;;- Machine description for Hitachi / SuperH SH.
-;; Copyright (C) 1993, 1994, 1995, 1996, 1997, 1998, 1999, 2000, 2001, 2002
+;; Copyright (C) 1993, 1994, 1995, 1996, 1997, 1998, 1999, 2000, 2001, 2002, 2003
;; Free Software Foundation, Inc.
;; Contributed by Steve Chamberlain (sac@cygnus.com).
;; Improved by Jim Wilson (wilson@cygnus.com).
(UNSPEC_FSINA 16)
(UNSPEC_NSB 17)
(UNSPEC_ALLOCO 18)
+ (UNSPEC_EH_RETURN 19)
+ (UNSPEC_TLSGD 20)
+ (UNSPEC_TLSLDM 21)
+ (UNSPEC_TLSIE 22)
+ (UNSPEC_DTPOFF 23)
+ (UNSPEC_GOTTPOFF 24)
+ (UNSPEC_TPOFF 25)
;; These are used with unspec_volatile.
(UNSPECV_BLOCKAGE 0)
;; Target CPU.
(define_attr "cpu"
- "sh1,sh2,sh3,sh3e,sh4,sh5"
+ "sh1,sh2,sh2e,sh3,sh3e,sh4,sh5"
(const (symbol_ref "sh_cpu_attr")))
(define_attr "endian" "big,little"
;; arith3b like above, but might end with a redirected branch
;; load from memory
;; load_si Likewise, SImode variant for general register.
+;; fload Likewise, but load to fp register.
;; store to memory
-;; move register to register
+;; move general purpose register to register
+;; mt_group other sh4 mt instructions
;; fmove register to register, floating point
;; smpy word precision integer multiply
;; dmpy longword or doublelongword precision integer multiply
;; pstore store of pr reg, which can't be put into delay slot of jsr
;; prget copy pr to register, ditto
;; pcload pc relative load of constant value
+;; pcfload Likewise, but load to fp register.
;; pcload_si Likewise, SImode variant for general register.
;; rte return from exception
;; sfunc special function call with known used registers
;; call function call
;; fp floating point
;; fdiv floating point divide (or square root)
-;; gp_fpul move between general purpose register and fpul
+;; gp_fpul move from general purpose register to fpul
+;; fpul_gp move from fpul to general purpose register
+;; mac_gp move from mac[lh] to general purpose register
;; dfp_arith, dfp_cmp,dfp_conv
+;; ftrc_s fix_truncsfsi2_i4
;; dfdiv double precision floating point divide (or square root)
+;; cwb ic_invalidate_line_i
+;; tls_load load TLS related address
;; arith_media SHmedia arithmetic, logical, and shift instructions
;; cbranch_media SHmedia conditional branch instructions
;; cmp_media SHmedia compare instructions
;; fpconv_media SHmedia single precision floating point conversions
;; fstore_media SHmedia floating point register store instructions
;; gettr_media SHmedia gettr instruction
-;; invalidate_line_media SHmedia invaldiate_line sequence
+;; invalidate_line_media SHmedia invalidate_line sequence
;; jump_media SHmedia unconditional branch instructions
;; load_media SHmedia general register load instructions
;; pt_media SHmedia pt instruction (expanded by assembler)
;; mcmp_media SHmedia multimedia compare, absolute, saturating ops
;; mac_media SHmedia mac-style fixed point operations
;; d2mpy_media SHmedia: two 32 bit integer multiplies
-;; atrans SHmedia approximate transcendential functions
+;; atrans SHmedia approximate transcendental functions
;; ustore_media SHmedia unaligned stores
;; nil no-op move, will be deleted.
(define_attr "type"
- "cbranch,jump,jump_ind,arith,arith3,arith3b,dyn_shift,load,load_si,store,move,fmove,smpy,dmpy,return,pload,prset,pstore,prget,pcload,pcload_si,rte,sfunc,call,fp,fdiv,dfp_arith,dfp_cmp,dfp_conv,dfdiv,gp_fpul,arith_media,cbranch_media,cmp_media,dfdiv_media,dfmul_media,dfparith_media,dfpconv_media,dmpy_media,fcmp_media,fdiv_media,fload_media,fmove_media,fparith_media,fpconv_media,fstore_media,gettr_media,invalidate_line_media,jump_media,load_media,pt_media,ptabs_media,store_media,mcmp_media,mac_media,d2mpy_media,atrans_media,ustore_media,nil,other"
+ "mt_group,cbranch,jump,jump_ind,arith,arith3,arith3b,dyn_shift,load,load_si,fload,store,move,fmove,smpy,dmpy,return,pload,prset,pstore,prget,pcload,pcload_si,pcfload,rte,sfunc,call,fp,fdiv,ftrc_s,dfp_arith,dfp_cmp,dfp_conv,dfdiv,gp_fpul,fpul_gp,mac_gp,mem_fpscr,gp_fpscr,cwb,tls_load,arith_media,cbranch_media,cmp_media,dfdiv_media,dfmul_media,dfparith_media,dfpconv_media,dmpy_media,fcmp_media,fdiv_media,fload_media,fmove_media,fparith_media,fpconv_media,fstore_media,gettr_media,invalidate_line_media,jump_media,load_media,pt_media,ptabs_media,store_media,mcmp_media,mac_media,d2mpy_media,atrans_media,ustore_media,nil,other"
(const_string "other"))
;; We define a new attribute namely "insn_class".We use
-;; this for DFA based pipeline description.
-;; Although the "type" attribute covers almost all insn
-;; classes,it is more convenient to define new attribute
-;; for certain reservations.
+;; this for the DFA based pipeline description.
;;
;; mt_group SH4 "mt" group instructions.
;;
-;; ex_group SH4 "ex" group instructions.They mostly
-;; overlap with arithmetic instructions but
-;; new attribute defined to distinguish from
-;; mt group instructions.
+;; ex_group SH4 "ex" group instructions.
+;;
+;; ls_group SH4 "ls" group instructions.
;;
-;; lds_to_fpscr The "type" attribute couldn't sufficiently
-;; distinguish it from others.It is part of
-;; new attribute.Similar case with ldsmem_to_fpscr
-;; and cwb.
(define_attr "insn_class"
- "mt_group,ex_group,lds_to_fpscr,ldsmem_to_fpscr,cwb,none"
- (const_string "none"))
+ "mt_group,ex_group,ls_group,br_group,fe_group,co_group,none"
+ (cond [(eq_attr "type" "move,mt_group") (const_string "mt_group")
+ (eq_attr "type" "arith,dyn_shift") (const_string "ex_group")
+ (eq_attr "type" "fmove,load,pcload,load_si,pcload_si,fload,pcfload,store,gp_fpul,fpul_gp") (const_string "ls_group")
+ (eq_attr "type" "cbranch,jump") (const_string "br_group")
+ (eq_attr "type" "fp,fdiv,ftrc_s,dfp_arith,dfp_conv,dfdiv")
+ (const_string "fe_group")
+ (eq_attr "type" "jump_ind,smpy,dmpy,mac_gp,return,pload,prset,pstore,prget,rte,sfunc,call,dfp_cmp,mem_fpscr,gp_fpscr,cwb") (const_string "co_group")]
+ (const_string "none")))
+;; nil are zero instructions, and arith3 / arith3b are multiple instructions,
+;; so these do not belong in an insn group, although they are modeled
+;; with their own define_insn_reservations.
;; Indicate what precision must be selected in fpscr for this insn, if any.
(eq_attr "type" "jump")
(cond [(eq_attr "med_branch_p" "yes")
(const_int 2)
- (and (eq (symbol_ref "GET_CODE (PREV_INSN (insn))")
+ (and (eq (symbol_ref "GET_CODE (prev_nonnote_insn (insn))")
(symbol_ref "INSN"))
- (eq (symbol_ref "INSN_CODE (PREV_INSN (insn))")
+ (eq (symbol_ref "INSN_CODE (prev_nonnote_insn (insn))")
(symbol_ref "code_for_indirect_jump_scratch")))
(if_then_else (eq_attr "braf_branch_p" "yes")
(const_int 6)
(and (eq_attr "pipe_model" "sh1") (eq_attr "type" "fdiv")) 13 12)
-;; SH4 scheduling
-;; The SH4 is a dual-issue implementation, thus we have to multiply all
-;; costs by at least two.
-;; There will be single increments of the modeled that don't correspond
-;; to the actual target ;; whenever two insns to be issued depend one a
-;; single resource, and the scheduler picks to be the first one.
-;; If we multiplied the costs just by two, just two of these single
-;; increments would amount to an actual cycle. By picking a larger
-;; factor, we can ameliorate the effect; However, we then have to make sure
-;; that only two insns are modeled as issued per actual cycle.
-;; Moreover, we need a way to specify the latency of insns that don't
-;; use an actual function unit.
-;; We use an 'issue' function unit to do that, and a cost factor of 10.
-
-(define_function_unit "issue" 2 0
- (and (eq_attr "pipe_model" "sh4") (eq_attr "type" "!nil,arith3"))
- 10 10)
-
-(define_function_unit "issue" 2 0
- (and (eq_attr "pipe_model" "sh4") (eq_attr "type" "arith3"))
- 30 30)
-
-;; There is no point in providing exact scheduling information about branches,
-;; because they are at the starts / ends of basic blocks anyways.
-
-;; Some insns cannot be issued before/after another insn in the same cycle,
-;; irrespective of the type of the other insn.
-
-;; default is dual-issue, but can't be paired with an insn that
-;; uses multiple function units.
-(define_function_unit "single_issue" 1 0
- (and (eq_attr "pipe_model" "sh4")
- (eq_attr "type" "!smpy,dmpy,pload,pstore,dfp_cmp,gp_fpul,call,sfunc,arith3,arith3b"))
- 1 10
- [(eq_attr "type" "smpy,dmpy,pload,pstore,dfp_cmp,gp_fpul")])
-
-(define_function_unit "single_issue" 1 0
- (and (eq_attr "pipe_model" "sh4")
- (eq_attr "type" "smpy,dmpy,pload,pstore,dfp_cmp,gp_fpul"))
- 10 10
- [(const_int 1)])
-
-;; arith3 insns are always pairable at the start, but not inecessarily at
-;; the end; however, there doesn't seem to be a way to express that.
-(define_function_unit "single_issue" 1 0
- (and (eq_attr "pipe_model" "sh4")
- (eq_attr "type" "arith3"))
- 30 20
- [(const_int 1)])
-
-;; arith3b insn are pairable at the end and have latency that prevents pairing
-;; with the following branch, but we don't want this latency be respected;
-;; When the following branch is immediately adjacent, we can redirect the
-;; internal branch, which is likly to be a larger win.
-(define_function_unit "single_issue" 1 0
- (and (eq_attr "pipe_model" "sh4")
- (eq_attr "type" "arith3b"))
- 20 20
- [(const_int 1)])
-
-;; calls introduce a longisch delay that is likely to flush the pipelines.
-(define_function_unit "single_issue" 1 0
- (and (eq_attr "pipe_model" "sh4")
- (eq_attr "type" "call,sfunc"))
- 160 160
- [(eq_attr "type" "!call") (eq_attr "type" "call")])
-
-;; Load and store instructions have no alignment peculiarities for the SH4,
-;; but they use the load-store unit, which they share with the fmove type
-;; insns (fldi[01]; fmov frn,frm; flds; fsts; fabs; fneg) .
-;; Loads have a latency of two.
-;; However, call insns can only paired with a preceding insn, and have
-;; a delay slot, so that we want two more insns to be scheduled between the
-;; load of the function address and the call. This is equivalent to a
-;; latency of three.
-;; We cannot use a conflict list for this, because we need to distinguish
-;; between the actual call address and the function arguments.
-;; ADJUST_COST can only properly handle reductions of the cost, so we
-;; use a latency of three here, which gets multiplied by 10 to yield 30.
-;; We only do this for SImode loads of general registers, to make the work
-;; for ADJUST_COST easier.
-
-;; When specifying different latencies for different insns using the
-;; the same function unit, genattrtab.c assumes a 'FIFO constraint'
-;; so that the blockage is at least READY-COST (E) + 1 - READY-COST (C)
-;; for an executing insn E and a candidate insn C.
-;; Therefore, we define three different function units for load_store:
-;; load_store, load and load_si.
-
-(define_function_unit "load_si" 1 0
- (and (eq_attr "pipe_model" "sh4")
- (eq_attr "type" "load_si,pcload_si")) 30 10)
-(define_function_unit "load" 1 0
- (and (eq_attr "pipe_model" "sh4")
- (eq_attr "type" "load,pcload,pload")) 20 10)
-(define_function_unit "load_store" 1 0
- (and (eq_attr "pipe_model" "sh4")
- (eq_attr "type" "load_si,pcload_si,load,pcload,pload,store,pstore,fmove"))
- 10 10)
-
-(define_function_unit "int" 1 0
- (and (eq_attr "pipe_model" "sh4") (eq_attr "type" "arith,dyn_shift")) 10 10)
-
-;; Again, we have to pretend a lower latency for the "int" unit to avoid a
-;; spurious FIFO constraint; the multiply instructions use the "int"
-;; unit actually only for two cycles.
-(define_function_unit "int" 1 0
- (and (eq_attr "pipe_model" "sh4") (eq_attr "type" "smpy,dmpy")) 20 20)
-
-;; We use a fictous "mpy" unit to express the actual latency.
-(define_function_unit "mpy" 1 0
- (and (eq_attr "pipe_model" "sh4") (eq_attr "type" "smpy,dmpy")) 40 20)
-
-;; Again, we have to pretend a lower latency for the "int" unit to avoid a
-;; spurious FIFO constraint.
-(define_function_unit "int" 1 0
- (and (eq_attr "pipe_model" "sh4") (eq_attr "type" "gp_fpul")) 10 10)
-
-;; We use a fictous "gp_fpul" unit to express the actual latency.
-(define_function_unit "gp_fpul" 1 0
- (and (eq_attr "pipe_model" "sh4") (eq_attr "type" "gp_fpul")) 20 10)
-
-;; ??? multiply uses the floating point unit, but with a two cycle delay.
-;; Thus, a simple single-precision fp operation could finish if issued in
-;; the very next cycle, but stalls when issued two or three cycles later.
-;; Similarily, a divide / sqrt can work without stalls if issued in
-;; the very next cycle, while it would have to block if issued two or
-;; three cycles later.
-;; There is no way to model this with gcc's function units. This problem is
-;; actually mentioned in md.texi. Tackling this problem requires first that
-;; it is possible to speak about the target in an open discussion.
-;;
-;; However, simple double-precision operations always conflict.
