k6.md: Rewrite using the DFA model.

	* config/i386/k6.md: Rewrite using the DFA model.
	* config/i386/i386.c (ix86_adjust_cost): Don't increase the
	cost of load-operation insns for the K6.
	(ia32_use_dfa_pipeline_interface): Add TARGET_K6.
	(ia32_multipass_dfa_lookahead): Likewise.

From-SVN: r82859

3 files changed, 267 insertions, 131 deletions

diff --git a/gcc/ChangeLog b/gcc/ChangeLog
index 07072a3..d9e3267 100644
--- a/gcc/ChangeLog
+++ b/gcc/ChangeLog
@@ -1,3 +1,11 @@
+2004-06-09  Steven Bosscher  <stevenb@suse.de>
+
+	* config/i386/k6.md: Rewrite using the DFA model.
+	* config/i386/i386.c (ix86_adjust_cost): Don't increase the
+	cost of load-operation insns for the K6.
+	(ia32_use_dfa_pipeline_interface): Add TARGET_K6.
+	(ia32_multipass_dfa_lookahead): Likewise.
+
 2004-06-09  Richard Henderson  <rth@redhat.com>
 
 	* config/alpha/alpha.c (alpha_gimplify_va_arg_1,
@@ -131,8 +139,8 @@
 
 2004-06-08  Jeff Law  <law@redhat.com>
 
-        * doc/contrib.texi: Add entries for Stefan Olsson and
-        Ola Ronnerup.
+	* doc/contrib.texi: Add entries for Stefan Olsson and
+	Ola Ronnerup.
 
 2004-06-08  DJ Delorie  <dj@redhat.com>
 
@@ -155,13 +163,13 @@
 
 2004-06-08  Richard Henderson  <rth@redhat.com>
 
-        * gimple-low.c (struct lower_data): Replace the_return_label and
-        one_return_stmt with return_statements.
-        (lower_function_body): Process the entire list of return_statements.
-        (lower_return_expr): Check source value before unifying return_exprs.   
-        * gimplify.c (gimplify_return_expr): Force the use of a temporary
-        for !aggregate_value_p.
-        * tree-gimple.c: Update RETURN_EXPR grammer.
+	* gimple-low.c (struct lower_data): Replace the_return_label and
+	one_return_stmt with return_statements.
+	(lower_function_body): Process the entire list of return_statements.
+	(lower_return_expr): Check source value before unifying return_exprs.   
+	* gimplify.c (gimplify_return_expr): Force the use of a temporary
+	for !aggregate_value_p.
+	* tree-gimple.c: Update RETURN_EXPR grammer.
 
 2004-06-08  Vladimir Makarov  <vmakarov@redhat.com>
 
diff --git a/gcc/config/i386/i386.c b/gcc/config/i386/i386.c
index 6e2463b..75e80a6 100644
--- a/gcc/config/i386/i386.c
+++ b/gcc/config/i386/i386.c
@@ -12482,7 +12482,7 @@ static int
 ix86_adjust_cost (rtx insn, rtx link, rtx dep_insn, int cost)
 {
   enum attr_type insn_type, dep_insn_type;
-  enum attr_memory memory, dep_memory;
+  enum attr_memory memory;
   rtx set, set2;
   int dep_insn_code_number;
 
@@ -12550,18 +12550,13 @@ ix86_adjust_cost (rtx insn, rtx link, rtx dep_insn, int cost)
 
     case PROCESSOR_K6:
       memory = get_attr_memory (insn);
-      dep_memory = get_attr_memory (dep_insn);
+
       /* The esp dependency is resolved before the instruction is really
          finished.  */
       if ((insn_type == TYPE_PUSH || insn_type == TYPE_POP)
 	  && (dep_insn_type == TYPE_PUSH || dep_insn_type == TYPE_POP))
 	return 1;
 
-      /* Since we can't represent delayed latencies of load+operation,
-	 increase the cost here for non-imov insns.  */
-      if (dep_memory == MEMORY_LOAD || dep_memory == MEMORY_BOTH)
-	cost += (dep_insn_type != TYPE_IMOV) ? 2 : 1;
-
       /* INT->FP conversion is expensive.  */
       if (get_attr_fp_int_src (dep_insn))
 	cost += 5;
@@ -12587,7 +12582,6 @@ ix86_adjust_cost (rtx insn, rtx link, rtx dep_insn, int cost)
     case PROCESSOR_ATHLON:
     case PROCESSOR_K8:
       memory = get_attr_memory (insn);
-      dep_memory = get_attr_memory (dep_insn);
 
