dt_cg.c revision 250812
1/*
2 * CDDL HEADER START
3 *
4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License, Version 1.0 only
6 * (the "License").  You may not use this file except in compliance
7 * with the License.
8 *
9 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
10 * or http://www.opensolaris.org/os/licensing.
11 * See the License for the specific language governing permissions
12 * and limitations under the License.
13 *
14 * When distributing Covered Code, include this CDDL HEADER in each
15 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
16 * If applicable, add the following below this CDDL HEADER, with the
17 * fields enclosed by brackets "[]" replaced with your own identifying
18 * information: Portions Copyright [yyyy] [name of copyright owner]
19 *
20 * CDDL HEADER END
21 */
22
23/*
24 * Copyright 2005 Sun Microsystems, Inc.  All rights reserved.
25 * Use is subject to license terms.
26 */
27
28/*
29 * Copyright (c) 2012 by Delphix. All rights reserved.
30 */
31
32#include <sys/types.h>
33#include <sys/sysmacros.h>
34#include <sys/isa_defs.h>
35
36#include <strings.h>
37#include <stdlib.h>
38#include <setjmp.h>
39#include <assert.h>
40#include <errno.h>
41
42#include <dt_impl.h>
43#include <dt_grammar.h>
44#include <dt_parser.h>
45#include <dt_provider.h>
46
47static void dt_cg_node(dt_node_t *, dt_irlist_t *, dt_regset_t *);
48
49static dt_irnode_t *
50dt_cg_node_alloc(uint_t label, dif_instr_t instr)
51{
52	dt_irnode_t *dip = malloc(sizeof (dt_irnode_t));
53
54	if (dip == NULL)
55		longjmp(yypcb->pcb_jmpbuf, EDT_NOMEM);
56
57	dip->di_label = label;
58	dip->di_instr = instr;
59	dip->di_extern = NULL;
60	dip->di_next = NULL;
61
62	return (dip);
63}
64
65/*
66 * Code generator wrapper function for ctf_member_info.  If we are given a
67 * reference to a forward declaration tag, search the entire type space for
68 * the actual definition and then call ctf_member_info on the result.
69 */
70static ctf_file_t *
71dt_cg_membinfo(ctf_file_t *fp, ctf_id_t type, const char *s, ctf_membinfo_t *mp)
72{
73	while (ctf_type_kind(fp, type) == CTF_K_FORWARD) {
74		char n[DT_TYPE_NAMELEN];
75		dtrace_typeinfo_t dtt;
76
77		if (ctf_type_name(fp, type, n, sizeof (n)) == NULL ||
78		    dt_type_lookup(n, &dtt) == -1 || (
79		    dtt.dtt_ctfp == fp && dtt.dtt_type == type))
80			break; /* unable to improve our position */
81
82		fp = dtt.dtt_ctfp;
83		type = ctf_type_resolve(fp, dtt.dtt_type);
84	}
85
86	if (ctf_member_info(fp, type, s, mp) == CTF_ERR)
87		return (NULL); /* ctf_errno is set for us */
88
89	return (fp);
90}
91
92static void
93dt_cg_xsetx(dt_irlist_t *dlp, dt_ident_t *idp, uint_t lbl, int reg, uint64_t x)
94{
95	int flag = idp != NULL ? DT_INT_PRIVATE : DT_INT_SHARED;
96	int intoff = dt_inttab_insert(yypcb->pcb_inttab, x, flag);
97	dif_instr_t instr = DIF_INSTR_SETX((uint_t)intoff, reg);
98
99	if (intoff == -1)
100		longjmp(yypcb->pcb_jmpbuf, EDT_NOMEM);
101
102	if (intoff > DIF_INTOFF_MAX)
103		longjmp(yypcb->pcb_jmpbuf, EDT_INT2BIG);
104
105	dt_irlist_append(dlp, dt_cg_node_alloc(lbl, instr));
106
107	if (idp != NULL)
108		dlp->dl_last->di_extern = idp;
109}
110
111static void
112dt_cg_setx(dt_irlist_t *dlp, int reg, uint64_t x)
113{
114	dt_cg_xsetx(dlp, NULL, DT_LBL_NONE, reg, x);
115}
116
117/*
118 * When loading bit-fields, we want to convert a byte count in the range
119 * 1-8 to the closest power of 2 (e.g. 3->4, 5->8, etc).  The clp2() function
120 * is a clever implementation from "Hacker's Delight" by Henry Warren, Jr.
121 */
122static size_t
123clp2(size_t x)
124{
125	x--;
126
127	x |= (x >> 1);
128	x |= (x >> 2);
129	x |= (x >> 4);
130	x |= (x >> 8);
131	x |= (x >> 16);
132
133	return (x + 1);
134}
135
136/*
137 * Lookup the correct load opcode to use for the specified node and CTF type.
138 * We determine the size and convert it to a 3-bit index.  Our lookup table
139 * is constructed to use a 5-bit index, consisting of the 3-bit size 0-7, a
140 * bit for the sign, and a bit for userland address.  For example, a 4-byte
141 * signed load from userland would be at the following table index:
142 * user=1 sign=1 size=4 => binary index 11011 = decimal index 27
143 */
144static uint_t
145dt_cg_load(dt_node_t *dnp, ctf_file_t *ctfp, ctf_id_t type)
146{
147	static const uint_t ops[] = {
148		DIF_OP_LDUB,	DIF_OP_LDUH,	0,	DIF_OP_LDUW,
149		0,		0,		0,	DIF_OP_LDX,
150		DIF_OP_LDSB,	DIF_OP_LDSH,	0,	DIF_OP_LDSW,
151		0,		0,		0,	DIF_OP_LDX,
152		DIF_OP_ULDUB,	DIF_OP_ULDUH,	0,	DIF_OP_ULDUW,
153		0,		0,		0,	DIF_OP_ULDX,
154		DIF_OP_ULDSB,	DIF_OP_ULDSH,	0,	DIF_OP_ULDSW,
155		0,		0,		0,	DIF_OP_ULDX,
156	};
157
158	ctf_encoding_t e;
159	ssize_t size;
160
161	/*
162	 * If we're loading a bit-field, the size of our load is found by
163	 * rounding cte_bits up to a byte boundary and then finding the
164	 * nearest power of two to this value (see clp2(), above).
165	 */
166	if ((dnp->dn_flags & DT_NF_BITFIELD) &&
167	    ctf_type_encoding(ctfp, type, &e) != CTF_ERR)
168		size = clp2(P2ROUNDUP(e.cte_bits, NBBY) / NBBY);
169	else
170		size = ctf_type_size(ctfp, type);
171
172	if (size < 1 || size > 8 || (size & (size - 1)) != 0) {
173		xyerror(D_UNKNOWN, "internal error -- cg cannot load "
174		    "size %ld when passed by value\n", (long)size);
175	}
176
177	size--; /* convert size to 3-bit index */
178
179	if (dnp->dn_flags & DT_NF_SIGNED)
180		size |= 0x08;
181	if (dnp->dn_flags & DT_NF_USERLAND)
182		size |= 0x10;
183
184	return (ops[size]);
185}
186
187static void
188dt_cg_ptrsize(dt_node_t *dnp, dt_irlist_t *dlp, dt_regset_t *drp,
189    uint_t op, int dreg)
190{
191	ctf_file_t *ctfp = dnp->dn_ctfp;
192	ctf_arinfo_t r;
193	dif_instr_t instr;
194	ctf_id_t type;
195	uint_t kind;
196	ssize_t size;
197	int sreg;
198
199	type = ctf_type_resolve(ctfp, dnp->dn_type);
200	kind = ctf_type_kind(ctfp, type);
201	assert(kind == CTF_K_POINTER || kind == CTF_K_ARRAY);
202
203	if (kind == CTF_K_ARRAY) {
204		if (ctf_array_info(ctfp, type, &r) != 0) {
205			yypcb->pcb_hdl->dt_ctferr = ctf_errno(ctfp);
206			longjmp(yypcb->pcb_jmpbuf, EDT_CTF);
207		}
208		type = r.ctr_contents;
209	} else
210		type = ctf_type_reference(ctfp, type);
211
212	if ((size = ctf_type_size(ctfp, type)) == 1)
213		return; /* multiply or divide by one can be omitted */
214
215	sreg = dt_regset_alloc(drp);
216	dt_cg_setx(dlp, sreg, size);
217	instr = DIF_INSTR_FMT(op, dreg, sreg, dreg);
218	dt_irlist_append(dlp, dt_cg_node_alloc(DT_LBL_NONE, instr));
219	dt_regset_free(drp, sreg);
220}
221
222/*
223 * If the result of a "." or "->" operation is a bit-field, we use this routine
224 * to generate an epilogue to the load instruction that extracts the value.  In
225 * the diagrams below the "ld??" is the load instruction that is generated to
226 * load the containing word that is generating prior to calling this function.
