Home My Page Projects Code Snippets Project Openings SML/NJ
Summary Tracker SCM

SCM Repository

[smlnj] Annotation of /sml/trunk/src/MLRISC/sparc/mltree/sparc.sml
ViewVC logotype

Annotation of /sml/trunk/src/MLRISC/sparc/mltree/sparc.sml

Parent Directory Parent Directory | Revision Log Revision Log


Revision 761 - (view) (download)

1 : monnier 245 (*
2 : monnier 411 * This is a new instruction selection module for Sparc,
3 :     * using the new instruction representation and the new MLTREE representation.
4 :     * Support for V9 has been added.
5 : monnier 245 *
6 : monnier 411 * The cc bit in arithmetic op are now embedded within the arithmetic
7 :     * opcode. This should save some space.
8 : monnier 245 *
9 : monnier 411 * -- Allen
10 : monnier 245 *)
11 :    
12 :     functor Sparc
13 :     (structure SparcInstr : SPARCINSTR
14 : monnier 411 structure SparcMLTree : MLTREE
15 :     structure PseudoInstrs : SPARC_PSEUDO_INSTR
16 : george 555 structure ExtensionComp : MLTREE_EXTENSION_COMP
17 :     where I = SparcInstr and T = SparcMLTree
18 : monnier 475 sharing SparcMLTree.Region = SparcInstr.Region
19 : george 545 sharing SparcMLTree.LabelExp = SparcInstr.LabelExp
20 : monnier 475 sharing PseudoInstrs.I = SparcInstr
21 : monnier 411 (*
22 :     * The client should also specify these parameters.
23 :     * These are the estimated cost of these instructions.
24 :     * The code generator will use alternative sequences that are
25 :     * cheaper when their costs are lower.
26 :     *)
27 : george 545 val muluCost : int ref (* cost of unsigned multiplication in cycles *)
28 : monnier 411 val divuCost : int ref (* cost of unsigned division in cycles *)
29 :     val multCost : int ref (* cost of trapping/signed multiplication in cycles *)
30 :     val divtCost : int ref (* cost of trapping/signed division in cycles *)
31 :    
32 :     (*
33 :     * If you don't want to use register windows at all, set this to false.
34 :     *)
35 :     val registerwindow : bool ref (* should we use register windows? *)
36 :    
37 :     val V9 : bool (* should we use V9 instruction set? *)
38 :     val useBR : bool ref
39 :     (* should we use the BR instruction (when in V9)?
40 :     * I think it is a good idea to use it.
41 :     *)
42 : monnier 245 ) : MLTREECOMP =
43 :     struct
44 : monnier 411 structure T = SparcMLTree
45 : monnier 429 structure S = T.Stream
46 : monnier 411 structure R = SparcMLTree.Region
47 :     structure I = SparcInstr
48 :     structure C = I.C
49 : george 545 structure LE = I.LabelExp
50 : monnier 245 structure W = Word32
51 :     structure P = PseudoInstrs
52 : george 545 structure A = MLRiscAnnotations
53 : monnier 245
54 : leunga 744 type instrStream = (I.instruction,C.cellset) T.stream
55 :     type mltreeStream = (T.stm,T.mlrisc list) T.stream
56 : george 545
57 : george 761 val int_0 = T.I.int_0
58 :     fun toInt n = T.I.toInt(32, n)
59 :     fun LI i = T.LI(T.I.fromInt(32, i))
60 :     fun LT (n,m) = T.I.LT(32, n, m)
61 :     fun LE (n,m) = T.I.LE(32, n, m)
62 :    
63 : leunga 624 val intTy = if V9 then 64 else 32
64 : monnier 411 structure Gen = MLTreeGen(structure T = T
65 : leunga 624 val intTy = intTy
66 : monnier 411 val naturalWidths = if V9 then [32,64] else [32]
67 : monnier 429 datatype rep = SE | ZE | NEITHER
68 :     val rep = NEITHER
69 : monnier 411 )
70 : monnier 245
71 : monnier 411 functor Multiply32 = MLTreeMult
72 :     (structure I = I
73 :     structure T = T
74 : monnier 429 type arg = {r1:C.cell,r2:C.cell,d:C.cell}
75 :     type argi = {r:C.cell,i:int,d:C.cell}
76 : monnier 411
77 :     val intTy = 32
78 :     fun mov{r,d} = I.COPY{dst=[d],src=[r],tmp=NONE,impl=ref NONE}
79 :     fun add{r1,r2,d} = I.ARITH{a=I.ADD,r=r1,i=I.REG r2,d=d}
80 :     fun slli{r,i,d} = [I.SHIFT{s=I.SLL,r=r,i=I.IMMED i,d=d}]
81 :     fun srli{r,i,d} = [I.SHIFT{s=I.SRL,r=r,i=I.IMMED i,d=d}]
82 :     fun srai{r,i,d} = [I.SHIFT{s=I.SRA,r=r,i=I.IMMED i,d=d}]
83 :     )
84 : monnier 245
85 : monnier 411 functor Multiply64 = MLTreeMult
86 :     (structure I = I
87 :     structure T = T
88 : monnier 429 type arg = {r1:C.cell,r2:C.cell,d:C.cell}
89 :     type argi = {r:C.cell,i:int,d:C.cell}
90 : monnier 411
91 :     val intTy = 64
92 :     fun mov{r,d} = I.COPY{dst=[d],src=[r],tmp=NONE,impl=ref NONE}
93 :     fun add{r1,r2,d} = I.ARITH{a=I.ADD,r=r1,i=I.REG r2,d=d}