       /* Show ability of reorder buffer to hide latency of load by executing
 	 in parallel with previous instruction in case
@@ -12626,6 +12620,7 @@ ia32_use_dfa_pipeline_interface (void)
 {
   if (TARGET_PENTIUM
       || TARGET_PENTIUMPRO
+      || TARGET_K6
       || TARGET_ATHLON_K8)
     return 1;
   return 0;
@@ -12641,7 +12636,8 @@ ia32_multipass_dfa_lookahead (void)
   if (ix86_tune == PROCESSOR_PENTIUM)
     return 2;
 
-  if (ix86_tune == PROCESSOR_PENTIUMPRO)
+  if (ix86_tune == PROCESSOR_PENTIUMPRO
+      || ix86_tune == PROCESSOR_K6)
     return 1;
 
   else
diff --git a/gcc/config/i386/k6.md b/gcc/config/i386/k6.md
index edd12f6..33d4f1e 100644
--- a/gcc/config/i386/k6.md
+++ b/gcc/config/i386/k6.md
@@ -1,5 +1,6 @@
 ;; AMD K6/K6-2 Scheduling
-;; Copyright (C) 2002 ;; Free Software Foundation, Inc.
+;; Copyright (C) 2002, 2004
+;; Free Software Foundation, Inc.
 ;;
 ;; This file is part of GCC.
 ;;
@@ -16,121 +17,252 @@
 ;; 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, 59 Temple Place - Suite 330,
-;; Boston, MA 02111-1307, USA.  */
+;; Boston, MA 02111-1307, USA.
 ;;
-;; The K6 has similar architecture to PPro.  Important difference is, that
-;; there are only two decoders and they seems to be much slower than execution
-;; units.  So we have to pay much more attention to proper decoding for
-;; schedulers.  We share most of scheduler code for PPro in i386.c
+;; The K6 architecture is quite similar to PPro.  Important difference is
+;; that there are only two decoders and they seems to be much slower than
+;; any of the execution units.  So we have to pay much more attention to
+;; proper scheduling for the decoders.
+;; FIXME: We don't do that right now.  A good start would be to sort the
+;;        instructions based on length.
 ;;
-;; The fp unit is not pipelined and do one operation per two cycles including
-;; the FXCH.
+;; This description is based on data from the following documents:
 ;;
-;; alu	  describes both ALU units (ALU-X and ALU-Y).
-;; alux   describes X alu unit
-;; fpu    describes FPU unit
-;; load   describes load unit.
-;; branch describes branch unit.
-;; store  describes store unit.  This unit is not modelled completely and only
-;;        used to model lea operation.  Otherwise it lie outside of the critical
-;;        path.
+;;    "AMD-K6 Processor Data Sheet (Preliminary information)"
+;;    Advanced Micro Devices, Inc., 1998.
+;;
+;;    "AMD-K6 Processor Code Optimization Application Note"
+;;    Advanced Micro Devices, Inc., 2000.
+;;
+;; CPU execution units of the K6:
+;;
+;; store	describes the Store unit.  This unit is not modelled
+;;		completely and it is only used to model lea operation.
+;;		Otherwise it lies outside of any critical path.
+;; load		describes the Load unit
+;; alux		describes the Integer X unit
+;; mm		describes the Multimedia unit, which shares a pipe
+;;		with the Integer X unit.  This unit is used for MMX,
+;;		which is not implemented for K6.
+;; aluy		describes the Integer Y unit
+;; fpu		describes the FPU unit
+;; branch	describes the Branch unit
+;;
+;; The fp unit is not pipelined, and it can only do one operation per two
+;; cycles, including fxcg.
+;;
+;; Generally this is a very poor description, but at least no worse than
+;; the old description, and a lot easier to extend to something more
+;; reasonable if anyone still cares enough about this architecture in 2004.
 ;;
 ;; ??? fxch isn't handled; not an issue until sched3 after reg-stack is real.
 
