1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
|
{
Copyright (c) 1998-2002 by Jonas Maebe, member of the Free Pascal
Development Team
This unit implements the PowerPC optimizer object
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License
along with this program; if not, write to the Free Software
Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
****************************************************************************
}
Unit aoptcpu;
Interface
{$i fpcdefs.inc}
uses cpubase, aoptobj, aoptcpub, aopt, aasmtai,aasmdata, aasmcpu;
Type
TCpuAsmOptimizer = class(TAsmOptimizer)
{ uses the same constructor as TAopObj }
function PeepHoleOptPass1Cpu(var p: tai): boolean; override;
function PostPeepHoleOptsCpu(var p: tai): boolean; override;
private
function cmpi_mfcr_opt(p, next1, next2: taicpu): boolean;
End;
Implementation
uses
cutils, verbose, cgbase, cgcpu, cgobj;
function TCpuAsmOptimizer.cmpi_mfcr_opt(p, next1, next2: taicpu): boolean;
var
next3, prev: tai;
inverse, prevrlwinm: boolean;
begin
result := true;
inverse :=
getnextinstruction(next2,next3) and
(next3.typ = ait_instruction) and
(taicpu(next3).opcode = A_XORI) and
(taicpu(next3).oper[0]^.reg = taicpu(next3).oper[1]^.reg) and
(taicpu(next3).oper[0]^.reg = taicpu(next2).oper[0]^.reg) and
(taicpu(next3).oper[2]^.val = 1);
case taicpu(next2).oper[2]^.val of
1:
begin
// less than zero or greater/equal than zero (the xori remains in
// in the latter case). Doesn't make sense for unsigned comparisons.
if (p.opcode = A_CMPWI) then
begin
p.opcode := A_SRWI;
p.ops := 3;
p.loadreg(1,p.oper[0]^.reg);
p.loadreg(0,next1.oper[0]^.reg);
p.loadconst(2,31);
asml.remove(next1);
next1.free;
asml.remove(next2);
next2.free;
end
else
result := false;
end;
{
needs two registers to work with
2:
begin
// greater or less/equal to zero
end;
}
3:
begin
prevrlwinm :=
getlastinstruction(p,prev) and
(prev.typ = ait_instruction) and
((taicpu(prev).opcode = A_RLWINM) or
(taicpu(prev).opcode = A_RLWINM_)) and
(taicpu(prev).oper[0]^.reg = p.oper[0]^.reg) and
(taicpu(prev).oper[3]^.val = taicpu(prev).oper[4]^.val);
if (prevrlwinm) then
begin
// isolate the bit we need
if (taicpu(prev).oper[3]^.val <> 31) then
begin
p.opcode := A_RLWINM;
p.ops := 5;
p.loadreg(1,p.oper[0]^.reg);
p.loadreg(0,next1.oper[0]^.reg);
p.loadconst(2,taicpu(prev).oper[3]^.val + 1);
p.loadconst(3,31);
p.loadconst(4,31);
end
else { if (taicpu(prev).oper[0]^.reg <> next1.oper[0]^.reg) then }
begin
p.opcode := A_MR;
p.loadreg(1,p.oper[0]^.reg);
p.loadreg(0,next1.oper[0]^.reg);
end;
if not inverse then
begin
next1.ops := 3;
next1.opcode := A_XORI;
next1.loadreg(1,next1.oper[0]^.reg);
next1.loadconst(2,1);
end
else
begin
asml.remove(next1);
next1.free;
asml.remove(next3);
next3.free;
end;
asml.remove(next2);
next2.free;
end
else
begin