94 :     fun slli{r,i,d} = [I.SHIFT{s=I.SLLX,r=r,i=I.IMMED i,d=d}]
95 :     fun srli{r,i,d} = [I.SHIFT{s=I.SRLX,r=r,i=I.IMMED i,d=d}]
96 :     fun srai{r,i,d} = [I.SHIFT{s=I.SRAX,r=r,i=I.IMMED i,d=d}]
97 :     )
98 : monnier 245
99 : monnier 411 (* signed, trapping version of multiply and divide *)
100 :     structure Mult32 = Multiply32
101 :     (val trapping = true
102 :     val multCost = multCost
103 :     fun addv{r1,r2,d} =
104 :     I.ARITH{a=I.ADDCC,r=r1,i=I.REG r2,d=d}::PseudoInstrs.overflowtrap32
105 :     fun subv{r1,r2,d} =
106 :     I.ARITH{a=I.SUBCC,r=r1,i=I.REG r2,d=d}::PseudoInstrs.overflowtrap32
107 :     val sh1addv = NONE
108 :     val sh2addv = NONE
109 :     val sh3addv = NONE
110 :     )
111 : monnier 429 (val signed = true)
112 : monnier 245
113 : monnier 411 (* unsigned, non-trapping version of multiply and divide *)
114 : leunga 657 functor Mul32 = Multiply32
115 : monnier 411 (val trapping = false
116 :     val multCost = muluCost
117 :     fun addv{r1,r2,d} = [I.ARITH{a=I.ADD,r=r1,i=I.REG r2,d=d}]
118 :     fun subv{r1,r2,d} = [I.ARITH{a=I.SUB,r=r1,i=I.REG r2,d=d}]
119 :     val sh1addv = NONE
120 :     val sh2addv = NONE
121 :     val sh3addv = NONE
122 :     )
123 : leunga 657 structure Mulu32 = Mul32(val signed = false)
124 : monnier 245
125 : leunga 657 structure Muls32 = Mul32(val signed = true)
126 :    
127 : monnier 411 (* signed, trapping version of multiply and divide *)
128 :     structure Mult64 = Multiply64
129 :     (val trapping = true
130 :     val multCost = multCost
131 :     fun addv{r1,r2,d} =
132 :     I.ARITH{a=I.ADDCC,r=r1,i=I.REG r2,d=d}::PseudoInstrs.overflowtrap64
133 :     fun subv{r1,r2,d} =
134 :     I.ARITH{a=I.SUBCC,r=r1,i=I.REG r2,d=d}::PseudoInstrs.overflowtrap64
135 :     val sh1addv = NONE
136 :     val sh2addv = NONE
137 :     val sh3addv = NONE
138 :     )
139 : monnier 429 (val signed = true)
140 : monnier 245
141 : monnier 411 (* unsigned, non-trapping version of multiply and divide *)
142 : leunga 657 functor Mul64 = Multiply64
143 : monnier 411 (val trapping = false
144 :     val multCost = muluCost
145 :     fun addv{r1,r2,d} = [I.ARITH{a=I.ADD,r=r1,i=I.REG r2,d=d}]
146 :     fun subv{r1,r2,d} = [I.ARITH{a=I.SUB,r=r1,i=I.REG r2,d=d}]
147 :     val sh1addv = NONE
148 :     val sh2addv = NONE
149 :     val sh3addv = NONE
150 :     )
151 : leunga 657 structure Mulu64 = Mul64(val signed = false)
152 : monnier 245
153 : leunga 657 structure Muls64 = Mul64(val signed = true)
154 :    
155 : monnier 411 datatype commutative = COMMUTE | NOCOMMUTE
156 :     datatype cc = REG (* write to register *)
157 :     | CC (* set condition code *)
158 :     | CC_REG (* do both *)
159 : monnier 245
160 : monnier 411 fun error msg = MLRiscErrorMsg.error("Sparc",msg)
161 : monnier 245
162 : leunga 744
163 :    
164 : monnier 411 fun selectInstructions
165 : george 545 (instrStream as
166 :     S.STREAM{emit,defineLabel,entryLabel,pseudoOp,annotation,
167 : leunga 744 beginCluster,endCluster,exitBlock,comment,...}) =
168 : monnier 411 let
169 :     (* Flags *)
170 :     val useBR = !useBR
171 :     val registerwindow = !registerwindow
172 : monnier 245
173 : leunga 744 val trap32 = PseudoInstrs.overflowtrap32
174 :     val trap64 = PseudoInstrs.overflowtrap64
175 :     val zeroR = C.r0
176 :     val newReg = C.newReg
177 : monnier 411 val newFreg = C.newFreg
178 : george 761 val int_m4096 = T.I.fromInt(32, ~4096)
179 :     val int_4096 = T.I.fromInt(32, 4096)
180 :     fun immed13 n = LE(int_m4096, n) andalso LT(n, int_4096)
181 : monnier 411 fun immed13w w = let val x = W.~>>(w,0w12)
182 :     in x = 0w0 orelse (W.notb x) = 0w0 end
183 :     fun splitw w = {hi=W.toInt(W.>>(w,0w10)),lo=W.toInt(W.andb(w,0wx3ff))}
184 : george 761 fun split n = splitw(T.I.toWord32(32, n))
185 : monnier 245
186 : monnier 411
187 : leunga 744 val zeroOpn = I.REG zeroR (* zero value operand *)
188 : monnier 245
189 : monnier 411 fun cond T.LT = I.BL
190 :     | cond T.LTU = I.BCS
191 :     | cond T.LE = I.BLE
192 :     | cond T.LEU = I.BLEU
193 :     | cond T.EQ = I.BE
194 :     | cond T.NE = I.BNE
195 :     | cond T.GE = I.BGE
196 :     | cond T.GEU = I.BCC
197 :     | cond T.GT = I.BG
198 :     | cond T.GTU = I.BGU
199 : leunga 744 | cond _ = error "cond"
200 : monnier 245
201 : monnier 411 fun rcond T.LT = I.RLZ
202 :     | rcond T.LE = I.RLEZ
203 :     | rcond T.EQ = I.RZ
204 :     | rcond T.NE = I.RNZ