227 *
228 * Epilogue for unsigned fields:	Epilogue for signed fields:
229 *
230 * ldu?	[r1], r1			lds? [r1], r1
231 * setx	USHIFT, r2			setx 64 - SSHIFT, r2
232 * srl	r1, r2, r1			sll  r1, r2, r1
233 * setx	(1 << bits) - 1, r2		setx 64 - bits, r2
234 * and	r1, r2, r1			sra  r1, r2, r1
235 *
236 * The *SHIFT constants above changes value depending on the endian-ness of our
237 * target architecture.  Refer to the comments below for more details.
238 */
239static void
240dt_cg_field_get(dt_node_t *dnp, dt_irlist_t *dlp, dt_regset_t *drp,
241    ctf_file_t *fp, const ctf_membinfo_t *mp)
242{
243	ctf_encoding_t e;
244	dif_instr_t instr;
245	uint64_t shift;
246	int r1, r2;
247
248	if (ctf_type_encoding(fp, mp->ctm_type, &e) != 0 || e.cte_bits > 64) {
249		xyerror(D_UNKNOWN, "cg: bad field: off %lu type <%ld> "
250		    "bits %u\n", mp->ctm_offset, mp->ctm_type, e.cte_bits);
251	}
252
253	assert(dnp->dn_op == DT_TOK_PTR || dnp->dn_op == DT_TOK_DOT);
254	r1 = dnp->dn_left->dn_reg;
255	r2 = dt_regset_alloc(drp);
256
257	/*
258	 * On little-endian architectures, ctm_offset counts from the right so
259	 * ctm_offset % NBBY itself is the amount we want to shift right to
260	 * move the value bits to the little end of the register to mask them.
261	 * On big-endian architectures, ctm_offset counts from the left so we
262	 * must subtract (ctm_offset % NBBY + cte_bits) from the size in bits
263	 * we used for the load.  The size of our load in turn is found by
264	 * rounding cte_bits up to a byte boundary and then finding the
265	 * nearest power of two to this value (see clp2(), above).  These
266	 * properties are used to compute shift as USHIFT or SSHIFT, below.
267	 */
268	if (dnp->dn_flags & DT_NF_SIGNED) {
269#if BYTE_ORDER == _BIG_ENDIAN
270		shift = clp2(P2ROUNDUP(e.cte_bits, NBBY) / NBBY) * NBBY -
271		    mp->ctm_offset % NBBY;
272#else
273		shift = mp->ctm_offset % NBBY + e.cte_bits;
274#endif
275		dt_cg_setx(dlp, r2, 64 - shift);
276		instr = DIF_INSTR_FMT(DIF_OP_SLL, r1, r2, r1);
277		dt_irlist_append(dlp, dt_cg_node_alloc(DT_LBL_NONE, instr));
278
279		dt_cg_setx(dlp, r2, 64 - e.cte_bits);
280		instr = DIF_INSTR_FMT(DIF_OP_SRA, r1, r2, r1);
281		dt_irlist_append(dlp, dt_cg_node_alloc(DT_LBL_NONE, instr));
282	} else {
283#if BYTE_ORDER == _BIG_ENDIAN
284		shift = clp2(P2ROUNDUP(e.cte_bits, NBBY) / NBBY) * NBBY -
285		    (mp->ctm_offset % NBBY + e.cte_bits);
286#else
287		shift = mp->ctm_offset % NBBY;
288#endif
289		dt_cg_setx(dlp, r2, shift);
290		instr = DIF_INSTR_FMT(DIF_OP_SRL, r1, r2, r1);
291		dt_irlist_append(dlp, dt_cg_node_alloc(DT_LBL_NONE, instr));
292
293		dt_cg_setx(dlp, r2, (1ULL << e.cte_bits) - 1);
294		instr = DIF_INSTR_FMT(DIF_OP_AND, r1, r2, r1);
295		dt_irlist_append(dlp, dt_cg_node_alloc(DT_LBL_NONE, instr));
296	}
297
298	dt_regset_free(drp, r2);
299}
300
301/*
302 * If the destination of a store operation is a bit-field, we use this routine
303 * to generate a prologue to the store instruction that loads the surrounding
304 * bits, clears the destination field, and ORs in the new value of the field.
305 * In the diagram below the "st?" is the store instruction that is generated to
306 * store the containing word that is generating after calling this function.
307 *
308 * ld	[dst->dn_reg], r1
309 * setx	~(((1 << cte_bits) - 1) << (ctm_offset % NBBY)), r2
310 * and	r1, r2, r1
311 *
312 * setx	(1 << cte_bits) - 1, r2
313 * and	src->dn_reg, r2, r2
314 * setx ctm_offset % NBBY, r3
315 * sll	r2, r3, r2
316 *
317 * or	r1, r2, r1
318 * st?	r1, [dst->dn_reg]
319 *
320 * This routine allocates a new register to hold the value to be stored and
321 * returns it.  The caller is responsible for freeing this register later.
322 */
323static int
324dt_cg_field_set(dt_node_t *src, dt_irlist_t *dlp,
325    dt_regset_t *drp, dt_node_t *dst)
326{
327	uint64_t cmask, fmask, shift;
328	dif_instr_t instr;
329	int r1, r2, r3;
330
331	ctf_membinfo_t m;
332	ctf_encoding_t e;
333	ctf_file_t *fp, *ofp;
334	ctf_id_t type;
335
336	assert(dst->dn_op == DT_TOK_PTR || dst->dn_op == DT_TOK_DOT);
337	assert(dst->dn_right->dn_kind == DT_NODE_IDENT);
338
339	fp = dst->dn_left->dn_ctfp;
340	type = ctf_type_resolve(fp, dst->dn_left->dn_type);
341
342	if (dst->dn_op == DT_TOK_PTR) {
343		type = ctf_type_reference(fp, type);
344		type = ctf_type_resolve(fp, type);
345	}
346
347	if ((fp = dt_cg_membinfo(ofp = fp, type,
348	    dst->dn_right->dn_string, &m)) == NULL) {
349		yypcb->pcb_hdl->dt_ctferr = ctf_errno(ofp);
350		longjmp(yypcb->pcb_jmpbuf, EDT_CTF);
351	}
352
353	if (ctf_type_encoding(fp, m.ctm_type, &e) != 0 || e.cte_bits > 64) {
354		xyerror(D_UNKNOWN, "cg: bad field: off %lu type <%ld> "
355		    "bits %u\n", m.ctm_offset, m.ctm_type, e.cte_bits);
356	}
357
358	r1 = dt_regset_alloc(drp);
359	r2 = dt_regset_alloc(drp);
360	r3 = dt_regset_alloc(drp);
361
362	/*
363	 * Compute shifts and masks.  We need to compute "shift" as the amount
364	 * we need to shift left to position our field in the containing word.
365	 * Refer to the comments in dt_cg_field_get(), above, for more info.
366	 * We then compute fmask as the mask that truncates the value in the
367	 * input register to width cte_bits, and cmask as the mask used to
368	 * pass through the containing bits and zero the field bits.
369	 */
370#if BYTE_ORDER == _BIG_ENDIAN
371	shift = clp2(P2ROUNDUP(e.cte_bits, NBBY) / NBBY) * NBBY -
372	    (m.ctm_offset % NBBY + e.cte_bits);
373#else
374	shift = m.ctm_offset % NBBY;
375#endif
376	fmask = (1ULL << e.cte_bits) - 1;
377	cmask = ~(fmask << shift);
378
379	instr = DIF_INSTR_LOAD(
380	    dt_cg_load(dst, fp, m.ctm_type), dst->dn_reg, r1);
381	dt_irlist_append(dlp, dt_cg_node_alloc(DT_LBL_NONE, instr));
382
383	dt_cg_setx(dlp, r2, cmask);
384	instr = DIF_INSTR_FMT(DIF_OP_AND, r1, r2, r1);
385	dt_irlist_append(dlp, dt_cg_node_alloc(DT_LBL_NONE, instr));
386
387	dt_cg_setx(dlp, r2, fmask);
388	instr = DIF_INSTR_FMT(DIF_OP_AND, src->dn_reg, r2, r2);
389	dt_irlist_append(dlp, dt_cg_node_alloc(DT_LBL_NONE, instr));
390
391	dt_cg_setx(dlp, r3, shift);
392	instr = DIF_INSTR_FMT(DIF_OP_SLL, r2, r3, r2);
393	dt_irlist_append(dlp, dt_cg_node_alloc(DT_LBL_NONE, instr));
394
395	instr = DIF_INSTR_FMT(DIF_OP_OR, r1, r2, r1);
396	dt_irlist_append(dlp, dt_cg_node_alloc(DT_LBL_NONE, instr));
397
398	dt_regset_free(drp, r3);
399	dt_regset_free(drp, r2);
400
401	return (r1);
402}
403
404static void
405dt_cg_store(dt_node_t *src, dt_irlist_t *dlp, dt_regset_t *drp, dt_node_t *dst)
406{
407	ctf_encoding_t e;
408	dif_instr_t instr;
409	size_t size;
410	int reg;
411
412	/*
413	 * If we're loading a bit-field, the size of our store is found by
414	 * rounding dst's cte_bits up to a byte boundary and then finding the
415	 * nearest power of two to this value (see clp2(), above).
