This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
[committed]: Fix VIS insn scheduling
- From: "David S. Miller" <davem at redhat dot com>
- To: gcc-patches at gcc dot gnu dot org
- Date: Mon, 20 Oct 2003 03:04:23 -0700
- Subject: [committed]: Fix VIS insn scheduling
VIS instructions were not being described accurately
to the instruction scheduler. Although we emit VIS
insns sparingly now, this will be more important when
I add VIS intrinsic support to the Sparc backend.
Committed to mainline.
2003-10-20 David S. Miller <davem@redhat.com>
* config/sparc/sparc.md (type attribute): Add new insn types
fpa, fpm_pack, fgm_mul, fgm_pdist, and fgm_cmp for VIS.
(patterns emitting VIS insns): Use them.
* config/sparc/ultra1_2.md: Add VIS scheduling rules.
* config/sparc/ultra3.md: Likewise.
--- ./config/sparc/sparc.md.~1~ Wed Oct 8 10:09:05 2003
+++ ./config/sparc/sparc.md Sun Oct 19 22:24:43 2003
@@ -106,6 +106,7 @@
fpcmp,
fpmul,fpdivs,fpdivd,
fpsqrts,fpsqrtd,
+ fga,fgm_pack,fgm_mul,fgm_pdist,fgm_cmp,
cmove,
ialuX,
multi,flushw,iflush,trap"
@@ -1920,7 +1921,7 @@
st\t%r1, %0
st\t%1, %0
fzeros\t%0"
- [(set_attr "type" "*,fpmove,*,*,load,fpload,store,fpstore,fpmove")])
+ [(set_attr "type" "*,fpmove,*,*,load,fpload,store,fpstore,fga")])
(define_insn "*movsi_lo_sum"
[(set (match_operand:SI 0 "register_operand" "=r")
@@ -2192,7 +2193,7 @@
ldd\t%1, %0
std\t%1, %0
fzero\t%0"
- [(set_attr "type" "*,*,*,load,store,fpmove,fpload,fpstore,fpmove")
+ [(set_attr "type" "*,*,*,load,store,fpmove,fpload,fpstore,fga")
(set_attr "fptype" "*,*,*,*,*,double,*,*,double")])
(define_expand "movdi_pic_label_ref"
@@ -2640,7 +2641,7 @@
abort();
}
}
- [(set_attr "type" "fpmove,fpmove,*,*,*,*,load,fpload,fpstore,store")])
+ [(set_attr "type" "fpmove,fga,*,*,*,*,load,fpload,fpstore,store")])
;; Exactly the same as above, except that all `f' cases are deleted.
;; This is necessary to prevent reload from ever trying to use a `f' reg
@@ -2957,7 +2958,7 @@
#
#
#"
- [(set_attr "type" "fpmove,fpmove,load,store,store,load,store,*,*,*")
+ [(set_attr "type" "fga,fpmove,load,store,store,load,store,*,*,*")
(set_attr "length" "*,*,*,*,*,*,*,2,2,2")
(set_attr "fptype" "double,double,*,*,*,*,*,*,*,*")])
@@ -3004,7 +3005,7 @@
ldx\t%1, %0
stx\t%r1, %0
#"
- [(set_attr "type" "fpmove,fpmove,load,store,*,load,store,*")
+ [(set_attr "type" "fga,fpmove,load,store,*,load,store,*")
(set_attr "length" "*,*,*,*,*,*,*,2")
(set_attr "fptype" "double,double,*,*,*,*,*,*")])
@@ -4976,7 +4977,7 @@
add\t%1, %2, %0
sub\t%1, -%2, %0
fpadd32s\t%1, %2, %0"
- [(set_attr "type" "*,*,fp")])
+ [(set_attr "type" "*,*,fga")])
(define_insn "*cmp_cc_plus"
[(set (reg:CC_NOOV 100)
@@ -5132,7 +5133,7 @@
sub\t%1, %2, %0
add\t%1, -%2, %0
fpsub32s\t%1, %2, %0"
- [(set_attr "type" "*,*,fp")])
+ [(set_attr "type" "*,*,fga")])
(define_insn "*cmp_minus_cc"
[(set (reg:CC_NOOV 100)
@@ -5856,7 +5857,7 @@
"@
#
fand\t%1, %2, %0"
- [(set_attr "type" "*,fp")
+ [(set_attr "type" "*,fga")
(set_attr "length" "2,*")
(set_attr "fptype" "double")])
@@ -5868,7 +5869,7 @@
"@
and\t%1, %2, %0
fand\t%1, %2, %0"
- [(set_attr "type" "*,fp")
+ [(set_attr "type" "*,fga")
(set_attr "fptype" "double")])
(define_insn "andsi3"
@@ -5879,7 +5880,7 @@