205 :     | rcond T.GE = I.RGEZ
206 :     | rcond T.GT = I.RGZ
207 :     | rcond _ = error "rcond"
208 : monnier 245
209 : monnier 411 fun signedCmp(T.LT | T.LE | T.EQ | T.NE | T.GE | T.GT) = true
210 :     | signedCmp _ = false
211 : monnier 245
212 : monnier 411 fun fcond T.== = I.FBE
213 :     | fcond T.?<> = I.FBNE
214 :     | fcond T.? = I.FBU
215 :     | fcond T.<=> = I.FBO
216 :     | fcond T.> = I.FBG
217 :     | fcond T.>= = I.FBGE
218 :     | fcond T.?> = I.FBUG
219 :     | fcond T.?>= = I.FBUGE
220 :     | fcond T.< = I.FBL
221 :     | fcond T.<= = I.FBLE
222 :     | fcond T.?< = I.FBUL
223 :     | fcond T.?<= = I.FBULE
224 :     | fcond T.<> = I.FBLG
225 :     | fcond T.?= = I.FBUE
226 : george 545 | fcond fc = error("fcond "^T.Basis.fcondToString fc)
227 : monnier 245
228 : monnier 411 fun mark'(i,[]) = i
229 :     | mark'(i,a::an) = mark'(I.ANNOTATION{i=i,a=a},an)
230 : monnier 245
231 : monnier 411 fun mark(i,an) = emit(mark'(i,an))
232 : monnier 245
233 : monnier 411 (* convert an operand into a register *)
234 :     fun reduceOpn(I.REG r) = r
235 : leunga 744 | reduceOpn(I.IMMED 0) = zeroR
236 : monnier 411 | reduceOpn i =
237 :     let val d = newReg()
238 : leunga 744 in emit(I.ARITH{a=I.OR,r=zeroR,i=i,d=d}); d end
239 : monnier 245
240 : monnier 411 (* emit parallel copies *)
241 :     fun copy(dst,src,an) =
242 :     mark(I.COPY{dst=dst,src=src,impl=ref NONE,
243 :     tmp=case dst of [_] => NONE
244 :     | _ => SOME(I.Direct(newReg()))},an)
245 :     fun fcopy(dst,src,an) =
246 :     mark(I.FCOPY{dst=dst,src=src,impl=ref NONE,
247 :     tmp=case dst of [_] => NONE
248 :     | _ => SOME(I.FDirect(newFreg()))},an)
249 : monnier 245
250 : monnier 411 (* move register s to register d *)
251 :     fun move(s,d,an) =
252 : leunga 744 if C.sameColor(s,d) orelse C.registerId d = 0 then ()
253 : monnier 411 else mark(I.COPY{dst=[d],src=[s],tmp=NONE,impl=ref NONE},an)
254 :    
255 :     (* move floating point register s to register d *)
256 :     fun fmoved(s,d,an) =
257 : leunga 744 if C.sameColor(s,d) then ()
258 : monnier 411 else mark(I.FCOPY{dst=[d],src=[s],tmp=NONE,impl=ref NONE},an)
259 : monnier 475 fun fmoves(s,d,an) = fmoved(s,d,an) (* error "fmoves" for now!!! XXX *)
260 : monnier 411 fun fmoveq(s,d,an) = error "fmoveq"
261 :    
262 :     (* load immediate *)
263 :     and loadImmed(n,d,cc,an) =
264 :     let val or = if cc <> REG then I.ORCC else I.OR
265 : george 761 in if immed13 n then mark(I.ARITH{a=or,r=zeroR,i=I.IMMED(toInt n),d=d},an)
266 : monnier 411 else let val {hi,lo} = split n
267 :     in if lo = 0 then
268 :     (mark(I.SETHI{i=hi,d=d},an); genCmp0(cc,d))
269 :     else let val t = newReg()
270 :     in emit(I.SETHI{i=hi,d=t});
271 :     mark(I.ARITH{a=or,r=t,i=I.IMMED lo,d=d},an)
272 :     end
273 :     end
274 :     end
275 : monnier 245
276 : monnier 411 (* load label expression *)
277 :     and loadLabel(lab,d,cc,an) =
278 :     let val or = if cc <> REG then I.ORCC else I.OR
279 : leunga 744 in mark(I.ARITH{a=or,r=zeroR,i=I.LAB lab,d=d},an) end
280 : monnier 245
281 : monnier 411 (* emit an arithmetic op *)
282 :     and arith(a,acc,e1,e2,d,cc,comm,trap,an) =
283 :     let val (a,d) = case cc of
284 :     REG => (a,d)
285 : leunga 744 | CC => (acc,zeroR)
286 : monnier 411 | CC_REG => (acc,d)
287 :     in case (opn e1,opn e2,comm) of
288 :     (i,I.REG r,COMMUTE)=> mark(I.ARITH{a=a,r=r,i=i,d=d},an)
289 :     | (I.REG r,i,_) => mark(I.ARITH{a=a,r=r,i=i,d=d},an)
290 :     | (r,i,_) => mark(I.ARITH{a=a,r=reduceOpn r,i=i,d=d},an)
291 :     ;
292 :     case trap of [] => () | _ => app emit trap
293 :     end
294 : monnier 245
295 : monnier 411 (* emit a shift op *)
296 :     and shift(s,e1,e2,d,cc,an) =
297 :     (mark(I.SHIFT{s=s,r=expr e1,i=opn e2,d=d},an);
298 :     genCmp0(cc,d)
299 :     )
300 : monnier 245
301 : monnier 411 (* emit externally defined multiply or division operation (V8) *)
302 :     and extarith(gen,genConst,e1,e2,d,cc,comm) =
303 :     let fun nonconst(e1,e2) =
304 :     case (opn e1,opn e2,comm) of
305 :     (i,I.REG r,COMMUTE) => gen({r=r,i=i,d=d},reduceOpn)
306 :     | (I.REG r,i,_) => gen({r=r,i=i,d=d},reduceOpn)
307 :     | (r,i,_) => gen({r=reduceOpn r,i=i,d=d},reduceOpn)
308 :     fun const(e,i) =
309 :     let val r = expr e
310 : george 761 in genConst{r=r,i=toInt i,d=d}
311 : monnier 411 handle _ => gen({r=r,i=opn(T.LI i),d=d},reduceOpn)
