[committed] Model copies between general and fp registers on PA
John David Anglin
dave@hiauly1.hia.nrc.ca
Sun Mar 18 20:29:00 GMT 2007
Sometime ago we had to stop using SECONDARY_MEMORY_NEEDED for copying
data between general and floating-point registers. The copies have to
go through memory and we can only use small offsets for this on PA 1.x
targets. The location at SP-16 is free in the frame marker, so it is
used for this purpose. However, exposing this location with
SECONDARY_MEMORY_NEEDED created a conflict in rare circumstances.
So, I created alternatives for these copies in the move patterns. This
effectively created two new instruction types which weren't scheduled
optimally.
This patch adds the necessary reservation, bypass and anti-bypass
descriptions for these new instructions.
Tested on hppa2.0w-hp-hpux11.11 and hppa64-hp-hpux11.11 with no
observed regressions. Committed to trunk.
Dave
--
J. David Anglin dave.anglin@nrc-cnrc.gc.ca
National Research Council of Canada (613) 990-0752 (FAX: 952-6602)
2007-03-18 John David Anglin <dave.anglin@nrc-cnrc.gc.ca>
* pa.md: Add fpstore_load and store_fpload instruction types. Provide
reservation, bypass and anti-bypass descriptions for these instructions.
Update move patterns.
* pa.c (hppa_fpstore_bypass_p): Check for both TYPE_FPSTORE_LOAD and
TYPE_FPSTORE.
Index: config/pa/pa.md
===================================================================
--- config/pa/pa.md (revision 123037)
+++ config/pa/pa.md (working copy)
@@ -68,7 +68,7 @@
;; type "binary" insns have two input operands (1,2) and one output (0)
(define_attr "type"
- "move,unary,binary,shift,nullshift,compare,load,store,uncond_branch,btable_branch,branch,cbranch,fbranch,call,dyncall,fpload,fpstore,fpalu,fpcc,fpmulsgl,fpmuldbl,fpdivsgl,fpdivdbl,fpsqrtsgl,fpsqrtdbl,multi,milli,parallel_branch"
+ "move,unary,binary,shift,nullshift,compare,load,store,uncond_branch,btable_branch,branch,cbranch,fbranch,call,dyncall,fpload,fpstore,fpalu,fpcc,fpmulsgl,fpmuldbl,fpdivsgl,fpdivdbl,fpsqrtsgl,fpsqrtdbl,multi,milli,parallel_branch,fpstore_load,store_fpload"
(const_string "binary"))
(define_attr "pa_combine_type"
@@ -258,21 +258,31 @@
(eq_attr "cpu" "700"))
"mem_700*3")
-(define_insn_reservation "W11" 1
- (and (eq_attr "type" "!fpcc,fpalu,fpmulsgl,fpmuldbl,fpdivsgl,fpdivdbl,fpsqrtsgl,fpsqrtdbl,load,fpload,store,fpstore")
+(define_insn_reservation "W11" 5
+ (and (eq_attr "type" "fpstore_load")
(eq_attr "cpu" "700"))
+ "mem_700*5")
+
+(define_insn_reservation "W12" 6
+ (and (eq_attr "type" "store_fpload")
+ (eq_attr "cpu" "700"))
+ "mem_700*6")
+
+(define_insn_reservation "W13" 1
+ (and (eq_attr "type" "!fpcc,fpalu,fpmulsgl,fpmuldbl,fpdivsgl,fpdivdbl,fpsqrtsgl,fpsqrtdbl,load,fpload,store,fpstore,fpstore_load,store_fpload")
+ (eq_attr "cpu" "700"))
"dummy_700")
;; We have a bypass for all computations in the FP unit which feed an
;; FP store as long as the sizes are the same.
-(define_bypass 2 "W1,W2" "W10" "hppa_fpstore_bypass_p")
-(define_bypass 9 "W3" "W10" "hppa_fpstore_bypass_p")
-(define_bypass 11 "W4" "W10" "hppa_fpstore_bypass_p")
-(define_bypass 13 "W5" "W10" "hppa_fpstore_bypass_p")
-(define_bypass 17 "W6" "W10" "hppa_fpstore_bypass_p")
+(define_bypass 2 "W1,W2" "W10,W11" "hppa_fpstore_bypass_p")
+(define_bypass 9 "W3" "W10,W11" "hppa_fpstore_bypass_p")
+(define_bypass 11 "W4" "W10,W11" "hppa_fpstore_bypass_p")
+(define_bypass 13 "W5" "W10,W11" "hppa_fpstore_bypass_p")
+(define_bypass 17 "W6" "W10,W11" "hppa_fpstore_bypass_p")
;; We have an "anti-bypass" for FP loads which feed an FP store.
