)
+{
+ return x != y ? -1 : 0;
+}
+
+int test_01 (int x, int y)
+{
+ return x == y ? -1 : 0;
+}
+
^ permalink raw reply [flat|nested] 2+ messages in thread
* Re: [SH] PR 51244 - Improve T bit to reg store
2012-07-29 22:53 [SH] PR 51244 - Improve T bit to reg store Oleg Endo
@ 2012-07-30 3:15 ` Kaz Kojima
0 siblings, 0 replies; 2+ messages in thread
From: Kaz Kojima @ 2012-07-30 3:15 UTC (permalink / raw)
To: oleg.endo; +Cc: gcc-patches
Oleg Endo <oleg.endo@t-online.de> wrote:
> This improves some cases where the T bit is stored as all ones or all
> zeros to a GP reg.
> Tested on rev 189916 with
> make -k check RUNTESTFLAGS="--target_board=sh-sim
> \{-m2/-ml,-m2/-mb,-m2a/-mb,-m2a-single/-mb,-m4/-ml,-m4/-mb,
> -m4-single/-ml,-m4-single/-mb,-m4a-single/-ml,-m4a-single/-mb}"
>
> and no new failures.
> OK?
OK.
Regards,
kaz
^ permalink raw reply [flat|nested] 2+ messages in thread
end of thread, other threads:[~2012-07-30 3:15 UTC | newest]
Thread overview: 2+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2012-07-29 22:53 [SH] PR 51244 - Improve T bit to reg store Oleg Endo
2012-07-30 3:15 ` Kaz Kojima
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).