-;; The decoder specification is in the PPro section above!
-
-;; Shift instructions and certain arithmetic are issued only to X pipe.
-(define_function_unit "k6_alux" 1 0
-  (and (eq_attr "cpu" "k6")
-       (eq_attr "type" "ishift,ishift1,rotate,rotate1,alu1,negnot,cld"))
-  1 1)
-
-;; The QI mode arithmetic is issued to X pipe only.
-(define_function_unit "k6_alux" 1 0
-  (and (eq_attr "cpu" "k6")
-       (and (eq_attr "type" "alu,alu1,negnot,icmp,test,imovx,incdec")
-	    (eq_attr "mode" "QI")))
-  1 1)
-
-(define_function_unit "k6_alu" 2 0
-  (and (eq_attr "cpu" "k6")
-       (eq_attr "type" "ishift,ishift1,rotate,rotate1,alu1,negnot,alu,icmp,test,imovx,incdec,setcc,lea"))
-  1 1)
-
-(define_function_unit "k6_alu" 2 0
-  (and (eq_attr "cpu" "k6")
-       (and (eq_attr "type" "imov")
-       	    (eq_attr "memory" "none")))
-  1 1)
-
-(define_function_unit "k6_branch" 1 0
-  (and (eq_attr "cpu" "k6")
-       (eq_attr "type" "call,callv,ibr"))
-  1 1)
-
-;; Load unit have two cycle latency, but we take care for it in adjust_cost
-(define_function_unit "k6_load" 1 0
-  (and (eq_attr "cpu" "k6")
-       (ior (eq_attr "type" "pop,leave")
-	    (eq_attr "memory" "load,both")))
-  1 1)
-
-(define_function_unit "k6_load" 1 0
-  (and (eq_attr "cpu" "k6")
-       (and (eq_attr "type" "str")
-	    (eq_attr "memory" "load,both")))
-  10 10)
-
-;; Lea have two instructions, so latency is probably 2
-(define_function_unit "k6_store" 1 0
-  (and (eq_attr "cpu" "k6")
-       (eq_attr "type" "lea"))
-  2 1)
-
-(define_function_unit "k6_store" 1 0
-  (and (eq_attr "cpu" "k6")
-       (eq_attr "type" "str"))
-  10 10)
-
-(define_function_unit "k6_store" 1 0
-  (and (eq_attr "cpu" "k6")
-       (ior (eq_attr "type" "push")
-	    (eq_attr "memory" "store,both")))
-  1 1)
-
-(define_function_unit "k6_fpu" 1 1
-  (and (eq_attr "cpu" "k6")
-       (eq_attr "type" "fop,fmov,fcmp,fistp"))
-  2 2)
-
-(define_function_unit "k6_fpu" 1 1
-  (and (eq_attr "cpu" "k6")
-       (eq_attr "type" "fmul"))
-  2 2)
-
-;; ??? Guess
-(define_function_unit "k6_fpu" 1 1
-  (and (eq_attr "cpu" "k6")
-       (eq_attr "type" "fdiv,fpspc"))
-  56 56)
-
-(define_function_unit "k6_alu" 2 0
-  (and (eq_attr "cpu" "k6")
-       (eq_attr "type" "imul"))
-  2 2)
-
-(define_function_unit "k6_alux" 1 0
-  (and (eq_attr "cpu" "k6")
-       (eq_attr "type" "imul"))
-  2 2)
-
-;; ??? Guess
-(define_function_unit "k6_alu" 2 0
-  (and (eq_attr "cpu" "k6")
-       (eq_attr "type" "idiv"))
-  17 17)
-
-(define_function_unit "k6_alux" 1 0
-  (and (eq_attr "cpu" "k6")
-       (eq_attr "type" "idiv"))
-  17 17)
+(define_automaton "k6_decoder,k6_load_unit,k6_store_unit,k6_integer_units,k6_fpu_unit,k6_branch_unit")
+
+;; The K6 instruction decoding begins before the on-chip instruction cache is
+;; filled.  Depending on the length of the instruction, two simple instructions
+;; can be decoded in two parallel short decoders, or one complex instruction can
+;; be decoded in either the long or the vector decoder.  For all practical
+;; purposes, the long and vector decoder can be modelled as one decoder.
+(define_cpu_unit "k6_decode_short0" "k6_decoder")
+(define_cpu_unit "k6_decode_short1" "k6_decoder")
+(define_cpu_unit "k6_decode_long" "k6_decoder")
+(exclusion_set "k6_decode_long" "k6_decode_short0,k6_decode_short1")
+(define_reservation "k6_decode_short" "k6_decode_short0|k6_decode_short1")
+(define_reservation "k6_decode_vector" "k6_decode_long")
+
+(define_cpu_unit "k6_store" "k6_store_unit")