-(define_bypass 4 "W8" "W10" "hppa_fpstore_bypass_p")
+(define_bypass 4 "W8,W12" "W10,W11" "hppa_fpstore_bypass_p")
;; Function units for the 7100 and 7150. The 7100/7150 can dual-issue
;; floating point computations with non-floating point computations (fp loads
@@ -344,19 +354,29 @@
(eq_attr "cpu" "7100"))
"i_7100+mem_7100,mem_7100")
-(define_insn_reservation "X7" 1
- (and (eq_attr "type" "!fpcc,fpalu,fpmulsgl,fpmuldbl,fpdivsgl,fpsqrtsgl,fpdivdbl,fpsqrtdbl,load,fpload,store,fpstore")
+(define_insn_reservation "X7" 4
+ (and (eq_attr "type" "fpstore_load")
(eq_attr "cpu" "7100"))
+ "i_7100+mem_7100,mem_7100*3")
+
+(define_insn_reservation "X8" 4
+ (and (eq_attr "type" "store_fpload")
+ (eq_attr "cpu" "7100"))
+ "i_7100+mem_7100,mem_7100*3")
+
+(define_insn_reservation "X9" 1
+ (and (eq_attr "type" "!fpcc,fpalu,fpmulsgl,fpmuldbl,fpdivsgl,fpsqrtsgl,fpdivdbl,fpsqrtdbl,load,fpload,store,fpstore,fpstore_load,store_fpload")
+ (eq_attr "cpu" "7100"))
"i_7100")
;; We have a bypass for all computations in the FP unit which feed an
;; FP store as long as the sizes are the same.
-(define_bypass 1 "X0" "X6" "hppa_fpstore_bypass_p")
-(define_bypass 7 "X1" "X6" "hppa_fpstore_bypass_p")
-(define_bypass 14 "X2" "X6" "hppa_fpstore_bypass_p")
+(define_bypass 1 "X0" "X6,X7" "hppa_fpstore_bypass_p")
+(define_bypass 7 "X1" "X6,X7" "hppa_fpstore_bypass_p")
+(define_bypass 14 "X2" "X6,X7" "hppa_fpstore_bypass_p")
;; We have an "anti-bypass" for FP loads which feed an FP store.
-(define_bypass 3 "X4" "X6" "hppa_fpstore_bypass_p")
+(define_bypass 3 "X4,X8" "X6,X7" "hppa_fpstore_bypass_p")
;; The 7100LC has three floating-point units: ALU, MUL, and DIV.
;; There's no value in modeling the ALU and MUL separately though
@@ -449,40 +469,70 @@
(eq_attr "cpu" "7100LC"))
"i1_7100lc+mem_7100lc,mem_7100lc")
-(define_insn_reservation "Y6" 1
+(define_insn_reservation "Y6" 4
+ (and (eq_attr "type" "fpstore_load")
+ (eq_attr "cpu" "7100LC"))
+ "i1_7100lc+mem_7100lc,mem_7100lc*3")
+
+(define_insn_reservation "Y7" 4
+ (and (eq_attr "type" "store_fpload")
+ (eq_attr "cpu" "7100LC"))
+ "i1_7100lc+mem_7100lc,mem_7100lc*3")
+
+(define_insn_reservation "Y8" 1
(and (eq_attr "type" "shift,nullshift")
(eq_attr "cpu" "7100LC,7200,7300"))
"i1_7100lc")
-(define_insn_reservation "Y7" 1
+(define_insn_reservation "Y9" 1
(and (eq_attr "type" "!fpcc,fpalu,fpmulsgl,fpmuldbl,fpdivsgl,fpsqrtsgl,fpdivdbl,fpsqrtdbl,load,fpload,store,fpstore,shift,nullshift")
(eq_attr "cpu" "7100LC,7200,7300"))
"(i0_7100lc|i1_7100lc)")
;; The 7200 has a store-load penalty
-(define_insn_reservation "Y8" 2
+(define_insn_reservation "Y10" 2
(and (eq_attr "type" "store")
(eq_attr "cpu" "7200"))
"i1_7100lc,mem_7100lc")
-(define_insn_reservation "Y9" 2
+(define_insn_reservation "Y11" 2
(and (eq_attr "type" "fpstore")
(eq_attr "cpu" "7200"))
"i1_7100lc,mem_7100lc")
+(define_insn_reservation "Y12" 4
