summary refs log tree commit diff
path: root/lisc/isel.c
blob: 2c05b22973098fedd6c4ab5599d57c79fba311d3 (plain) (blame)
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
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
#include "lisc.h"
#include <limits.h>

/* For x86_64, do the following:
 *
 * - lower calls
 * - check that constants are used only in
 *   places allowed
 * - ensure immediates always fit in 32b
 * - explicit machine register contraints
 *   on instructions like division.
 * - implement fast locals (the streak of
 *   constant allocX in the first basic block)
 * - recognize complex addressing modes
 *
 * Invariant: the use counts that are used
 *            in sel() must be sound.  This
 *            is not so trivial, maybe the
 *            dce should be moved out...
 */

typedef struct ANum ANum;
typedef struct AClass AClass;

struct ANum {
	char n, l, r;
	Ins *i;
	Ref mem;
};

static void amatch(Addr *, Ref, ANum *, Fn *, int);

static int
fcmptoi(int fc)
{
	switch (fc) {
	default:   diag("isel: fcmptoi defaulted");
	case FCle: return ICule;
	case FClt: return ICult;
	case FCgt: return ICugt;
	case FCge: return ICuge;
	case FCne: return ICne;
	case FCeq: return ICeq;
	case FCo:  return ICXnp;
	case FCuo: return ICXp;
	}
}

static int
iscmp(int op, int *pk, int *pc)
{
	int k, c;

	if (OCmpw <= op && op <= OCmpw1) {
		c = op - OCmpw;
		k = Kw;
	}
	else if (OCmpl <= op && op <= OCmpl1) {
		c = op - OCmpl;
		k = Kl;
	}
	else if (OCmps <= op && op <= OCmps1) {
		c = fcmptoi(op - OCmps);
		k = Ks;
	}
	else if (OCmpd <= op && op <= OCmpd1) {
		c = fcmptoi(op - OCmpd);
		k = Kd;
	}
	else
		return 0;
	if (pk)
		*pk = k;
	if (pc)
		*pc = c;
	return 1;
}

static int
noimm(Ref r, Fn *fn)
{
	int64_t val;

	if (rtype(r) != RCon)
		return 0;
	switch (fn->con[r.val].type) {
	default:
		diag("isel: invalid constant");
	case CAddr:
		/* we only support the 'small'
		 * code model of the ABI, this
		 * means that we can always
		 * address data with 32bits
		 */
		return 0;
	case CBits:
		val = fn->con[r.val].bits.i;
		return (val < INT32_MIN || val > INT32_MAX);
	}
}

static int
rslot(Ref r, Fn *fn)
{
	if (rtype(r) != RTmp)
		return -1;
	return fn->tmp[r.val].slot;
}

static int
argcls(Ins *i, int n)
{
	switch (i->op) {
	case OStores:
		return n == 0 ? Ks : Kl;
	case OStored:
		return n == 0 ? Kd : Kl;
	case OStoreb:
	case OStoreh:
	case OStorew:
		return n == 0 ? Kw : Kl;
	case OStorel:
		return Kl;
	case OExts:
		return Ks;
	case OTruncd:
		return Kd;
	case OFtosi:
		return KWIDE(i->cls) ? Kl : Kw;
	case OSitof:
		return KWIDE(i->cls) ? Kd : Ks;
	case OCast:
		switch (i->cls) {
		case Kw: return Ks;
		case Kl: return Kd;
		case Ks: return Kw;
		case Kd: return Kl;
		}
	default:
		if (OCmpw <= i->op && i->op <= OCmpd1)
			diag("isel: invalid call to argcls");
		if (isload(i->op))
			return Kl;
		if (isext(i->op))
			return Kw;
		return i->cls;
	}
}

static void
fixarg(Ref *r, int k, int phi, Fn *fn)
{
	Addr a;
	Ref r0, r1;
	int s, n;