312 :     end
313 :     val instrs =
314 :     case (comm,e1,e2) of
315 :     (_,e1,T.LI i) => const(e1,i)
316 :     | (COMMUTE,T.LI i,e2) => const(e2,i)
317 :     | _ => nonconst(e1,e2)
318 :     in app emit instrs;
319 :     genCmp0(cc,d)
320 :     end
321 : monnier 245
322 : monnier 411 (* emit 64-bit multiply or division operation (V9) *)
323 :     and muldiv64(a,genConst,e1,e2,d,cc,comm,an) =
324 :     let fun nonconst(e1,e2) =
325 :     [mark'(
326 :     case (opn e1,opn e2,comm) of
327 :     (i,I.REG r,COMMUTE) => I.ARITH{a=a,r=r,i=i,d=d}
328 :     | (I.REG r,i,_) => I.ARITH{a=a,r=r,i=i,d=d}
329 :     | (r,i,_) => I.ARITH{a=a,r=reduceOpn r,i=i,d=d},an)
330 :     ]
331 :     fun const(e,i) =
332 :     let val r = expr e
333 : george 761 in genConst{r=r,i=toInt i,d=d}
334 : monnier 411 handle _ => [mark'(I.ARITH{a=a,r=r,i=opn(T.LI i),d=d},an)]
335 :     end
336 :     val instrs =
337 :     case (comm,e1,e2) of
338 :     (_,e1,T.LI i) => const(e1,i)
339 :     | (COMMUTE,T.LI i,e2) => const(e2,i)
340 :     | _ => nonconst(e1,e2)
341 :     in app emit instrs;
342 :     genCmp0(cc,d)
343 :     end
344 :    
345 :     (* divisions *)
346 : george 545 and divu32 x = Mulu32.divide{mode=T.TO_ZERO,stm=doStmt} x
347 : leunga 657 and divs32 x = Muls32.divide{mode=T.TO_ZERO,stm=doStmt} x
348 : george 545 and divt32 x = Mult32.divide{mode=T.TO_ZERO,stm=doStmt} x
349 :     and divu64 x = Mulu64.divide{mode=T.TO_ZERO,stm=doStmt} x
350 : leunga 657 and divs64 x = Muls64.divide{mode=T.TO_ZERO,stm=doStmt} x
351 : george 545 and divt64 x = Mult64.divide{mode=T.TO_ZERO,stm=doStmt} x
352 : monnier 411
353 :     (* emit an unary floating point op *)
354 :     and funary(a,e,d,an) = mark(I.FPop1{a=a,r=fexpr e,d=d},an)
355 :    
356 :     (* emit a binary floating point op *)
357 :     and farith(a,e1,e2,d,an) =
358 :     mark(I.FPop2{a=a,r1=fexpr e1,r2=fexpr e2,d=d},an)
359 :    
360 :     (* convert an expression into an addressing mode *)
361 :     and addr(T.ADD(_,e,T.LI n)) =
362 : george 761 if immed13 n then (expr e,I.IMMED(toInt n))
363 : monnier 411 else let val d = newReg()
364 :     in loadImmed(n,d,REG,[]); (d,opn e) end
365 : george 545 | addr(T.ADD(_,e,T.CONST c)) = (expr e,I.LAB(LE.CONST c))
366 : monnier 411 | addr(T.ADD(_,e,T.LABEL l)) = (expr e,I.LAB l)
367 :     | addr(T.ADD(ty,i as T.LI _,e)) = addr(T.ADD(ty,e,i))
368 : george 545 | addr(T.ADD(_,T.CONST c,e)) = (expr e,I.LAB(LE.CONST c))
369 : monnier 411 | addr(T.ADD(_,T.LABEL l,e)) = (expr e,I.LAB l)
370 :     | addr(T.ADD(_,e1,e2)) = (expr e1,I.REG(expr e2))
371 : george 761 | addr(T.SUB(ty,e,T.LI n)) = addr(T.ADD(ty,e,T.LI(T.I.NEG(32,n))))
372 : leunga 744 | addr(T.LABEL l) = (zeroR,I.LAB l)
373 : monnier 411 | addr a = (expr a,zeroOpn)
374 :    
375 :     (* emit an integer load *)
376 :     and load(l,a,d,mem,cc,an) =
377 :     let val (r,i) = addr a
378 :     in mark(I.LOAD{l=l,r=r,i=i,d=d,mem=mem},an);
379 :     genCmp0(cc,d)
380 :     end
381 :    
382 :     (* emit an integer store *)
383 :     and store(s,a,d,mem,an) =
384 :     let val (r,i) = addr a
385 :     in mark(I.STORE{s=s,r=r,i=i,d=expr d,mem=mem},an) end
386 :    
387 :     (* emit a floating point load *)
388 :     and fload(l,a,d,mem,an) =
389 :     let val (r,i) = addr a
390 :     in mark(I.FLOAD{l=l,r=r,i=i,d=d,mem=mem},an) end
391 :    
392 :     (* emit a floating point store *)
393 :     and fstore(s,a,d,mem,an) =
394 :     let val (r,i) = addr a
395 :     in mark(I.FSTORE{s=s,r=r,i=i,d=fexpr d,mem=mem},an) end
396 :    
397 :     (* emit a jump *)
398 :     and jmp(a,labs,an) =
399 :     let val (r,i) = addr a
400 :     in mark(I.JMP{r=r,i=i,labs=labs,nop=true},an) end
401 :    
402 : george 545 (* convert mlrisc to cellset *)
403 :     and cellset mlrisc =
404 :     let fun g([],set) = set
405 : leunga 744 | g(T.GPR(T.REG(_,r))::regs,set) = g(regs,C.CellSet.add(r,set))
406 :     | g(T.FPR(T.FREG(_,f))::regs,set) = g(regs,C.CellSet.add(f,set))
407 :     | g(T.CCR(T.CC(_,cc))::regs,set) = g(regs,C.CellSet.add(cc,set))
408 : george 545 | g(_::regs, set) = g(regs,set)
409 :     in g(mlrisc, C.empty) end
410 :    
411 : monnier 411 (* emit a function call *)
412 : george 545 and call(a,flow,defs,uses,mem,an) =
413 : monnier 411 let val (r,i) = addr a
414 : george 545 val defs=cellset(defs)
415 :     val uses=cellset(uses)
416 : leunga 744 in case (C.registerId r,i) of