-
-(define_function_unit "fp" 1 0
- (and (eq_attr "pipe_model" "sh4") (eq_attr "type" "smpy,dmpy")) 40 40
- [(eq_attr "type" "dfp_cmp,dfp_conv,dfp_arith")])
-
-;; The "fp" unit is for pipeline stages F1 and F2.
-
-(define_function_unit "fp" 1 0
- (and (eq_attr "pipe_model" "sh4") (eq_attr "type" "fp")) 30 10)
-
-;; Again, we have to pretend a lower latency for the "fp" unit to avoid a
-;; spurious FIFO constraint; the bulk of the fdiv type insns executes in
-;; the F3 stage.
-(define_function_unit "fp" 1 0
- (and (eq_attr "pipe_model" "sh4") (eq_attr "type" "fdiv")) 30 10)
-
-;; The "fdiv" function unit models the aggregate effect of the F1, F2 and F3
-;; pipeline stages on the pipelining of fdiv/fsqrt insns.
-;; We also use it to give the actual latency here.
-;; fsqrt is actually one cycle faster than fdiv (and the value used here),
-;; but that will hardly matter in practice for scheduling.
-(define_function_unit "fdiv" 1 0
- (and (eq_attr "pipe_model" "sh4") (eq_attr "type" "fdiv")) 120 100)
-
-;; There is again a late use of the "fp" unit by [d]fdiv type insns
-;; that we can't express.
-
-(define_function_unit "fp" 1 0
- (and (eq_attr "pipe_model" "sh4") (eq_attr "type" "dfp_cmp,dfp_conv")) 40 20)
-
-(define_function_unit "fp" 1 0
- (and (eq_attr "pipe_model" "sh4") (eq_attr "type" "dfp_arith")) 80 60)
-
-(define_function_unit "fp" 1 0
- (and (eq_attr "pipe_model" "sh4") (eq_attr "type" "dfdiv")) 230 10)
-
-(define_function_unit "fdiv" 1 0
- (and (eq_attr "pipe_model" "sh4") (eq_attr "type" "dfdiv")) 230 210)
-
;; SH-5 SHmedia scheduling
;; When executing SHmedia code, the SH-5 is a fairly straightforward
;; single-issue machine. It has four pipelines, the branch unit (br),
(eq_attr "length" "2") (const_string "yes")
] (const_string "no")))
+(define_attr "cond_delay_slot" "yes,no"
+ (cond [(eq_attr "in_delay_slot" "yes") (const_string "yes")
+ ] (const_string "no")))
+
(define_attr "is_sfunc" ""
(if_then_else (eq_attr "type" "sfunc") (const_int 1) (const_int 0)))
(define_attr "is_mac_media" ""
(if_then_else (eq_attr "type" "mac_media") (const_int 1) (const_int 0)))
+(define_attr "branch_zero" "yes,no"
+ (cond [(eq_attr "type" "!cbranch") (const_string "no")
+ (ne (symbol_ref "(next_active_insn (insn)\
+ == (prev_active_insn\
+ (XEXP (SET_SRC (PATTERN (insn)), 1))))\
+ && get_attr_length (next_active_insn (insn)) == 2")
+ (const_int 0))
+ (const_string "yes")]
+ (const_string "no")))
+
+;; SH4 Double-precision computation with double-precision result -
+;; the two halves are ready at different times.
+(define_attr "dfp_comp" "yes,no"
+ (cond [(eq_attr "type" "dfp_arith,dfp_conv,dfdiv") (const_string "yes")]
+ (const_string "no")))
+
+;; Insns for which the latency of a preceding fp insn is decreased by one.
+(define_attr "late_fp_use" "yes,no" (const_string "no"))
+;; And feeding insns for which this relevant.
+(define_attr "any_fp_comp" "yes,no"
+ (cond [(eq_attr "type" "fp,fdiv,ftrc_s,dfp_arith,dfp_conv,dfdiv")
+ (const_string "yes")]
+ (const_string "no")))
+
+(define_attr "any_int_load" "yes,no"
+ (cond [(eq_attr "type" "load,load_si,pcload,pcload_si")
+ (const_string "yes")]
+ (const_string "no")))
+
(define_delay
(eq_attr "needs_delay_slot" "yes")
[(eq_attr "in_delay_slot" "yes") (nil) (nil)])
(define_delay
(and (eq_attr "type" "cbranch")
(ne (symbol_ref "TARGET_SH2") (const_int 0)))
- [(eq_attr "in_delay_slot" "yes") (eq_attr "in_delay_slot" "yes") (nil)])
+ ;; SH2e has a hardware bug that pretty much prohibits the use of
+ ;; annuled delay slots.
+ [(eq_attr "in_delay_slot" "yes") (and (eq_attr "cond_delay_slot" "yes")
+ (not (eq_attr "cpu" "sh2e"))) (nil)])
\f
;; -------------------------------------------------------------------------
;; SImode signed integer comparisons
(const_int 0)))]
"TARGET_SH1"
"tst %1,%0"
- [(set_attr "insn_class" "mt_group")])
+ [(set_attr "type" "mt_group")])
;; ??? Perhaps should only accept reg/constant if the register is reg 0.
;; That would still allow reload to create cmpi instructions, but would
tst %0,%0
cmp/eq %1,%0
cmp/eq %1,%0"
- [(set_attr "insn_class" "mt_group,mt_group,mt_group")])
+ [(set_attr "type" "mt_group")])
(define_insn "cmpgtsi_t"
[(set (reg:SI T_REG)
"@
cmp/gt %1,%0
cmp/pl %0"
- [(set_attr "insn_class" "mt_group,mt_group")])
+ [(set_attr "type" "mt_group")])
(define_insn "cmpgesi_t"
[(set (reg:SI T_REG)
"@
cmp/ge %1,%0
cmp/pz %0"
- [(set_attr "insn_class" "mt_group,mt_group")])
+ [(set_attr "type" "mt_group")])
;; -------------------------------------------------------------------------
;; SImode unsigned integer comparisons
(match_operand:SI 1 "arith_reg_operand" "r")))]
"TARGET_SH1"
"cmp/hs %1,%0"
- [(set_attr "insn_class" "mt_group")])
+ [(set_attr "type" "mt_group")])
(define_insn "cmpgtusi_t"
[(set (reg:SI T_REG)
(match_operand:SI 1 "arith_reg_operand" "r")))]
"TARGET_SH1"
"cmp/hi %1,%0"
- [(set_attr "insn_class" "mt_group")])
+ [(set_attr "type" "mt_group")])
;; We save the compare operands in the cmpxx patterns and use them when
;; we generate the branch.
cmp/eq\\t%S1,%S0\;bf{.|/}s\\t%,Ldi%=\;cmp/ge\\t%S1,%S0\;cmp/hs\\t%R1,%R0\\n%,Ldi%=:
cmp/pz\\t%S0"
[(set_attr "length" "8,2")
- (set_attr "type" "arith3,arith")])
+ (set_attr "type" "arith3,mt_group")])
\f
;; -------------------------------------------------------------------------
;; DImode unsigned integer comparisons
(ltu:SI (plus:SI (match_dup 1) (match_dup 2)) (match_dup 1)))]
"TARGET_SH1"
"addc %2,%0"
- [(set_attr "type" "arith")
- (set_attr "insn_class" "ex_group")])
+ [(set_attr "type" "arith")])
(define_insn "addc1"
[(set (match_operand:SI 0 "arith_reg_operand" "=r")
(clobber (reg:SI T_REG))]
"TARGET_SH1"
"addc %2,%0"
- [(set_attr "type" "arith")
- (set_attr "insn_class" "ex_group")])
+ [(set_attr "type" "arith")])
(define_expand "addsi3"
[(set (match_operand:SI 0 "arith_reg_operand" "")
(match_operand:SI 2 "arith_operand" "rI")))]
"TARGET_SH1"
"add %2,%0"
- [(set_attr "type" "arith")
- (set_attr "insn_class" "ex_group")])
+ [(set_attr "type" "arith")])
;; -------------------------------------------------------------------------
;; Subtraction instructions
(gtu:SI (minus:SI (match_dup 1) (match_dup 2)) (match_dup 1)))]
"TARGET_SH1"
"subc %2,%0"
- [(set_attr "type" "arith")
- (set_attr "insn_class" "ex_group")])
+ [(set_attr "type" "arith")])
(define_insn "subc1"
[(set (match_operand:SI 0 "arith_reg_operand" "=r")
(clobber (reg:SI T_REG))]
"TARGET_SH1"
"subc %2,%0"
- [(set_attr "type" "arith")
- (set_attr "insn_class" "ex_group")])
+ [(set_attr "type" "arith")])
(define_insn "*subsi3_internal"
[(set (match_operand:SI 0 "arith_reg_operand" "=r")
(match_operand:SI 2 "arith_reg_operand" "r")))]
"TARGET_SH1"
"sub %2,%0"
- [(set_attr "type" "arith")
- (set_attr "insn_class" "ex_group")])
+ [(set_attr "type" "arith")])
(define_insn "*subsi3_media"
[(set (match_operand:SI 0 "arith_reg_operand" "=r")
; the udivsi3 libcall has the same name, we must consider all registers
; clobbered that are in the union of the registers clobbered by the
; shmedia and the shcompact implementation. Note, if the shcompact
-; implemenation actually used shcompact code, we'd need to clobber
+; implementation actually used shcompact code, we'd need to clobber
; also r23 and fr23.
(define_insn "udivsi3_i1_media"
[(set (match_operand:SI 0 "register_operand" "=z")
""
"
{
- rtx first = 0, last;
+ rtx first, last;
operands[3] = gen_reg_rtx (Pmode);
/* Emit the move of the address to a pseudo outside of the libcall. */
- if (TARGET_HARD_SH4 && TARGET_SH3E)
+ if (TARGET_HARD_SH4 && TARGET_SH2E)
{
emit_move_insn (operands[3],
gen_rtx_SYMBOL_REF (SImode, \"__udivsi3_i4\"));
{
operands[1] = force_reg (SImode, operands[1]);
operands[2] = force_reg (SImode, operands[2]);
- last = gen_udivsi3_i4_media (operands[0], operands[1], operands[2]);
- first = last;
+ emit_insn (gen_udivsi3_i4_media (operands[0], operands[1], operands[2]));
+ DONE;
}
else if (TARGET_SH5)
{
gen_rtx_SYMBOL_REF (SImode, \"__udivsi3\"));
last = gen_udivsi3_i1 (operands[0], operands[3]);
}
- if (! first)
- {
- first = emit_move_insn (gen_rtx_REG (SImode, 4), operands[1]);
- emit_move_insn (gen_rtx_REG (SImode, 5), operands[2]);
- }
+ first = emit_move_insn (gen_rtx_REG (SImode, 4), operands[1]);
+ emit_move_insn (gen_rtx_REG (SImode, 5), operands[2]);
last = emit_insn (last);
/* Wrap the sequence in REG_LIBCALL / REG_RETVAL notes so that loop
invariant code motion can move it. */
; the sdivsi3 libcall has the same name, we must consider all registers
; clobbered that are in the union of the registers clobbered by the
; shmedia and the shcompact implementation. Note, if the shcompact
-; implemenation actually used shcompact code, we'd need to clobber
+; implementation actually used shcompact code, we'd need to clobber
; also r22, r23 and fr23.
(define_insn "divsi3_i1_media"
[(set (match_operand:SI 0 "register_operand" "=z")
""
"
{
- rtx first = 0, last;
+ rtx first, last;
operands[3] = gen_reg_rtx (Pmode);
/* Emit the move of the address to a pseudo outside of the libcall. */
- if (TARGET_HARD_SH4 && TARGET_SH3E)
+ if (TARGET_HARD_SH4 && TARGET_SH2E)
{
emit_move_insn (operands[3],
gen_rtx_SYMBOL_REF (SImode, \"__sdivsi3_i4\"));
{
operands[1] = force_reg (SImode, operands[1]);
operands[2] = force_reg (SImode, operands[2]);
- last = gen_divsi3_i4_media (operands[0], operands[1], operands[2]);
- first = last;
+ emit_insn (gen_divsi3_i4_media (operands[0], operands[1], operands[2]));
+ DONE;
}
else if (TARGET_SH5)
{
emit_move_insn (operands[3], gen_rtx_SYMBOL_REF (SImode, \"__sdivsi3\"));
last = gen_divsi3_i1 (operands[0], operands[3]);
}
- if (! first)
- {
- first = emit_move_insn (gen_rtx_REG (SImode, 4), operands[1]);
- emit_move_insn (gen_rtx_REG (SImode, 5), operands[2]);
- }
+ first = emit_move_insn (gen_rtx_REG (SImode, 4), operands[1]);
+ emit_move_insn (gen_rtx_REG (SImode, 5), operands[2]);
last = emit_insn (last);
/* Wrap the sequence in REG_LIBCALL / REG_RETVAL notes so that loop
invariant code motion can move it. */
invariant code motion can move it. */
REG_NOTES (first) = gen_rtx_INSN_LIST (REG_LIBCALL, last, REG_NOTES (first));
REG_NOTES (last) = gen_rtx_INSN_LIST (REG_RETVAL, first, REG_NOTES (last));
+ /* expand_binop can't find a suitable code in umul_widen_optab to
+ make a REG_EQUAL note from, so make one here.
+ See also smulsi3_highpart.
+ ??? Alternatively, we could put this at the calling site of expand_binop,
+ i.e. expand_expr. */
+ REG_NOTES (last)
+ = gen_rtx_EXPR_LIST (REG_EQUAL, copy_rtx (SET_SRC (single_set (first))),
+ REG_NOTES (last));
DONE;
}")
invariant code motion can move it. */
REG_NOTES (first) = gen_rtx_INSN_LIST (REG_LIBCALL, last, REG_NOTES (first));
REG_NOTES (last) = gen_rtx_INSN_LIST (REG_RETVAL, first, REG_NOTES (last));
+ /* expand_binop can't find a suitable code in umul_widen_optab to
+ make a REG_EQUAL note from, so make one here.
+ See also smulsi3_highpart.
+ ??? Alternatively, we could put this at the calling site of expand_binop,
+ i.e. expand_expr. */
+ REG_NOTES (last)
+ = gen_rtx_EXPR_LIST (REG_EQUAL, copy_rtx (SET_SRC (single_set (first))),
+ REG_NOTES (last));
DONE;
}")
REG_NOTES (last) = gen_rtx_INSN_LIST (REG_RETVAL, first, REG_NOTES (last));
/* expand_binop can't find a suitable code in mul_highpart_optab to
make a REG_EQUAL note from, so make one here.
+ See also {,u}mulhisi.
