cortex-r4.md: New.

2008-03-10  Paul Brook  <paul@codesourcery.com>
	Mark Shinwell  <shinwell@codesourcery.com>

	gcc/
	* config/arm/cortex-r4.md: New.
	* config/arm/thumb2.md (divsi3, udivsi3): Annotate with
	insn attributes.
	* config/arm/arm.md: Include cortex-r4.md.
	(insn): Add smmls, sdiv and udiv values.
	(generic_sched): Don't use generic scheduling for Cortex-R4.
	(arm_issue_rate): New function.
	(TARGET_SCHED_ISSUE_RATE): Define.


Co-Authored-By: Mark Shinwell <shinwell@codesourcery.com>

From-SVN: r133078

5 files changed, 331 insertions, 6 deletions

diff --git a/gcc/ChangeLog b/gcc/ChangeLog
index 052d616..5eab262 100644
--- a/gcc/ChangeLog
+++ b/gcc/ChangeLog
@@ -1,3 +1,16 @@
+2008-03-10  Paul Brook  <paul@codesourcery.com>
+	Mark Shinwell  <shinwell@codesourcery.com>
+
+	gcc/
+	* config/arm/cortex-r4.md: New.
+	* config/arm/thumb2.md (divsi3, udivsi3): Annotate with
+	insn attributes.
+	* config/arm/arm.md: Include cortex-r4.md.
+	(insn): Add smmls, sdiv and udiv values.
+	(generic_sched): Don't use generic scheduling for Cortex-R4.
+	(arm_issue_rate): New function.
+	(TARGET_SCHED_ISSUE_RATE): Define.
+
 2008-03-10  Sebastian Pop  <sebastian.pop@amd.com>
 
 	* doc/invoke.texi (-ftree-loop-distribution): Add an example.
diff --git a/gcc/config/arm/arm.c b/gcc/config/arm/arm.c
index fc11cb1..bddb0e2 100644
--- a/gcc/config/arm/arm.c
+++ b/gcc/config/arm/arm.c
@@ -188,6 +188,7 @@ static void arm_target_help (void);
 static unsigned HOST_WIDE_INT arm_shift_truncation_mask (enum machine_mode);
 static bool arm_cannot_copy_insn_p (rtx);
 static bool arm_tls_symbol_p (rtx x);
+static int arm_issue_rate (void);
 static void arm_output_dwarf_dtprel (FILE *, int, rtx) ATTRIBUTE_UNUSED;
 
 
@@ -358,6 +359,9 @@ static void arm_output_dwarf_dtprel (FILE *, int, rtx) ATTRIBUTE_UNUSED;
 #undef TARGET_CANNOT_FORCE_CONST_MEM
 #define TARGET_CANNOT_FORCE_CONST_MEM arm_cannot_force_const_mem
 
+#undef TARGET_SCHED_ISSUE_RATE
+#define TARGET_SCHED_ISSUE_RATE arm_issue_rate
+
 #undef TARGET_MANGLE_TYPE
 #define TARGET_MANGLE_TYPE arm_mangle_type
 
@@ -18710,6 +18714,22 @@ thumb2_output_casesi (rtx *operands)
     }
 }
 
+/* Most ARM cores are single issue, but some newer ones can dual issue.
+   The scheduler descriptions rely on this being correct.  */
+static int
+arm_issue_rate (void)
+{
+  switch (arm_tune)
+    {
+    case cortexr4:
+    case cortexa8:
+      return 2;
+
+    default:
+      return 1;
+    }
+}
+
 /* A table and a function to perform ARM-specific name mangling for
    NEON vector types in order to conform to the AAPCS (see "Procedure
    Call Standard for the ARM Architecture", Appendix A).  To qualify
diff --git a/gcc/config/arm/arm.md b/gcc/config/arm/arm.md
index 765b89b..6eef6502 100644
--- a/gcc/config/arm/arm.md
+++ b/gcc/config/arm/arm.md
@@ -1,6 +1,7 @@
 ;;- Machine description for ARM for GNU compiler
 ;;  Copyright 1991, 1993, 1994, 1995, 1996, 1996, 1997, 1998, 1999, 2000,
-;;  2001, 2002, 2003, 2004, 2005, 2006, 2007 Free Software Foundation, Inc.
+;;  2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008
+;;  Free Software Foundation, Inc.
 ;;  Contributed by Pieter `Tiggr' Schoenmakers (rcpieter@win.tue.nl)
 ;;  and Martin Simmons (@harleqn.co.uk).
 ;;  More major hacks by Richard Earnshaw (rearnsha@arm.com).
@@ -183,7 +184,7 @@
 ;; scheduling information.
 