417 : monnier 245 (0,I.LAB(LE.LABEL l)) =>
418 : monnier 411 mark(I.CALL{label=l,defs=C.addReg(C.linkReg,defs),uses=uses,
419 :     mem=mem,nop=true},an)
420 :     | _ => mark(I.JMPL{r=r,i=i,d=C.linkReg,defs=defs,uses=uses,mem=mem,
421 :     nop=true},an)
422 : monnier 245 end
423 :    
424 : monnier 411 (* emit an integer branch instruction *)
425 : leunga 744 and branch(T.CMP(ty,cond,a,b),lab,an) =
426 : monnier 411 let val (cond,a,b) =
427 :     case a of
428 : george 761 (T.LI _ | T.CONST _ | T.LABEL _) =>
429 : george 545 (T.Basis.swapCond cond,b,a)
430 : monnier 411 | _ => (cond,a,b)
431 :     in if V9 then
432 :     branchV9(cond,a,b,lab,an)
433 :     else
434 :     (doExpr(T.SUB(ty,a,b),newReg(),CC,[]); br(cond,lab,an))
435 :     end
436 : leunga 744 | branch(T.CC(cond,r),lab,an) =
437 :     if C.sameCell(r, C.psr) then br(cond,lab,an)
438 :     else (genCmp0(CC,r); br(cond,lab,an))
439 :     | branch(T.FCMP(fty,cond,a,b),lab,an) =
440 : george 545 let val cmp = case fty of
441 :     32 => I.FCMPs
442 :     | 64 => I.FCMPd
443 :     | _ => error "fbranch"
444 :     in emit(I.FCMP{cmp=cmp,r1=fexpr a,r2=fexpr b,nop=true});
445 :     mark(I.FBfcc{b=fcond cond,a=false,label=lab,nop=true},an)
446 :     end
447 : monnier 411 | branch _ = error "branch"
448 : monnier 245
449 : monnier 411 and branchV9(cond,a,b,lab,an) =
450 : leunga 624 let val size = Gen.Size.size a
451 : monnier 411 in if useBR andalso signedCmp cond then
452 :     let val r = newReg()
453 :     in doExpr(T.SUB(size,a,b),r,REG,[]);
454 :     brcond(cond,r,lab,an)
455 :     end
456 :     else
457 :     let val cc = case size of 32 => I.ICC
458 :     | 64 => I.XCC
459 :     | _ => error "branchV9"
460 :     in doExpr(T.SUB(size,a,b),newReg(),CC,[]);
461 :     bp(cond,cc,lab,an)
462 :     end
463 :     end
464 : monnier 245
465 : monnier 411 and br(c,lab,an) = mark(I.Bicc{b=cond c,a=true,label=lab,nop=true},an)
466 : monnier 245
467 : monnier 411 and brcond(c,r,lab,an) =
468 :     mark(I.BR{rcond=rcond c,r=r,p=I.PT,a=true,label=lab,nop=true},an)
469 : monnier 245
470 : monnier 411 and bp(c,cc,lab,an) =
471 :     mark(I.BP{b=cond c,cc=cc,p=I.PT,a=true,label=lab,nop=true},an)
472 : monnier 245
473 : monnier 411 (* generate code for a statement *)
474 :     and stmt(T.MV(_,d,e),an) = doExpr(e,d,REG,an)
475 :     | stmt(T.FMV(_,d,e),an) = doFexpr(e,d,an)
476 :     | stmt(T.CCMV(d,e),an) = doCCexpr(e,d,an)
477 :     | stmt(T.COPY(_,dst,src),an) = copy(dst,src,an)
478 : monnier 475 | stmt(T.FCOPY(_,dst,src),an) = fcopy(dst,src,an)
479 : leunga 744 | stmt(T.JMP(T.LABEL(LE.LABEL l),_),an) =
480 : monnier 411 mark(I.Bicc{b=I.BA,a=true,label=l,nop=false},an)
481 : leunga 744 | stmt(T.JMP(e,labs),an) = jmp(e,labs,an)
482 :     | stmt(T.CALL{funct,targets,defs,uses,region,...},an) =
483 : leunga 591 call(funct,targets,defs,uses,region,an)
484 : george 545 | stmt(T.RET _,an) = mark(I.RET{leaf=not registerwindow,nop=true},an)
485 : monnier 411 | stmt(T.STORE(8,a,d,mem),an) = store(I.STB,a,d,mem,an)
486 :     | stmt(T.STORE(16,a,d,mem),an) = store(I.STH,a,d,mem,an)
487 :     | stmt(T.STORE(32,a,d,mem),an) = store(I.ST,a,d,mem,an)
488 :     | stmt(T.STORE(64,a,d,mem),an) =
489 :     store(if V9 then I.STX else I.STD,a,d,mem,an)
490 :     | stmt(T.FSTORE(32,a,d,mem),an) = fstore(I.STF,a,d,mem,an)
491 :     | stmt(T.FSTORE(64,a,d,mem),an) = fstore(I.STDF,a,d,mem,an)
492 : leunga 744 | stmt(T.BCC(cc,lab),an) = branch(cc,lab,an)
493 : george 545 | stmt(T.DEFINE l,_) = defineLabel l
494 : monnier 411 | stmt(T.ANNOTATION(s,a),an) = stmt(s,a::an)
495 : george 555 | stmt(T.EXT s,an) = ExtensionComp.compileSext(reducer()) {stm=s, an=an}
496 : george 545 | stmt(s,an) = doStmts(Gen.compileStm s)
497 : monnier 245
498 : monnier 411 and doStmt s = stmt(s,[])
499 : monnier 245
500 : george 545 and doStmts ss = app doStmt ss
501 : monnier 245
502 : monnier 411 (* convert an expression into a register *)
503 : george 761 and expr e = let
504 :     fun comp() = let
505 :     val d = newReg()
506 :     in doExpr(e, d, REG, []); d
507 :     end
508 :     in case e
509 :     of T.REG(_,r) => r
510 :     | T.LI z => if T.I.isZero z then zeroR else comp()
511 :     | _ => comp()
512 :     end
513 : monnier 245
514 : monnier 411 (* compute an integer expression and put the result in register d
515 :     * If cc is set then set the condition code with the result.