// equal/not equal to zero (the xori remains in the latter case;
// there's a more optimal sequence without it, but needs extra
// register)
p.opcode := A_CNTLZW;
p.loadreg(1,p.oper[0]^.reg);
p.loadreg(0,next1.oper[0]^.reg);
next1.ops := 3;
next1.opcode := A_SRWI;
next1.loadreg(1,next1.oper[0]^.reg);
next1.loadconst(2,5);
asml.remove(next2);
next2.free;
end;
end;
else
result := false;
end;
end;
function rlwinm2mask(l1,l2: longint): longint;
begin
// 1 shl 32 = 1 instead of 0 on x86
if (l1 <> 0) then
result := longint(cardinal(1) shl (32 - l1) - 1) xor (cardinal(1) shl (31 - l2) - 1)
else
result := longint(not(cardinal(1) shl (31 - l2) - 1));
if (l1 > l2) then
result := not(result);
end;
function TCpuAsmOptimizer.PeepHoleOptPass1Cpu(var p: tai): boolean;
var
next1, next2: tai;
l1, l2, shlcount: longint;
begin
result := false;
case p.typ of
ait_instruction:
begin
case taicpu(p).opcode of
A_CMPWI,
A_CMPLWI:
begin
if (taicpu(p).oper[1]^.typ = top_const) and
(taicpu(p).oper[1]^.val = 0) and
getnextinstruction(p,next1) and
(next1.typ = ait_instruction) and
(taicpu(next1).opcode = A_MFCR) and
getnextinstruction(next1,next2) and
(taicpu(next2).opcode = A_RLWINM) and
(taicpu(next2).oper[0]^.reg = taicpu(next2).oper[1]^.reg) and
(taicpu(next2).oper[0]^.reg = taicpu(next1).oper[0]^.reg) and
(taicpu(next2).oper[3]^.val = 31) and
(taicpu(next2).oper[4]^.val = 31) and
cmpi_mfcr_opt(taicpu(p),taicpu(next1),taicpu(next2)) then
result := true;
end;
{ seems the register allocator doesn't generate superfluous fmr's }
{ A_FMR, }
A_MR:
begin
if getnextinstruction(p,next1) and
(next1.typ = ait_instruction) and
(taicpu(next1).ops >= 1) and
{ spilling_get_operation_type does not support lmw/stmw }
(taicpu(next1).opcode <> A_LMW) and
(taicpu(next1).opcode <> A_STMW) and
(taicpu(next1).spilling_get_operation_type(0) = operand_write) and
(taicpu(next1).oper[0]^.reg = taicpu(p).oper[0]^.reg) then
begin
for l1 := 1 to taicpu(next1).ops - 1 do
case taicpu(next1).oper[l1]^.typ of
top_reg:
if taicpu(next1).oper[l1]^.reg = taicpu(p).oper[0]^.reg then
taicpu(next1).loadreg(l1,taicpu(p).oper[1]^.reg);
top_ref:
begin
if taicpu(next1).oper[l1]^.ref^.base = taicpu(p).oper[0]^.reg then
taicpu(next1).oper[l1]^.ref^.base := taicpu(p).oper[1]^.reg;
if taicpu(next1).oper[l1]^.ref^.index = taicpu(p).oper[0]^.reg then
taicpu(next1).oper[l1]^.ref^.index := taicpu(p).oper[1]^.reg;
end;
end;
asml.remove(p);
p.free;
p := next1;
result := true;
end;
end;
A_SLWI:
begin
if getnextinstruction(p,next1) and
(next1.typ = ait_instruction) and
((taicpu(next1).opcode = A_RLWINM) or
(taicpu(next1).opcode = A_SLWI) or
(taicpu(next1).opcode = A_SRWI)) and
(taicpu(next1).oper[0]^.reg = taicpu(p).oper[0]^.reg) and
(taicpu(next1).oper[1]^.reg = taicpu(p).oper[0]^.reg) then
begin
{ convert slwi to rlwinm and see if the rlwinm }