	r1 = r0 = *r;
	s = rslot(r0, fn);
	if (KBASE(k) == 1 && rtype(r0) == RCon) {
		/* load floating points from memory
		 * slots, they can't be used as
		 * immediates
		 */
		r1 = MEM(fn->nmem);
		vgrow(&fn->mem, ++fn->nmem);
		memset(&a, 0, sizeof a);
		a.offset.type = CAddr;
		n = stashfp(fn->con[r0.val].bits.i, KWIDE(k));
		sprintf(a.offset.label, ".Lfp%d", n);
		fn->mem[fn->nmem-1] = a;
	}
	else if (!phi && k == Kl && noimm(r0, fn)) {
		/* load constants that do not fit in
		 * a 32bit signed integer into a
		 * long temporary
		 */
		r1 = newtmp("isel", fn);
		emit(OCopy, Kl, r1, r0, R);
	}
	else if (s != -1) {
		/* load fast locals' addresses into
		 * temporaries right before the
		 * instruction
		 */
		r1 = newtmp("isel", fn);
		emit(OAddr, Kl, r1, SLOT(s), R);
	}
	*r = r1;
}

static void
chuse(Ref r, int du, Fn *fn)
{
	if (rtype(r) == RTmp)
		fn->tmp[r.val].nuse += du;
}

static void
seladdr(Ref *r, ANum *an, Fn *fn)
{
	Addr a;
	Ref r0, r1;

	r0 = *r;
	if (rtype(r0) == RTmp) {
		chuse(r0, -1, fn);
		r1 = an[r0.val].mem;
		if (req(r1, R)) {
			amatch(&a, r0, an, fn, 1);
			vgrow(&fn->mem, ++fn->nmem);
			fn->mem[fn->nmem-1] = a;
			r1 = MEM(fn->nmem-1);
			chuse(a.base, +1, fn);
			chuse(a.index, +1, fn);
			if (rtype(a.base) != RTmp)
			if (rtype(a.index) != RTmp)
				an[r0.val].mem = r1;
		}
		*r = r1;
	}
}

static void
selcmp(Ref arg[2], int k, Fn *fn)
{
	Ref r;

	if (rtype(arg[0]) == RCon) {
		r = arg[1];
		arg[1] = arg[0];
		arg[0] = r;
	}
	assert(rtype(arg[0]) != RCon);
	emit(OXCmp, k, R, arg[1], arg[0]);
	fixarg(&curi->arg[0], k, 0, fn);
}

static void
sel(Ins i, ANum *an, Fn *fn)
{
	Ref r0, r1;
	int x, k, kc;
	int64_t val;
	Ins *i0;

	if (rtype(i.to) == RTmp)
	if (!isreg(i.to) && !isreg(i.arg[0]) && !isreg(i.arg[1]))
	if (fn->tmp[i.to.val].nuse == 0) {
		chuse(i.arg[0], -1, fn);
		chuse(i.arg[1], -1, fn);
		return;
	}
	i0 = curi;
	k = i.cls;
	switch (i.op) {
	case ODiv:
	case ORem:
		if (i.op == ODiv)
			r0 = TMP(RAX), r1 = TMP(RDX);
		else
			r0 = TMP(RDX), r1 = TMP(RAX);
		emit(OCopy, k, i.to, r0, R);
		emit(OCopy, k, R, r1, R);
		if (rtype(i.arg[1]) == RCon) {
			/* immediates not allowed for
			 * divisions in x86
			 */
			r0 = newtmp("isel", fn);
		} else
			r0 = i.arg[1];
		emit(OXDiv, k, R, r0, R);
		emit(OSign, k, TMP(RDX), TMP(RAX), R);
		emit(OCopy, k, TMP(RAX), i.arg[0], R);
		if (rtype(i.arg[1]) == RCon)
			emit(OCopy, k, r0, i.arg[1], R);
		break;
	case ONop:
		break;
	case OStored:
		if (rtype(i.arg[0]) == RCon)
			i.op = OStorel;
	case OStores:
		if (rtype(i.arg[0]) == RCon)
			i.op = OStorew;
	case OStorel:
	case OStorew:
	case OStoreh:
	case OStoreb:
		seladdr(&i.arg[1], an, fn);
		goto Emit;
	case_OLoad:
		seladdr(&i.arg[0], an, fn);
		goto Emit;
	case OXPush:
	case OCall:
	case OSAlloc:
	case OCopy:
	case OAdd:
	case OSub:
	case OMul:
	case OAnd:
	case OOr:
	case OXor:
	case OXTest:
	case OFtosi:
	case OSitof:
	case OCast:
	case_OExt:
Emit:
		emiti(i);
		fixarg(&curi->arg[0], argcls(curi, 0), 0, fn);
		fixarg(&curi->arg[1], argcls(curi, 1), 0, fn);
		break;
	case OAlloc:
	case OAlloc+1:
	case OAlloc+2: /* == OAlloc1 */
		/* we need to make sure
		 * the stack remains aligned
		 * (rsp = 0) mod 16
		 */
		if (rtype(i.arg[0]) == RCon) {
			assert(fn->con[i.arg[0].val].type == CBits);
			val = fn->con[i.arg[0].val].bits.i;
			val = (val + 15)  & ~INT64_C(15);
			if (val < 0 || val > INT32_MAX)
				diag("isel: alloc too large");
			emit(OAlloc, Kl, i.to, getcon(val, fn), R);
		} else {
			/* r0 = (i.arg[0] + 15) & -16 */
			r0 = newtmp("isel", fn);
			r1 = newtmp("isel", fn);
			emit(OSAlloc, Kl, i.to, r0, R);
			emit(OAnd, 1, r0, r1, getcon(-16, fn));
			emit(OAdd, 1, r1, i.arg[0], getcon(15, fn));
		}
		break;
	default:
		if (isext(i.op))
			goto case_OExt;
		if (isload(i.op))
			goto case_OLoad;
		if (iscmp(i.op, &kc, &x)) {
			if (rtype(i.arg[0]) == RCon)
				x = icmpop(x);
			emit(OXSet+x, k, i.to, R, R);
			selcmp(i.arg, kc, fn);
			break;
		}
		diag("isel: non-exhaustive implementation");
	}