516 :     *)
517 :     and doExpr(e,d,cc,an) =
518 :     case e of
519 :     T.REG(_,r) => (move(r,d,an); genCmp0(cc,r))
520 :     | T.LI n => loadImmed(n,d,cc,an)
521 :     | T.LABEL l => loadLabel(l,d,cc,an)
522 : george 545 | T.CONST c => loadLabel(LE.CONST c,d,cc,an)
523 : monnier 245
524 : monnier 411 (* generic 32/64 bit support *)
525 :     | T.ADD(_,a,b) => arith(I.ADD,I.ADDCC,a,b,d,cc,COMMUTE,[],an)
526 : george 761 | T.SUB(_,a,b) => let
527 :     fun default() = arith(I.SUB,I.SUBCC,a,b,d,cc,NOCOMMUTE,[],an)
528 :     in
529 :     case b
530 :     of T.LI z =>
531 :     if T.I.isZero(z) then doExpr(a,d,cc,an) else default()
532 :     | _ => default()
533 :     (*esac*)
534 :     end
535 :    
536 : monnier 411 | T.ANDB(_,a,T.NOTB(_,b)) =>
537 :     arith(I.ANDN,I.ANDNCC,a,b,d,cc,NOCOMMUTE,[],an)
538 :     | T.ORB(_,a,T.NOTB(_,b)) =>
539 :     arith(I.ORN,I.ORNCC,a,b,d,cc,NOCOMMUTE,[],an)
540 :     | T.XORB(_,a,T.NOTB(_,b)) =>
541 :     arith(I.XNOR,I.XNORCC,a,b,d,cc,COMMUTE,[],an)
542 :     | T.ANDB(_,T.NOTB(_,a),b) =>
543 :     arith(I.ANDN,I.ANDNCC,b,a,d,cc,NOCOMMUTE,[],an)
544 :     | T.ORB(_,T.NOTB(_,a),b) =>
545 :     arith(I.ORN,I.ORNCC,b,a,d,cc,NOCOMMUTE,[],an)
546 :     | T.XORB(_,T.NOTB(_,a),b) =>
547 :     arith(I.XNOR,I.XNORCC,b,a,d,cc,COMMUTE,[],an)
548 :     | T.NOTB(_,T.XORB(_,a,b)) =>
549 :     arith(I.XNOR,I.XNORCC,a,b,d,cc,COMMUTE,[],an)
550 : monnier 245
551 : monnier 411 | T.ANDB(_,a,b) => arith(I.AND,I.ANDCC,a,b,d,cc,COMMUTE,[],an)
552 :     | T.ORB(_,a,b) => arith(I.OR,I.ORCC,a,b,d,cc,COMMUTE,[],an)
553 :     | T.XORB(_,a,b) => arith(I.XOR,I.XORCC,a,b,d,cc,COMMUTE,[],an)
554 : george 761 | T.NOTB(_,a) => arith(I.XNOR,I.XNORCC,a,LI 0,d,cc,COMMUTE,[],an)
555 : monnier 245
556 : monnier 411 (* 32 bit support *)
557 :     | T.SRA(32,a,b) => shift(I.SRA,a,b,d,cc,an)
558 :     | T.SRL(32,a,b) => shift(I.SRL,a,b,d,cc,an)
559 :     | T.SLL(32,a,b) => shift(I.SLL,a,b,d,cc,an)
560 :     | T.ADDT(32,a,b)=>
561 :     arith(I.ADDCC,I.ADDCC,a,b,d,CC_REG,COMMUTE,trap32,an)
562 :     | T.SUBT(32,a,b)=>
563 :     arith(I.SUBCC,I.SUBCC,a,b,d,CC_REG,NOCOMMUTE,trap32,an)
564 : leunga 657 | T.MULU(32,a,b) => extarith(P.umul32,
565 :     Mulu32.multiply,a,b,d,cc,COMMUTE)
566 :     | T.MULS(32,a,b) => extarith(P.smul32,
567 :     Muls32.multiply,a,b,d,cc,COMMUTE)
568 :     | T.MULT(32,a,b) => extarith(P.smul32trap,
569 :     Mult32.multiply,a,b,d,cc,COMMUTE)
570 :     | T.DIVU(32,a,b) => extarith(P.udiv32,divu32,a,b,d,cc,NOCOMMUTE)
571 :     | T.DIVS(32,a,b) => extarith(P.sdiv32,divs32,a,b,d,cc,NOCOMMUTE)
572 :     | T.DIVT(32,a,b) => extarith(P.sdiv32trap,divt32,a,b,d,cc,NOCOMMUTE)
573 : monnier 245
574 : monnier 411 (* 64 bit support *)
575 :     | T.SRA(64,a,b) => shift(I.SRAX,a,b,d,cc,an)
576 :     | T.SRL(64,a,b) => shift(I.SRLX,a,b,d,cc,an)
577 :     | T.SLL(64,a,b) => shift(I.SLLX,a,b,d,cc,an)
578 :     | T.ADDT(64,a,b)=>
579 :     arith(I.ADDCC,I.ADDCC,a,b,d,CC_REG,COMMUTE,trap64,an)
580 :     | T.SUBT(64,a,b)=>
581 :     arith(I.SUBCC,I.SUBCC,a,b,d,CC_REG,NOCOMMUTE,trap64,an)
582 :     | T.MULU(64,a,b) =>
583 :     muldiv64(I.MULX,Mulu64.multiply,a,b,d,cc,COMMUTE,an)
584 : leunga 657 | T.MULS(64,a,b) =>
585 :     muldiv64(I.MULX,Muls64.multiply,a,b,d,cc,COMMUTE,an)
586 : monnier 411 | T.MULT(64,a,b) =>
587 :     (muldiv64(I.MULX,Mult64.multiply,a,b,d,CC_REG,COMMUTE,an);
588 :     app emit trap64)
589 :     | T.DIVU(64,a,b) => muldiv64(I.UDIVX,divu64,a,b,d,cc,NOCOMMUTE,an)
590 : leunga 657 | T.DIVS(64,a,b) => muldiv64(I.SDIVX,divs64,a,b,d,cc,NOCOMMUTE,an)
591 : monnier 411 | T.DIVT(64,a,b) => muldiv64(I.SDIVX,divt64,a,b,d,cc,NOCOMMUTE,an)
592 : monnier 245
593 : monnier 411 (* loads *)
594 :     | T.LOAD(8,a,mem) => load(I.LDUB,a,d,mem,cc,an)