+(define_cpu_unit "k6_load" "k6_load_unit")
+(define_cpu_unit "k6_alux,k6_aluy" "k6_integer_units")
+(define_cpu_unit "k6_fpu" "k6_fpu_unit")
+(define_cpu_unit "k6_branch" "k6_branch_unit")
+
+;; Shift instructions and certain arithmetic are issued only on Integer X.
+(define_insn_reservation "k6_alux_only" 1
+			 (and (eq_attr "cpu" "k6")
+			      (and (eq_attr "type" "ishift,ishift1,rotate,rotate1,alu1,negnot,cld")
+				   (eq_attr "memory" "none")))
+			 "k6_decode_short,k6_alux")
+
+(define_insn_reservation "k6_alux_only_load" 3
+			 (and (eq_attr "cpu" "k6")
+			       (and (eq_attr "type" "ishift,ishift1,rotate,rotate1,alu1,negnot,cld")
+				    (eq_attr "memory" "load")))
+			 "k6_decode_short,k6_load,k6_alux")
+
+(define_insn_reservation "k6_alux_only_store" 3
+			 (and (eq_attr "cpu" "k6")
+			       (and (eq_attr "type" "ishift,ishift1,rotate,rotate1,alu1,negnot,cld")
+				    (eq_attr "memory" "store,both,unknown")))
+			 "k6_decode_long,k6_load,k6_alux,k6_store")
+
+;; Integer divide and multiply can only be issued on Integer X, too.
+(define_insn_reservation "k6_alu_imul" 2
+			 (and (eq_attr "cpu" "k6")
+			      (eq_attr "type" "imul"))
+			 "k6_decode_vector,k6_alux*3")
+
+(define_insn_reservation "k6_alu_imul_load" 4
+			 (and (eq_attr "cpu" "k6")
+			      (and (eq_attr "type" "imul")
+				   (eq_attr "memory" "load")))
+			 "k6_decode_vector,k6_load,k6_alux*3")
+
+(define_insn_reservation "k6_alu_imul_store" 4
+			 (and (eq_attr "cpu" "k6")
+			      (and (eq_attr "type" "imul")
+				   (eq_attr "memory" "store,both,unknown")))
+			 "k6_decode_vector,k6_load,k6_alux*3,k6_store")
+
+;; ??? Guessed latencies based on the old pipeline description.
+(define_insn_reservation "k6_alu_idiv" 17
+			 (and (eq_attr "cpu" "k6")
+			      (and (eq_attr "type" "idiv")
+				   (eq_attr "memory" "none")))
+			 "k6_decode_vector,k6_alux*17")
+
+(define_insn_reservation "k6_alu_idiv_mem" 19
+			 (and (eq_attr "cpu" "k6")
+			      (and (eq_attr "type" "idiv")
+				   (eq_attr "memory" "!none")))
+			 "k6_decode_vector,k6_load,k6_alux*17")
+
+;; Basic word and doubleword ALU ops can be issued on both Integer units.
+(define_insn_reservation "k6_alu" 1
+			 (and (eq_attr "cpu" "k6")
+			      (and (eq_attr "type" "alu,alu1,negnot,icmp,test,imovx,incdec,setcc")
+				   (eq_attr "memory" "none")))
+			 "k6_decode_short,k6_alux|k6_aluy")
+
+(define_insn_reservation "k6_alu_load" 3
+			 (and (eq_attr "cpu" "k6")
+			      (and (eq_attr "type" "alu,alu1,negnot,icmp,test,imovx,incdec,setcc")
+				   (eq_attr "memory" "load")))
+			 "k6_decode_short,k6_load,k6_alux|k6_aluy")
+
+(define_insn_reservation "k6_alu_store" 3
+			 (and (eq_attr "cpu" "k6")
+			      (and (eq_attr "type" "alu,alu1,negnot,icmp,test,imovx,incdec,setcc")
+				   (eq_attr "memory" "store,both,unknown")))
+			 "k6_decode_long,k6_load,k6_alux|k6_aluy,k6_store")
+
+;; A "load immediate" operation does not require execution at all,
+;; it is available immediately after decoding.  Special-case this.
+(define_insn_reservation "k6_alu_imov" 1
+			 (and (eq_attr "cpu" "k6")
+			      (and (eq_attr "type" "imov")
+				   (and (eq_attr "memory" "none")
+					(match_operand 1 "nonimmediate_operand"))))
+			 "k6_decode_short,k6_alux|k6_aluy")
+
+(define_insn_reservation "k6_alu_imov_imm" 0
+			 (and (eq_attr "cpu" "k6")
+			      (and (eq_attr "type" "imov")
+				   (and (eq_attr "memory" "none")
+					(match_operand 1 "immediate_operand"))))
+			 "k6_decode_short")