"@
and\t%1, %2, %0
fands\t%1, %2, %0"
- [(set_attr "type" "*,fp")])
+ [(set_attr "type" "*,fga")])
(define_split
[(set (match_operand:SI 0 "register_operand" "")
@@ -5951,7 +5952,7 @@
operands[6] = gen_lowpart (SImode, operands[0]);
operands[7] = gen_lowpart (SImode, operands[1]);
operands[8] = gen_lowpart (SImode, operands[2]);"
- [(set_attr "type" "*,fp")
+ [(set_attr "type" "*,fga")
(set_attr "length" "2,*")
(set_attr "fptype" "double")])
@@ -5963,7 +5964,7 @@
"@
andn\t%2, %1, %0
fandnot1\t%1, %2, %0"
- [(set_attr "type" "*,fp")
+ [(set_attr "type" "*,fga")
(set_attr "fptype" "double")])
(define_insn "*and_not_si"
@@ -5974,7 +5975,7 @@
"@
andn\t%2, %1, %0
fandnot1s\t%1, %2, %0"
- [(set_attr "type" "*,fp")])
+ [(set_attr "type" "*,fga")])
(define_expand "iordi3"
[(set (match_operand:DI 0 "register_operand" "")
@@ -5991,7 +5992,7 @@
"@
#
for\t%1, %2, %0"
- [(set_attr "type" "*,fp")
+ [(set_attr "type" "*,fga")
(set_attr "length" "2,*")
(set_attr "fptype" "double")])
@@ -6003,7 +6004,7 @@
"@
or\t%1, %2, %0
for\t%1, %2, %0"
- [(set_attr "type" "*,fp")
+ [(set_attr "type" "*,fga")
(set_attr "fptype" "double")])
(define_insn "iorsi3"
@@ -6014,7 +6015,7 @@
"@
or\t%1, %2, %0
fors\t%1, %2, %0"
- [(set_attr "type" "*,fp")])
+ [(set_attr "type" "*,fga")])
(define_split
[(set (match_operand:SI 0 "register_operand" "")
@@ -6052,7 +6053,7 @@
operands[6] = gen_lowpart (SImode, operands[0]);
operands[7] = gen_lowpart (SImode, operands[1]);
operands[8] = gen_lowpart (SImode, operands[2]);"
- [(set_attr "type" "*,fp")
+ [(set_attr "type" "*,fga")
(set_attr "length" "2,*")
(set_attr "fptype" "double")])
@@ -6064,7 +6065,7 @@
"@
orn\t%2, %1, %0
fornot1\t%1, %2, %0"
- [(set_attr "type" "*,fp")
+ [(set_attr "type" "*,fga")
(set_attr "fptype" "double")])
(define_insn "*or_not_si"
@@ -6075,7 +6076,7 @@
"@
orn\t%2, %1, %0
fornot1s\t%1, %2, %0"
- [(set_attr "type" "*,fp")])
+ [(set_attr "type" "*,fga")])
(define_expand "xordi3"
[(set (match_operand:DI 0 "register_operand" "")
@@ -6092,7 +6093,7 @@
"@
#
fxor\t%1, %2, %0"
- [(set_attr "type" "*,fp")
+ [(set_attr "type" "*,fga")
(set_attr "length" "2,*")
(set_attr "fptype" "double")])
@@ -6104,7 +6105,7 @@
"@
xor\t%r1, %2, %0
fxor\t%1, %2, %0"
- [(set_attr "type" "*,fp")
+ [(set_attr "type" "*,fga")
(set_attr "fptype" "double")])
(define_insn "*xordi3_sp64_dbl"
@@ -6123,7 +6124,7 @@
"@
xor\t%r1, %2, %0
fxors\t%1, %2, %0"
- [(set_attr "type" "*,fp")])
+ [(set_attr "type" "*,fga")])
(define_split
[(set (match_operand:SI 0 "register_operand" "")
@@ -6177,7 +6178,7 @@
operands[6] = gen_lowpart (SImode, operands[0]);
operands[7] = gen_lowpart (SImode, operands[1]);
operands[8] = gen_lowpart (SImode, operands[2]);"
- [(set_attr "type" "*,fp")
+ [(set_attr "type" "*,fga")
(set_attr "length" "2,*")
(set_attr "fptype" "double")])
@@ -6189,7 +6190,7 @@
"@
xnor\t%r1, %2, %0
fxnor\t%1, %2, %0"
- [(set_attr "type" "*,fp")
+ [(set_attr "type" "*,fga")
(set_attr "fptype" "double")])
(define_insn "*xor_not_si"
@@ -6200,7 +6201,7 @@
"@
xnor\t%r1, %2, %0
fxnors\t%1, %2, %0"
- [(set_attr "type" "*,fp")])
+ [(set_attr "type" "*,fga")])
;; These correspond to the above in the case where we also (or only)
;; want to set the condition code.