595 : leunga 744 | T.SX(_,_,T.LOAD(8,a,mem)) => load(I.LDSB,a,d,mem,cc,an)
596 : monnier 411 | T.LOAD(16,a,mem) => load(I.LDUH,a,d,mem,cc,an)
597 : leunga 744 | T.SX(_,_,T.LOAD(16,a,mem)) => load(I.LDSH,a,d,mem,cc,an)
598 : monnier 411 | T.LOAD(32,a,mem) => load(I.LD,a,d,mem,cc,an)
599 : george 545 | T.LOAD(64,a,mem) =>
600 :     load(if V9 then I.LDX else I.LDD,a,d,mem,cc,an)
601 : monnier 245
602 : monnier 411 (* conditional expression *)
603 : george 545 | T.COND exp => doStmts (Gen.compileCond{exp=exp,rd=d,an=an})
604 : monnier 411
605 :     (* misc *)
606 : george 545 | T.LET(s,e) => (doStmt s; doExpr(e, d, cc, an))
607 :     | T.MARK(e,A.MARKREG f) => (f d; doExpr(e,d,cc,an))
608 :     | T.MARK(e,a) => doExpr(e,d,cc,a::an)
609 :     | T.PRED(e,c) => doExpr(e,d,cc,A.CTRLUSE c::an)
610 : george 555 | T.REXT e => ExtensionComp.compileRext (reducer()) {e=e, rd=d, an=an}
611 : george 545 | e => doExpr(Gen.compileRexp e,d,cc,an)
612 : monnier 411
613 :     (* generate a comparison with zero *)
614 :     and genCmp0(REG,_) = ()
615 : leunga 744 | genCmp0(_,d) = emit(I.ARITH{a=I.SUBCC,r=d,i=zeroOpn,d=zeroR})
616 : monnier 411
617 :     (* convert an expression into a floating point register *)
618 :     and fexpr(T.FREG(_,r)) = r
619 :     | fexpr e = let val d = newFreg() in doFexpr(e,d,[]); d end
620 :    
621 :     (* compute a floating point expression and put the result in d *)
622 :     and doFexpr(e,d,an) =
623 :     case e of
624 :     (* single precision *)
625 :     T.FREG(32,r) => fmoves(r,d,an)
626 :     | T.FLOAD(32,ea,mem) => fload(I.LDF,ea,d,mem,an)
627 :     | T.FADD(32,a,b) => farith(I.FADDs,a,b,d,an)
628 :     | T.FSUB(32,a,b) => farith(I.FSUBs,a,b,d,an)
629 :     | T.FMUL(32,a,b) => farith(I.FMULs,a,b,d,an)
630 :     | T.FDIV(32,a,b) => farith(I.FDIVs,a,b,d,an)
631 :     | T.FABS(32,a) => funary(I.FABSs,a,d,an)
632 :     | T.FNEG(32,a) => funary(I.FNEGs,a,d,an)
633 :     | T.FSQRT(32,a) => funary(I.FSQRTs,a,d,an)
634 :    
635 :     (* double precision *)
636 :     | T.FREG(64,r) => fmoved(r,d,an)
637 :     | T.FLOAD(64,ea,mem) => fload(I.LDDF,ea,d,mem,an)
638 :     | T.FADD(64,a,b) => farith(I.FADDd,a,b,d,an)
639 :     | T.FSUB(64,a,b) => farith(I.FSUBd,a,b,d,an)
640 :     | T.FMUL(64,a,b) => farith(I.FMULd,a,b,d,an)
641 :     | T.FDIV(64,a,b) => farith(I.FDIVd,a,b,d,an)
642 :     | T.FABS(64,a) => funary(I.FABSd,a,d,an)
643 :     | T.FNEG(64,a) => funary(I.FNEGd,a,d,an)
644 :     | T.FSQRT(64,a) => funary(I.FSQRTd,a,d,an)
645 :    
646 :     (* quad precision *)
647 :     | T.FREG(128,r) => fmoveq(r,d,an)
648 :     | T.FADD(128,a,b) => farith(I.FADDq,a,b,d,an)
649 :     | T.FSUB(128,a,b) => farith(I.FSUBq,a,b,d,an)
650 :     | T.FMUL(128,a,b) => farith(I.FMULq,a,b,d,an)
651 :     | T.FDIV(128,a,b) => farith(I.FDIVq,a,b,d,an)
652 :     | T.FABS(128,a) => funary(I.FABSq,a,d,an)
653 :     | T.FNEG(128,a) => funary(I.FNEGq,a,d,an)
654 :     | T.FSQRT(128,a) => funary(I.FSQRTq,a,d,an)
655 :    
656 :     (* floating point to floating point *)
657 : george 545 | T.CVTF2F(ty,ty',e) =>
658 : monnier 475 (case (ty,ty') of
659 :     (32,32) => doFexpr(e,d,an)
660 :     | (64,32) => funary(I.FsTOd,e,d,an)
661 : monnier 411 | (128,32) => funary(I.FsTOq,e,d,an)
662 : monnier 475 | (32,64) => funary(I.FdTOs,e,d,an)
663 :     | (64,64) => doFexpr(e,d,an)
664 : monnier 411 | (128,64) => funary(I.FdTOq,e,d,an)
665 :     | (32,128) => funary(I.FqTOs,e,d,an)
666 :     | (64,128) => funary(I.FqTOd,e,d,an)
667 :     | (128,128) => doFexpr(e,d,an)
668 :     | _ => error "CVTF2F"
669 :     )
670 :    
671 :     (* integer to floating point *)
672 : george 545 | T.CVTI2F(32,32,e) => app emit (P.cvti2s({i=opn e,d=d},reduceOpn))