??? Alternatively, we could put this at the calling site of expand_binop,
i.e. expand_mult_highpart. */
REG_NOTES (last)
(match_operand:SI 2 "logical_operand" "r,L")))]
"TARGET_SH1"
"and %2,%0"
- [(set_attr "type" "arith")
- (set_attr "insn_class" "ex_group")])
+ [(set_attr "type" "arith")])
-;; If the constant is 255, then emit a extu.b instruction instead of an
+;; If the constant is 255, then emit an extu.b instruction instead of an
;; and, since that will give better code.
(define_expand "andsi3"
(define_insn_and_split "anddi3"
[(set (match_operand:DI 0 "arith_reg_operand" "=r,r,r")
(and:DI (match_operand:DI 1 "arith_reg_operand" "%r,r,r")
- (match_operand:DI 2 "and_operand" "r,P,n")))]
+ (match_operand:DI 2 "and_operand" "r,P,Z")))]
"TARGET_SHMEDIA"
"@
and %1, %2, %0
(match_operand:SI 2 "logical_operand" "r,L")))]
"TARGET_SH1"
"or %2,%0"
- [(set_attr "type" "arith")
- (set_attr "insn_class" "ex_group")])
+ [(set_attr "type" "arith")])
(define_insn "iordi3"
[(set (match_operand:DI 0 "arith_reg_operand" "=r,r")
(match_operand:SI 2 "logical_operand" "L,r")))]
"TARGET_SH1"
"xor %2,%0"
- [(set_attr "type" "arith")
- (set_attr "insn_class" "ex_group")])
+ [(set_attr "type" "arith")])
(define_insn "xordi3"
[(set (match_operand:DI 0 "arith_reg_operand" "=r,r")
xor %1, %2, %0
xori %1, %2, %0"
[(set_attr "type" "arith_media")])
+
+;; Combiner bridge pattern for 2 * sign extend -> logical op -> truncate.
+;; converts 2 * sign extend -> logical op into logical op -> sign extend
+(define_split
+ [(set (match_operand:DI 0 "arith_reg_operand" "")
+ (sign_extend:DI (match_operator 4 "binary_logical_operator"
+ [(match_operand 1 "any_register_operand" "")
+ (match_operand 2 "any_register_operand" "")])))]
+ "TARGET_SHMEDIA"
+ [(set (match_dup 5) (match_dup 4))
+ (set (match_dup 0) (sign_extend:DI (match_dup 5)))]
+"
+{
+ enum machine_mode inmode = GET_MODE (operands[1]);
+ int regno, offset = 0;
+
+ if (GET_CODE (operands[0]) == SUBREG)
+ {
+ offset = SUBREG_BYTE (operands[0]);
+ operands[0] = SUBREG_REG (operands[0]);
+ }
+ if (GET_CODE (operands[0]) != REG)
+ abort ();
+ if (! TARGET_LITTLE_ENDIAN)
+ offset += 8 - GET_MODE_SIZE (inmode);
+ operands[5] = gen_rtx_SUBREG (inmode, operands[0], offset);
+}")
\f
;; -------------------------------------------------------------------------
;; Shifts and rotates
(lshiftrt:SI (match_dup 1) (const_int 31)))]
"TARGET_SH1"
"rotl %0"
- [(set_attr "type" "arith")
- (set_attr "insn_class" "ex_group")])
+ [(set_attr "type" "arith")])
(define_insn "rotlsi3_31"
[(set (match_operand:SI 0 "arith_reg_operand" "=r")
(clobber (reg:SI T_REG))]
"TARGET_SH1"
"rotr %0"
- [(set_attr "type" "arith")
- (set_attr "insn_class" "ex_group")])
+ [(set_attr "type" "arith")])
(define_insn "rotlsi3_16"
[(set (match_operand:SI 0 "arith_reg_operand" "=r")
(const_int 16)))]
"TARGET_SH1"
"swap.w %1,%0"
- [(set_attr "type" "arith")
- (set_attr "insn_class" "ex_group")])
+ [(set_attr "type" "arith")])
(define_expand "rotlsi3"
[(set (match_operand:SI 0 "arith_reg_operand" "")
(const_int 8)))]
"TARGET_SH1"
"swap.b %1,%0"
- [(set_attr "type" "arith")
- (set_attr "insn_class" "ex_group")])
+ [(set_attr "type" "arith")])
(define_expand "rotlhi3"
[(set (match_operand:HI 0 "arith_reg_operand" "")
(clobber (match_dup 4))])]
"operands[4] = gen_rtx_SCRATCH (SImode);"
[(set_attr "length" "*,*,*,4")
- (set_attr "type" "dyn_shift,arith,arith,arith")
- (set_attr "insn_class" "ex_group,ex_group,ex_group,ex_group")])
+ (set_attr "type" "dyn_shift,arith,arith,arith")])
(define_insn "ashlhi3_k"
[(set (match_operand:HI 0 "arith_reg_operand" "=r,r")
"@
add %0,%0
shll%O2 %0"
- [(set_attr "type" "arith")
- (set_attr "insn_class" "ex_group")])
+ [(set_attr "type" "arith")])
(define_insn "ashlsi3_n"
[(set (match_operand:SI 0 "arith_reg_operand" "=r")
(eq (symbol_ref "shift_insns_rtx (insn)") (const_int 3))
(const_string "6")]
(const_string "8")))
- (set_attr "type" "arith")
- (set_attr "insn_class" "ex_group")])
+ (set_attr "type" "arith")])
(define_split
[(set (match_operand:SI 0 "arith_reg_operand" "")
(clobber (reg:SI T_REG))]
"TARGET_SH1 && INTVAL (operands[2]) == 1"
"shar %0"
- [(set_attr "type" "arith")
- (set_attr "insn_class" "ex_group")])
+ [(set_attr "type" "arith")])
;; We can't do HImode right shifts correctly unless we start out with an
;; explicit zero / sign extension; doing that would result in worse overall
(lt:SI (match_dup 1) (const_int 0)))]
"TARGET_SH1"
"shll %0"
- [(set_attr "type" "arith")
- (set_attr "insn_class" "ex_group")])
+ [(set_attr "type" "arith")])
(define_insn "ashrsi3_d"
[(set (match_operand:SI 0 "arith_reg_operand" "=r")
(neg:SI (match_operand:SI 2 "arith_reg_operand" "r"))))]
"TARGET_SH3"
"shad %2,%0"
- [(set_attr "type" "dyn_shift")
- (set_attr "insn_class" "ex_group")])
+ [(set_attr "type" "dyn_shift")])
(define_insn "ashrsi3_n"
[(set (reg:SI R4_REG)
(neg:SI (match_operand:SI 2 "arith_reg_operand" "r"))))]
"TARGET_SH3"
"shld %2,%0"
- [(set_attr "type" "dyn_shift")
- (set_attr "insn_class" "ex_group")])
+ [(set_attr "type" "dyn_shift")])
;; Only the single bit shift clobbers the T bit.
(clobber (reg:SI T_REG))]
"TARGET_SH1 && CONST_OK_FOR_M (INTVAL (operands[2]))"
"shlr %0"
- [(set_attr "type" "arith")
- (set_attr "insn_class" "ex_group")])
+ [(set_attr "type" "arith")])
(define_insn "lshrsi3_k"
[(set (match_operand:SI 0 "arith_reg_operand" "=r")
"TARGET_SH1 && CONST_OK_FOR_K (INTVAL (operands[2]))
&& ! CONST_OK_FOR_M (INTVAL (operands[2]))"
"shlr%O2 %0"
- [(set_attr "type" "arith")
- (set_attr "insn_class" "ex_group")])
+ [(set_attr "type" "arith")])
(define_insn "lshrsi3_n"
[(set (match_operand:SI 0 "arith_reg_operand" "=r")
"TARGET_SH1"
"shll %R0\;rotcl %S0"
[(set_attr "length" "4")
- (set_attr "type" "arith")
- (set_attr "insn_class" "ex_group")])
+ (set_attr "type" "arith")])
(define_insn "ashldi3_media"
[(set (match_operand:DI 0 "arith_reg_operand" "=r,r")
"TARGET_SH1"
"shlr %S0\;rotcr %R0"
[(set_attr "length" "4")
- (set_attr "type" "arith")
- (set_attr "insn_class" "ex_group")])
+ (set_attr "type" "arith")])
(define_insn "lshrdi3_media"
[(set (match_operand:DI 0 "arith_reg_operand" "=r,r")
"TARGET_SH1"
"shar %S0\;rotcr %R0"
[(set_attr "length" "4")
- (set_attr "type" "arith")
- (set_attr "insn_class" "ex_group")])
+ (set_attr "type" "arith")])
(define_insn "ashrdi3_media"
[(set (match_operand:DI 0 "arith_reg_operand" "=r,r")
(and:SI (ashift:SI (match_operand:SI 1 "register_operand" "")
(match_operand:SI 2 "const_int_operand" ""))
(match_operand:SI 3 "const_int_operand" "")))]
- "TARGET_SH1 && (unsigned)INTVAL (operands[2]) < 32"
+ "TARGET_SH1 && reload_completed && (unsigned)INTVAL (operands[2]) < 32"
[(use (reg:SI R0_REG))]
"if (gen_shl_and (operands[0], operands[2], operands[3], operands[1])) FAIL;
DONE;")
(match_operand:SI 2 "const_int_operand" ""))
(match_operand:SI 3 "const_int_operand" "")))
(clobber (reg:SI T_REG))]
- "TARGET_SH1 && (unsigned)INTVAL (operands[2]) < 32"
+ "TARGET_SH1 && reload_completed && (unsigned)INTVAL (operands[2]) < 32"
[(use (reg:SI R0_REG))]
"if (gen_shl_and (operands[0], operands[2], operands[3], operands[1])) FAIL;
DONE;")
(const_int 16))))]
"TARGET_SH1"
"xtrct %1,%0"
- [(set_attr "type" "arith")
- (set_attr "insn_class" "ex_group")])
+ [(set_attr "type" "arith")])
(define_insn "xtrct_right"
[(set (match_operand:SI 0 "arith_reg_operand" "=r")
(const_int 16))))]
"TARGET_SH1"
"xtrct %2,%0"
- [(set_attr "type" "arith")
- (set_attr "insn_class" "ex_group")])
+ [(set_attr "type" "arith")])
;; -------------------------------------------------------------------------
;; Unary arithmetic
(const_int 0)))]
"TARGET_SH1"
"negc %1,%0"
- [(set_attr "type" "arith")
- (set_attr "insn_class" "ex_group")])
+ [(set_attr "type" "arith")])
(define_insn "*negdi_media"
[(set (match_operand:DI 0 "arith_reg_operand" "=r")
(neg:SI (match_operand:SI 1 "arith_reg_operand" "r")))]
"TARGET_SH1"
"neg %1,%0"
- [(set_attr "type" "arith")
- (set_attr "insn_class" "ex_group")])
+ [(set_attr "type" "arith")])
(define_insn "one_cmplsi2"
[(set (match_operand:SI 0 "arith_reg_operand" "=r")
(not:SI (match_operand:SI 1 "arith_reg_operand" "r")))]
"TARGET_SH1"
"not %1,%0"
- [(set_attr "type" "arith")
- (set_attr "insn_class" "ex_group")])
+ [(set_attr "type" "arith")])
(define_expand "one_cmpldi2"
[(set (match_operand:DI 0 "arith_reg_operand" "")
(zero_extend:SI (match_operand:HI 1 "arith_reg_operand" "r")))]
"TARGET_SH1"
"extu.w %1,%0"
- [(set_attr "type" "arith")
- (set_attr "insn_class" "ex_group")])
+ [(set_attr "type" "arith")])
(define_insn "*zero_extendhisi2_media"
[(set (match_operand:SI 0 "register_operand" "=r,r")
(zero_extend:SI (match_operand:QI 1 "arith_reg_operand" "r")))]
"TARGET_SH1"
"extu.b %1,%0"
- [(set_attr "type" "arith")
- (set_attr "insn_class" "ex_group")])
+ [(set_attr "type" "arith")])
(define_insn "*zero_extendqisi2_media"
[(set (match_operand:SI 0 "register_operand" "=r,r")
(zero_extend:HI (match_operand:QI 1 "arith_reg_operand" "r")))]
"TARGET_SH1"
"extu.b %1,%0"
- [(set_attr "type" "arith")
- (set_attr "insn_class" "ex_group")])
+ [(set_attr "type" "arith")])
;; -------------------------------------------------------------------------
;; Sign extension instructions
"@
exts.w %1,%0
mov.w %1,%0"
- [(set_attr "type" "arith,load")
- (set_attr "insn_class" "ex_group,*")])
+ [(set_attr "type" "arith,load")])
(define_insn "*extendhisi2_media"
[(set (match_operand:SI 0 "register_operand" "=r,r")
"@
exts.b %1,%0
mov.b %1,%0"
- [(set_attr "type" "arith,load")
- (set_attr "insn_class" "ex_group,*")])
+ [(set_attr "type" "arith,load")])
(define_insn "*extendqisi2_media"
[(set (match_operand:SI 0 "register_operand" "=r,r")
"@
exts.b %1,%0
mov.b %1,%0"
- [(set_attr "type" "arith,load")
- (set_attr "insn_class" "ex_group,*")])
+ [(set_attr "type" "arith,load")])
/* It would seem useful to combine the truncXi patterns into the movXi
patterns, but unary operators are ignored when matching constraints,
(define_insn "push_fpul"
[(set (mem:SF (pre_dec:SI (reg:SI SP_REG))) (reg:SF FPUL_REG))]
- "TARGET_SH3E && ! TARGET_SH5"
+ "TARGET_SH2E && ! TARGET_SH5"
"sts.l fpul,@-r15"
[(set_attr "type" "store")
+ (set_attr "late_fp_use" "yes")
(set_attr "hit_stack" "yes")])
;; DFmode pushes for sh4 require a lot of what is defined for movdf_i4,
(define_insn "pop_fpul"
[(set (reg:SF FPUL_REG) (mem:SF (post_inc:SI (reg:SI SP_REG))))]
- "TARGET_SH3E && ! TARGET_SH5"
+ "TARGET_SH2E && ! TARGET_SH5"
"lds.l @r15+,fpul"
[(set_attr "type" "load")
(set_attr "hit_stack" "yes")])
"TARGET_SH1 && ! TARGET_SH5"
"")
+(define_expand "push_fpscr"
+ [(const_int 0)]
+ "TARGET_SH3E"
+ "
+{
+ rtx insn = emit_insn (gen_fpu_switch (gen_rtx (MEM, PSImode,
+ gen_rtx (PRE_DEC, Pmode,
+ stack_pointer_rtx)),
+ get_fpscr_rtx ()));
+ REG_NOTES (insn) = gen_rtx (EXPR_LIST, REG_INC, stack_pointer_rtx, NULL_RTX);
+ DONE;
+}")
+
+(define_expand "pop_fpscr"
+ [(const_int 0)]
+ "TARGET_SH3E"
+ "
+{
+ rtx insn = emit_insn (gen_fpu_switch (get_fpscr_rtx (),
+ gen_rtx (MEM, PSImode,
+ gen_rtx (POST_INC, Pmode,
+ stack_pointer_rtx))));
+ REG_NOTES (insn) = gen_rtx (EXPR_LIST, REG_INC, stack_pointer_rtx, NULL_RTX);
+ DONE;
+}")
+
;; These two patterns can happen as the result of optimization, when
;; comparisons get simplified to a move of zero or 1 into the T reg.