 (define_attr "insn"
-        "mov,mvn,smulxy,smlaxy,smlalxy,smulwy,smlawx,mul,muls,mla,mlas,umull,umulls,umlal,umlals,smull,smulls,smlal,smlals,smlawy,smuad,smuadx,smlad,smladx,smusd,smusdx,smlsd,smlsdx,smmul,smmulr,smmla,umaal,smlald,smlsld,clz,mrs,msr,xtab,other"
+        "mov,mvn,smulxy,smlaxy,smlalxy,smulwy,smlawx,mul,muls,mla,mlas,umull,umulls,umlal,umlals,smull,smulls,smlal,smlals,smlawy,smuad,smuadx,smlad,smladx,smusd,smusdx,smlsd,smlsdx,smmul,smmulr,smmla,umaal,smlald,smlsld,clz,mrs,msr,xtab,sdiv,udiv,other"
         (const_string "other"))
 
 ; TYPE attribute is used to detect floating point instructions which, if
@@ -332,7 +333,7 @@
 
 (define_attr "generic_sched" "yes,no"
   (const (if_then_else 
-          (eq_attr "tune" "arm926ejs,arm1020e,arm1026ejs,arm1136js,arm1136jfs,cortexa8")
+          (eq_attr "tune" "arm926ejs,arm1020e,arm1026ejs,arm1136js,arm1136jfs,cortexa8,cortexr4")
           (const_string "no")
           (const_string "yes"))))
 
@@ -349,6 +350,7 @@
 (include "arm1026ejs.md")
 (include "arm1136jfs.md")
 (include "cortex-a8.md")
+(include "cortex-r4.md")
 
