;; AMD K6/K6-2 Scheduling
;; Copyright (C) 2002, 2004
;; Free Software Foundation, Inc.
;;
;; This file is part of GCC.
;;
;; GCC is free software; you can redistribute it and/or modify
;; it under the terms of the GNU General Public License as published by
;; the Free Software Foundation; either version 2, or (at your option)
;; any later version.
;;
;; 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.
;;
;; 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.
;;
;; This description is based on data from the following documents:
;;
;; "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.
(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")