;; They don't disappear completely, because the T reg is a fixed hard reg.
[(set (match_operand:SI 0 "general_movdst_operand" "=r,r,t,r,r,r,r,m,<,<,x,l,x,l,r")
(match_operand:SI 1 "general_movsrc_operand" "Q,rI,r,mr,x,l,t,r,x,l,r,r,>,>,i"))]
"TARGET_SH1
- && ! TARGET_SH3E
+ && ! TARGET_SH2E
&& (register_operand (operands[0], SImode)
|| register_operand (operands[1], SImode))"
"@
lds.l %1,%0
lds.l %1,%0
fake %1,%0"
- [(set_attr "type" "pcload_si,move,*,load_si,move,prget,move,store,store,pstore,move,prset,load,pload,pcload_si")
- (set_attr "insn_class" "*,*,mt_group,*,*,*,*,*,*,*,*,*,*,*,*")
+ [(set_attr "type" "pcload_si,move,mt_group,load_si,mac_gp,prget,move,store,store,pstore,move,prset,load,pload,pcload_si")
(set_attr "length" "*,*,*,*,*,*,*,*,*,*,*,*,*,*,*")])
;; t/r must come after r/r, lest reload will try to reload stuff like
;; (subreg:SI (reg:SF FR14_REG) 0) into T (compiling stdlib/strtod.c -m3e -O2)
;; ??? This allows moves from macl to fpul to be recognized, but these moves
;; will require a reload.
+;; ??? We can't include f/f because we need the proper FPSCR setting when
+;; TARGET_FMOVD is in effect, and mode switching is done before reload.
(define_insn "movsi_ie"
- [(set (match_operand:SI 0 "general_movdst_operand" "=r,r,t,r,r,r,r,m,<,<,x,l,x,l,y,<,r,y,r,y")
- (match_operand:SI 1 "general_movsrc_operand" "Q,rI,r,mr,x,l,t,r,x,l,r,r,>,>,>,y,i,r,y,y"))]
- "TARGET_SH3E
+ [(set (match_operand:SI 0 "general_movdst_operand" "=r,r,t,r,r,r,r,m,<,<,x,l,x,l,y,<,r,y,r,*f,y,*f,y")
+ (match_operand:SI 1 "general_movsrc_operand" "Q,rI,r,mr,x,l,t,r,x,l,r,r,>,>,>,y,i,r,y,y,*f,*f,y"))]
+ "TARGET_SH2E
&& (register_operand (operands[0], SImode)
|| register_operand (operands[1], SImode))"
"@
fake %1,%0
lds %1,%0
sts %1,%0
+ fsts fpul,%0
+ flds %1,fpul
+ fmov %1,%0
! move optimized away"
- [(set_attr "type" "pcload_si,move,*,load_si,move,prget,move,store,store,pstore,move,prset,load,pload,load,store,pcload_si,gp_fpul,gp_fpul,nil")
- (set_attr "length" "*,*,*,*,*,*,*,*,*,*,*,*,*,*,*,*,*,*,*,0")])
+ [(set_attr "type" "pcload_si,move,*,load_si,mac_gp,prget,move,store,store,pstore,move,prset,load,pload,load,store,pcload_si,gp_fpul,fpul_gp,fmove,fmove,fmove,nil")
+ (set_attr "late_fp_use" "*,*,*,*,*,*,*,*,*,*,*,*,*,*,*,yes,*,*,yes,*,*,*,*")
+ (set_attr "length" "*,*,*,*,*,*,*,*,*,*,*,*,*,*,*,*,*,*,*,*,*,*,0")])
(define_insn "movsi_i_lowpart"
[(set (strict_low_part (match_operand:SI 0 "general_movdst_operand" "+r,r,r,r,r,r,m,r"))
(define_insn "*movsi_media"
[(set (match_operand:SI 0 "general_movdst_operand" "=r,r,r,r,m,f,m,f,r,f,*b,r,b")
- (match_operand:SI 1 "general_movsrc_operand" "r,JS,ns,m,r,m,f,rU,f,f,r,*b,T"))]
+ (match_operand:SI 1 "general_movsrc_operand" "r,JS,ns,m,rU,m,f,rU,f,f,r,*b,T"))]
"TARGET_SHMEDIA_FPU
&& (register_operand (operands[0], SImode)
- || register_operand (operands[1], SImode))"
+ || sh_register_operand (operands[1], SImode))"
"@
add.l %1, r63, %0
movi %1, %0
#
ld%M1.l %m1, %0
- st%M0.l %m0, %1
+ st%M0.l %m0, %N1
fld%M1.s %m1, %0
fst%M0.s %m0, %1
fmov.ls %N1, %0
(define_insn "*movsi_media_nofpu"
[(set (match_operand:SI 0 "general_movdst_operand" "=r,r,r,r,m,*b,r,b")
- (match_operand:SI 1 "general_movsrc_operand" "r,JS,ns,m,r,r,*b,T"))]
+ (match_operand:SI 1 "general_movsrc_operand" "r,JS,ns,m,rU,r,*b,T"))]
"TARGET_SHMEDIA
&& (register_operand (operands[0], SImode)
- || register_operand (operands[1], SImode))"
+ || sh_register_operand (operands[1], SImode))"
"@
add.l %1, r63, %0
movi %1, %0
#
ld%M1.l %m1, %0
- st%M0.l %m0, %1
+ st%M0.l %m0, %N1
ptabs %1, %0
gettr %1, %0
pt %1, %0"
"TARGET_HARD_SH4"
"ocbwb\\t@%0\;extu.w\\t%0,%2\;or\\t%1,%2\;mov.l\\t%0,@%2"
[(set_attr "length" "8")
- (set_attr "insn_class" "cwb")])
+ (set_attr "type" "cwb")])
;; ??? could make arg 0 an offsettable memory operand to allow to save
;; an add in the code that calculates the address.
{
rtx sfun, tramp;
+ tramp = force_reg (Pmode, operands[0]);
sfun = force_reg (Pmode, gen_rtx_SYMBOL_REF (Pmode, \"__init_trampoline\"));
- tramp = gen_rtx_REG (SImode, R0_REG);
- emit_move_insn (tramp, operands[0]);
emit_move_insn (gen_rtx_REG (SImode, R2_REG), operands[1]);
emit_move_insn (gen_rtx_REG (SImode, R3_REG), operands[2]);
(define_insn "*movqi_media"
[(set (match_operand:QI 0 "general_movdst_operand" "=r,r,r,m")
- (match_operand:QI 1 "general_movsrc_operand" "r,JS,m,r"))]
+ (match_operand:QI 1 "general_movsrc_operand" "r,JS,m,rU"))]
"TARGET_SHMEDIA
&& (arith_reg_operand (operands[0], QImode)
- || arith_reg_operand (operands[1], QImode))"
+ || arith_reg_or_0_operand (operands[1], QImode))"
"@
add.l %1, r63, %0
movi %1, %0
ld%M1.ub %m1, %0
- st%M0.b %m0, %1"
+ st%M0.b %m0, %N1"
[(set_attr "type" "arith_media,arith_media,load_media,store_media")])
(define_expand "movqi"
(define_insn "*movhi_media"
[(set (match_operand:HI 0 "general_movdst_operand" "=r,r,r,r,m")
- (match_operand:HI 1 "general_movsrc_operand" "r,JS,n,m,r"))]
+ (match_operand:HI 1 "general_movsrc_operand" "r,JS,n,m,rU"))]
"TARGET_SHMEDIA
&& (arith_reg_operand (operands[0], HImode)
- || arith_reg_operand (operands[1], HImode))"
+ || arith_reg_or_0_operand (operands[1], HImode))"
"@
add.l %1, r63, %0
movi %1, %0
#
ld%M1.w %m1, %0
- st%M0.w %m0, %1"
+ st%M0.w %m0, %N1"
[(set_attr "type" "arith_media,arith_media,*,load_media,store_media")])
(define_split
operands[3] = gen_rtx_REG (DImode, REGNO (operands[2]));
}")
-;; ??? This should be a define expand.
-
;; x/r can be created by inlining/cse, e.g. for execute/961213-1.c
;; compiled with -m2 -ml -O3 -funroll-loops
-(define_insn ""
+(define_insn "*movdi_i"
[(set (match_operand:DI 0 "general_movdst_operand" "=r,r,r,m,r,r,r,*!x")
(match_operand:DI 1 "general_movsrc_operand" "Q,r,m,r,I,i,x,r"))]
"TARGET_SH1
(define_insn "*movdi_media"
[(set (match_operand:DI 0 "general_movdst_operand" "=r,r,r,rl,m,f,m,f,r,f,*b,r,b")
- (match_operand:DI 1 "general_movsrc_operand" "r,JS,iF,m,rl,m,f,rU,f,f,r,*b,T"))]
+ (match_operand:DI 1 "general_movsrc_operand" "r,JS,iF,m,rlU,m,f,rU,f,f,r,*b,T"))]
"TARGET_SHMEDIA_FPU
&& (register_operand (operands[0], DImode)
- || register_operand (operands[1], DImode))"
+ || sh_register_operand (operands[1], DImode))"
"@
add %1, r63, %0
movi %1, %0
#
ld%M1.q %m1, %0
- st%M0.q %m0, %1
+ st%M0.q %m0, %N1
fld%M1.d %m1, %0
fst%M0.d %m0, %1
fmov.qd %N1, %0
(define_insn "*movdi_media_nofpu"
[(set (match_operand:DI 0 "general_movdst_operand" "=r,r,r,rl,m,*b,r,b")
- (match_operand:DI 1 "general_movsrc_operand" "r,JS,iF,m,rl,r,*b,T"))]
+ (match_operand:DI 1 "general_movsrc_operand" "r,JS,iF,m,rlU,r,*b,T"))]
"TARGET_SHMEDIA
&& (register_operand (operands[0], DImode)
- || register_operand (operands[1], DImode))"
+ || sh_register_operand (operands[1], DImode))"
"@
add %1, r63, %0
movi %1, %0
#
ld%M1.q %m1, %0
- st%M0.q %m0, %1
+ st%M0.q %m0, %N1
ptabs %1, %0
gettr %1, %0
pt %1, %0"
&& MOVI_SHORI_BASE_OPERAND_P (operands[1])"
"
{
- if (GET_CODE (operands[1]) == LABEL_REF
- && GET_CODE (XEXP (operands[1], 0)) == CODE_LABEL)
- LABEL_NUSES (XEXP (operands[1], 0)) += 4;
- else if (GOTOFF_P (operands[1])
- && GET_CODE (XVECEXP (XEXP (operands[1], 0), 0, 0)) == LABEL_REF
- && (GET_CODE (XEXP (XVECEXP (XEXP (operands[1], 0), 0, 0), 0))
- == CODE_LABEL))
- LABEL_NUSES (XEXP (XVECEXP (XEXP (operands[1], 0), 0, 0), 0)) += 4;
+ sh_mark_label (operands[1], 4);
}")
(define_expand "movdi_const_32bit"
&& MOVI_SHORI_BASE_OPERAND_P (operands[1])"
"
{
- if (GET_CODE (operands[1]) == LABEL_REF
- && GET_CODE (XEXP (operands[1], 0)) == CODE_LABEL)
- LABEL_NUSES (XEXP (operands[1], 0)) += 2;
- else if (GOTOFF_P (operands[1])
- && GET_CODE (XVECEXP (XEXP (operands[1], 0), 0, 0)) == LABEL_REF
- && (GET_CODE (XEXP (XVECEXP (XEXP (operands[1], 0), 0, 0), 0))
- == CODE_LABEL))
- LABEL_NUSES (XEXP (XVECEXP (XEXP (operands[1], 0), 0, 0), 0)) += 2;
+ sh_mark_label (operands[1], 2);
}")
(define_expand "movdi_const_16bit"
(define_insn "movdf_media"
[(set (match_operand:DF 0 "general_movdst_operand" "=f,f,r,r,r,f,m,r,m")
- (match_operand:DF 1 "general_movsrc_operand" "f,rU,f,r,F,m,f,m,r"))]
+ (match_operand:DF 1 "general_movsrc_operand" "f,rU,f,r,F,m,f,m,rU"))]
"TARGET_SHMEDIA_FPU
&& (register_operand (operands[0], DFmode)
- || register_operand (operands[1], DFmode))"
+ || sh_register_operand (operands[1], DFmode))"
"@
fmov.d %1, %0
fmov.qd %N1, %0
fld%M1.d %m1, %0
fst%M0.d %m0, %1
ld%M1.q %m1, %0
- st%M0.q %m0, %1"
+ st%M0.q %m0, %N1"
[(set_attr "type" "fmove_media,fload_media,dfpconv_media,arith_media,*,fload_media,fstore_media,load_media,store_media")])
(define_insn "movdf_media_nofpu"
[(set (match_operand:DF 0 "general_movdst_operand" "=r,r,r,m")
- (match_operand:DF 1 "general_movsrc_operand" "r,F,m,r"))]
+ (match_operand:DF 1 "general_movsrc_operand" "r,F,m,rU"))]
"TARGET_SHMEDIA
&& (register_operand (operands[0], DFmode)
- || register_operand (operands[1], DFmode))"
+ || sh_register_operand (operands[1], DFmode))"
"@
add %1, r63, %0
#
ld%M1.q %m1, %0
- st%M0.q %m0, %1"
+ st%M0.q %m0, %N1"
[(set_attr "type" "arith_media,*,load_media,store_media")])
(define_split
(if_then_else
(ne (symbol_ref "TARGET_SHCOMPACT") (const_int 0))
(const_int 10) (const_int 8))])
- (set_attr "type" "fmove,move,pcload,load,store,pcload,load,store,load,load")
+ (set_attr "type" "fmove,move,pcfload,fload,store,pcload,load,store,load,fload")
+ (set_attr "late_fp_use" "*,*,*,*,yes,*,*,*,*,*")
(set (attr "fp_mode") (if_then_else (eq_attr "fmovd" "yes")
(const_string "double")
(const_string "none")))])
(match_operand:SF 1 "register_operand" ""))