 
 ;;---------------------------------------------------------------------------
diff --git a/gcc/config/arm/cortex-r4.md b/gcc/config/arm/cortex-r4.md
new file mode 100644
index 0000000..d912f20
--- /dev/null
+++ b/gcc/config/arm/cortex-r4.md
@@ -0,0 +1,288 @@
+;; ARM Cortex-R4 scheduling description.
+;; Copyright (C) 2007, 2008 Free Software Foundation, Inc.
+;; Contributed by CodeSourcery.
+
+;; This file is part of GCC.
+
+;; GCC is distributed in the hope that it will be useful, but WITHOUT
+;; ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
+;; or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public
+;; License for more details.
+
+;; You should have received a copy of the GNU General Public License
+;; along with GCC; see the file COPYING.  If not, write to
+;; the Free Software Foundation, 51 Franklin Street, Fifth Floor,
+;; Boston, MA 02110-1301, USA.
+
+(define_automaton "cortex_r4")
+
+;; We approximate the dual-issue constraints of this core using four
+;; "issue units" and a reservation matrix as follows.  The numbers indicate
+;; the instruction groups' preferences in order.  Multiple entries for
+;; the same numbered preference indicate units that must be reserved
+;; together.
+;;
+;; Issue unit:		A	B	C	ALU
+;;
+;; ALU w/o reg shift	1st	2nd		1st and 2nd
+;; ALU w/ reg shift	1st	2nd	2nd	1st and 2nd
+;; Moves		1st	2nd		2nd
+;; Multiplication	1st			1st
+;; Division		1st			1st
+;; Load/store single	1st		1st
+;; Other load/store	1st	1st
+;; Branches			1st
+
+(define_cpu_unit "cortex_r4_issue_a" "cortex_r4")
+(define_cpu_unit "cortex_r4_issue_b" "cortex_r4")
+(define_cpu_unit "cortex_r4_issue_c" "cortex_r4")
+(define_cpu_unit "cortex_r4_issue_alu" "cortex_r4")
+
+(define_reservation "cortex_r4_alu"
+                    "(cortex_r4_issue_a+cortex_r4_issue_alu)|\
+                     (cortex_r4_issue_b+cortex_r4_issue_alu)")
+(define_reservation "cortex_r4_alu_shift_reg"
+                    "(cortex_r4_issue_a+cortex_r4_issue_alu)|\
+                     (cortex_r4_issue_b+cortex_r4_issue_c+\
+                      cortex_r4_issue_alu)")
+(define_reservation "cortex_r4_mov"
+                    "cortex_r4_issue_a|(cortex_r4_issue_b+\
+                     cortex_r4_issue_alu)")
+(define_reservation "cortex_r4_mul" "cortex_r4_issue_a+cortex_r4_issue_alu")
+(define_reservation "cortex_r4_mul_2"
+                    "(cortex_r4_issue_a+cortex_r4_issue_alu)*2")
+;; Division instructions execute out-of-order with respect to the
+;; rest of the pipeline and only require reservations on their first and
+;; final cycles.
+(define_reservation "cortex_r4_div_9"
+                    "cortex_r4_issue_a+cortex_r4_issue_alu,\
+                     nothing*7,\
+                     cortex_r4_issue_a+cortex_r4_issue_alu")
+(define_reservation "cortex_r4_div_10"
+                    "cortex_r4_issue_a+cortex_r4_issue_alu,\
+                     nothing*8,\
+                     cortex_r4_issue_a+cortex_r4_issue_alu")
+(define_reservation "cortex_r4_load_store"
+                    "cortex_r4_issue_a+cortex_r4_issue_c")
+(define_reservation "cortex_r4_load_store_2"
+                    "(cortex_r4_issue_a+cortex_r4_issue_b)*2")
+(define_reservation "cortex_r4_branch" "cortex_r4_issue_b")
+
+;; We assume that all instructions are unconditional.
+
+;; Data processing instructions.  Moves without shifts are kept separate
+;; for the purposes of the dual-issue constraints above.
+(define_insn_reservation "cortex_r4_alu" 2
+  (and (eq_attr "tune" "cortexr4")
+       (and (eq_attr "type" "alu")
+            (not (eq_attr "insn" "mov"))))
+  "cortex_r4_alu")
+
+(define_insn_reservation "cortex_r4_mov" 2
+  (and (eq_attr "tune" "cortexr4")
+       (and (eq_attr "type" "alu")
+            (eq_attr "insn" "mov")))
+  "cortex_r4_mov")
+
+(define_insn_reservation "cortex_r4_alu_shift" 2
+  (and (eq_attr "tune" "cortexr4")
+       (eq_attr "type" "alu_shift"))
+  "cortex_r4_alu")
+
+(define_insn_reservation "cortex_r4_alu_shift_reg" 2
+  (and (eq_attr "tune" "cortexr4")
+       (eq_attr "type" "alu_shift_reg"))
+  "cortex_r4_alu_shift_reg")
+
+;; An ALU instruction followed by an ALU instruction with no early dep.