416	 */
417	if ((dst->dn_flags & DT_NF_BITFIELD) &&
418	    ctf_type_encoding(dst->dn_ctfp, dst->dn_type, &e) != CTF_ERR)
419		size = clp2(P2ROUNDUP(e.cte_bits, NBBY) / NBBY);
420	else
421		size = dt_node_type_size(src);
422
423	if (src->dn_flags & DT_NF_REF) {
424		reg = dt_regset_alloc(drp);
425		dt_cg_setx(dlp, reg, size);
426		instr = DIF_INSTR_COPYS(src->dn_reg, reg, dst->dn_reg);
427		dt_irlist_append(dlp, dt_cg_node_alloc(DT_LBL_NONE, instr));
428		dt_regset_free(drp, reg);
429	} else {
430		if (dst->dn_flags & DT_NF_BITFIELD)
431			reg = dt_cg_field_set(src, dlp, drp, dst);
432		else
433			reg = src->dn_reg;
434
435		switch (size) {
436		case 1:
437			instr = DIF_INSTR_STORE(DIF_OP_STB, reg, dst->dn_reg);
438			break;
439		case 2:
440			instr = DIF_INSTR_STORE(DIF_OP_STH, reg, dst->dn_reg);
441			break;
442		case 4:
443			instr = DIF_INSTR_STORE(DIF_OP_STW, reg, dst->dn_reg);
444			break;
445		case 8:
446			instr = DIF_INSTR_STORE(DIF_OP_STX, reg, dst->dn_reg);
447			break;
448		default:
449			xyerror(D_UNKNOWN, "internal error -- cg cannot store "
450			    "size %lu when passed by value\n", (ulong_t)size);
451		}
452		dt_irlist_append(dlp, dt_cg_node_alloc(DT_LBL_NONE, instr));
453
454		if (dst->dn_flags & DT_NF_BITFIELD)
455			dt_regset_free(drp, reg);
456	}
457}
458
459/*
460 * Generate code for a typecast or for argument promotion from the type of the
461 * actual to the type of the formal.  We need to generate code for casts when
462 * a scalar type is being narrowed or changing signed-ness.  We first shift the
463 * desired bits high (losing excess bits if narrowing) and then shift them down
464 * using logical shift (unsigned result) or arithmetic shift (signed result).
465 */
466static void
467dt_cg_typecast(const dt_node_t *src, const dt_node_t *dst,
468    dt_irlist_t *dlp, dt_regset_t *drp)
469{
470	size_t srcsize = dt_node_type_size(src);
471	size_t dstsize = dt_node_type_size(dst);
472
473	dif_instr_t instr;
474	int rg;
475
476	if (!dt_node_is_scalar(dst))
477		return; /* not a scalar */
478	if (dstsize == srcsize &&
479	    ((src->dn_flags ^ dst->dn_flags) & DT_NF_SIGNED) != 0)
480		return; /* not narrowing or changing signed-ness */
481	if (dstsize > srcsize && (src->dn_flags & DT_NF_SIGNED) == 0)
482		return; /* nothing to do in this case */
483
484	rg = dt_regset_alloc(drp);
485
486	if (dstsize > srcsize) {
487		int n = sizeof (uint64_t) * NBBY - srcsize * NBBY;
488		int s = (dstsize - srcsize) * NBBY;
489
490		dt_cg_setx(dlp, rg, n);
491
492		instr = DIF_INSTR_FMT(DIF_OP_SLL, src->dn_reg, rg, dst->dn_reg);
493		dt_irlist_append(dlp, dt_cg_node_alloc(DT_LBL_NONE, instr));
494
495		if ((dst->dn_flags & DT_NF_SIGNED) || n == s) {
496			instr = DIF_INSTR_FMT(DIF_OP_SRA,
497			    dst->dn_reg, rg, dst->dn_reg);
498			dt_irlist_append(dlp,
499			    dt_cg_node_alloc(DT_LBL_NONE, instr));
500		} else {
501			dt_cg_setx(dlp, rg, s);
502			instr = DIF_INSTR_FMT(DIF_OP_SRA,
503			    dst->dn_reg, rg, dst->dn_reg);
504			dt_irlist_append(dlp,
505			    dt_cg_node_alloc(DT_LBL_NONE, instr));
506			dt_cg_setx(dlp, rg, n - s);
507			instr = DIF_INSTR_FMT(DIF_OP_SRL,
508			    dst->dn_reg, rg, dst->dn_reg);
509			dt_irlist_append(dlp,
510			    dt_cg_node_alloc(DT_LBL_NONE, instr));
511		}
512	} else if (dstsize != sizeof (uint64_t)) {
513		int n = sizeof (uint64_t) * NBBY - dstsize * NBBY;
514
515		dt_cg_setx(dlp, rg, n);
516
517		instr = DIF_INSTR_FMT(DIF_OP_SLL, src->dn_reg, rg, dst->dn_reg);
518		dt_irlist_append(dlp, dt_cg_node_alloc(DT_LBL_NONE, instr));
519
520		instr = DIF_INSTR_FMT((dst->dn_flags & DT_NF_SIGNED) ?
521		    DIF_OP_SRA : DIF_OP_SRL, dst->dn_reg, rg, dst->dn_reg);
522		dt_irlist_append(dlp, dt_cg_node_alloc(DT_LBL_NONE, instr));
523	}
524
525	dt_regset_free(drp, rg);
526}
527
528/*
529 * Generate code to push the specified argument list on to the tuple stack.
530 * We use this routine for handling subroutine calls and associative arrays.
531 * We must first generate code for all subexpressions before loading the stack
532 * because any subexpression could itself require the use of the tuple stack.
533 * This holds a number of registers equal to the number of arguments, but this
534 * is not a huge problem because the number of arguments can't exceed the
535 * number of tuple register stack elements anyway.  At most one extra register
536 * is required (either by dt_cg_typecast() or for dtdt_size, below).  This
537 * implies that a DIF implementation should offer a number of general purpose
538 * registers at least one greater than the number of tuple registers.
539 */
540static void
541dt_cg_arglist(dt_ident_t *idp, dt_node_t *args,
542    dt_irlist_t *dlp, dt_regset_t *drp)
543{
544	const dt_idsig_t *isp = idp->di_data;
545	dt_node_t *dnp;
546	int i = 0;
547
548	for (dnp = args; dnp != NULL; dnp = dnp->dn_list)
549		dt_cg_node(dnp, dlp, drp);
550
551	dt_irlist_append(dlp, dt_cg_node_alloc(DT_LBL_NONE, DIF_INSTR_FLUSHTS));
552
553	for (dnp = args; dnp != NULL; dnp = dnp->dn_list, i++) {
554		dtrace_diftype_t t;
555		dif_instr_t instr;
556		uint_t op;
557		int reg;
558
559		dt_node_diftype(yypcb->pcb_hdl, dnp, &t);
560
561		isp->dis_args[i].dn_reg = dnp->dn_reg; /* re-use register */
562		dt_cg_typecast(dnp, &isp->dis_args[i], dlp, drp);
563		isp->dis_args[i].dn_reg = -1;
564
565		if (t.dtdt_flags & DIF_TF_BYREF) {
566			op = DIF_OP_PUSHTR;
567			if (t.dtdt_size != 0) {
568				reg = dt_regset_alloc(drp);
569				dt_cg_setx(dlp, reg, t.dtdt_size);
570			} else {
571				reg = DIF_REG_R0;
572			}
573		} else {
574			op = DIF_OP_PUSHTV;
575			reg = DIF_REG_R0;
576		}
577
578		instr = DIF_INSTR_PUSHTS(op, t.dtdt_kind, reg, dnp->dn_reg);
579		dt_irlist_append(dlp, dt_cg_node_alloc(DT_LBL_NONE, instr));
580		dt_regset_free(drp, dnp->dn_reg);
581
582		if (reg != DIF_REG_R0)
583			dt_regset_free(drp, reg);
584	}
585
586	if (i > yypcb->pcb_hdl->dt_conf.dtc_diftupregs)
587		longjmp(yypcb->pcb_jmpbuf, EDT_NOTUPREG);
588}
589
590static void
591dt_cg_arithmetic_op(dt_node_t *dnp, dt_irlist_t *dlp,
592    dt_regset_t *drp, uint_t op)
593{
594	int is_ptr_op = (dnp->dn_op == DT_TOK_ADD || dnp->dn_op == DT_TOK_SUB ||
595	    dnp->dn_op == DT_TOK_ADD_EQ || dnp->dn_op == DT_TOK_SUB_EQ);
596
597	int lp_is_ptr = dt_node_is_pointer(dnp->dn_left);
598	int rp_is_ptr = dt_node_is_pointer(dnp->dn_right);
599
600	dif_instr_t instr;
601
602	if (lp_is_ptr && rp_is_ptr) {
603		assert(dnp->dn_op == DT_TOK_SUB);
604		is_ptr_op = 0;
605	}
606
607	dt_cg_node(dnp->dn_left, dlp, drp);
608	if (is_ptr_op && rp_is_ptr)
609		dt_cg_ptrsize(dnp, dlp, drp, DIF_OP_MUL, dnp->dn_left->dn_reg);
610
611	dt_cg_node(dnp->dn_right, dlp, drp);
612	if (is_ptr_op && lp_is_ptr)
613		dt_cg_ptrsize(dnp, dlp, drp, DIF_OP_MUL, dnp->dn_right->dn_reg);
614
615	instr = DIF_INSTR_FMT(op, dnp->dn_left->dn_reg,
616	    dnp->dn_right->dn_reg, dnp->dn_left->dn_reg);
617
618	dt_irlist_append(dlp, dt_cg_node_alloc(DT_LBL_NONE, instr));
619	dt_regset_free(drp, dnp->dn_right->dn_reg);
620	dnp->dn_reg = dnp->dn_left->dn_reg;
621