(use (match_operand:PSI 2 "fpscr_operand" ""))
(clobber (match_scratch:SI 3 "X"))]
- "TARGET_SH3E && reload_completed
+ "TARGET_SH2E && reload_completed
&& true_regnum (operands[0]) == true_regnum (operands[1])"
[(set (match_dup 0) (match_dup 0))]
"")
(define_insn_and_split "*movv4sf_i"
[(set (match_operand:V4SF 0 "nonimmediate_operand" "=f,f,m")
- (match_operand:V4SF 1 "general_operand" "fU,m,f"))]
+ (match_operand:V4SF 1 "general_operand" "fU,m,fU"))]
"TARGET_SHMEDIA_FPU"
"#"
"&& reload_completed"
(define_insn "movsf_media"
[(set (match_operand:SF 0 "general_movdst_operand" "=f,f,r,r,r,f,m,r,m")
- (match_operand:SF 1 "general_movsrc_operand" "f,rU,f,r,F,m,f,m,r"))]
+ (match_operand:SF 1 "general_movsrc_operand" "f,rU,f,r,F,m,f,m,rU"))]
"TARGET_SHMEDIA_FPU
&& (register_operand (operands[0], SFmode)
- || register_operand (operands[1], SFmode))"
+ || sh_register_operand (operands[1], SFmode))"
"@
fmov.s %1, %0
fmov.ls %N1, %0
fld%M1.s %m1, %0
fst%M0.s %m0, %1
ld%M1.l %m1, %0
- st%M0.l %m0, %1"
+ st%M0.l %m0, %N1"
[(set_attr "type" "fmove_media,fload_media,fpconv_media,arith_media,*,fload_media,fstore_media,load_media,store_media")])
(define_insn "movsf_media_nofpu"
[(set (match_operand:SF 0 "general_movdst_operand" "=r,r,r,m")
- (match_operand:SF 1 "general_movsrc_operand" "r,F,m,r"))]
+ (match_operand:SF 1 "general_movsrc_operand" "r,F,m,rU"))]
"TARGET_SHMEDIA
&& (register_operand (operands[0], SFmode)
- || register_operand (operands[1], SFmode))"
+ || sh_register_operand (operands[1], SFmode))"
"@
add.l %1, r63, %0
#
ld%M1.l %m1, %0
- st%M0.l %m0, %1"
+ st%M0.l %m0, %N1"
[(set_attr "type" "arith_media,*,load_media,store_media")])
(define_split
[(set (match_operand:SF 0 "general_movdst_operand" "=r,r,r,r,m,l,r")
(match_operand:SF 1 "general_movsrc_operand" "r,I,FQ,mr,r,r,l"))]
"TARGET_SH1
- && (! TARGET_SH3E
+ && (! TARGET_SH2E
/* ??? We provide some insn so that direct_{load,store}[SFmode] get set */
|| (GET_CODE (operands[0]) == REG && REGNO (operands[0]) == 3)
|| (GET_CODE (operands[1]) == REG && REGNO (operands[1]) == 3))
(use (match_operand:PSI 2 "fpscr_operand" "c,c,c,c,c,c,c,c,c,c,c,c,c,c,c,c,c,c,c"))
(clobber (match_scratch:SI 3 "=X,X,X,X,&z,X,X,X,X,X,X,X,X,y,X,X,X,X,X"))]
- "TARGET_SH3E
+ "TARGET_SH2E
&& (arith_reg_operand (operands[0], SFmode)
|| arith_reg_operand (operands[1], SFmode)
|| arith_reg_operand (operands[3], SImode)
sts.l %1,%0
lds.l %1,%0
! move optimized away"
- [(set_attr "type" "fmove,move,fmove,fmove,pcload,load,store,pcload,load,store,fmove,fmove,load,*,gp_fpul,gp_fpul,store,load,nil")
+ [(set_attr "type" "fmove,move,fmove,fmove,pcfload,fload,store,pcload,load,store,fmove,fmove,load,*,fpul_gp,gp_fpul,store,load,nil")
+ (set_attr "late_fp_use" "*,*,*,*,*,*,yes,*,*,*,*,*,*,*,yes,*,yes,*,*")
(set_attr "length" "*,*,*,*,4,*,*,*,*,*,2,2,2,4,2,2,2,2,0")
(set (attr "fp_mode") (if_then_else (eq_attr "fmovd" "yes")
(const_string "single")
emit_insn (gen_movsf_media_nofpu (operands[0], operands[1]));
DONE;
}
- if (TARGET_SH3E)
+ if (TARGET_SH2E)
{
emit_sf_insn (gen_movsf_ie (operands[0], operands[1], get_fpscr_rtx ()));
DONE;
(define_insn "mov_nop"
[(set (match_operand 0 "any_register_operand" "") (match_dup 0))]
- "TARGET_SH3E"
+ "TARGET_SH2E"
""
[(set_attr "length" "0")
(set_attr "type" "nil")])
[(set (match_operand:SI 0 "register_operand" "=y,y")
(match_operand:SI 1 "immediate_operand" "Qi,I"))
(clobber (match_scratch:SI 2 "=&z,r"))]
- "TARGET_SH3E
+ "TARGET_SH2E
&& (reload_in_progress || reload_completed)"
"#"
[(set_attr "length" "4")
;; This one has the additional purpose to record a possible scratch register
;; for the following branch.
+;; ??? Unfortunately, just setting the scratch register is not good enough,
+;; because the insn then might be deemed dead and deleted. And we can't
+;; make the use in the jump insn explicit because that would disable
+;; delay slot scheduling from the target.
(define_insn "indirect_jump_scratch"
[(set (match_operand:SI 0 "register_operand" "=r")
- (unspec:SI [(match_operand 1 "const_int_operand" "")] UNSPEC_BBR))]
+ (unspec:SI [(match_operand 1 "const_int_operand" "")] UNSPEC_BBR))
+ (set (pc) (unspec [(const_int 0)] UNSPEC_BBR))]
"TARGET_SH1"
""
[(set_attr "length" "0")])
+
+;; This one is used to preemt an insn from beyond the bra / braf / jmp
+;; being pulled into the delay slot of a condbranch that has been made to
+;; jump around the unconditional jump because it was out of range.
+(define_insn "stuff_delay_slot"
+ [(set (pc)
+ (unspec [(match_operand 0 "const_int_operand" "") (pc)] UNSPEC_BBR))
+ (set (reg:SI T_REG) (match_operand 1 "const_int_operand" ""))]
+ "TARGET_SH1"
+ ""
+ [(set_attr "length" "0")
+ (set_attr "cond_delay_slot" "yes")])
\f
;; Conditional branch insns
DONE;
}
- if (TARGET_SH3E
+ if (TARGET_SH2E
&& TARGET_IEEE
&& GET_MODE_CLASS (GET_MODE (sh_compare_op0)) == MODE_FLOAT)
{
DONE;
}
- if (TARGET_SH3E
+ if (TARGET_SH2E
&& ! TARGET_IEEE
&& GET_MODE_CLASS (GET_MODE (sh_compare_op0)) == MODE_FLOAT)
{
[(set_attr "type" "jump")
(set_attr "needs_delay_slot" "yes")])
+;; ??? It would be much saner to explicitly use the scratch register
+;; in the jump insn, and have indirect_jump_scratch only set it,
+;; but fill_simple_delay_slots would refuse to do delay slot filling
+;; from the target then, as it uses simplejump_p.
+;;(define_insn "jump_compact_far"
+;; [(set (pc)
+;; (label_ref (match_operand 0 "" "")))
+;; (use (match_operand 1 "register_operand" "r")]
+;; "TARGET_SH1"
+;; "* return output_far_jump(insn, operands[0], operands[1]);"
+;; [(set_attr "type" "jump")
+;; (set_attr "needs_delay_slot" "yes")])
+
(define_insn "jump_media"
[(set (pc)
(match_operand:DI 0 "target_operand" "b"))]
(const_int 0))
(match_operand 1 "" "")
(match_operand 2 "" "")])]
- "TARGET_SH3E || TARGET_SHMEDIA"
+ "TARGET_SH2E || TARGET_SHMEDIA"
"
{
int i;
""
"")
+;; TLS code generation.
+;; ??? this should be a define_insn_and_split
+;; See the thread [PATCH/RFA] SH TLS support on gcc-patches
+;; <http://gcc.gnu.org/ml/gcc-patches/2003-02/msg01898.html>
+;; for details.
+
+(define_insn "tls_global_dynamic"
+ [(set (match_operand:SI 0 "register_operand" "=&z")
+ (unspec:SI [(match_operand:SI 1 "" "")]
+ UNSPEC_TLSGD))
+ (use (reg:PSI FPSCR_REG))
+ (use (reg:SI PIC_REG))
+ (clobber (reg:SI PR_REG))
+ (clobber (scratch:SI))]
+ "TARGET_SH1"
+ "*
+{
+ return \"\\
+mov.l\\t1f,r4\\n\\
+\\tmova\\t2f,r0\\n\\
+\\tmov.l\\t2f,r1\\n\\
+\\tadd\\tr0,r1\\n\\
+\\tjsr\\t@r1\\n\\
+\\tadd\\tr12,r4\\n\\
+\\tbra\\t3f\\n\\
+\\tnop\\n\\
+\\t.align\\t2\\n\\
+1:\\t.long\\t%a1@TLSGD\\n\\
+2:\\t.long\\t__tls_get_addr@PLT\\n\\
+3:\";
+}"
+ [(set_attr "type" "tls_load")
+ (set_attr "length" "26")])
+
+(define_insn "tls_local_dynamic"
+ [(set (match_operand:SI 0 "register_operand" "=&z")
+ (unspec:SI [(match_operand:SI 1 "" "")]
+ UNSPEC_TLSLDM))
+ (use (reg:PSI FPSCR_REG))
+ (use (reg:SI PIC_REG))
+ (clobber (reg:SI PR_REG))
+ (clobber (scratch:SI))]
+ "TARGET_SH1"
+ "*
+{
+ return \"\\
+mov.l\\t1f,r4\\n\\
+\\tmova\\t2f,r0\\n\\
+\\tmov.l\\t2f,r1\\n\\
+\\tadd\\tr0,r1\\n\\
+\\tjsr\\t@r1\\n\\
+\\tadd\\tr12,r4\\n\\
+\\tbra\\t3f\\n\\
+\\tnop\\n\\
+\\t.align\\t2\\n\\
+1:\\t.long\\t%a1@TLSLDM\\n\\
+2:\\t.long\\t__tls_get_addr@PLT\\n\\
+3:\";
+}"
+ [(set_attr "type" "tls_load")
+ (set_attr "length" "26")])
+
+(define_expand "sym2DTPOFF"
+ [(const (unspec [(match_operand 0 "" "")] UNSPEC_DTPOFF))]
+ ""
+ "")
+
+(define_expand "symDTPOFF2reg"
+ [(match_operand 0 "" "") (match_operand 1 "" "") (match_operand 2 "" "")]
+ ""
+ "
+{
+ rtx dtpoffsym, insn;
+ rtx t = no_new_pseudos ? operands[0] : gen_reg_rtx (GET_MODE (operands[0]));
+
+ dtpoffsym = gen_sym2DTPOFF (operands[1]);
+ PUT_MODE (dtpoffsym, Pmode);
+ emit_move_insn (t, dtpoffsym);
+ insn = emit_move_insn (operands[0],
+ gen_rtx_PLUS (Pmode, t, operands[2]));
+ DONE;
+}")
+
+(define_expand "sym2GOTTPOFF"
+ [(const (unspec [(match_operand 0 "" "")] UNSPEC_GOTTPOFF))]
+ ""
+ "")
+
+(define_insn "tls_initial_exec"
+ [(set (match_operand:SI 0 "register_operand" "=&r")
+ (unspec:SI [(match_operand:SI 1 "" "")]
+ UNSPEC_TLSIE))
+ (use (reg:SI GBR_REG))
+ (use (reg:SI PIC_REG))
+ (clobber (reg:SI R0_REG))]
+ ""
+ "*
+{
+ return \"\\
+mov.l\\t1f,r0\\n\\
+\\tstc\\tgbr,%0\\n\\
+\\tmov.l\\t@(r0,r12),r0\\n\\
+\\tbra\\t2f\\n\\
+\\tadd\\tr0,%0\\n\\
+\\t.align\\t2\\n\\
+1:\\t.long\\t%a1\\n\\
+2:\";
+}"
+ [(set_attr "type" "tls_load")
+ (set_attr "length" "16")])
+
+(define_expand "sym2TPOFF"
+ [(const (unspec [(match_operand 0 "" "")] UNSPEC_TPOFF))]
+ ""
+ "")
+
+(define_expand "symTPOFF2reg"
+ [(match_operand 0 "" "") (match_operand 1 "" "")]
+ ""
+ "
+{
+ rtx tpoffsym, insn;
+
+ tpoffsym = gen_sym2TPOFF (operands[1]);
+ PUT_MODE (tpoffsym, Pmode);
+ insn = emit_move_insn (operands[0], tpoffsym);
+ DONE;
+}")
+
+(define_insn "load_gbr"
+ [(set (match_operand:SI 0 "register_operand" "") (reg:SI GBR_REG))
+ (use (reg:SI GBR_REG))]
+ ""
+ "stc gbr,%0"
+ [(set_attr "type" "tls_load")])
+
;; case instruction for switch statements.
;; Operand 0 is index
DONE;
}")
+(define_expand "eh_return"
+ [(use (match_operand 0 "register_operand" ""))
+ (use (match_operand 1 "register_operand" ""))]
+ ""
+{
+ rtx tmp, sa = operands[0], ra = operands[1];
+
+ if (TARGET_SHMEDIA64)
+ emit_insn (gen_eh_set_ra_di (ra));
+ else
+ emit_insn (gen_eh_set_ra_si (ra));
+
+ emit_move_insn (EH_RETURN_STACKADJ_RTX, sa);
+ DONE;
+})
+
+;; Clobber the return address on the stack. We can't expand this
+;; until we know where it will be put in the stack frame.