+ (and (eq_attr "type" "fpstore_load")
+ (eq_attr "cpu" "7200"))
+ "i1_7100lc,mem_7100lc,i1_7100lc+mem_7100lc")
+
+(define_insn_reservation "Y13" 4
+ (and (eq_attr "type" "store_fpload")
+ (eq_attr "cpu" "7200"))
+ "i1_7100lc,mem_7100lc,i1_7100lc+mem_7100lc")
+
;; The 7300 has no penalty for store-store or store-load
-(define_insn_reservation "Y10" 2
+(define_insn_reservation "Y14" 2
(and (eq_attr "type" "store")
(eq_attr "cpu" "7300"))
"i1_7100lc")
-(define_insn_reservation "Y11" 2
+(define_insn_reservation "Y15" 2
(and (eq_attr "type" "fpstore")
(eq_attr "cpu" "7300"))
"i1_7100lc")
+(define_insn_reservation "Y16" 4
+ (and (eq_attr "type" "fpstore_load")
+ (eq_attr "cpu" "7300"))
+ "i1_7100lc,i1_7100lc+mem_7100lc")
+
+(define_insn_reservation "Y17" 4
+ (and (eq_attr "type" "store_fpload")
+ (eq_attr "cpu" "7300"))
+ "i1_7100lc,i1_7100lc+mem_7100lc")
+
;; We have an "anti-bypass" for FP loads which feed an FP store.
-(define_bypass 3 "Y3" "Y5,Y9,Y11" "hppa_fpstore_bypass_p")
+(define_bypass 3 "Y3,Y7,Y13,Y17" "Y5,Y6,Y11,Y12,Y15,Y16" "hppa_fpstore_bypass_p")
;; Scheduling for the PA8000 is somewhat different than scheduling for a
;; traditional architecture.
@@ -536,18 +586,23 @@
(eq_attr "cpu" "8000"))
"im_8000,rm_8000+store_8000")
+(define_insn_reservation "Z2" 0
+ (and (eq_attr "type" "fpstore_load,store_fpload")
+ (eq_attr "cpu" "8000"))
+ "im_8000,rm_8000+store_8000,im_8000,rm_8000")
+
;; We can issue and retire two non-memory operations per cycle with
;; a few exceptions (branches). This group catches those we want
;; to assume have zero latency.
-(define_insn_reservation "Z2" 0
+(define_insn_reservation "Z3" 0
(and
- (eq_attr "type" "!load,fpload,store,fpstore,uncond_branch,btable_branch,branch,cbranch,fbranch,call,dyncall,multi,milli,parallel_branch,fpcc,fpalu,fpmulsgl,fpmuldbl,fpsqrtsgl,fpsqrtdbl,fpdivsgl,fpdivdbl")
+ (eq_attr "type" "!load,fpload,store,fpstore,uncond_branch,btable_branch,branch,cbranch,fbranch,call,dyncall,multi,milli,parallel_branch,fpcc,fpalu,fpmulsgl,fpmuldbl,fpsqrtsgl,fpsqrtdbl,fpdivsgl,fpdivdbl,fpstore_load,store_fpload")
(eq_attr "cpu" "8000"))
"inm_8000,rnm_8000")
;; Branches use both slots in the non-memory issue and
;; retirement unit.
-(define_insn_reservation "Z3" 0
+(define_insn_reservation "Z4" 0
(and
(eq_attr "type" "uncond_branch,btable_branch,branch,cbranch,fbranch,call,dyncall,multi,milli,parallel_branch")
(eq_attr "cpu" "8000"))
@@ -557,7 +612,7 @@
;; They can issue/retire two at a time in the non-memory
;; units. We fix their latency at 2 cycles and they
;; are fully pipelined.
-(define_insn_reservation "Z4" 1
+(define_insn_reservation "Z5" 1
(and
(eq_attr "type" "fpcc,fpalu,fpmulsgl,fpmuldbl")
(eq_attr "cpu" "8000"))
@@ -566,13 +621,13 @@
;; The fdivsqrt units are not pipelined and have a very long latency.
;; To keep the DFA from exploding, we do not show all the
;; reservations for the divsqrt unit.