622	if (lp_is_ptr && rp_is_ptr)
623		dt_cg_ptrsize(dnp->dn_right,
624		    dlp, drp, DIF_OP_UDIV, dnp->dn_reg);
625}
626
627static uint_t
628dt_cg_stvar(const dt_ident_t *idp)
629{
630	static const uint_t aops[] = { DIF_OP_STGAA, DIF_OP_STTAA, DIF_OP_NOP };
631	static const uint_t sops[] = { DIF_OP_STGS, DIF_OP_STTS, DIF_OP_STLS };
632
633	uint_t i = (((idp->di_flags & DT_IDFLG_LOCAL) != 0) << 1) |
634	    ((idp->di_flags & DT_IDFLG_TLS) != 0);
635
636	return (idp->di_kind == DT_IDENT_ARRAY ? aops[i] : sops[i]);
637}
638
639static void
640dt_cg_prearith_op(dt_node_t *dnp, dt_irlist_t *dlp, dt_regset_t *drp, uint_t op)
641{
642	ctf_file_t *ctfp = dnp->dn_ctfp;
643	dif_instr_t instr;
644	ctf_id_t type;
645	ssize_t size = 1;
646	int reg;
647
648	if (dt_node_is_pointer(dnp)) {
649		type = ctf_type_resolve(ctfp, dnp->dn_type);
650		assert(ctf_type_kind(ctfp, type) == CTF_K_POINTER);
651		size = ctf_type_size(ctfp, ctf_type_reference(ctfp, type));
652	}
653
654	dt_cg_node(dnp->dn_child, dlp, drp);
655	dnp->dn_reg = dnp->dn_child->dn_reg;
656
657	reg = dt_regset_alloc(drp);
658	dt_cg_setx(dlp, reg, size);
659
660	instr = DIF_INSTR_FMT(op, dnp->dn_reg, reg, dnp->dn_reg);
661	dt_irlist_append(dlp, dt_cg_node_alloc(DT_LBL_NONE, instr));
662	dt_regset_free(drp, reg);
663
664	/*
665	 * If we are modifying a variable, generate an stv instruction from
666	 * the variable specified by the identifier.  If we are storing to a
667	 * memory address, generate code again for the left-hand side using
668	 * DT_NF_REF to get the address, and then generate a store to it.
669	 * In both paths, we store the value in dnp->dn_reg (the new value).
670	 */
671	if (dnp->dn_child->dn_kind == DT_NODE_VAR) {
672		dt_ident_t *idp = dt_ident_resolve(dnp->dn_child->dn_ident);
673
674		idp->di_flags |= DT_IDFLG_DIFW;
675		instr = DIF_INSTR_STV(dt_cg_stvar(idp),
676		    idp->di_id, dnp->dn_reg);
677		dt_irlist_append(dlp, dt_cg_node_alloc(DT_LBL_NONE, instr));
678	} else {
679		uint_t rbit = dnp->dn_child->dn_flags & DT_NF_REF;
680
681		assert(dnp->dn_child->dn_flags & DT_NF_WRITABLE);
682		assert(dnp->dn_child->dn_flags & DT_NF_LVALUE);
683
684		dnp->dn_child->dn_flags |= DT_NF_REF; /* force pass-by-ref */
685		dt_cg_node(dnp->dn_child, dlp, drp);
686
687		dt_cg_store(dnp, dlp, drp, dnp->dn_child);
688		dt_regset_free(drp, dnp->dn_child->dn_reg);
689
690		dnp->dn_left->dn_flags &= ~DT_NF_REF;
691		dnp->dn_left->dn_flags |= rbit;
692	}
693}
694
695static void
696dt_cg_postarith_op(dt_node_t *dnp, dt_irlist_t *dlp,
697    dt_regset_t *drp, uint_t op)
698{
699	ctf_file_t *ctfp = dnp->dn_ctfp;
700	dif_instr_t instr;
701	ctf_id_t type;
702	ssize_t size = 1;
703	int nreg;
704
705	if (dt_node_is_pointer(dnp)) {
706		type = ctf_type_resolve(ctfp, dnp->dn_type);
707		assert(ctf_type_kind(ctfp, type) == CTF_K_POINTER);
708		size = ctf_type_size(ctfp, ctf_type_reference(ctfp, type));
709	}
710
711	dt_cg_node(dnp->dn_child, dlp, drp);
712	dnp->dn_reg = dnp->dn_child->dn_reg;
713
714	nreg = dt_regset_alloc(drp);
715	dt_cg_setx(dlp, nreg, size);
716	instr = DIF_INSTR_FMT(op, dnp->dn_reg, nreg, nreg);
717	dt_irlist_append(dlp, dt_cg_node_alloc(DT_LBL_NONE, instr));
718
719	/*
720	 * If we are modifying a variable, generate an stv instruction from
721	 * the variable specified by the identifier.  If we are storing to a
722	 * memory address, generate code again for the left-hand side using
723	 * DT_NF_REF to get the address, and then generate a store to it.
724	 * In both paths, we store the value from 'nreg' (the new value).
725	 */
726	if (dnp->dn_child->dn_kind == DT_NODE_VAR) {
727		dt_ident_t *idp = dt_ident_resolve(dnp->dn_child->dn_ident);
728
729		idp->di_flags |= DT_IDFLG_DIFW;
730		instr = DIF_INSTR_STV(dt_cg_stvar(idp), idp->di_id, nreg);
731		dt_irlist_append(dlp, dt_cg_node_alloc(DT_LBL_NONE, instr));
732	} else {
733		uint_t rbit = dnp->dn_child->dn_flags & DT_NF_REF;
734		int oreg = dnp->dn_reg;
735
736		assert(dnp->dn_child->dn_flags & DT_NF_WRITABLE);
737		assert(dnp->dn_child->dn_flags & DT_NF_LVALUE);
738
739		dnp->dn_child->dn_flags |= DT_NF_REF; /* force pass-by-ref */
740		dt_cg_node(dnp->dn_child, dlp, drp);
741
742		dnp->dn_reg = nreg;
743		dt_cg_store(dnp, dlp, drp, dnp->dn_child);
744		dnp->dn_reg = oreg;
745
746		dt_regset_free(drp, dnp->dn_child->dn_reg);
747		dnp->dn_left->dn_flags &= ~DT_NF_REF;
748		dnp->dn_left->dn_flags |= rbit;
749	}
750
751	dt_regset_free(drp, nreg);
752}
753
754/*
755 * Determine if we should perform signed or unsigned comparison for an OP2.
756 * If both operands are of arithmetic type, perform the usual arithmetic
757 * conversions to determine the common real type for comparison [ISOC 6.5.8.3].
758 */
759static int
760dt_cg_compare_signed(dt_node_t *dnp)
761{
762	dt_node_t dn;
763
764	if (dt_node_is_string(dnp->dn_left) ||
765	    dt_node_is_string(dnp->dn_right))
766		return (1); /* strings always compare signed */
767	else if (!dt_node_is_arith(dnp->dn_left) ||
768	    !dt_node_is_arith(dnp->dn_right))
769		return (0); /* non-arithmetic types always compare unsigned */
770
771	bzero(&dn, sizeof (dn));
772	dt_node_promote(dnp->dn_left, dnp->dn_right, &dn);
773	return (dn.dn_flags & DT_NF_SIGNED);
774}
775
776static void
777dt_cg_compare_op(dt_node_t *dnp, dt_irlist_t *dlp, dt_regset_t *drp, uint_t op)
778{
779	uint_t lbl_true = dt_irlist_label(dlp);
780	uint_t lbl_post = dt_irlist_label(dlp);
781
782	dif_instr_t instr;
783	uint_t opc;
784
785	dt_cg_node(dnp->dn_left, dlp, drp);
786	dt_cg_node(dnp->dn_right, dlp, drp);
787
788	if (dt_node_is_string(dnp->dn_left) || dt_node_is_string(dnp->dn_right))
789		opc = DIF_OP_SCMP;
790	else
791		opc = DIF_OP_CMP;
792
793	instr = DIF_INSTR_CMP(opc, dnp->dn_left->dn_reg, dnp->dn_right->dn_reg);
794	dt_irlist_append(dlp, dt_cg_node_alloc(DT_LBL_NONE, instr));
795	dt_regset_free(drp, dnp->dn_right->dn_reg);
796	dnp->dn_reg = dnp->dn_left->dn_reg;
797
798	instr = DIF_INSTR_BRANCH(op, lbl_true);
799	dt_irlist_append(dlp, dt_cg_node_alloc(DT_LBL_NONE, instr));
800
801	instr = DIF_INSTR_MOV(DIF_REG_R0, dnp->dn_reg);
802	dt_irlist_append(dlp, dt_cg_node_alloc(DT_LBL_NONE, instr));
803
804	instr = DIF_INSTR_BRANCH(DIF_OP_BA, lbl_post);
805	dt_irlist_append(dlp, dt_cg_node_alloc(DT_LBL_NONE, instr));
806
807	dt_cg_xsetx(dlp, NULL, lbl_true, dnp->dn_reg, 1);
808	dt_irlist_append(dlp, dt_cg_node_alloc(lbl_post, DIF_INSTR_NOP));
809}
810
811/*
812 * Code generation for the ternary op requires some trickery with the assembler
813 * in order to conserve registers.  We generate code for dn_expr and dn_left
814 * and free their registers so they do not have be consumed across codegen for
815 * dn_right.  We insert a dummy MOV at the end of dn_left into the destination
816 * register, which is not yet known because we haven't done dn_right yet, and
817 * save the pointer to this instruction node.  We then generate code for
818 * dn_right and use its register as our output.  Finally, we reach back and
819 * patch the instruction for dn_left to move its output into this register.