{ optimization can do something with it }
taicpu(p).opcode := A_RLWINM;
taicpu(p).ops := 5;
taicpu(p).loadconst(2,taicpu(p).oper[2]^.val);
taicpu(p).loadconst(3,0);
taicpu(p).loadconst(4,31-taicpu(p).oper[2]^.val);
result := true;
end;
end;
A_SRWI:
begin
if getnextinstruction(p,next1) and
(next1.typ = ait_instruction) and
((taicpu(next1).opcode = A_SLWI) or
(taicpu(next1).opcode = A_RLWINM) or
(taicpu(next1).opcode = A_SRWI)) and
(taicpu(next1).oper[0]^.reg = taicpu(p).oper[0]^.reg) and
(taicpu(next1).oper[1]^.reg = taicpu(p).oper[0]^.reg) then
case taicpu(next1).opcode of
A_SLWI:
begin
taicpu(p).opcode := A_RLWINM;
taicpu(p).ops := 5;
taicpu(p).loadconst(2,taicpu(next1).oper[2]^.val-taicpu(p).oper[2]^.val);
if (taicpu(p).oper[2]^.val < 0) then
begin
taicpu(p).loadconst(3,-taicpu(p).oper[2]^.val);
taicpu(p).loadconst(4,31-taicpu(next1).oper[2]^.val);
inc(taicpu(p).oper[2]^.val,32);
end
else
begin
taicpu(p).loadconst(3,0);
taicpu(p).loadconst(4,31-taicpu(next1).oper[2]^.val);
end;
asml.remove(next1);
next1.free;
result := true;
end;
A_RLWINM:
begin
{ convert srwi to rlwinm and see if the rlwinm }
{ optimization can do something with it }
taicpu(p).opcode := A_RLWINM;
taicpu(p).ops := 5;
taicpu(p).loadconst(3,taicpu(p).oper[2]^.val);
taicpu(p).loadconst(4,31);
taicpu(p).loadconst(2,(32-taicpu(p).oper[2]^.val) and 31);
result := true;
end;
end;
end;
A_RLWINM:
begin
if getnextinstruction(p,next1) and
(next1.typ = ait_instruction) and
((taicpu(next1).opcode = A_RLWINM) or
(taicpu(next1).opcode = A_SRWI) or
(taicpu(next1).opcode = A_SLWI)) and
(taicpu(next1).oper[0]^.reg = taicpu(p).oper[0]^.reg) and
// both source and target of next1 must equal target of p
(taicpu(next1).oper[1]^.reg = taicpu(p).oper[0]^.reg) then
begin
case taicpu(next1).opcode of
A_RLWINM:
begin
shlcount := taicpu(next1).oper[2]^.val;
l2 := rlwinm2mask(taicpu(next1).oper[3]^.val,taicpu(next1).oper[4]^.val);
end;
A_SLWI:
begin
shlcount := taicpu(next1).oper[2]^.val;
l2 := (-1) shl shlcount;
end;
A_SRWI:
begin
shlcount := 32-taicpu(next1).oper[2]^.val;
l2 := (-1) shr taicpu(next1).oper[2]^.val;
end;
else
internalerror(2013113008);
end;
l1 := rlwinm2mask((taicpu(p).oper[3]^.val-shlcount) and 31,(taicpu(p).oper[4]^.val-shlcount) and 31);
l1 := l1 and l2;
case l1 of
-1:
begin
taicpu(p).oper[2]^.val := (taicpu(p).oper[2]^.val + shlcount) and 31;
asml.remove(next1);
next1.free;
if (taicpu(p).oper[2]^.val = 0) then
begin
next1 := tai(p.next);
asml.remove(p);
p.free;
p := next1;
result := true;
end;
end;
0:
begin
// masks have no bits in common
taicpu(p).opcode := A_LI;
taicpu(p).loadconst(1,0);
taicpu(p).freeop(2);
taicpu(p).freeop(3);
taicpu(p).freeop(4);
taicpu(p).ops := 2;
taicpu(p).opercnt := 2;
asml.remove(next1);
next1.free;
result := true;
end
else if tcgppc(cg).get_rlwi_const(l1,l1,l2) then