	while (i0 > curi && --i0)
		if (rslot(i0->arg[0], fn) != -1
		||  rslot(i0->arg[1], fn) != -1)
			diag("isel: usupported address argument");
}

static Ins *
flagi(Ins *i0, Ins *i)
{
	while (i>i0)
		switch ((--i)->op) {
		default:
			if (iscmp(i->op, 0, 0))
				return i;
			if (isext(i->op) || isload(i->op))
				continue;
			return 0;
		case OAdd:  /* flag-setting */
		case OSub:
		case OAnd:
		case OOr:
		case OXor:
			return i;
		case OCopy: /* flag-transparent */
		case OStored:
		case OStores:
		case OStorel:
		case OStorew:
		case OStoreh:
		case OStoreb:
		case OExts:
		case OTruncd:
		case OFtosi:
		case OSitof:
		case OCast:
			;
		}
	return 0;
}

static void
seljmp(Blk *b, Fn *fn)
{
	Ref r;
	int c, k;
	Ins *fi;

	switch (b->jmp.type) {
	default:
		return;
	case JRetc:
		assert(!"retc todo");
	case JRetw:
	case JRetl:
	case JRets:
	case JRetd:
		k = b->jmp.type - JRetw;
		b->jmp.type = JRet0;
		r = b->jmp.arg;
		b->jmp.arg = R;
		if (KBASE(k) == 0)
			emit(OCopy, k, TMP(RAX), r, R);
		else
			emit(OCopy, k, TMP(XMM0), r, R);
		return;
	case JJnz:;
	}
	r = b->jmp.arg;
	b->jmp.arg = R;
	assert(!req(r, R));
	if (rtype(r) == RCon) {
		b->jmp.type = JJmp;
		if (req(r, CON_Z))
			b->s1 = b->s2;
		b->s2 = 0;
		return;
	}
	fi = flagi(b->ins, &b->ins[b->nins]);
	if (fi && req(fi->to, r)) {
		if (iscmp(fi->op, &k, &c)) {
			if (rtype(fi->arg[0]) == RCon)
				c = icmpop(c);
			b->jmp.type = JXJc + c;
			if (fn->tmp[r.val].nuse == 1) {
				assert(fn->tmp[r.val].ndef == 1);
				selcmp(fi->arg, k, fn);
				*fi = (Ins){.op = ONop};
			}
			return;
		}
		if (fi->op == OAnd && fn->tmp[r.val].nuse == 1
		&& (rtype(fi->arg[0]) == RTmp ||
		    rtype(fi->arg[1]) == RTmp)) {
			fi->op = OXTest;
			fi->to = R;
			b->jmp.type = JXJc + ICne;
			if (rtype(fi->arg[1]) == RCon) {
				r = fi->arg[1];
				fi->arg[1] = fi->arg[0];
				fi->arg[0] = r;
			}
			return;
		}
		if (fn->tmp[r.val].nuse > 1) {
			b->jmp.type = JXJc + ICne;
			return;
		}
	}
	selcmp((Ref[2]){r, CON_Z}, Kw, fn); /* todo, add long branch if non-zero */
	b->jmp.type = JXJc + ICne;
}

struct AClass {
	int inmem;
	int align;
	uint size;
	int cls[2];
};

static void
aclass(AClass *a, Typ *t)
{
	int e, s, n, cls;
	uint sz, al;

	sz = t->size;
	al = 1u << t->align;