820 */
821static void
822dt_cg_ternary_op(dt_node_t *dnp, dt_irlist_t *dlp, dt_regset_t *drp)
823{
824	uint_t lbl_false = dt_irlist_label(dlp);
825	uint_t lbl_post = dt_irlist_label(dlp);
826
827	dif_instr_t instr;
828	dt_irnode_t *dip;
829
830	dt_cg_node(dnp->dn_expr, dlp, drp);
831	instr = DIF_INSTR_TST(dnp->dn_expr->dn_reg);
832	dt_irlist_append(dlp, dt_cg_node_alloc(DT_LBL_NONE, instr));
833	dt_regset_free(drp, dnp->dn_expr->dn_reg);
834
835	instr = DIF_INSTR_BRANCH(DIF_OP_BE, lbl_false);
836	dt_irlist_append(dlp, dt_cg_node_alloc(DT_LBL_NONE, instr));
837
838	dt_cg_node(dnp->dn_left, dlp, drp);
839	instr = DIF_INSTR_MOV(dnp->dn_left->dn_reg, DIF_REG_R0);
840	dip = dt_cg_node_alloc(DT_LBL_NONE, instr); /* save dip for below */
841	dt_irlist_append(dlp, dip);
842	dt_regset_free(drp, dnp->dn_left->dn_reg);
843
844	instr = DIF_INSTR_BRANCH(DIF_OP_BA, lbl_post);
845	dt_irlist_append(dlp, dt_cg_node_alloc(DT_LBL_NONE, instr));
846
847	dt_irlist_append(dlp, dt_cg_node_alloc(lbl_false, DIF_INSTR_NOP));
848	dt_cg_node(dnp->dn_right, dlp, drp);
849	dnp->dn_reg = dnp->dn_right->dn_reg;
850
851	/*
852	 * Now that dn_reg is assigned, reach back and patch the correct MOV
853	 * instruction into the tail of dn_left.  We know dn_reg was unused
854	 * at that point because otherwise dn_right couldn't have allocated it.
855	 */
856	dip->di_instr = DIF_INSTR_MOV(dnp->dn_left->dn_reg, dnp->dn_reg);
857	dt_irlist_append(dlp, dt_cg_node_alloc(lbl_post, DIF_INSTR_NOP));
858}
859
860static void
861dt_cg_logical_and(dt_node_t *dnp, dt_irlist_t *dlp, dt_regset_t *drp)
862{
863	uint_t lbl_false = dt_irlist_label(dlp);
864	uint_t lbl_post = dt_irlist_label(dlp);
865
866	dif_instr_t instr;
867
868	dt_cg_node(dnp->dn_left, dlp, drp);
869	instr = DIF_INSTR_TST(dnp->dn_left->dn_reg);
870	dt_irlist_append(dlp, dt_cg_node_alloc(DT_LBL_NONE, instr));
871	dt_regset_free(drp, dnp->dn_left->dn_reg);
872
873	instr = DIF_INSTR_BRANCH(DIF_OP_BE, lbl_false);
874	dt_irlist_append(dlp, dt_cg_node_alloc(DT_LBL_NONE, instr));
875
876	dt_cg_node(dnp->dn_right, dlp, drp);
877	instr = DIF_INSTR_TST(dnp->dn_right->dn_reg);
878	dt_irlist_append(dlp, dt_cg_node_alloc(DT_LBL_NONE, instr));
879	dnp->dn_reg = dnp->dn_right->dn_reg;
880
881	instr = DIF_INSTR_BRANCH(DIF_OP_BE, lbl_false);
882	dt_irlist_append(dlp, dt_cg_node_alloc(DT_LBL_NONE, instr));
883
884	dt_cg_setx(dlp, dnp->dn_reg, 1);
885
886	instr = DIF_INSTR_BRANCH(DIF_OP_BA, lbl_post);
887	dt_irlist_append(dlp, dt_cg_node_alloc(DT_LBL_NONE, instr));
888
889	instr = DIF_INSTR_MOV(DIF_REG_R0, dnp->dn_reg);
890	dt_irlist_append(dlp, dt_cg_node_alloc(lbl_false, instr));
891
892	dt_irlist_append(dlp, dt_cg_node_alloc(lbl_post, DIF_INSTR_NOP));
893}
894
895static void
896dt_cg_logical_xor(dt_node_t *dnp, dt_irlist_t *dlp, dt_regset_t *drp)
897{
898	uint_t lbl_next = dt_irlist_label(dlp);
899	uint_t lbl_tail = dt_irlist_label(dlp);
900
901	dif_instr_t instr;
902
903	dt_cg_node(dnp->dn_left, dlp, drp);
904	instr = DIF_INSTR_TST(dnp->dn_left->dn_reg);
905	dt_irlist_append(dlp, dt_cg_node_alloc(DT_LBL_NONE, instr));
906
907	instr = DIF_INSTR_BRANCH(DIF_OP_BE, lbl_next);
908	dt_irlist_append(dlp, dt_cg_node_alloc(DT_LBL_NONE, instr));
909	dt_cg_setx(dlp, dnp->dn_left->dn_reg, 1);
910
911	dt_irlist_append(dlp, dt_cg_node_alloc(lbl_next, DIF_INSTR_NOP));
912	dt_cg_node(dnp->dn_right, dlp, drp);
913
914	instr = DIF_INSTR_TST(dnp->dn_right->dn_reg);
915	dt_irlist_append(dlp, dt_cg_node_alloc(DT_LBL_NONE, instr));
916
917	instr = DIF_INSTR_BRANCH(DIF_OP_BE, lbl_tail);
918	dt_irlist_append(dlp, dt_cg_node_alloc(DT_LBL_NONE, instr));
919	dt_cg_setx(dlp, dnp->dn_right->dn_reg, 1);
920
921	instr = DIF_INSTR_FMT(DIF_OP_XOR, dnp->dn_left->dn_reg,
922	    dnp->dn_right->dn_reg, dnp->dn_left->dn_reg);
923
924	dt_irlist_append(dlp, dt_cg_node_alloc(lbl_tail, instr));
925
926	dt_regset_free(drp, dnp->dn_right->dn_reg);
927	dnp->dn_reg = dnp->dn_left->dn_reg;
928}
929
930static void
931dt_cg_logical_or(dt_node_t *dnp, dt_irlist_t *dlp, dt_regset_t *drp)
932{
933	uint_t lbl_true = dt_irlist_label(dlp);
934	uint_t lbl_false = dt_irlist_label(dlp);
935	uint_t lbl_post = dt_irlist_label(dlp);
936
937	dif_instr_t instr;
938
939	dt_cg_node(dnp->dn_left, dlp, drp);
940	instr = DIF_INSTR_TST(dnp->dn_left->dn_reg);
941	dt_irlist_append(dlp, dt_cg_node_alloc(DT_LBL_NONE, instr));
942	dt_regset_free(drp, dnp->dn_left->dn_reg);
943
944	instr = DIF_INSTR_BRANCH(DIF_OP_BNE, lbl_true);
945	dt_irlist_append(dlp, dt_cg_node_alloc(DT_LBL_NONE, instr));
946
947	dt_cg_node(dnp->dn_right, dlp, drp);
948	instr = DIF_INSTR_TST(dnp->dn_right->dn_reg);
949	dt_irlist_append(dlp, dt_cg_node_alloc(DT_LBL_NONE, instr));
950	dnp->dn_reg = dnp->dn_right->dn_reg;
951
952	instr = DIF_INSTR_BRANCH(DIF_OP_BE, lbl_false);
953	dt_irlist_append(dlp, dt_cg_node_alloc(DT_LBL_NONE, instr));
954
955	dt_cg_xsetx(dlp, NULL, lbl_true, dnp->dn_reg, 1);
956
957	instr = DIF_INSTR_BRANCH(DIF_OP_BA, lbl_post);
958	dt_irlist_append(dlp, dt_cg_node_alloc(DT_LBL_NONE, instr));
959
960	instr = DIF_INSTR_MOV(DIF_REG_R0, dnp->dn_reg);
961	dt_irlist_append(dlp, dt_cg_node_alloc(lbl_false, instr));
962
963	dt_irlist_append(dlp, dt_cg_node_alloc(lbl_post, DIF_INSTR_NOP));
964}
965
966static void
967dt_cg_logical_neg(dt_node_t *dnp, dt_irlist_t *dlp, dt_regset_t *drp)
968{
969	uint_t lbl_zero = dt_irlist_label(dlp);
970	uint_t lbl_post = dt_irlist_label(dlp);
971
972	dif_instr_t instr;
973
974	dt_cg_node(dnp->dn_child, dlp, drp);
975	dnp->dn_reg = dnp->dn_child->dn_reg;
976
977	instr = DIF_INSTR_TST(dnp->dn_reg);
978	dt_irlist_append(dlp, dt_cg_node_alloc(DT_LBL_NONE, instr));
979
980	instr = DIF_INSTR_BRANCH(DIF_OP_BE, lbl_zero);
981	dt_irlist_append(dlp, dt_cg_node_alloc(DT_LBL_NONE, instr));
982
983	instr = DIF_INSTR_MOV(DIF_REG_R0, dnp->dn_reg);
984	dt_irlist_append(dlp, dt_cg_node_alloc(DT_LBL_NONE, instr));
985
986	instr = DIF_INSTR_BRANCH(DIF_OP_BA, lbl_post);
987	dt_irlist_append(dlp, dt_cg_node_alloc(DT_LBL_NONE, instr));
988
989	dt_cg_xsetx(dlp, NULL, lbl_zero, dnp->dn_reg, 1);
990	dt_irlist_append(dlp, dt_cg_node_alloc(lbl_post, DIF_INSTR_NOP));
991}
992
993static void
994dt_cg_asgn_op(dt_node_t *dnp, dt_irlist_t *dlp, dt_regset_t *drp)
995{
996	dif_instr_t instr;
997	dt_ident_t *idp;
998
999	/*
1000	 * If we are performing a structure assignment of a translated type,
1001	 * we must instantiate all members and create a snapshot of the object
1002	 * in scratch space.  We allocs a chunk of memory, generate code for
1003	 * each member, and then set dnp->dn_reg to the scratch object address.