-(define_insn_reservation "Z5" 17
+(define_insn_reservation "Z6" 17
(and
(eq_attr "type" "fpdivsgl,fpsqrtsgl")
(eq_attr "cpu" "8000"))
"inm_8000,fdivsqrt_8000*6,rnm_8000")
-(define_insn_reservation "Z6" 31
+(define_insn_reservation "Z7" 31
(and
(eq_attr "type" "fpdivdbl,fpsqrtdbl")
(eq_attr "cpu" "8000"))
@@ -2503,7 +2558,7 @@
fstw%F0 %1,%0
{fstws|fstw} %1,-16(%%sp)\n\t{ldws|ldw} -16(%%sp),%0
{stws|stw} %1,-16(%%sp)\n\t{fldws|fldw} -16(%%sp),%0"
- [(set_attr "type" "load,move,move,move,shift,load,store,move,move,fpalu,fpload,fpstore,move,move")
+ [(set_attr "type" "load,move,move,move,shift,load,store,move,move,fpalu,fpload,fpstore,fpstore_load,store_fpload")
(set_attr "pa_combine_type" "addmove")
(set_attr "length" "4,4,4,4,4,4,4,4,4,4,4,4,8,8")])
@@ -3142,7 +3197,7 @@
fcpy,sgl %f1,%0
{fstws|fstw} %1,-16(%%sp)\n\t{ldws|ldw} -16(%%sp),%0
{stws|stw} %1,-16(%%sp)\n\t{fldws|fldw} -16(%%sp),%0"
- [(set_attr "type" "move,move,move,shift,load,store,move,move,move,move,move")
+ [(set_attr "type" "move,move,move,shift,load,store,move,move,move,fpstore_load,store_fpload")
(set_attr "pa_combine_type" "addmove")
(set_attr "length" "4,4,4,4,4,4,4,4,4,8,8")])
@@ -3315,7 +3370,7 @@
fcpy,sgl %f1,%0
{fstws|fstw} %1,-16(%%sp)\n\t{ldws|ldw} -16(%%sp),%0
{stws|stw} %1,-16(%%sp)\n\t{fldws|fldw} -16(%%sp),%0"
- [(set_attr "type" "move,move,move,shift,load,store,move,move,move,move,move")
+ [(set_attr "type" "move,move,move,shift,load,store,move,move,move,fpstore_load,store_fpload")
(set_attr "pa_combine_type" "addmove")
(set_attr "length" "4,4,4,4,4,4,4,4,4,8,8")])
@@ -4171,7 +4226,7 @@
return output_fp_move_double (operands);
return output_move_double (operands);
}"
- [(set_attr "type" "fpalu,move,fpstore,store,store,fpload,load,load,move,move")
+ [(set_attr "type" "fpalu,move,fpstore,store,store,fpload,load,load,fpstore_load,store_fpload")
(set_attr "length" "4,8,4,8,16,4,8,16,12,12")])
(define_insn ""
@@ -4505,7 +4560,7 @@
return output_move_double (operands);
}"
[(set_attr "type"
- "move,store,store,load,load,multi,fpalu,fpload,fpstore,move,move")
+ "move,store,store,load,load,multi,fpalu,fpload,fpstore,fpstore_load,store_fpload")
(set_attr "length" "8,8,16,8,16,16,4,4,4,12,12")])
(define_insn ""
@@ -4744,7 +4799,7 @@
stw%M0 %r1,%0
{fstws|fstw} %1,-16(%%sp)\n\t{ldws|ldw} -16(%%sp),%0
{stws|stw} %1,-16(%%sp)\n\t{fldws|fldw} -16(%%sp),%0"
- [(set_attr "type" "fpalu,move,fpload,load,fpstore,store,move,move")
+ [(set_attr "type" "fpalu,move,fpload,load,fpstore,store,fpstore_load,store_fpload")
(set_attr "pa_combine_type" "addmove")
(set_attr "length" "4,4,4,4,4,4,8,8")])
Index: config/pa/pa.c
===================================================================
--- config/pa/pa.c (revision 123037)
+++ config/pa/pa.c (working copy)
@@ -58,7 +58,8 @@
rtx set;
if (recog_memoized (in_insn) < 0
- || get_attr_type (in_insn) != TYPE_FPSTORE
+ || (get_attr_type (in_insn) != TYPE_FPSTORE
+ && get_attr_type (in_insn) != TYPE_FPSTORE_LOAD)
|| recog_memoized (out_insn) < 0)
return 0;
More information about the Gcc-patches
mailing list