+(define_bypass 1 "cortex_r4_alu,cortex_r4_alu_shift,cortex_r4_alu_shift_reg,\
+                  cortex_r4_mov"
+               "cortex_r4_alu")
+(define_bypass 1 "cortex_r4_alu,cortex_r4_alu_shift,cortex_r4_alu_shift_reg,\
+                  cortex_r4_mov"
+               "cortex_r4_alu_shift"
+               "arm_no_early_alu_shift_dep")
+(define_bypass 1 "cortex_r4_alu,cortex_r4_alu_shift,cortex_r4_alu_shift_reg,\
+                  cortex_r4_mov"
+               "cortex_r4_alu_shift_reg"
+               "arm_no_early_alu_shift_value_dep")
+
+;; In terms of availabilities, a consumer mov could theoretically be
+;; issued together with a producer ALU instruction, without stalls.
+;; In practice this cannot happen because mov;add (in that order) is not
+;; eligible for dual issue and furthermore dual issue is not permitted
+;; when a dependency is involved.  We therefore note it as latency one.
+;; A mov followed by another of the same is also latency one.
+(define_bypass 1 "cortex_r4_alu,cortex_r4_alu_shift,cortex_r4_alu_shift_reg,\
+                  cortex_r4_mov"
+               "cortex_r4_mov")
+
+;; qadd, qdadd, qsub and qdsub are not currently emitted, and neither are
+;; media data processing instructions nor sad instructions.
+
+;; Multiplication instructions.
+
+(define_insn_reservation "cortex_r4_mul_4" 4
+  (and (eq_attr "tune" "cortexr4")
+       (eq_attr "insn" "mul,smmul"))
+  "cortex_r4_mul_2")
+
+(define_insn_reservation "cortex_r4_mul_3" 3
+  (and (eq_attr "tune" "cortexr4")
+       (eq_attr "insn" "smulxy,smulwy,smuad,smusd"))
+  "cortex_r4_mul")
+
+(define_insn_reservation "cortex_r4_mla_4" 4
+  (and (eq_attr "tune" "cortexr4")
+       (eq_attr "insn" "mla,smmla"))
+  "cortex_r4_mul_2")
+
+(define_insn_reservation "cortex_r4_mla_3" 3
+  (and (eq_attr "tune" "cortexr4")
+       (eq_attr "insn" "smlaxy,smlawy,smlad,smlsd"))
+  "cortex_r4_mul")
+
+(define_insn_reservation "cortex_r4_smlald" 3
+  (and (eq_attr "tune" "cortexr4")
+       (eq_attr "insn" "smlald,smlsld"))
+  "cortex_r4_mul")
+
+(define_insn_reservation "cortex_r4_mull" 4
+  (and (eq_attr "tune" "cortexr4")
+       (eq_attr "insn" "smull,umull,umlal,umaal"))
+  "cortex_r4_mul_2")
+
+;; A multiply or an MLA with a single-register result, followed by an
+;; MLA with an accumulator dependency, has its result forwarded.
+(define_bypass 2 "cortex_r4_mul_3,cortex_r4_mla_3"
+               "cortex_r4_mla_3,cortex_r4_mla_4"
+               "arm_mac_accumulator_is_mul_result")
+
+(define_bypass 3 "cortex_r4_mul_4,cortex_r4_mla_4"
+               "cortex_r4_mla_3,cortex_r4_mla_4"
+               "arm_mac_accumulator_is_mul_result")
+
+;; A multiply followed by an ALU instruction needing the multiply
+;; result only at ALU has lower latency than one needing it at Shift.
+(define_bypass 2 "cortex_r4_mul_3,cortex_r4_mla_3,cortex_r4_smlald"
+               "cortex_r4_alu")
+(define_bypass 2 "cortex_r4_mul_3,cortex_r4_mla_3,cortex_r4_smlald"
+               "cortex_r4_alu_shift"
+               "arm_no_early_alu_shift_dep")
+(define_bypass 2 "cortex_r4_mul_3,cortex_r4_mla_3,cortex_r4_smlald"
+               "cortex_r4_alu_shift_reg"
+               "arm_no_early_alu_shift_value_dep")
+(define_bypass 3 "cortex_r4_mul_4,cortex_r4_mla_4,cortex_r4_mull"
+               "cortex_r4_alu")
+(define_bypass 3 "cortex_r4_mul_4,cortex_r4_mla_4,cortex_r4_mull"
+               "cortex_r4_alu_shift"
+               "arm_no_early_alu_shift_dep")
+(define_bypass 3 "cortex_r4_mul_4,cortex_r4_mla_4,cortex_r4_mull"
+               "cortex_r4_alu_shift_reg"
+               "arm_no_early_alu_shift_value_dep")
+
+;; A multiply followed by a mov has one cycle lower latency again.
+(define_bypass 1 "cortex_r4_mul_3,cortex_r4_mla_3,cortex_r4_smlald"
+               "cortex_r4_mov")
+(define_bypass 2 "cortex_r4_mul_4,cortex_r4_mla_4,cortex_r4_mull"
+               "cortex_r4_mov")
+
+;; We guess that division of A/B using sdiv or udiv, on average, 
+;; is performed with B having ten more leading zeros than A.
+;; This gives a latency of nine for udiv and ten for sdiv.
+(define_insn_reservation "cortex_r4_udiv" 9
+  (and (eq_attr "tune" "cortexr4")
+       (eq_attr "insn" "udiv"))