1004	 */
1005	if ((idp = dt_node_resolve(dnp->dn_right, DT_IDENT_XLSOU)) != NULL) {
1006		ctf_membinfo_t ctm;
1007		dt_xlator_t *dxp = idp->di_data;
1008		dt_node_t *mnp, dn, mn;
1009		int r1, r2;
1010
1011		/*
1012		 * Create two fake dt_node_t's representing operator "." and a
1013		 * right-hand identifier child node.  These will be repeatedly
1014		 * modified according to each instantiated member so that we
1015		 * can pass them to dt_cg_store() and effect a member store.
1016		 */
1017		bzero(&dn, sizeof (dt_node_t));
1018		dn.dn_kind = DT_NODE_OP2;
1019		dn.dn_op = DT_TOK_DOT;
1020		dn.dn_left = dnp;
1021		dn.dn_right = &mn;
1022
1023		bzero(&mn, sizeof (dt_node_t));
1024		mn.dn_kind = DT_NODE_IDENT;
1025		mn.dn_op = DT_TOK_IDENT;
1026
1027		/*
1028		 * Allocate a register for our scratch data pointer.  First we
1029		 * set it to the size of our data structure, and then replace
1030		 * it with the result of an allocs of the specified size.
1031		 */
1032		r1 = dt_regset_alloc(drp);
1033		dt_cg_setx(dlp, r1,
1034		    ctf_type_size(dxp->dx_dst_ctfp, dxp->dx_dst_base));
1035
1036		instr = DIF_INSTR_ALLOCS(r1, r1);
1037		dt_irlist_append(dlp, dt_cg_node_alloc(DT_LBL_NONE, instr));
1038
1039		/*
1040		 * When dt_cg_asgn_op() is called, we have already generated
1041		 * code for dnp->dn_right, which is the translator input.  We
1042		 * now associate this register with the translator's input
1043		 * identifier so it can be referenced during our member loop.
1044		 */
1045		dxp->dx_ident->di_flags |= DT_IDFLG_CGREG;
1046		dxp->dx_ident->di_id = dnp->dn_right->dn_reg;
1047
1048		for (mnp = dxp->dx_members; mnp != NULL; mnp = mnp->dn_list) {
1049			/*
1050			 * Generate code for the translator member expression,
1051			 * and then cast the result to the member type.
1052			 */
1053			dt_cg_node(mnp->dn_membexpr, dlp, drp);
1054			mnp->dn_reg = mnp->dn_membexpr->dn_reg;
1055			dt_cg_typecast(mnp->dn_membexpr, mnp, dlp, drp);
1056
1057			/*
1058			 * Ask CTF for the offset of the member so we can store
1059			 * to the appropriate offset.  This call has already
1060			 * been done once by the parser, so it should succeed.
1061			 */
1062			if (ctf_member_info(dxp->dx_dst_ctfp, dxp->dx_dst_base,
1063			    mnp->dn_membname, &ctm) == CTF_ERR) {
1064				yypcb->pcb_hdl->dt_ctferr =
1065				    ctf_errno(dxp->dx_dst_ctfp);
1066				longjmp(yypcb->pcb_jmpbuf, EDT_CTF);
1067			}
1068
1069			/*
1070			 * If the destination member is at offset 0, store the
1071			 * result directly to r1 (the scratch buffer address).
1072			 * Otherwise allocate another temporary for the offset
1073			 * and add r1 to it before storing the result.
1074			 */
1075			if (ctm.ctm_offset != 0) {
1076				r2 = dt_regset_alloc(drp);
1077
1078				/*
1079				 * Add the member offset rounded down to the
1080				 * nearest byte.  If the offset was not aligned
1081				 * on a byte boundary, this member is a bit-
1082				 * field and dt_cg_store() will handle masking.
1083				 */
1084				dt_cg_setx(dlp, r2, ctm.ctm_offset / NBBY);
1085				instr = DIF_INSTR_FMT(DIF_OP_ADD, r1, r2, r2);
1086				dt_irlist_append(dlp,
1087				    dt_cg_node_alloc(DT_LBL_NONE, instr));
1088
1089				dt_node_type_propagate(mnp, &dn);
1090				dn.dn_right->dn_string = mnp->dn_membname;
1091				dn.dn_reg = r2;
1092
1093				dt_cg_store(mnp, dlp, drp, &dn);
1094				dt_regset_free(drp, r2);
1095
1096			} else {
1097				dt_node_type_propagate(mnp, &dn);
1098				dn.dn_right->dn_string = mnp->dn_membname;
1099				dn.dn_reg = r1;
1100
1101				dt_cg_store(mnp, dlp, drp, &dn);
1102			}
1103
1104			dt_regset_free(drp, mnp->dn_reg);
1105		}
1106
1107		dxp->dx_ident->di_flags &= ~DT_IDFLG_CGREG;
1108		dxp->dx_ident->di_id = 0;
1109
1110		if (dnp->dn_right->dn_reg != -1)
1111			dt_regset_free(drp, dnp->dn_right->dn_reg);
1112
1113		assert(dnp->dn_reg == dnp->dn_right->dn_reg);
1114		dnp->dn_reg = r1;
1115	}
1116
1117	/*
1118	 * If we are storing to a variable, generate an stv instruction from
1119	 * the variable specified by the identifier.  If we are storing to a
1120	 * memory address, generate code again for the left-hand side using
1121	 * DT_NF_REF to get the address, and then generate a store to it.
1122	 * In both paths, we assume dnp->dn_reg already has the new value.
1123	 */
1124	if (dnp->dn_left->dn_kind == DT_NODE_VAR) {
1125		idp = dt_ident_resolve(dnp->dn_left->dn_ident);
1126
1127		if (idp->di_kind == DT_IDENT_ARRAY)
1128			dt_cg_arglist(idp, dnp->dn_left->dn_args, dlp, drp);
1129
1130		idp->di_flags |= DT_IDFLG_DIFW;
1131		instr = DIF_INSTR_STV(dt_cg_stvar(idp),
1132		    idp->di_id, dnp->dn_reg);
1133		dt_irlist_append(dlp, dt_cg_node_alloc(DT_LBL_NONE, instr));
1134	} else {
1135		uint_t rbit = dnp->dn_left->dn_flags & DT_NF_REF;
1136
1137		assert(dnp->dn_left->dn_flags & DT_NF_WRITABLE);
1138		assert(dnp->dn_left->dn_flags & DT_NF_LVALUE);
1139
1140		dnp->dn_left->dn_flags |= DT_NF_REF; /* force pass-by-ref */
1141
1142		dt_cg_node(dnp->dn_left, dlp, drp);
1143		dt_cg_store(dnp, dlp, drp, dnp->dn_left);
1144		dt_regset_free(drp, dnp->dn_left->dn_reg);
1145
1146		dnp->dn_left->dn_flags &= ~DT_NF_REF;
1147		dnp->dn_left->dn_flags |= rbit;
1148	}
1149}
1150
1151static void
1152dt_cg_assoc_op(dt_node_t *dnp, dt_irlist_t *dlp, dt_regset_t *drp)
1153{
1154	dif_instr_t instr;
1155	uint_t op;
1156
1157	assert(dnp->dn_kind == DT_NODE_VAR);
1158	assert(!(dnp->dn_ident->di_flags & DT_IDFLG_LOCAL));
1159	assert(dnp->dn_args != NULL);
1160
1161	dt_cg_arglist(dnp->dn_ident, dnp->dn_args, dlp, drp);
1162
1163	dnp->dn_reg = dt_regset_alloc(drp);
1164
1165	if (dnp->dn_ident->di_flags & DT_IDFLG_TLS)
1166		op = DIF_OP_LDTAA;
1167	else
1168		op = DIF_OP_LDGAA;
1169
1170	dnp->dn_ident->di_flags |= DT_IDFLG_DIFR;
1171	instr = DIF_INSTR_LDV(op, dnp->dn_ident->di_id, dnp->dn_reg);
1172	dt_irlist_append(dlp, dt_cg_node_alloc(DT_LBL_NONE, instr));
1173
1174	/*
1175	 * If the associative array is a pass-by-reference type, then we are
1176	 * loading its value as a pointer to either load or store through it.