673 :     | T.CVTI2F(64,32,e) => app emit (P.cvti2d({i=opn e,d=d},reduceOpn))
674 :     | T.CVTI2F(128,32,e) => app emit (P.cvti2q({i=opn e,d=d},reduceOpn))
675 : monnier 411
676 : george 545 | T.FMARK(e,A.MARKREG f) => (f d; doFexpr(e,d,an))
677 :     | T.FMARK(e,a) => doFexpr(e,d,a::an)
678 :     | T.FPRED(e,c) => doFexpr(e,d,A.CTRLUSE c::an)
679 : george 555 | T.FEXT e => ExtensionComp.compileFext (reducer()) {e=e, fd=d, an=an}
680 : george 545 | e => doFexpr(Gen.compileFexp e,d,an)
681 : monnier 411
682 : leunga 744 and doCCexpr(T.CMP(ty,cond,e1,e2),cc,an) =
683 :     if C.sameCell(cc,C.psr) then
684 :     doExpr(T.SUB(ty,e1,e2),newReg(),CC,an)
685 :     else error "doCCexpr"
686 :     | doCCexpr(T.CC(_,r),d,an) =
687 :     if C.sameColor(r,C.psr) then error "doCCexpr"
688 :     else move(r,d,an)
689 : george 545 | doCCexpr(T.CCMARK(e,A.MARKREG f),d,an) = (f d; doCCexpr(e,d,an))
690 :     | doCCexpr(T.CCMARK(e,a),d,an) = doCCexpr(e,d,a::an)
691 :     | doCCexpr(T.CCEXT e,d,an) =
692 : george 555 ExtensionComp.compileCCext (reducer()) {e=e, ccd=d, an=an}
693 : monnier 411 | doCCexpr e = error "doCCexpr"
694 :    
695 :     and ccExpr e = let val d = newReg() in doCCexpr(e,d,[]); d end
696 :    
697 :     (* convert an expression into an operand *)
698 : george 761 and opn(T.CONST c) = I.LAB(LE.CONST c)
699 : monnier 411 | opn(T.LABEL l) = I.LAB l
700 : george 761 | opn(e as T.LI n) =
701 :     if T.I.isZero(n) then zeroOpn
702 :     else if immed13 n then I.IMMED(toInt n)
703 :     else I.REG(expr e)
704 : monnier 411 | opn e = I.REG(expr e)
705 :    
706 : george 545 and reducer() =
707 :     T.REDUCER{reduceRexp = expr,
708 :     reduceFexp = fexpr,
709 :     reduceCCexp = ccExpr,
710 :     reduceStm = stmt,
711 :     operand = opn,
712 :     reduceOperand = reduceOpn,
713 :     addressOf = addr,
714 :     emit = mark,
715 :     instrStream = instrStream,
716 :     mltreeStream = self()
717 :     }
718 :     and self() =
719 :     S.STREAM
720 :     { beginCluster= beginCluster,
721 :     endCluster = endCluster,
722 :     emit = doStmt,
723 :     pseudoOp = pseudoOp,
724 :     defineLabel = defineLabel,
725 :     entryLabel = entryLabel,
726 :     comment = comment,
727 :     annotation = annotation,
728 : leunga 744 exitBlock = fn regs => exitBlock(cellset regs)
729 : george 545 }
730 :     in self()
731 : monnier 245 end
732 :    
733 :     end
734 :    
735 : monnier 411 (*
736 :     * Machine code generator for SPARC.
737 : monnier 245 *
738 : monnier 411 * The SPARC architecture has 32 general purpose registers (%g0 is always 0)
739 :     * and 32 single precision floating point registers.
740 : monnier 245 *
741 : monnier 411 * Some Ugliness: double precision floating point registers are
742 :     * register pairs. There are no double precision moves, negation and absolute
743 :     * values. These require two single precision operations. I've created
744 :     * composite instructions FMOVd, FNEGd and FABSd to stand for these.
745 : monnier 245 *
746 : monnier 411 * All integer arithmetic instructions can optionally set the condition
747 :     * code register. We use this to simplify certain comparisons with zero.
748 : monnier 245 *
749 : monnier 411 * Integer multiplication, division and conversion from integer to floating
750 :     * go thru the pseudo instruction interface, since older sparcs do not
751 :     * implement these instructions in hardware.
752 : monnier 245 *
753 : monnier 411 * In addition, the trap instruction for detecting overflow is a parameter.
754 :     * This allows different trap vectors to be used.
755 : monnier 245 *
756 : monnier 411 * -- Allen
757 :     *)

root@smlnj-gforge.cs.uchicago.edu
ViewVC Help
Powered by ViewVC 1.0.0