	/* the ABI requires sizes to be rounded
	 * up to the nearest multiple of 8, moreover
	 * it makes it easy load and store structures
	 * in registers
	 */
	if (al < 8)
		al = 8;
	sz = (sz + al-1) & -al;

	a->size = sz;
	a->align = t->align;

	if (t->dark || sz > 16) {
		/* large or unaligned structures are
		 * required to be passed in memory
		 */
		a->inmem = 1;
		return;
	}

	for (e=0, s=0; e<2; e++) {
		cls = -1;
		for (n=0; n<8 && t->seg[s].len; s++) {
			if (t->seg[s].flt) {
				if (cls == -1)
					cls = Kd;
			} else
				cls = Kl;
			n += t->seg[s].len;
		}
		assert(n <= 8);
		a->cls[e] = cls;
	}
}

static int
classify(Ins *i0, Ins *i1, AClass *ac, int op)
{
	int nint, ni, nsse, ns, n, *pn;
	AClass *a;
	Ins *i;

	nint = 6;
	nsse = 8;
	for (i=i0, a=ac; i<i1; i++, a++) {
		if (i->op == op) {
			if (KBASE(i->cls) == 0)
				pn = &nint;
			else
				pn = &nsse;
			if (*pn > 0) {
				--*pn;
				a->inmem = 0;
			} else
				a->inmem = 2;
			a->align = 3;
			a->size = 8;
			a->cls[0] = i->cls;
		} else {
			n = i->arg[0].val & AMask;
			aclass(a, &typ[n]);
			if (a->inmem)
				continue;
			ni = ns = 0;
			for (n=0; n<2; n++)
				if (KBASE(a->cls[n]) == 0)
					ni++;
				else
					ns++;
			if (nint >= ni && nsse >= ns) {
				nint -= ni;
				nsse -= ns;
			} else
				a->inmem = 1;
		}
	}

	return ((6-nint) << 4) | ((8-nsse) << 8);
}

int rsave[] = {
	RDI, RSI, RDX, RCX, R8, R9, R10, R11, RAX,
	XMM0, XMM1, XMM2, XMM3, XMM4, XMM5, XMM6, XMM7,
	XMM8, XMM9, XMM10, XMM11, XMM12, XMM13, XMM14
};
int rclob[] = {RBX, R12, R13, R14, R15};

MAKESURE(rsave_has_correct_size, sizeof rsave == NRSave * sizeof(int));
MAKESURE(rclob_has_correct_size, sizeof rclob == NRClob * sizeof(int));

bits
calldef(Ins i, int p[2])
{
	bits b;
	int ni, nf;

	b = 0;
	ni = i.arg[1].val & 3;
	nf = (i.arg[1].val >> 2) & 3;
	if (ni >= 1)
		b |= BIT(RAX);
	if (ni >= 2)
		b |= BIT(RDX);
	if (nf >= 1)
		b |= BIT(XMM0);
	if (nf >= 2)
		b |= BIT(XMM1);
	if (p) {
		p[0] = ni;
		p[1] = nf;
	}
	return b;
}

bits
calluse(Ins i, int p[2])
{
	bits b;
	int j, ni, nf;

	b = 0;
	ni = (i.arg[1].val >> 4) & 15;
	nf = (i.arg[1].val >> 8) & 15;
	for (j=0; j<ni; j++)
		b |= BIT(rsave[j]);
	for (j=0; j<nf; j++)
		b |= BIT(XMM0+j);
	if (p) {
		p[0] = ni;
		p[1] = nf;
	}
	return b;
}

static void
blit(Ref rstk, uint soff, Ref rsrc, uint sz, Fn *fn)
{
	Ref r, r1;
	uint boff;