begin
taicpu(p).oper[2]^.val := (taicpu(p).oper[2]^.val + shlcount) and 31;
taicpu(p).oper[3]^.val := l1;
taicpu(p).oper[4]^.val := l2;
asml.remove(next1);
next1.free;
result := true;
end;
end;
end;
end;
end;
end;
end;
end;
const
modifyflags: array[tasmop] of tasmop =
(a_none, a_add_, a_add_, a_addo_, a_addo_, a_addc_, a_addc_, a_addco_, a_addco_,
a_adde_, a_adde_, a_addeo_, a_addeo_, {a_addi could be addic_ if sure doesn't disturb carry} a_none, a_addic_, a_addic_, a_none,
a_addme_, a_addme_, a_addmeo_, a_addmeo_, a_addze_, a_addze_, a_addzeo_,
a_addzeo_, a_and_, a_and_, a_andc_, a_andc_, a_andi_, a_andis_, a_none,
a_none, a_none, a_none, a_none, a_none, a_none, a_none, a_none, a_none, a_none,
a_none, a_none, a_none, a_none, a_none, a_cntlzw_, a_cntlzw_, a_none,
a_none, a_none, a_none, a_none, a_none, a_none, a_none, a_none,
a_none, a_none, a_none, a_none, a_none, a_none, a_divw_, a_divw_, a_divwo_, a_divwo_,
a_divwu_, a_divwu_, a_divwuo_, a_divwuo_, a_none, a_none, a_none, a_eqv_,
a_eqv_, a_extsb_, a_extsb_, a_extsh_, a_extsh_, a_none, a_none, a_none,
a_none, a_none, a_none, a_none, a_none, a_none, a_none, a_none,
a_none, a_none, a_none, a_none, a_none, a_none, a_none, a_none,
a_none, a_none, a_none, a_none, a_none, a_none, a_none, a_none,
a_none, a_none, a_none, a_none, a_none, a_none, a_none,
a_none, a_none, a_none, a_none, a_none, a_none,
a_none, a_none, a_none, a_none,
a_none, a_none, a_none, a_none, a_none, a_none, a_none,
a_none, a_none, a_none, a_none, a_none, a_none, a_none, a_none,
a_none, a_none, a_none, a_none, a_none, a_none, a_none, a_none,
a_none, a_none, a_none, a_none, a_none, a_none, a_none, a_none, a_none,
a_none, a_none, a_none, a_none, a_none, a_none, a_none, a_none, a_none,
a_none, a_none, a_none, a_none, a_none, a_none, a_none, a_none, a_none,
a_none, a_none, a_none, a_mffs, a_mffs_, a_mfmsr, a_mfspr, a_mfsr,
a_mfsrin, a_mftb, a_mtcrf, a_none, a_none, a_none, a_none,
a_none, a_none, a_none, a_none, a_none, a_none, a_mulhw_,
a_mulhw_, a_mulhwu_, a_mulhwu_, a_none, a_mullw_, a_mullw_, a_mullwo_,
a_mullwo_, a_nand_, a_nand_, a_neg_, a_neg_, a_nego_, a_nego_, a_nor_, a_nor_,
a_or_, a_or_, a_orc_, a_orc_, a_none, a_none, a_none, a_rlwimi_, a_rlwimi_,
a_rlwinm_, a_rlwinm_, a_rlwnm_, a_rlwnm_, a_none, a_slw_, a_slw_, a_sraw_, a_sraw_,
a_srawi_, a_srawi_,a_srw_, a_srw_, a_none, a_none, a_none, a_none, a_none,
a_none, a_none, a_none, a_none, a_none, a_none, a_none, a_none,
a_none, a_none, a_none, a_none, a_none, a_none, a_none, a_none, a_none,
a_none, a_none, a_none, a_none, a_none, a_subf_, a_subf_, a_subfo_,
a_subfo_, a_subfc_, a_subfc_, a_subfco_, a_subfco_, a_subfe_, a_subfe_,
a_subfeo_, a_subfeo_, a_none, a_subfme_, a_subfme_, a_subfmeo_, a_subfmeo_,
a_subfze_, a_subfze_, a_subfzeo_, a_subfzeo_, a_none, a_none, a_none,