@@ -6463,7 +6464,7 @@
operands[3] = gen_highpart (SImode, operands[1]);
operands[4] = gen_lowpart (SImode, operands[0]);
operands[5] = gen_lowpart (SImode, operands[1]);"
- [(set_attr "type" "*,fp")
+ [(set_attr "type" "*,fga")
(set_attr "length" "2,*")
(set_attr "fptype" "double")])
@@ -6474,7 +6475,7 @@
"@
xnor\t%%g0, %1, %0
fnot1\t%1, %0"
- [(set_attr "type" "*,fp")
+ [(set_attr "type" "*,fga")
(set_attr "fptype" "double")])
(define_insn "one_cmplsi2"
@@ -6484,7 +6485,7 @@
"@
xnor\t%%g0, %1, %0
fnot1s\t%1, %0"
- [(set_attr "type" "*,fp")])
+ [(set_attr "type" "*,fga")])
(define_insn "*cmp_cc_not"
[(set (reg:CC 100)
--- ./config/sparc/ultra1_2.md.~1~ Wed Jul 16 01:38:37 2003
+++ ./config/sparc/ultra1_2.md Fri Oct 17 22:37:55 2003
@@ -250,3 +250,53 @@
;; An integer branch may execute in the same cycle as the compare
;; creating the condition codes.
(define_bypass 0 "us1_simple_ieu1" "us1_branch")
+
+;; VIS scheduling
+(define_insn_reservation "us1_fga_single"
+ 2
+ (and (and
+ (eq_attr "cpu" "ultrasparc")
+ (eq_attr "type" "fga"))
+ (eq_attr "fptype" "single"))
+ "us1_fpa + us1_fp_single + us1_slotany, nothing")
+
+(define_bypass 1 "us1_fga_single" "us1_fga_single")
+
+(define_insn_reservation "us1_fga_double"
+ 2
+ (and (and
+ (eq_attr "cpu" "ultrasparc")
+ (eq_attr "type" "fga"))
+ (eq_attr "fptype" "double"))
+ "us1_fpa + us1_fp_double + us1_slotany, nothing")
+
+(define_bypass 1 "us1_fga_double" "us1_fga_double")
+
+(define_insn_reservation "us1_fgm_single"
+ 4
+ (and (and
+ (eq_attr "cpu" "ultrasparc")
+ (eq_attr "type" "fgm_pack,fgm_mul,fgm_cmp"))
+ (eq_attr "fptype" "single"))
+ "us1_fpm + us1_fp_single + us1_slotany, nothing*3")
+
+(define_bypass 3 "us1_fgm_single" "us1_fga_single")
+
+(define_insn_reservation "us1_fgm_double"
+ 4
+ (and (and
+ (eq_attr "cpu" "ultrasparc")
+ (eq_attr "type" "fgm_pack,fgm_mul,fgm_cmp"))
+ (eq_attr "fptype" "double"))
+ "us1_fpm + us1_fp_double + us1_slotany, nothing*3")
+
+(define_bypass 3 "us1_fgm_double" "us1_fga_double")
+
+(define_insn_reservation "us1_pdist"
+ 4
+ (and (eq_attr "cpu" "ultrasparc")
+ (eq_attr "type" "fgm_pdist"))
+ "us1_fpm + us1_fp_double + us1_slotany, nothing*3")
+
+(define_bypass 3 "us1_pdist" "us1_fga_double,us1_fga_single")
+(define_bypass 1 "us1_pdist" "us1_pdist")
--- ./config/sparc/ultra3.md.~1~ Wed Jul 16 01:38:37 2003
+++ ./config/sparc/ultra3.md Fri Oct 17 22:49:20 2003
@@ -167,3 +167,24 @@
;; If FMOVfcc is user of FPCMP, latency is only 1 cycle.
(define_bypass 1 "us3_fpcmp" "us3_fcmov")
+
+;; VIS scheduling
+(define_insn_reservation "us3_fga"
+ 3
+ (and (eq_attr "cpu" "ultrasparc3")
+ (eq_attr "type" "fga"))
+ "us3_fpa + us3_slotany, nothing*2")
+
+(define_insn_reservation "us3_fgm"
+ 4
+ (and (eq_attr "cpu" "ultrasparc3")
+ (eq_attr "type" "fgm_pack,fgm_mul,fgm_cmp"))
+ "us3_fpm + us3_slotany, nothing*3")
+
+(define_insn_reservation "us3_pdist"
+ 4
+ (and (eq_attr "cpu" "ultrasparc3")
+ (eq_attr "type" "fgm_pdist"))
+ "us3_fpm + us3_slotany, nothing*3")
+
+(define_bypass 1 "us3_pdist" "us3_pdist")