1177	 * The array element in question may not have been faulted in yet, in
1178	 * which case DIF_OP_LD*AA will return zero.  We append an epilogue
1179	 * of instructions similar to the following:
1180	 *
1181	 *	  ld?aa	 id, %r1	! base ld?aa instruction above
1182	 *	  tst	 %r1		! start of epilogue
1183	 *   +--- bne	 label
1184	 *   |    setx	 size, %r1
1185	 *   |    allocs %r1, %r1
1186	 *   |    st?aa	 id, %r1
1187	 *   |    ld?aa	 id, %r1
1188	 *   v
1189	 * label: < rest of code >
1190	 *
1191	 * The idea is that we allocs a zero-filled chunk of scratch space and
1192	 * do a DIF_OP_ST*AA to fault in and initialize the array element, and
1193	 * then reload it to get the faulted-in address of the new variable
1194	 * storage.  This isn't cheap, but pass-by-ref associative array values
1195	 * are (thus far) uncommon and the allocs cost only occurs once.  If
1196	 * this path becomes important to DTrace users, we can improve things
1197	 * by adding a new DIF opcode to fault in associative array elements.
1198	 */
1199	if (dnp->dn_flags & DT_NF_REF) {
1200		uint_t stvop = op == DIF_OP_LDTAA ? DIF_OP_STTAA : DIF_OP_STGAA;
1201		uint_t label = dt_irlist_label(dlp);
1202
1203		instr = DIF_INSTR_TST(dnp->dn_reg);
1204		dt_irlist_append(dlp, dt_cg_node_alloc(DT_LBL_NONE, instr));
1205
1206		instr = DIF_INSTR_BRANCH(DIF_OP_BNE, label);
1207		dt_irlist_append(dlp, dt_cg_node_alloc(DT_LBL_NONE, instr));
1208
1209		dt_cg_setx(dlp, dnp->dn_reg, dt_node_type_size(dnp));
1210		instr = DIF_INSTR_ALLOCS(dnp->dn_reg, dnp->dn_reg);
1211		dt_irlist_append(dlp, dt_cg_node_alloc(DT_LBL_NONE, instr));
1212
1213		dnp->dn_ident->di_flags |= DT_IDFLG_DIFW;
1214		instr = DIF_INSTR_STV(stvop, dnp->dn_ident->di_id, dnp->dn_reg);
1215		dt_irlist_append(dlp, dt_cg_node_alloc(DT_LBL_NONE, instr));
1216
1217		instr = DIF_INSTR_LDV(op, dnp->dn_ident->di_id, dnp->dn_reg);
1218		dt_irlist_append(dlp, dt_cg_node_alloc(DT_LBL_NONE, instr));
1219
1220		dt_irlist_append(dlp, dt_cg_node_alloc(label, DIF_INSTR_NOP));
1221	}
1222}
1223
1224static void
1225dt_cg_array_op(dt_node_t *dnp, dt_irlist_t *dlp, dt_regset_t *drp)
1226{
1227	dt_probe_t *prp = yypcb->pcb_probe;
1228	uintmax_t saved = dnp->dn_args->dn_value;
1229	dt_ident_t *idp = dnp->dn_ident;
1230
1231	dif_instr_t instr;
1232	uint_t op;
1233	size_t size;
1234	int reg, n;
1235
1236	assert(dnp->dn_kind == DT_NODE_VAR);
1237	assert(!(idp->di_flags & DT_IDFLG_LOCAL));
1238
1239	assert(dnp->dn_args->dn_kind == DT_NODE_INT);
1240	assert(dnp->dn_args->dn_list == NULL);
1241
1242	/*
1243	 * If this is a reference in the args[] array, temporarily modify the
1244	 * array index according to the static argument mapping (if any),
1245	 * unless the argument reference is provided by a dynamic translator.
1246	 * If we're using a dynamic translator for args[], then just set dn_reg
1247	 * to an invalid reg and return: DIF_OP_XLARG will fetch the arg later.
1248	 */
1249	if (idp->di_id == DIF_VAR_ARGS) {
1250		if ((idp->di_kind == DT_IDENT_XLPTR ||
1251		    idp->di_kind == DT_IDENT_XLSOU) &&
1252		    dt_xlator_dynamic(idp->di_data)) {
1253			dnp->dn_reg = -1;
1254			return;
1255		}
1256		dnp->dn_args->dn_value = prp->pr_mapping[saved];
1257	}
1258
1259	dt_cg_node(dnp->dn_args, dlp, drp);
1260	dnp->dn_args->dn_value = saved;
1261
1262	dnp->dn_reg = dnp->dn_args->dn_reg;
1263
1264	if (idp->di_flags & DT_IDFLG_TLS)
1265		op = DIF_OP_LDTA;
1266	else
1267		op = DIF_OP_LDGA;
1268
1269	idp->di_flags |= DT_IDFLG_DIFR;
1270
1271	instr = DIF_INSTR_LDA(op, idp->di_id,
1272	    dnp->dn_args->dn_reg, dnp->dn_reg);
1273
1274	dt_irlist_append(dlp, dt_cg_node_alloc(DT_LBL_NONE, instr));
1275
1276	/*
1277	 * If this is a reference to the args[] array, we need to take the
1278	 * additional step of explicitly eliminating any bits larger than the
1279	 * type size: the DIF interpreter in the kernel will always give us
1280	 * the raw (64-bit) argument value, and any bits larger than the type
1281	 * size may be junk.  As a practical matter, this arises only on 64-bit
1282	 * architectures and only when the argument index is larger than the
1283	 * number of arguments passed directly to DTrace: if a 8-, 16- or
1284	 * 32-bit argument must be retrieved from the stack, it is possible
1285	 * (and it some cases, likely) that the upper bits will be garbage.
1286	 */
1287	if (idp->di_id != DIF_VAR_ARGS || !dt_node_is_scalar(dnp))
1288		return;
1289
1290	if ((size = dt_node_type_size(dnp)) == sizeof (uint64_t))
1291		return;
1292
1293	reg = dt_regset_alloc(drp);
1294	assert(size < sizeof (uint64_t));
1295	n = sizeof (uint64_t) * NBBY - size * NBBY;
1296
1297	dt_cg_setx(dlp, reg, n);
1298
1299	instr = DIF_INSTR_FMT(DIF_OP_SLL, dnp->dn_reg, reg, dnp->dn_reg);
1300	dt_irlist_append(dlp, dt_cg_node_alloc(DT_LBL_NONE, instr));
1301
1302	instr = DIF_INSTR_FMT((dnp->dn_flags & DT_NF_SIGNED) ?
1303	    DIF_OP_SRA : DIF_OP_SRL, dnp->dn_reg, reg, dnp->dn_reg);
1304
1305	dt_irlist_append(dlp, dt_cg_node_alloc(DT_LBL_NONE, instr));
1306	dt_regset_free(drp, reg);
1307}
1308
1309/*
1310 * Generate code for an inlined variable reference.  Inlines can be used to
1311 * define either scalar or associative array substitutions.  For scalars, we
1312 * simply generate code for the parse tree saved in the identifier's din_root,
1313 * and then cast the resulting expression to the inline's declaration type.
1314 * For arrays, we take the input parameter subtrees from dnp->dn_args and
1315 * temporarily store them in the din_root of each din_argv[i] identifier,
1316 * which are themselves inlines and were set up for us by the parser.  The
1317 * result is that any reference to the inlined parameter inside the top-level
1318 * din_root will turn into a recursive call to dt_cg_inline() for a scalar
1319 * inline whose din_root will refer to the subtree pointed to by the argument.