	/* it's an impolite blit, we might go across the end
	 * of the source object a little bit... */
	for (boff=0; sz>0; sz-=8, soff+=8, boff+=8) {
		r = newtmp("abi", fn);
		r1 = newtmp("abi", fn);
		emit(OStorel, 0, R, r, r1);
		emit(OAdd, Kl, r1, rstk, getcon(soff, fn));
		r1 = newtmp("abi", fn);
		emit(OLoad, Kl, r, r1, R);
		emit(OAdd, Kl, r1, rsrc, getcon(boff, fn));
	}
}

static Ref
rarg(int ty, int *ni, int *ns)
{
	if (KBASE(ty) == 0)
		return TMP(rsave[(*ni)++]);
	else
		return TMP(XMM0 + (*ns)++);
}

static void
selcall(Fn *fn, Ins *i0, Ins *i1)
{
	Ins *i;
	AClass *ac, *a;
	int ca, ni, ns;
	uint stk, off;
	Ref r, r1, r2;

	ac = alloc((i1-i0) * sizeof ac[0]);
	ca = classify(i0, i1, ac, OArg);

	for (stk=0, a=&ac[i1-i0]; a>ac;)
		if ((--a)->inmem) {
			assert(a->align <= 4);
			stk += a->size;
			if (a->align == 4)
				stk += stk & 15;
		}
	stk += stk & 15;

	if (!req(i1->arg[1], R))
		diag("struct-returning function not implemented");
	if (stk) {
		r = getcon(-(int64_t)stk, fn);
		emit(OSAlloc, Kl, R, r, R);
	}
	emit(OCopy, i1->cls, i1->to, TMP(RAX), R);
	emit(OCall, i1->cls, R, i1->arg[0], CALL(1 | ca));

	for (i=i0, a=ac, ni=ns=0; i<i1; i++, a++) {
		if (a->inmem)
			continue;
		r1 = rarg(a->cls[0], &ni, &ns);
		if (i->op == OArgc) {
			if (a->size > 8) {
				r2 = rarg(a->cls[1], &ni, &ns);
				r = newtmp("abi", fn);
				emit(OLoad, a->cls[1], r2, r, R);
				emit(OAdd, Kl, r, i->arg[1], getcon(8, fn));
			}
			emit(OLoad, a->cls[0], r1, i->arg[1], R);
		} else
			emit(OCopy, i->cls, r1, i->arg[0], R);
	}

	r = newtmp("abi", fn);
	for (i=i0, a=ac, off=0; i<i1; i++, a++) {
		if (!a->inmem)
			continue;
		if (i->op == OArgc) {
			if (a->align == 4)
				off += off & 15;
			blit(r, off, i->arg[1], a->size, fn);
		} else {
			r1 = newtmp("abi", fn);
			emit(OStorel, 0, R, i->arg[0], r1);
			emit(OAdd, Kl, r1, r, getcon(off, fn));
		}
		off += a->size;
	}
	emit(OSAlloc, Kl, r, getcon(stk, fn), R);
}

static void
selpar(Fn *fn, Ins *i0, Ins *i1)
{
	AClass *ac, *a;
	Ins *i;
	int ni, ns, s, al;
	Ref r, r1;

	ac = alloc((i1-i0) * sizeof ac[0]);
	classify(i0, i1, ac, OPar);

	curi = insb;
	ni = ns = 0;
	assert(NAlign == 3);
	s = 4;
	for (i=i0, a=ac; i<i1; i++, a++) {
		switch (a->inmem) {
		case 1:
			assert(a->align <= 4);
			if (a->align == 4)
				s = (s+3) & -4;
			fn->tmp[i->to.val].slot = -s; /* HACK! */
			s += a->size;
			continue;
		case 2:
			*curi++ = (Ins){OLoad, i->to, {SLOT(-s)}, i->cls};
			s += 2;
			continue;
		}
		r1 = rarg(a->cls[0], &ni, &ns);
		if (i->op == OParc) {
			r = newtmp("abi", fn);
			*curi++ = (Ins){OCopy, r, {r1}, Kl};
			a->cls[0] = r.val;
			if (a->size > 8) {
				r1 = rarg(a->cls[1], &ni, &ns);
				r = newtmp("abi", fn);
				*curi++ = (Ins){OCopy, r, {r1}, Kl};
				a->cls[1] = r.val;
			}
		} else
			*curi++ = (Ins){OCopy, i->to, {r1}, i->cls};
	}
	for (i=i0, a=ac; i<i1; i++, a++) {
		if (i->op != OParc || a->inmem)
			continue;
		assert(NAlign == 3);
		for (al=0; a->align >> (al+2); al++)
			;
		r = TMP(a->cls[0]);
		r1 = i->to;
		*curi++ = (Ins){OAlloc+al, r1, {getcon(a->size, fn)}, Kl};
		*curi++ = (Ins){OStorel, R, {r, r1}, 0};
		if (a->size > 8) {
			r = newtmp("abi", fn);
			*curi++ = (Ins){OAdd, r, {r1, getcon(8, fn)}, Kl};
			r1 = TMP(a->cls[1]);
			*curi++ = (Ins){OStorel, R, {r1, r}, 0};
		}
	}
}