+  "cortex_r4_div_9")
+
+(define_insn_reservation "cortex_r4_sdiv" 10
+  (and (eq_attr "tune" "cortexr4")
+       (eq_attr "insn" "sdiv"))
+  "cortex_r4_div_10")
+
+;; Branches.  We assume correct prediction.
+
+(define_insn_reservation "cortex_r4_branch" 0
+  (and (eq_attr "tune" "cortexr4")
+       (eq_attr "type" "branch"))
+  "cortex_r4_branch")
+
+;; Call latencies are not predictable.  A semi-arbitrary very large
+;; number is used as "positive infinity" so that everything should be
+;; finished by the time of return.
+(define_insn_reservation "cortex_r4_call" 32
+  (and (eq_attr "tune" "cortexr4")
+       (eq_attr "type" "call"))
+  "nothing")
+
+;; Status register access instructions are not currently emitted.
+
+;; Load instructions.
+;; We do not model the "addr_md_3cycle" cases and assume that
+;; accesses following are correctly aligned.
+
+(define_insn_reservation "cortex_r4_load_1_2" 3
+  (and (eq_attr "tune" "cortexr4")
+       (eq_attr "type" "load1,load2"))
+  "cortex_r4_load_store")
+
+(define_insn_reservation "cortex_r4_load_3_4" 4
+  (and (eq_attr "tune" "cortexr4")
+       (eq_attr "type" "load3,load4"))
+  "cortex_r4_load_store_2")
+
+;; If a producing load is followed by an instruction consuming only
+;; as a Normal Reg, there is one fewer cycle of latency.
+
+(define_bypass 2 "cortex_r4_load_1_2"
+               "cortex_r4_alu")
+(define_bypass 2 "cortex_r4_load_1_2"
+               "cortex_r4_alu_shift"
+               "arm_no_early_alu_shift_dep")
+(define_bypass 2 "cortex_r4_load_1_2"
+               "cortex_r4_alu_shift_reg"
+               "arm_no_early_alu_shift_value_dep")
+
+(define_bypass 3 "cortex_r4_load_3_4"
+               "cortex_r4_alu")
+(define_bypass 3 "cortex_r4_load_3_4"
+               "cortex_r4_alu_shift"
+               "arm_no_early_alu_shift_dep")
+(define_bypass 3 "cortex_r4_load_3_4"
+               "cortex_r4_alu_shift_reg"
+               "arm_no_early_alu_shift_value_dep")
+
+;; If a producing load is followed by an instruction consuming only
+;; as a Late Reg, there are two fewer cycles of latency.  Such consumer
+;; instructions are moves and stores.
+
+(define_bypass 1 "cortex_r4_load_1_2"
+               "cortex_r4_mov,cortex_r4_store_1_2,cortex_r4_store_3_4")
+(define_bypass 2 "cortex_r4_load_3_4"
+               "cortex_r4_mov,cortex_r4_store_1_2,cortex_r4_store_3_4")
+
+;; If a producer's result is required as the base or offset of a load,
+;; there is an extra cycle latency.
+
+(define_bypass 3 "cortex_r4_alu,cortex_r4_mov,cortex_r4_alu_shift,\
+                  cortex_r4_alu_shift_reg"
+               "cortex_r4_load_1_2,cortex_r4_load_3_4")
+
+(define_bypass 4 "cortex_r4_mul_3,cortex_r4_mla_3,cortex_r4_smlald"
+               "cortex_r4_load_1_2,cortex_r4_load_3_4")
+
+(define_bypass 5 "cortex_r4_mul_4,cortex_r4_mla_4,cortex_r4_mull"
+               "cortex_r4_load_1_2,cortex_r4_load_3_4")
+
+;; Store instructions.
+
+(define_insn_reservation "cortex_r4_store_1_2" 0
+  (and (eq_attr "tune" "cortexr4")
+       (eq_attr "type" "store1,store2"))
+  "cortex_r4_load_store")
+
+(define_insn_reservation "cortex_r4_store_3_4" 0
+  (and (eq_attr "tune" "cortexr4")
+       (eq_attr "type" "store3,store4"))
+  "cortex_r4_load_store_2")
+
diff --git a/gcc/config/arm/thumb2.md b/gcc/config/arm/thumb2.md
index 0efe31f..ba45c88 100644
--- a/gcc/config/arm/thumb2.md
+++ b/gcc/config/arm/thumb2.md
@@ -1,5 +1,5 @@
 ;; ARM Thumb-2 Machine Description
-;; Copyright (C) 2007 Free Software Foundation, Inc.
+;; Copyright (C) 2007, 2008 Free Software Foundation, Inc.
 ;; Written by CodeSourcery, LLC.
 ;;
 ;; This file is part of GCC.
@@ -1131,7 +1131,8 @@
 		(match_operand:SI 2 "s_register_operand"  "r")))]
   "TARGET_THUMB2 && arm_arch_hwdiv"
   "sdiv%?\t%0, %1, %2"
-  [(set_attr "predicable" "yes")]
+  [(set_attr "predicable" "yes")
+   (set_attr "insn" "sdiv")]
 )
 
 (define_insn "udivsi3"
@@ -1140,7 +1141,8 @@
 		 (match_operand:SI 2 "s_register_operand"  "r")))]
   "TARGET_THUMB2 && arm_arch_hwdiv"
   "udiv%?\t%0, %1, %2"
-  [(set_attr "predicable" "yes")]
+  [(set_attr "predicable" "yes")
+   (set_attr "insn" "udiv")]
 )
 
 (define_insn "*thumb2_cbz"