+
+(define_insn "eh_set_ra_si"
+ [(unspec [(match_operand:SI 0 "register_operand" "r")] UNSPEC_EH_RETURN)
+ (clobber (match_scratch:SI 1 "=&r"))]
+ "! TARGET_SHMEDIA64"
+ "#")
+
+(define_insn "eh_set_ra_di"
+ [(unspec [(match_operand:DI 0 "register_operand" "r")] UNSPEC_EH_RETURN)
+ (clobber (match_scratch:DI 1 "=&r"))]
+ "TARGET_SHMEDIA64"
+ "#")
+
+(define_split
+ [(unspec [(match_operand 0 "register_operand" "")] UNSPEC_EH_RETURN)
+ (clobber (match_scratch 1 ""))]
+ "reload_completed"
+ [(const_int 0)]
+ "
+{
+ sh_set_return_address (operands[0], operands[1]);
+ DONE;
+}")
+
(define_insn "blockage"
[(unspec_volatile [(const_int 0)] UNSPECV_BLOCKAGE)]
""
;; GO_IF_LEGITIMATE_ADDRESS guards about bogus addresses before reload,
;; SECONDARY_INPUT_RELOAD_CLASS does this during reload, and the insn's
;; predicate after reload.
-;; The gp_fpul type for r/!c might look a bit odd, but it actually schedules
-;; like a gpr <-> fpul move.
+;; The mac_gp type for r/!c might look a bit odd, but it actually schedules
+;; like a mac -> gpr move.
(define_insn "fpu_switch"
- [(set (match_operand:PSI 0 "register_operand" "=c,c,r,c,c,r,m,r")
- (match_operand:PSI 1 "general_movsrc_operand" "c,>,m,m,r,r,r,!c"))]
- "TARGET_SH4
+ [(set (match_operand:PSI 0 "general_movdst_operand" "=c,c,r,c,c,r,m,r,<")
+ (match_operand:PSI 1 "general_movsrc_operand" "c,>,m,m,r,r,r,!c,c"))]
+ "TARGET_SH3E
&& (! reload_completed
|| true_regnum (operands[0]) != FPSCR_REG
|| GET_CODE (operands[1]) != MEM
lds %1,fpscr
mov %1,%0
mov.l %1,%0
- sts fpscr,%0"
- [(set_attr "length" "0,2,2,4,2,2,2,2")
- (set_attr "type" "dfp_conv,dfp_conv,load,dfp_conv,dfp_conv,move,store,gp_fpul")
- (set_attr "insn_class" "ldsmem_to_fpscr,*,*,lds_to_fpscr,*,*,*,*")])
+ sts fpscr,%0
+ sts.l fpscr,%0"
+ [(set_attr "length" "0,2,2,4,2,2,2,2,2")
+ (set_attr "type" "nil,mem_fpscr,load,mem_fpscr,gp_fpscr,move,store,mac_gp,store")])
(define_split
[(set (reg:PSI FPSCR_REG)
[(set (match_operand:SF 0 "arith_reg_operand" "")
(plus:SF (match_operand:SF 1 "arith_reg_operand" "")
(match_operand:SF 2 "arith_reg_operand" "")))]
- "TARGET_SH3E || TARGET_SHMEDIA_FPU"
+ "TARGET_SH2E || TARGET_SHMEDIA_FPU"
"
{
- if (TARGET_SH3E)
+ if (TARGET_SH2E)
{
expand_sf_binop (&gen_addsf3_i, operands);
DONE;
(vec_concat:V2SF
(vec_select:SF
(match_dup 0)
- (parallel [(not:BI (match_operand 4 "const_int_operand" "n"))]))
+ (parallel [(match_operand 7 "const_int_operand" "n")]))
(match_operator:SF 3 "binary_float_operator"
[(vec_select:SF (match_operand:V2SF 1 "fp_arith_reg_operand" "f")
(parallel [(match_operand 5
(vec_select:SF (match_operand:V2SF 2 "fp_arith_reg_operand" "f")
(parallel [(match_operand 6
"const_int_operand" "n")]))]))
- (parallel [(not:BI (match_dup 4)) (match_dup 4)])))]
- "TARGET_SHMEDIA_FPU"
+ (parallel [(match_dup 7) (match_operand 4 "const_int_operand" "n")])))]
+ "TARGET_SHMEDIA_FPU && INTVAL (operands[4]) != INTVAL (operands[7])"
"#"
- "TARGET_SHMEDIA_FPU && reload_completed"
- [(set (match_dup 7) (match_dup 8))]
+ "&& reload_completed"
+ [(set (match_dup 8) (match_dup 9))]
"
{
int endian = TARGET_LITTLE_ENDIAN ? 0 : 1;
(true_regnum (operands[2])
+ (INTVAL (operands[6]) ^ endian)));
- operands[7] = gen_rtx_REG (SFmode,
+ operands[8] = gen_rtx_REG (SFmode,
(true_regnum (operands[0])
+ (INTVAL (operands[4]) ^ endian)));
- operands[8] = gen_rtx (GET_CODE (operands[3]), SFmode, op1, op2);
+ operands[9] = gen_rtx (GET_CODE (operands[3]), SFmode, op1, op2);
}"
[(set_attr "type" "fparith_media")])
(plus:SF (match_operand:SF 1 "arith_reg_operand" "%0")
(match_operand:SF 2 "arith_reg_operand" "f")))
(use (match_operand:PSI 3 "fpscr_operand" "c"))]
- "TARGET_SH3E"
+ "TARGET_SH2E"
"fadd %2,%0"
[(set_attr "type" "fp")
(set_attr "fp_mode" "single")])
[(set (match_operand:SF 0 "fp_arith_reg_operand" "")
(minus:SF (match_operand:SF 1 "fp_arith_reg_operand" "")
(match_operand:SF 2 "fp_arith_reg_operand" "")))]
- "TARGET_SH3E || TARGET_SHMEDIA_FPU"
+ "TARGET_SH2E || TARGET_SHMEDIA_FPU"
"
{
- if (TARGET_SH3E)
+ if (TARGET_SH2E)
{
expand_sf_binop (&gen_subsf3_i, operands);
DONE;
(minus:SF (match_operand:SF 1 "fp_arith_reg_operand" "0")
(match_operand:SF 2 "fp_arith_reg_operand" "f")))
(use (match_operand:PSI 3 "fpscr_operand" "c"))]
- "TARGET_SH3E"
+ "TARGET_SH2E"
"fsub %2,%0"
[(set_attr "type" "fp")
(set_attr "fp_mode" "single")])
[(set (match_operand:SF 0 "fp_arith_reg_operand" "")
(mult:SF (match_operand:SF 1 "fp_arith_reg_operand" "")
(match_operand:SF 2 "fp_arith_reg_operand" "")))]
- "TARGET_SH3E || TARGET_SHMEDIA_FPU"
+ "TARGET_SH2E || TARGET_SHMEDIA_FPU"
"
{
if (TARGET_SH4)
expand_sf_binop (&gen_mulsf3_i4, operands);
- else if (TARGET_SH3E)
+ else if (TARGET_SH2E)
emit_insn (gen_mulsf3_ie (operands[0], operands[1], operands[2]));
if (! TARGET_SHMEDIA)
DONE;
(mult:SF (match_operand:SF 1 "fp_arith_reg_operand" "%0")
(match_operand:SF 2 "fp_arith_reg_operand" "f")))
(use (match_operand:PSI 3 "fpscr_operand" "c"))]
- "TARGET_SH3E"
+ "TARGET_SH2E"
"fmul %2,%0"
[(set_attr "type" "fp")
(set_attr "fp_mode" "single")])
[(set (match_operand:SF 0 "fp_arith_reg_operand" "=f")
(mult:SF (match_operand:SF 1 "fp_arith_reg_operand" "%0")
(match_operand:SF 2 "fp_arith_reg_operand" "f")))]
- "TARGET_SH3E && ! TARGET_SH4"
+ "TARGET_SH2E && ! TARGET_SH4"
"fmul %2,%0"
[(set_attr "type" "fp")])
(match_operand:SF 2 "fp_arith_reg_operand" "f"))
(match_operand:SF 3 "arith_reg_operand" "0")))
(use (match_operand:PSI 4 "fpscr_operand" "c"))]
- "TARGET_SH3E && ! TARGET_SH4"
+ "TARGET_SH2E && ! TARGET_SH4"
"fmac fr0,%2,%0"
[(set_attr "type" "fp")
(set_attr "fp_mode" "single")])
[(set (match_operand:SF 0 "arith_reg_operand" "")
(div:SF (match_operand:SF 1 "arith_reg_operand" "")
(match_operand:SF 2 "arith_reg_operand" "")))]
- "TARGET_SH3E || TARGET_SHMEDIA_FPU"
+ "TARGET_SH2E || TARGET_SHMEDIA_FPU"
"
{
- if (TARGET_SH3E)
+ if (TARGET_SH2E)
{
expand_sf_binop (&gen_divsf3_i, operands);
DONE;
(div:SF (match_operand:SF 1 "arith_reg_operand" "0")
(match_operand:SF 2 "arith_reg_operand" "f")))
(use (match_operand:PSI 3 "fpscr_operand" "c"))]
- "TARGET_SH3E"
+ "TARGET_SH2E"
"fdiv %2,%0"
[(set_attr "type" "fdiv")
(set_attr "fp_mode" "single")])
(define_expand "floatsisf2"
[(set (match_operand:SF 0 "fp_arith_reg_operand" "")
(float:SF (match_operand:SI 1 "fpul_operand" "")))]
- "TARGET_SH3E || TARGET_SHMEDIA_FPU"
+ "TARGET_SH2E || TARGET_SHMEDIA_FPU"
"
{
if (TARGET_SH4)
(define_insn "*floatsisf2_ie"
[(set (match_operand:SF 0 "fp_arith_reg_operand" "=f")
(float:SF (match_operand:SI 1 "fpul_operand" "y")))]
- "TARGET_SH3E && ! TARGET_SH4"
+ "TARGET_SH2E && ! TARGET_SH4"
"float %1,%0"
[(set_attr "type" "fp")])
(define_expand "fix_truncsfsi2"
[(set (match_operand:SI 0 "fpul_operand" "=y")
(fix:SI (match_operand:SF 1 "fp_arith_reg_operand" "f")))]
- "TARGET_SH3E || TARGET_SHMEDIA_FPU"
+ "TARGET_SH2E || TARGET_SHMEDIA_FPU"
"
{
if (TARGET_SH4)
(use (match_operand:PSI 2 "fpscr_operand" "c"))]
"TARGET_SH4"
"ftrc %1,%0"
- [(set_attr "type" "fp")
+ [(set_attr "type" "ftrc_s")
(set_attr "fp_mode" "single")])
;; ??? This pattern is used nowhere. fix_truncsfsi2 always expands to
(define_insn "*fixsfsi"
[(set (match_operand:SI 0 "fpul_operand" "=y")
(fix:SI (match_operand:SF 1 "fp_arith_reg_operand" "f")))]
- "TARGET_SH3E && ! TARGET_SH4"
+ "TARGET_SH2E && ! TARGET_SH4"
"ftrc %1,%0"
[(set_attr "type" "fp")])
[(set (reg:SI T_REG)
(gt:SI (match_operand:SF 0 "fp_arith_reg_operand" "f")
(match_operand:SF 1 "fp_arith_reg_operand" "f")))]
- "TARGET_SH3E && ! TARGET_SH4"
+ "TARGET_SH2E && ! TARGET_SH4"
"fcmp/gt %1,%0"
[(set_attr "type" "fp")
(set_attr "fp_mode" "single")])
[(set (reg:SI T_REG)
(eq:SI (match_operand:SF 0 "fp_arith_reg_operand" "f")
(match_operand:SF 1 "fp_arith_reg_operand" "f")))]
- "TARGET_SH3E && ! TARGET_SH4"
+ "TARGET_SH2E && ! TARGET_SH4"
"fcmp/eq %1,%0"
[(set_attr "type" "fp")
(set_attr "fp_mode" "single")])
(ior:SI (reg:SI T_REG)
(eq:SI (match_operand:SF 0 "fp_arith_reg_operand" "f")
(match_operand:SF 1 "fp_arith_reg_operand" "f"))))]
- "TARGET_SH3E && TARGET_IEEE && ! TARGET_SH4"
+ "TARGET_SH2E && TARGET_IEEE && ! TARGET_SH4"
"* return output_ieee_ccmpeq (insn, operands);"
[(set_attr "length" "4")])
[(set (reg:SI T_REG)
(compare (match_operand:SF 0 "arith_operand" "")
(match_operand:SF 1 "arith_operand" "")))]
- "TARGET_SH3E || TARGET_SHMEDIA_FPU"
+ "TARGET_SH2E || TARGET_SHMEDIA_FPU"
"
{
sh_compare_op0 = operands[0];
(define_expand "negsf2"
[(set (match_operand:SF 0 "fp_arith_reg_operand" "")
(neg:SF (match_operand:SF 1 "fp_arith_reg_operand" "")))]
- "TARGET_SH3E || TARGET_SHMEDIA_FPU"
+ "TARGET_SH2E || TARGET_SHMEDIA_FPU"
"
{
- if (TARGET_SH3E)
+ if (TARGET_SH2E)
{
expand_sf_unop (&gen_negsf2_i, operands);
DONE;
[(set (match_operand:SF 0 "fp_arith_reg_operand" "=f")
(neg:SF (match_operand:SF 1 "fp_arith_reg_operand" "0")))
(use (match_operand:PSI 2 "fpscr_operand" "c"))]
- "TARGET_SH3E"
+ "TARGET_SH2E"
"fneg %0"
[(set_attr "type" "fmove")
(set_attr "fp_mode" "single")])
(define_expand "abssf2"
[(set (match_operand:SF 0 "fp_arith_reg_operand" "")
(abs:SF (match_operand:SF 1 "fp_arith_reg_operand" "")))]
- "TARGET_SH3E || TARGET_SHMEDIA_FPU"
+ "TARGET_SH2E || TARGET_SHMEDIA_FPU"
"
{
- if (TARGET_SH3E)
+ if (TARGET_SH2E)
{
expand_sf_unop (&gen_abssf2_i, operands);
DONE;
[(set (match_operand:SF 0 "fp_arith_reg_operand" "=f")
(abs:SF (match_operand:SF 1 "fp_arith_reg_operand" "0")))
(use (match_operand:PSI 2 "fpscr_operand" "c"))]
- "TARGET_SH3E"
+ "TARGET_SH2E"
"fabs %0"
[(set_attr "type" "fmove")
(set_attr "fp_mode" "single")])
"TARGET_SH4"
"ftrc %1,%0"
[(set_attr "type" "dfp_conv")
+ (set_attr "dfp_comp" "no")
(set_attr "fp_mode" "double")])
;; ??? This pattern is used nowhere. fix_truncdfsi2 always expands to
(plus:SI (match_dup 0) (match_operand:SI 1 "register_operand" "r")))
(set (mem:SF (match_dup 0))
(match_operand:SF 2 "general_movsrc_operand" ""))]
- "TARGET_SH3E && REGNO (operands[0]) == 0
+ "TARGET_SH2E && REGNO (operands[0]) == 0
&& ((GET_CODE (operands[2]) == REG
&& FP_OR_XD_REGISTER_P (REGNO (operands[2])))
|| (GET_CODE (operands[2]) == SUBREG
(set (match_operand:SF 2 "general_movdst_operand" "")
(mem:SF (match_dup 0)))]
- "TARGET_SH3E && REGNO (operands[0]) == 0
+ "TARGET_SH2E && REGNO (operands[0]) == 0
&& ((GET_CODE (operands[2]) == REG
&& FP_OR_XD_REGISTER_P (REGNO (operands[2])))
|| (GET_CODE (operands[2]) == SUBREG
(define_insn "movv8qi_i"
[(set (match_operand:V8QI 0 "general_movdst_operand" "=r,r,r,rl,m")
- (match_operand:V8QI 1 "general_movsrc_operand" "r,JSU,nW,m,rl"))]