static int
aref(Ref r, ANum *ai)
{
	switch (rtype(r)) {
	default:
		diag("isel: aref defaulted");
	case RCon:
		return 2;
	case RTmp:
		return ai[r.val].n;
	}
}

static int
ascale(Ref r, Con *con)
{
	int64_t n;

	if (rtype(r) != RCon)
		return 0;
	if (con[r.val].type != CBits)
		return 0;
	n = con[r.val].bits.i;
	return n == 1 || n == 2 || n == 4 || n == 8;
}

static void
anumber(ANum *ai, Blk *b, Con *con)
{
	/* This should be made obsolete by a proper
	 * reassoc pass.
	 *
	 * Rules:
	 *
	 *   RTmp(_) -> 0    tmp
	 *   ( RTmp(_) -> 1    slot )
	 *   RCon(_) -> 2    con
	 *   0 * 2   -> 3    s * i (when constant is 1,2,4,8)
	 */
	static char add[10][10] = {
		[2] [2] = 2,              /* folding */
		[2] [5] = 5, [5] [2] = 5,
		[2] [6] = 6, [6] [2] = 6,
		[2] [7] = 7, [7] [2] = 7,
		[0] [0] = 4,              /* 4: b + s * i */
		[0] [3] = 4, [3] [0] = 4,
		[2] [3] = 5, [3] [2] = 5, /* 5: o + s * i */
		[0] [2] = 6, [2] [0] = 6, /* 6: o + b */
		[2] [4] = 7, [4] [2] = 7, /* 7: o + b + s * i */
		[0] [5] = 7, [5] [0] = 7,
		[6] [3] = 7, [3] [6] = 7,

	};
	int a, a1, a2, n1, n2, t1, t2;
	Ins *i;

	for (i=b->ins; i-b->ins < b->nins; i++) {
		if (rtype(i->to) == RTmp)
			ai[i->to.val].i = i;
		if (i->op != OAdd && i->op != OMul)
			continue;
		a1 = aref(i->arg[0], ai);
		a2 = aref(i->arg[1], ai);
		t1 = a1 != 1 && a1 != 2;
		t2 = a2 != 1 && a2 != 2;
		if (i->op == OAdd) {
			a = add[n1 = a1][n2 = a2];
			if (t1 && a < add[0][a2])
				a = add[n1 = 0][n2 = a2];
			if (t2 && a < add[a1][0])
				a = add[n1 = a1][n2 = 0];
			if (t1 && t2 && a < add[0][0])
				a = add[n1 = 0][n2 = 0];
		} else {
			n1 = n2 = a = 0;
			if (ascale(i->arg[0], con) && t2)
				a = 3, n1 = 2, n2 = 0;
			if (t1 && ascale(i->arg[1], con))
				a = 3, n1 = 0, n2 = 2;
		}
		ai[i->to.val].n = a;
		ai[i->to.val].l = n1;
		ai[i->to.val].r = n2;
	}
}

static void
amatch(Addr *a, Ref r, ANum *ai, Fn *fn, int top)
{
	Ins *i;
	int nl, nr, t, s;
	Ref al, ar;