a_none, a_none, a_none, a_xor_, a_xor_, a_none, a_none,
{ simplified mnemonics }
a_none, a_none, a_subic_, a_subic_, a_sub_, a_sub_, a_subo_, a_subo_,
a_subc_, a_subc_, a_subco_, a_subco_, a_none, a_none, a_none, a_none,
a_extlwi_, a_extlwi_, a_extrwi_, a_extrwi_, a_inslwi_, a_inslwi_, a_insrwi_,
a_insrwi_, a_rotlwi_, a_rotlwi_, a_rotlw_, a_rotlw_, a_slwi_, a_slwi_,
a_srwi_, a_srwi_, a_clrlwi_, a_clrlwi_, a_clrrwi_, a_clrrwi_, a_clrslwi_,
a_clrslwi_, a_none, a_none, a_none, a_none, a_none, a_none, a_none,
a_none, a_none {move to special purpose reg}, a_none {move from special purpose reg},
a_none, a_none, a_none, a_none, a_mr_, a_mr_, a_not_, a_not_, a_none, a_none, a_none,
a_none, a_none, a_none, a_none,
a_none, a_none, a_none, a_none, a_none);
function changetomodifyflags(p: taicpu): boolean;
begin
result := false;
if (modifyflags[p.opcode] <> a_none) then
begin
p.opcode := modifyflags[p.opcode];
result := true;
end;
end;
function TCpuAsmOptimizer.PostPeepHoleOptsCpu(var p: tai): boolean;
var
next1: tai;
begin
result := false;
case p.typ of
ait_instruction:
begin
case taicpu(p).opcode of
A_RLWINM_:
begin
// rlwinm_ is cracked on the G5, andi_/andis_ aren't
if (taicpu(p).oper[2]^.val = 0) then
if (taicpu(p).oper[3]^.val < 16) and
(taicpu(p).oper[4]^.val < 16) then
begin
taicpu(p).opcode := A_ANDIS_;
taicpu(p).oper[2]^.val := word(
((1 shl (16-taicpu(p).oper[3]^.val)) - 1) xor
((1 shl (15-taicpu(p).oper[4]^.val)) - 1));
taicpu(p).freeop(3);
taicpu(p).freeop(4);
taicpu(p).ops := 3;
taicpu(p).opercnt := 3;
end
else if (taicpu(p).oper[3]^.val >= 16) and
(taicpu(p).oper[4]^.val >= 16) then
begin
taicpu(p).opcode := A_ANDI_;
taicpu(p).oper[2]^.val := word(rlwinm2mask(taicpu(p).oper[3]^.val,taicpu(p).oper[4]^.val));
taicpu(p).freeop(3);
taicpu(p).freeop(4);
taicpu(p).ops := 3;
taicpu(p).opercnt := 3;
end;
end;
end;
// change "integer operation with destination reg" followed by a
// comparison to zero of that reg, with a variant of that integer
// operation which sets the flags (if it exists)
if not(result) and
(taicpu(p).ops >= 2) and
(taicpu(p).oper[0]^.typ = top_reg) and
(taicpu(p).oper[1]^.typ = top_reg) and
getnextinstruction(p,next1) and
(next1.typ = ait_instruction) and
(taicpu(next1).opcode = A_CMPWI) and
// make sure it the result goes to cr0
(((taicpu(next1).ops = 2) and
(taicpu(next1).oper[1]^.val = 0) and
(taicpu(next1).oper[0]^.reg = taicpu(p).oper[0]^.reg)) or
((taicpu(next1).ops = 3) and
(taicpu(next1).oper[2]^.val = 0) and
(taicpu(next1).oper[0]^.typ = top_reg) and
(getsupreg(taicpu(next1).oper[0]^.reg) = RS_CR0) and
(taicpu(next1).oper[1]^.reg = taicpu(p).oper[0]^.reg))) and
changetomodifyflags(taicpu(p)) then
begin
asml.remove(next1);
next1.free;
result := true;
end;
end;
end;
end;
begin
casmoptimizer:=TCpuAsmOptimizer;
End.
|