+ (match_operand:V8QI 1 "general_movsrc_operand" "r,JSU,nW,m,rlU"))]
"TARGET_SHMEDIA
&& (register_operand (operands[0], V8QImode)
- || register_operand (operands[1], V8QImode))"
+ || sh_register_operand (operands[1], V8QImode))"
"@
add %1, r63, %0
movi %1, %0
#
ld%M1.q %m1, %0
- st%M0.q %m0, %1"
+ st%M0.q %m0, %N1"
[(set_attr "type" "arith_media,arith_media,*,load_media,store_media")
(set_attr "length" "4,4,16,4,4")])
&& VECTOR_MODE_SUPPORTED_P (GET_MODE (operands[0]))
&& GET_MODE_SIZE (GET_MODE (operands[0])) == 8
&& (XVECEXP (operands[1], 0, 0) != const0_rtx
- || XVECEXP (operands[1], 0, 1) != const0_rtx)"
+ || XVECEXP (operands[1], 0, 1) != const0_rtx)
+ && (XVECEXP (operands[1], 0, 0) != constm1_rtx
+ || XVECEXP (operands[1], 0, 1) != constm1_rtx)"
[(set (match_dup 0) (match_dup 1))
(match_dup 2)]
"
if (unit_size > 2)
operands[2] = gen_mshflo_l (operands[0], operands[0], operands[0]);
else
- operands[2] = gen_mperm_w0 (operands[0], operands[0]);
+ {
+ if (unit_size < 2)
+ operands[0] = gen_rtx_REG (V4HImode, true_regnum (operands[0]));
+ operands[2] = gen_mperm_w0 (operands[0], operands[0]);
+ }
operands[0] = gen_rtx_REG (DImode, true_regnum (operands[0]));
operands[1] = XVECEXP (operands[1], 0, 0);
if (unit_size < 2)
(define_insn "movv2hi_i"
[(set (match_operand:V2HI 0 "general_movdst_operand" "=r,r,r,rl,m")
- (match_operand:V2HI 1 "general_movsrc_operand" "r,JSU,nW,m,rl"))]
+ (match_operand:V2HI 1 "general_movsrc_operand" "r,JSU,nW,m,rlU"))]
"TARGET_SHMEDIA
&& (register_operand (operands[0], V2HImode)
- || register_operand (operands[1], V2HImode))"
+ || sh_register_operand (operands[1], V2HImode))"
"@
addz.l %1, r63, %0
movi %1, %0
#
ld%M1.l %m1, %0
- st%M0.l %m0, %1"
+ st%M0.l %m0, %N1"
[(set_attr "type" "arith_media,arith_media,*,load_media,store_media")
(set_attr "length" "4,4,16,4,4")])
(define_insn "movv4hi_i"
[(set (match_operand:V4HI 0 "general_movdst_operand" "=r,r,r,rl,m")
- (match_operand:V4HI 1 "general_movsrc_operand" "r,JSU,nW,m,rl"))]
+ (match_operand:V4HI 1 "general_movsrc_operand" "r,JSU,nW,m,rlU"))]
"TARGET_SHMEDIA
&& (register_operand (operands[0], V4HImode)
- || register_operand (operands[1], V4HImode))"
+ || sh_register_operand (operands[1], V4HImode))"
"@
add %1, r63, %0
movi %1, %0
#
ld%M1.q %m1, %0
- st%M0.q %m0, %1"
+ st%M0.q %m0, %N1"
[(set_attr "type" "arith_media,arith_media,*,load_media,store_media")
(set_attr "length" "4,4,16,4,4")])
(define_insn "movv2si_i"
[(set (match_operand:V2SI 0 "general_movdst_operand" "=r,r,r,rl,m")
- (match_operand:V2SI 1 "general_movsrc_operand" "r,JSU,nW,m,rl"))]
+ (match_operand:V2SI 1 "general_movsrc_operand" "r,JSU,nW,m,rlU"))]
"TARGET_SHMEDIA
&& (register_operand (operands[0], V2SImode)
- || register_operand (operands[1], V2SImode))"
+ || sh_register_operand (operands[1], V2SImode))"
"@
add %1, r63, %0
#
#
ld%M1.q %m1, %0
- st%M0.q %m0, %1"
+ st%M0.q %m0, %N1"
[(set_attr "type" "arith_media,arith_media,*,load_media,store_media")
(set_attr "length" "4,4,16,4,4")])
{
emit_insn ((TARGET_LITTLE_ENDIAN ? gen_mperm_w_little : gen_mperm_w_big)
(operands[0], operands[1], operands[2]));
+ DONE;
}")
; This use of vec_select isn't exactly correct according to rtl.texi
(vec_select:V4HI
(match_operand:V4HI 1 "arith_reg_operand" "r")
(parallel
- [(zero_extract (match_operand:QI 2 "extend_reg_or_0_operand" "rU")
- (const_int 2) (const_int 0))
- (zero_extract (match_dup 2) (const_int 2) (const_int 2))
- (zero_extract (match_dup 2) (const_int 2) (const_int 4))
- (zero_extract (match_dup 2) (const_int 2) (const_int 6))])))]
+ [(zero_extract:QI (match_operand:QI 2 "extend_reg_or_0_operand" "rU")
+ (const_int 2) (const_int 0))
+ (zero_extract:QI (match_dup 2) (const_int 2) (const_int 2))
+ (zero_extract:QI (match_dup 2) (const_int 2) (const_int 4))
+ (zero_extract:QI (match_dup 2) (const_int 2) (const_int 6))])))]
"TARGET_SHMEDIA && TARGET_LITTLE_ENDIAN"
"mperm.w %1, %N2, %0"
[(set_attr "type" "arith_media")])
(vec_select:V4HI
(match_operand:V4HI 1 "arith_reg_operand" "r")
(parallel
- [(zero_extract (not:QI (match_operand:QI 2
- "extend_reg_or_0_operand" "rU"))
- (const_int 2) (const_int 0))
- (zero_extract (not:QI (match_dup 2)) (const_int 2) (const_int 2))
- (zero_extract (not:QI (match_dup 2)) (const_int 2) (const_int 4))
- (zero_extract (not:QI (match_dup 2)) (const_int 2) (const_int 6))])))]
+ [(zero_extract:QI (not:QI (match_operand:QI 2
+ "extend_reg_or_0_operand" "rU"))
+ (const_int 2) (const_int 0))
+ (zero_extract:QI (not:QI (match_dup 2)) (const_int 2) (const_int 2))
+ (zero_extract:QI (not:QI (match_dup 2)) (const_int 2) (const_int 4))
+ (zero_extract:QI (not:QI (match_dup 2))
+ (const_int 2) (const_int 6))])))]
"TARGET_SHMEDIA && ! TARGET_LITTLE_ENDIAN"
"mperm.w %1, %N2, %0"
[(set_attr "type" "arith_media")])
(vec_select:V4HI
(vec_concat:V4HI (match_operand:V2HI 1 "extend_reg_or_0_operand" "rU")
(match_operand:V2HI 2 "extend_reg_or_0_operand" "rU"))
- (parallel [(const_int 0) (const_int 2) (const_int 1) (const_int 3)])))]
+ (parallel [(const_int 2) (const_int 0) (const_int 3) (const_int 1)])))]
"TARGET_SHMEDIA"
"mshflo.w %N1, %N2, %0"
[(set_attr "type" "arith_media")])
rtx discratch = gen_reg_rtx (DImode);
rtx last;
- emit_insn (gen_adddi3z_media (discratch, operands[1],
- force_reg (SImode, GEN_INT (-1))));
- emit_insn (gen_andcdi3 (discratch, discratch,
- simplify_gen_subreg (DImode, operands[1],
- SImode, 0)));
+ emit_insn (gen_adddi3 (discratch,
+ simplify_gen_subreg (DImode, operands[1], SImode, 0),
+ GEN_INT (-1)));
+ emit_insn (gen_andcdi3 (discratch,
+ simplify_gen_subreg (DImode, operands[1], SImode, 0),
+ discratch));
emit_insn (gen_nsbsi (scratch, discratch));
last = emit_insn (gen_subsi3 (operands[0],
- force_reg (SImode, GEN_INT (-64)), scratch));
+ force_reg (SImode, GEN_INT (63)), scratch));
REG_NOTES (last)
= gen_rtx_EXPR_LIST (REG_EQUAL,
gen_rtx_FFS (SImode, operands[0]), REG_NOTES (last));
(define_cpu_unit "f1_1,f1_2" "fpu_pipe")
-;; The floating point units.
+;; The floating point units (except FS - F2 always precedes it.)
-(define_cpu_unit "F1,F2,F3,FS" "fpu_pipe")
+(define_cpu_unit "F0,F1,F2,F3" "fpu_pipe")
;; This is basically the MA unit of SH4
;; used in LOAD/STORE pipeline.
(define_cpu_unit "memory" "inst_pipeline")
+;; However, there are LS group insns that don't use it, even ones that
+;; complete in 0 cycles. So we use an extra unit for the issue of LS insns.
+(define_cpu_unit "load_store" "inst_pipeline")
+
;; The address calculator used for branch instructions.
-;; This will be reserved with "issue" of branch instructions
-;; and this is to make sure that no two branch instructions
+;; This will be reserved after "issue" of branch instructions
+;; and this is to make sure that no two branch instructions
;; can be issued in parallel.
(define_cpu_unit "pcr_addrcalc" "inst_pipeline")
(define_reservation "issue" "pipe_01|pipe_02")
;; This is to express the locking of D stage.
+;; Note that the issue of a CO group insn also effectively locks the D stage.
(define_reservation "d_lock" "pipe_01+pipe_02")
+;; Every FE instruction but fipr / ftrv starts with issue and this.
+(define_reservation "F01" "F0+F1")
+
;; This is to simplify description where F1,F2,FS
;; are used simultaneously.
-(define_reservation "fpu" "F1+F2+FS")
+(define_reservation "fpu" "F1+F2")
;; This is to highlight the fact that f1
;; cannot overlap with F1.
(exclusion_set "f1_1,f1_2" "F1")
+(define_insn_reservation "nil" 0 (eq_attr "type" "nil") "nothing")
+
;; Although reg moves have a latency of zero
;; we need to highlight that they use D stage
;; for one cycle.
+;; Group: MT
+
(define_insn_reservation "reg_mov" 0
- (eq_attr "type" "move,fmove")
- "issue")
+ (and (eq_attr "pipe_model" "sh4")
+ (eq_attr "type" "move"))
+ "issue")
+
+;; Group: LS
-;; Other MT group intructions(1 step operations)
+(define_insn_reservation "freg_mov" 0
+ (and (eq_attr "pipe_model" "sh4")
+ (eq_attr "type" "fmove"))
+ "issue+load_store")
+
+;; We don't model all pipeline stages; we model the issue ('D') stage
+;; inasmuch as we allow only two instructions to issue simultaneously,
+;; and CO instructions prevent any simultaneous issue of another instruction.
+;; (This uses pipe_01 and pipe_02).
+;; Double issue of EX insns is prevented by using the int unit in the EX stage.
+;; Double issue of EX / BR insns is prevented by using the int unit /
+;; pcr_addrcalc unit in the EX stage.
+;; Double issue of BR / LS instructions is prevented by using the
+;; pcr_addrcalc / load_store unit in the issue cycle.
+;; Double issue of FE instructions is prevented by using F0 in the first
+;; pipeline stage after the first D stage.
+;; There is no need to describe the [ES]X / [MN]A / S stages after a D stage
+;; (except in the cases outlined above), nor to describe the FS stage after
+;; the F2 stage.
+
+;; Other MT group instructions(1 step operations)
;; Group: MT
;; Latency: 1
;; Issue Rate: 1
(define_insn_reservation "mt" 1
- (eq_attr "insn_class" "mt_group")
- "issue,nothing")
+ (and (eq_attr "pipe_model" "sh4")
+ (eq_attr "type" "mt_group"))
+ "issue")
;; Fixed Point Arithmetic Instructions(1 step operations)
;; Group: EX
;; Latency: 1
;; Issue Rate: 1
-(define_insn_reservation "simple_arith" 1
- (eq_attr "insn_class" "ex_group")
- "issue,int")
+(define_insn_reservation "sh4_simple_arith" 1
+ (and (eq_attr "pipe_model" "sh4")
+ (eq_attr "insn_class" "ex_group"))
+ "issue,int")
+
+;; Load and store instructions have no alignment peculiarities for the SH4,
+;; but they use the load-store unit, which they share with the fmove type
+;; insns (fldi[01]; fmov frn,frm; flds; fsts; fabs; fneg) .
+;; Loads have a latency of two.
+;; However, call insns can only paired with a preceding insn, and have
+;; a delay slot, so that we want two more insns to be scheduled between the
+;; load of the function address and the call. This is equivalent to a
+;; latency of three.
+;; ADJUST_COST can only properly handle reductions of the cost, so we
+;; use a latency of three here, which gets multiplied by 10 to yield 30.
+;; We only do this for SImode loads of general registers, to make the work
+;; for ADJUST_COST easier.
;; Load Store instructions. (MOV.[BWL]@(d,GBR)
;; Group: LS
;; Latency: 2
;; Issue Rate: 1
-(define_insn_reservation "load_store" 2
- (eq_attr "type" "load,load_si,pcload,pcload_si,store")
- "issue,memory*2")
+(define_insn_reservation "sh4_load" 2
+ (and (eq_attr "pipe_model" "sh4")
+ (eq_attr "type" "load,pcload"))
+ "issue+load_store,nothing,memory")
+
+;; calls / sfuncs need an extra instruction for their delay slot.