	if (top)
		memset(a, 0, sizeof *a);
	if (rtype(r) == RCon) {
		addcon(&a->offset, &fn->con[r.val]);
		return;
	}
	assert(rtype(r) == RTmp);
	i = ai[r.val].i;
	nl = ai[r.val].l;
	nr = ai[r.val].r;
	if (i) {
		if (nl > nr) {
			al = i->arg[1];
			ar = i->arg[0];
			t = nl, nl = nr, nr = t;
		} else {
			al = i->arg[0];
			ar = i->arg[1];
		}
	}
	switch (ai[r.val].n) {
	default:
		diag("isel: amatch defaulted");
	case 3: /* s * i */
		if (!top) {
			a->index = al;
			a->scale = fn->con[ar.val].bits.i;
		} else
			a->base = r;
		break;
	case 4: /* b + s * i */
		switch (nr) {
		case 0:
			if (fn->tmp[ar.val].slot != -1) {
				al = i->arg[1];
				ar = i->arg[0];
			}
			a->index = ar;
			a->scale = 1;
			break;
		case 3:
			amatch(a, ar, ai, fn, 0);
			break;
		}
		r = al;
	case 0:
		s = fn->tmp[r.val].slot;
		if (s != -1)
			r = SLOT(s);
		a->base = r;
		break;
	case 2: /* constants */
	case 5: /* o + s * i */
	case 6: /* o + b */
	case 7: /* o + b + s * i */
		amatch(a, ar, ai, fn, 0);
		amatch(a, al, ai, fn, 0);
		break;
	}
}

/* instruction selection
 * requires use counts (as given by parsing)
 */
void
isel(Fn *fn)
{
	Blk *b, **sb;
	Ins *i, *i0, *ip;
	Phi *p;
	uint a;
	int n, al;
	int64_t sz;
	ANum *ainfo;

	for (n=0; n<fn->ntmp; n++)
		fn->tmp[n].slot = -1;
	fn->slot = 0;

	/* lower arguments */
	for (b=fn->start, i=b->ins; i-b->ins < b->nins; i++)
		if (i->op != OPar && i->op != OParc)
			break;
	selpar(fn, b->ins, i);
	n = b->nins - (i - b->ins) + (curi - insb);
	i0 = alloc(n * sizeof(Ins));
	ip = icpy(ip = i0, insb, curi - insb);
	ip = icpy(ip, i, &b->ins[b->nins] - i);
	b->nins = n;
	b->ins = i0;

	/* lower function calls */
	for (b=fn->start; b; b=b->link) {
		curi = &insb[NIns];
		for (i=&b->ins[b->nins]; i!=b->ins;) {
			if ((--i)->op == OCall) {
				for (i0=i; i0>b->ins; i0--)
					if ((i0-1)->op != OArg)
					if ((i0-1)->op != OArgc)
						break;
				selcall(fn, i0, i);
				i = i0;
				continue;
			}
			assert(i->op != OArg && i->op != OArgc);
			emiti(*i);
		}
		b->nins = &insb[NIns] - curi;
		idup(&b->ins, curi, b->nins);
	}

	if (debug['A']) {
		fprintf(stderr, "\n> After call lowering:\n");
		printfn(fn, stderr);
	}

	/* assign slots to fast allocs */
	b = fn->start;
	assert(NAlign == 3 && "change n=4 and sz /= 4 below");
	for (al=OAlloc, n=4; al<=OAlloc1; al++, n*=2)
		for (i=b->ins; i-b->ins < b->nins; i++)
			if (i->op == al) {
				if (rtype(i->arg[0]) != RCon)
					break;
				sz = fn->con[i->arg[0].val].bits.i;
				if (sz < 0 || sz >= INT_MAX-3)
					diag("isel: invalid alloc size");
				sz = (sz + n-1) & -n;
				sz /= 4;
				fn->tmp[i->to.val].slot = fn->slot;
				fn->slot += sz;
				*i = (Ins){.op = ONop};
			}

	/* process basic blocks */
	n = fn->ntmp;
	ainfo = emalloc(n * sizeof ainfo[0]);
	for (b=fn->start; b; b=b->link) {
		curi = &insb[NIns];
		for (sb=(Blk*[3]){b->s1, b->s2, 0}; *sb; sb++)
			for (p=(*sb)->phi; p; p=p->link) {
				for (a=0; p->blk[a] != b; a++)
					assert(a+1 < p->narg);
				fixarg(&p->arg[a], p->cls, 1, fn);
			}
		memset(ainfo, 0, n * sizeof ainfo[0]);
		anumber(ainfo, b, fn->con);
		seljmp(b, fn);
		for (i=&b->ins[b->nins]; i!=b->ins;)
			sel(*--i, ainfo, fn);
		b->nins = &insb[NIns] - curi;
		idup(&b->ins, curi, b->nins);
	}
	free(ainfo);

	if (debug['I']) {
		fprintf(stderr, "\n> After instruction selection:\n");
		printfn(fn, stderr);
	}
}