+
+(define_insn_reservation "k6_alu_imov_load" 2
+			 (and (eq_attr "cpu" "k6")
+			      (and (eq_attr "type" "imov")
+				   (eq_attr "memory" "load")))
+			 "k6_decode_short,k6_load")
+
+(define_insn_reservation "k6_alu_imov_store" 1
+			 (and (eq_attr "cpu" "k6")
+			      (and (eq_attr "type" "imov")
+				   (eq_attr "memory" "store")))
+			 "k6_decode_short,k6_store")
+
+(define_insn_reservation "k6_alu_imov_both" 2
+			 (and (eq_attr "cpu" "k6")
+			      (and (eq_attr "type" "imov")
+				   (eq_attr "memory" "both,unknown")))
+			 "k6_decode_long,k6_load,k6_alux|k6_aluy")
+
+;; The branch unit.
+(define_insn_reservation "k6_branch_call" 1
+			 (and (eq_attr "cpu" "k6")
+			      (eq_attr "type" "call,callv"))
+			 "k6_decode_vector,k6_branch")
+
+(define_insn_reservation "k6_branch_branch" 1
+			 (and (eq_attr "cpu" "k6")
+			      (eq_attr "type" "ibr"))
+			 "k6_decode_short,k6_branch")
+
+;; The load and units have two pipeline stages.  The load latency is
+;; two cycles.
+(define_insn_reservation "k6_load_pop" 3
+			 (and (eq_attr "cpu" "k6")
+			      (ior (eq_attr "type" "pop")
+				   (eq_attr "memory" "load,both")))
+			 "k6_decode_short,k6_load")
+
+(define_insn_reservation "k6_load_leave" 5
+			 (and (eq_attr "cpu" "k6")
+			      (eq_attr "type" "leave"))
+			 "k6_decode_long,k6_load,(k6_alux|k6_aluy)*2")
+
+;; ??? From the old pipeline description.  Egad!
+;; ??? Apparently we take care of this reservation in adjust_cost.
+(define_insn_reservation "k6_load_str" 10
+			 (and (eq_attr "cpu" "k6")
+			      (and (eq_attr "type" "str")
+				   (eq_attr "memory" "load,both")))
+			 "k6_decode_vector,k6_load*10")
+
+;; The store unit handles lea and push.  It is otherwise unmodelled.
+(define_insn_reservation "k6_store_lea" 2
+			 (and (eq_attr "cpu" "k6")
+			      (eq_attr "type" "lea"))
+			 "k6_decode_short,k6_store,k6_alux|k6_aluy")
+
+(define_insn_reservation "k6_store_push" 2
+			 (and (eq_attr "cpu" "k6")
+			      (ior (eq_attr "type" "push")
+				   (eq_attr "memory" "store,both")))
+			 "k6_decode_short,k6_store")
+
+(define_insn_reservation "k6_store_str" 10
+			 (and (eq_attr "cpu" "k6")
+			      (eq_attr "type" "str"))
+			 "k6_store*10")
+
+;; Most FPU instructions have latency 2 and throughput 2.
+(define_insn_reservation "k6_fpu" 2
+			 (and (eq_attr "cpu" "k6")
+			      (and (eq_attr "type" "fop,fmov,fcmp,fistp")
+				   (eq_attr "memory" "none")))
+			 "k6_decode_vector,k6_fpu*2")
+
+(define_insn_reservation "k6_fpu_load" 6
+			 (and (eq_attr "cpu" "k6")
+			      (and (eq_attr "type" "fop,fmov,fcmp,fistp")
+				   (eq_attr "memory" "load,both")))
+			 "k6_decode_short,k6_load,k6_fpu*2")
+
+(define_insn_reservation "k6_fpu_store" 6
+			 (and (eq_attr "cpu" "k6")
+			      (and (eq_attr "type" "fop,fmov,fcmp,fistp")
+				   (eq_attr "memory" "store")))
+			 "k6_decode_short,k6_store,k6_fpu*2")
+
+(define_insn_reservation "k6_fpu_fmul" 2
+			 (and (eq_attr "cpu" "k6")
+			      (and (eq_attr "type" "fmul")
+				   (eq_attr "memory" "none")))
+			 "k6_decode_short,k6_fpu*2")
+
+(define_insn_reservation "k6_fpu_fmul_load" 2
+			 (and (eq_attr "cpu" "k6")
+			      (and (eq_attr "type" "fmul")
+				   (eq_attr "memory" "load,both")))
+			 "k6_decode_short,k6_load,k6_fpu*2")
+
+;; ??? Guessed latencies from the old pipeline description.
+(define_insn_reservation "k6_fpu_expensive" 56
+			 (and (eq_attr "cpu" "k6")
+			      (eq_attr "type" "fdiv,fpspc"))
+			 "k6_decode_short,k6_fpu*56")
+