+;; Moreover, estimating the latency for SImode loads as 3 will also allow
+;; adjust_cost to meaningfully bump it back up to 3 if they load the shift
+;; count of a dynamic shift.
+(define_insn_reservation "sh4_load_si" 3
+ (and (eq_attr "pipe_model" "sh4")
+ (eq_attr "type" "load_si,pcload_si"))
+ "issue+load_store,nothing,memory")
+
+;; (define_bypass 2 "sh4_load_si" "!sh4_call")
+
+;; The load latency is upped to three higher if the dependent insn does
+;; double precision computation. We want the 'default' latency to reflect
+;; that increased latency because otherwise the insn priorities won't
+;; allow proper scheduling.
+(define_insn_reservation "sh4_fload" 3
+ (and (eq_attr "pipe_model" "sh4")
+ (eq_attr "type" "fload,pcfload"))
+ "issue+load_store,nothing,memory")
+
+;; (define_bypass 2 "sh4_fload" "!")
+
+(define_insn_reservation "sh4_store" 1
+ (and (eq_attr "pipe_model" "sh4")
+ (eq_attr "type" "store"))
+ "issue+load_store,nothing,memory")
+
+;; Load Store instructions.
+;; Group: LS
+;; Latency: 1
+;; Issue Rate: 1
+
+(define_insn_reservation "sh4_gp_fpul" 1
+ (and (eq_attr "pipe_model" "sh4")
+ (eq_attr "type" "gp_fpul"))
+ "issue+load_store")
+
+;; Load Store instructions.
+;; Group: LS
+;; Latency: 3
+;; Issue Rate: 1
+
+(define_insn_reservation "sh4_fpul_gp" 3
+ (and (eq_attr "pipe_model" "sh4")
+ (eq_attr "type" "fpul_gp"))
+ "issue+load_store")
;; Branch (BF,BF/S,BT,BT/S,BRA)
;; Group: BR
-;; Latency: 2 (or 1) Actually Observed to be 5/7
+;; Latency when taken: 2 (or 1)
;; Issue Rate: 1
;; The latency is 1 when displacement is 0.
-;; This reservation can be further broken into 2
-;; 1. branch_zero : One with latency 1 and in the TEST
-;; part it also checks for 0 (ZERO) displacement
-;; 2. branch: Latency 2.
+;; We can't really do much with the latency, even if we could express it,
+;; but the pairing restrictions are useful to take into account.
+;; ??? If the branch is likely, we might want to fill the delay slot;
+;; if the branch is likely, but not very likely, should we pretend to use
+;; a resource that CO instructions use, to get a pairable delay slot insn?
-(define_insn_reservation "branch_zero" 5
- (and (eq_attr "type" "cbranch")
- (eq_attr "length" "2"))
- "(issue+pcr_addrcalc),pcr_addrcalc,nothing")
-
-(define_insn_reservation "branch" 7
- (eq_attr "type" "cbranch")
- "(issue+pcr_addrcalc),pcr_addrcalc,nothing")
+(define_insn_reservation "sh4_branch" 1
+ (and (eq_attr "pipe_model" "sh4")
+ (eq_attr "type" "cbranch,jump"))
+ "issue+pcr_addrcalc")
;; Branch Far (JMP,RTS,BRAF)
;; Group: CO
;; Latency: 3
;; Issue Rate: 2
-;; Since issue stage (D stage) is blocked for 2nd cycle,
-;; cpu_unit int is reserved since it might be required for far
-;; address calculation.
+;; ??? Scheduling happens before branch shortening, and hence jmp and braf
+;; can't be distinguished from bra for the "jump" pattern.
-(define_insn_reservation "branch_far" 12
- (and (eq_attr "type" "jump,return")
- (eq_attr "length" "6"))
- "d_lock*2,int+pcr_addrcalc,pcr_addrcalc")
+(define_insn_reservation "sh4_return" 3
+ (and (eq_attr "pipe_model" "sh4")
+ (eq_attr "type" "return,jump_ind"))
+ "d_lock*2")
;; RTE
;; Group: CO
-;; atency: 5
+;; Latency: 5
;; Issue Rate: 5
;; this instruction can be executed in any of the pipelines
;; and blocks the pipeline for next 4 stages.
-(define_insn_reservation "return_from_exp" 5
- (eq_attr "type" "rte")
- "(issue+pcr_addrcalc),d_lock*4,int+pcr_addrcalc,nothing")
+(define_insn_reservation "sh4_return_from_exp" 5
+ (and (eq_attr "pipe_model" "sh4")
+ (eq_attr "type" "rte"))
+ "d_lock*5")
;; OCBP, OCBWB
;; Group: CO
-;; Latency: 5
+;; Latency: 1-5
;; Issue Rate: 1
-(define_insn_reservation "ocbwb" 5
- (eq_attr "insn_class" "cwb")
- "issue,(int+memory),memory*5")
+;; cwb is used for the sequence ocbwb @%0; extu.w %0,%2; or %1,%2; mov.l %0,@%2
+;; ocbwb on its own would be "d_lock,nothing,memory*5"
+(define_insn_reservation "ocbwb" 6
+ (and (eq_attr "pipe_model" "sh4")
+ (eq_attr "type" "cwb"))
+ "d_lock*2,(d_lock+memory)*3,issue+load_store+memory,memory*2")
;; LDS to PR,JSR
;; Group: CO
;; Latency: 3
;; Issue Rate: 2
;; The SX stage is blocked for last 2 cycles.
+;; OTOH, the only time that has an effect for insns generated by the compiler
+;; is when lds to PR is followed by sts from PR - and that is highly unlikely -
+;; or when we are doing a function call - and we don't do inter-function
+;; scheduling. For the function call case, it's really best that we end with
+;; something that models an rts.
-(define_insn_reservation "lds_to_pr" 3
- (eq_attr "type" "prset,call,sfunc")
- "(issue+pcr_addrcalc),(issue+int+pcr_addrcalc),(int+pcr_addrcalc)*2")
+(define_insn_reservation "sh4_lds_to_pr" 3
+ (and (eq_attr "pipe_model" "sh4")
+ (eq_attr "type" "prset") )
+ "d_lock*2")
+
+;; calls introduce a longisch delay that is likely to flush the pipelines
+;; of the caller's instructions. Ordinary functions tend to end with a
+;; load to restore a register (in the delay slot of rts), while sfuncs
+;; tend to end with an EX or MT insn. But that is not actually relevant,
+;; since there are no instructions that contend for memory access early.
+;; We could, of course, provide exact scheduling information for specific
+;; sfuncs, if that should prove useful.
+
+(define_insn_reservation "sh4_call" 16
+ (and (eq_attr "pipe_model" "sh4")
+ (eq_attr "type" "call,sfunc"))
+ "d_lock*16")
;; LDS.L to PR
;; Group: CO
;; The SX unit is blocked for last 2 cycles.
(define_insn_reservation "ldsmem_to_pr" 3
- (eq_attr "type" "pload")
- "(issue+pcr_addrcalc),(issue+int+pcr_addrcalc),(int+memory+pcr_addrcalc),(int+pcr_addrcalc)")
+ (and (eq_attr "pipe_model" "sh4")
+ (eq_attr "type" "pload"))
+ "d_lock*2")
;; STS from PR
;; Group: CO
;; The SX unit in second and third cycles.
(define_insn_reservation "sts_from_pr" 2
- (eq_attr "type" "prget")
- "(issue+pcr_addrcalc),(pipe_01+int+pcr_addrcalc),(int+pcr_addrcalc),nothing")
+ (and (eq_attr "pipe_model" "sh4")
+ (eq_attr "type" "prget"))
+ "d_lock*2")
;; STS.L from PR
;; Group: CO
;; Latency: 2
;; Issue Rate: 2
-(define_insn_reservation "prload_mem" 2
- (eq_attr "type" "pstore")
- "(issue+pcr_addrcalc),(pipe_01+int+pcr_addrcalc),(int+memory+pcr_addrcalc),memory")
+(define_insn_reservation "sh4_prstore_mem" 2
+ (and (eq_attr "pipe_model" "sh4")
+ (eq_attr "type" "pstore"))
+ "d_lock*2,nothing,memory")
;; LDS to FPSCR
;; Group: CO
;; Issue Rate: 1
;; F1 is blocked for last three cycles.
-(define_insn_reservation "fpscr_store" 4
- (eq_attr "insn_class" "lds_to_fpscr")
- "issue,int,F1*3")
+(define_insn_reservation "fpscr_load" 4
+ (and (eq_attr "pipe_model" "sh4")
+ (eq_attr "type" "gp_fpscr"))
+ "d_lock,nothing,F1*3")
;; LDS.L to FPSCR
;; Group: CO
;; Issue Rate: 1
;; F1 is blocked for last three cycles.
-(define_insn_reservation "fpscr_store_mem" 4
- (eq_attr "insn_class" "ldsmem_to_fpscr")
- "issue,(int+memory),(F1+memory),F1*2")
+(define_insn_reservation "fpscr_load_mem" 4
+ (and (eq_attr "pipe_model" "sh4")
+ (eq_attr "type" "mem_fpscr"))
+ "d_lock,nothing,(F1+memory),F1*2")
\f
;; Fixed point multiplication (DMULS.L DMULU.L MUL.L MULS.W,MULU.W)
;; Issue Rate: 1
(define_insn_reservation "multi" 4
- (eq_attr "type" "smpy,dmpy")
- "issue,(issue+int+f1_1),(int+f1_1),(f1_1|f1_2)*2,F2,FS")
+ (and (eq_attr "pipe_model" "sh4")
+ (eq_attr "type" "smpy,dmpy"))
+ "d_lock,(d_lock+f1_1),(f1_1|f1_2)*3,F2")
+
+;; Fixed STS from MACL / MACH
+;; Group: CO
+;; Latency: 3
+;; Issue Rate: 1
+
+(define_insn_reservation "sh4_mac_gp" 3
+ (and (eq_attr "pipe_model" "sh4")
+ (eq_attr "type" "mac_gp"))
+ "d_lock")
;; Single precision floating point computation FCMP/EQ,
-;; FCP/GT, FADD, FLOAT, FMAC, FMUL, FSUB, FTRC, FRVHG, FSCHG
+;; FCMP/GT, FADD, FLOAT, FMAC, FMUL, FSUB, FTRC, FRVHG, FSCHG
;; Group: FE
-;; Latency: 4
+;; Latency: 3/4
;; Issue Rate: 1
-(define_insn_reservation "fp_arith" 4
- (eq_attr "type" "fp")
- "issue,F1,F2,FS")
+(define_insn_reservation "fp_arith" 3
+ (and (eq_attr "pipe_model" "sh4")
+ (eq_attr "type" "fp"))
+ "issue,F01,F2")
+
+(define_insn_reservation "fp_arith_ftrc" 3
+ (and (eq_attr "pipe_model" "sh4")
+ (eq_attr "type" "ftrc_s"))
+ "issue,F01,F2")
+
+(define_bypass 1 "fp_arith_ftrc" "sh4_fpul_gp")
;; Single Precision FDIV/SQRT
;; Group: FE
-;; Latency: 12/13
+;; Latency: 12/13 (FDIV); 11/12 (FSQRT)
;; Issue Rate: 1
+;; We describe fdiv here; fsqrt is actually one cycle faster.
-(define_insn_reservation "fp_div" 13
- (eq_attr "type" "fdiv")
- "issue,F1+F3,F1+F2+F3,F3*7,F1+F3,F2,FS")
+(define_insn_reservation "fp_div" 12
+ (and (eq_attr "pipe_model" "sh4")
+ (eq_attr "type" "fdiv"))
+ "issue,F01+F3,F2+F3,F3*7,F1+F3,F2")
;; Double Precision floating point computation
;; (FCNVDS, FCNVSD, FLOAT, FTRC)
;; Latency: (3,4)/5
;; Issue Rate: 1
-(define_insn_reservation "dp_float" 5
- (eq_attr "type" "dfp_conv")
- "issue,F1,F1+F2,F2+FS,FS")
+(define_insn_reservation "dp_float" 4
+ (and (eq_attr "pipe_model" "sh4")
+ (eq_attr "type" "dfp_conv"))
+ "issue,F01,F1+F2,F2")
-;; Double-precision floating-point (FADD ,FMUL,FSUB)
+;; Double-precision floating-point (FADD,FMUL,FSUB)
;; Group: FE
;; Latency: (7,8)/9
;; Issue Rate: 1
-(define_insn_reservation "fp_double_arith" 9
- (eq_attr "type" "dfp_arith")
- "issue,F1,F1+F2,fpu*4,F2+FS,FS")
+(define_insn_reservation "fp_double_arith" 8
+ (and (eq_attr "pipe_model" "sh4")
+ (eq_attr "type" "dfp_arith"))
+ "issue,F01,F1+F2,fpu*4,F2")
;; Double-precision FCMP (FCMP/EQ,FCMP/GT)
-;; Group: FE
+;; Group: CO
;; Latency: 3/5
;; Issue Rate: 2
-(define_insn_reservation "fp_double_cmp" 5
- (eq_attr "type" "dfp_cmp")
- "issue,(issue+F1),F1+F2,F2+FS,FS")
+(define_insn_reservation "fp_double_cmp" 3
+ (and (eq_attr "pipe_model" "sh4")
+ (eq_attr "type" "dfp_cmp"))
+ "d_lock,(d_lock+F01),F1+F2,F2")
;; Double precision FDIV/SQRT
;; Group: FE
;; Latency: (24,25)/26
;; Issue Rate: 1
-(define_insn_reservation "dp_div" 26
- (eq_attr "type" "dfdiv")
- "issue,F1+F3,F1+F2+F3,F2+F3+FS,F3*16,F1+F3,F1+F2+F3,fpu+F3,F2+FS,FS")
+(define_insn_reservation "dp_div" 25
+ (and (eq_attr "pipe_model" "sh4")
+ (eq_attr "type" "dfdiv"))
+ "issue,F01+F3,F1+F2+F3,F2+F3,F3*16,F1+F3,(fpu+F3)*2,F2")
+
+;; Use the branch-not-taken case to model arith3 insns. For the branch taken
+;; case, we'd get a d_lock instead of issue at the end.
+(define_insn_reservation "arith3" 3
+ (and (eq_attr "pipe_model" "sh4")
+ (eq_attr "type" "arith3"))
+ "issue,d_lock+pcr_addrcalc,issue")
+
+;; arith3b insns schedule the same no matter if the branch is taken or not.
+(define_insn_reservation "arith3b" 2
+ (and (eq_attr "pipe_model" "sh4")
+ (eq_attr "type" "arith3"))
+ "issue,d_lock+pcr_addrcalc")