1320 */
1321static void
1322dt_cg_inline(dt_node_t *dnp, dt_irlist_t *dlp, dt_regset_t *drp)
1323{
1324	dt_ident_t *idp = dnp->dn_ident;
1325	dt_idnode_t *inp = idp->di_iarg;
1326
1327	dt_idnode_t *pinp;
1328	dt_node_t *pnp;
1329	int i;
1330
1331	assert(idp->di_flags & DT_IDFLG_INLINE);
1332	assert(idp->di_ops == &dt_idops_inline);
1333
1334	if (idp->di_kind == DT_IDENT_ARRAY) {
1335		for (i = 0, pnp = dnp->dn_args;
1336		    pnp != NULL; pnp = pnp->dn_list, i++) {
1337			if (inp->din_argv[i] != NULL) {
1338				pinp = inp->din_argv[i]->di_iarg;
1339				pinp->din_root = pnp;
1340			}
1341		}
1342	}
1343
1344	dt_cg_node(inp->din_root, dlp, drp);
1345	dnp->dn_reg = inp->din_root->dn_reg;
1346	dt_cg_typecast(inp->din_root, dnp, dlp, drp);
1347
1348	if (idp->di_kind == DT_IDENT_ARRAY) {
1349		for (i = 0; i < inp->din_argc; i++) {
1350			pinp = inp->din_argv[i]->di_iarg;
1351			pinp->din_root = NULL;
1352		}
1353	}
1354}
1355
1356static void
1357dt_cg_func_typeref(dtrace_hdl_t *dtp, dt_node_t *dnp)
1358{
1359	dtrace_typeinfo_t dtt;
1360	dt_node_t *addr = dnp->dn_args;
1361	dt_node_t *nelm = addr->dn_list;
1362	dt_node_t *strp = nelm->dn_list;
1363	dt_node_t *typs = strp->dn_list;
1364	char buf[DT_TYPE_NAMELEN];
1365	char *p;
1366
1367	ctf_type_name(addr->dn_ctfp, addr->dn_type, buf, sizeof (buf));
1368
1369	/*
1370	 * XXX Hack alert! XXX
1371	 * The prototype has two dummy args that we munge to represent
1372	 * the type string and the type size.
1373	 *
1374	 * Yes, I hear your grumble, but it works for now. We'll come
1375	 * up with a more elegant implementation later. :-)
1376	 */
1377	free(strp->dn_string);
1378
1379	if ((p = strchr(buf, '*')) != NULL)
1380		*p = '\0';
1381
1382	strp->dn_string = strdup(buf);
1383
1384	if (dtrace_lookup_by_type(dtp,  DTRACE_OBJ_EVERY, buf, &dtt) < 0)
1385		return;
1386
1387	typs->dn_value = ctf_type_size(dtt.dtt_ctfp, dtt.dtt_type);
1388}
1389
1390static void
1391dt_cg_node(dt_node_t *dnp, dt_irlist_t *dlp, dt_regset_t *drp)
1392{
1393	ctf_file_t *ctfp = dnp->dn_ctfp;
1394	ctf_file_t *octfp;
1395	ctf_membinfo_t m;
1396	ctf_id_t type;
1397
1398	dif_instr_t instr;
1399	dt_ident_t *idp;
1400	ssize_t stroff;
1401	uint_t op;
1402
1403	switch (dnp->dn_op) {
1404	case DT_TOK_COMMA:
1405		dt_cg_node(dnp->dn_left, dlp, drp);
1406		dt_regset_free(drp, dnp->dn_left->dn_reg);
1407		dt_cg_node(dnp->dn_right, dlp, drp);
1408		dnp->dn_reg = dnp->dn_right->dn_reg;
1409		break;
1410
1411	case DT_TOK_ASGN:
1412		dt_cg_node(dnp->dn_right, dlp, drp);
1413		dnp->dn_reg = dnp->dn_right->dn_reg;
1414		dt_cg_asgn_op(dnp, dlp, drp);
1415		break;
1416
1417	case DT_TOK_ADD_EQ:
1418		dt_cg_arithmetic_op(dnp, dlp, drp, DIF_OP_ADD);
1419		dt_cg_asgn_op(dnp, dlp, drp);
1420		break;
1421
1422	case DT_TOK_SUB_EQ:
1423		dt_cg_arithmetic_op(dnp, dlp, drp, DIF_OP_SUB);
1424		dt_cg_asgn_op(dnp, dlp, drp);
1425		break;
1426
1427	case DT_TOK_MUL_EQ:
1428		dt_cg_arithmetic_op(dnp, dlp, drp, DIF_OP_MUL);
1429		dt_cg_asgn_op(dnp, dlp, drp);
1430		break;
1431
1432	case DT_TOK_DIV_EQ:
1433		dt_cg_arithmetic_op(dnp, dlp, drp,
1434		    (dnp->dn_flags & DT_NF_SIGNED) ? DIF_OP_SDIV : DIF_OP_UDIV);
1435		dt_cg_asgn_op(dnp, dlp, drp);
1436		break;
1437
1438	case DT_TOK_MOD_EQ:
1439		dt_cg_arithmetic_op(dnp, dlp, drp,
1440		    (dnp->dn_flags & DT_NF_SIGNED) ? DIF_OP_SREM : DIF_OP_UREM);
1441		dt_cg_asgn_op(dnp, dlp, drp);
1442		break;
1443
1444	case DT_TOK_AND_EQ:
1445		dt_cg_arithmetic_op(dnp, dlp, drp, DIF_OP_AND);
1446		dt_cg_asgn_op(dnp, dlp, drp);
1447		break;
1448
1449	case DT_TOK_XOR_EQ:
1450		dt_cg_arithmetic_op(dnp, dlp, drp, DIF_OP_XOR);
1451		dt_cg_asgn_op(dnp, dlp, drp);
1452		break;
1453
1454	case DT_TOK_OR_EQ:
1455		dt_cg_arithmetic_op(dnp, dlp, drp, DIF_OP_OR);
1456		dt_cg_asgn_op(dnp, dlp, drp);
1457		break;
1458
1459	case DT_TOK_LSH_EQ:
1460		dt_cg_arithmetic_op(dnp, dlp, drp, DIF_OP_SLL);
1461		dt_cg_asgn_op(dnp, dlp, drp);
1462		break;
1463
1464	case DT_TOK_RSH_EQ:
1465		dt_cg_arithmetic_op(dnp, dlp, drp,
1466		    (dnp->dn_flags & DT_NF_SIGNED) ? DIF_OP_SRA : DIF_OP_SRL);
1467		dt_cg_asgn_op(dnp, dlp, drp);
1468		break;
1469
1470	case DT_TOK_QUESTION:
1471		dt_cg_ternary_op(dnp, dlp, drp);
1472		break;
1473
1474	case DT_TOK_LOR:
1475		dt_cg_logical_or(dnp, dlp, drp);
1476		break;
1477
1478	case DT_TOK_LXOR:
1479		dt_cg_logical_xor(dnp, dlp, drp);
1480		break;
1481
1482	case DT_TOK_LAND:
1483		dt_cg_logical_and(dnp, dlp, drp);
1484		break;
1485
1486	case DT_TOK_BOR:
1487		dt_cg_arithmetic_op(dnp, dlp, drp, DIF_OP_OR);
1488		break;
1489
1490	case DT_TOK_XOR:
1491		dt_cg_arithmetic_op(dnp, dlp, drp, DIF_OP_XOR);
1492		break;
1493
1494	case DT_TOK_BAND:
1495		dt_cg_arithmetic_op(dnp, dlp, drp, DIF_OP_AND);
1496		break;
1497
1498	case DT_TOK_EQU:
1499		dt_cg_compare_op(dnp, dlp, drp, DIF_OP_BE);
1500		break;
1501
1502	case DT_TOK_NEQ:
1503		dt_cg_compare_op(dnp, dlp, drp, DIF_OP_BNE);
1504		break;
1505
1506	case DT_TOK_LT:
1507		dt_cg_compare_op(dnp, dlp, drp,
1508		    dt_cg_compare_signed(dnp) ? DIF_OP_BL : DIF_OP_BLU);
1509		break;
1510
1511	case DT_TOK_LE:
1512		dt_cg_compare_op(dnp, dlp, drp,
1513		    dt_cg_compare_signed(dnp) ? DIF_OP_BLE : DIF_OP_BLEU);
1514		break;
1515
1516	case DT_TOK_GT:
1517		dt_cg_compare_op(dnp, dlp, drp,
1518		    dt_cg_compare_signed(dnp) ? DIF_OP_BG : DIF_OP_BGU);
1519		break;
1520
1521	case DT_TOK_GE:
1522		dt_cg_compare_op(dnp, dlp, drp,
1523		    dt_cg_compare_signed(dnp) ? DIF_OP_BGE : DIF_OP_BGEU);
1524		break;
1525
1526	case DT_TOK_LSH:
1527		dt_cg_arithmetic_op(dnp, dlp, drp, DIF_OP_SLL);
1528		break;
1529
1530	case DT_TOK_RSH:
1531		dt_cg_arithmetic_op(dnp, dlp, drp,
1532		    (dnp->dn_flags & DT_NF_SIGNED) ? DIF_OP_SRA : DIF_OP_SRL);
1533		break;
1534
1535	case DT_TOK_ADD:
1536		dt_cg_arithmetic_op(dnp, dlp, drp, DIF_OP_ADD);
1537		break;
1538
1539	case DT_TOK_SUB:
1540		dt_cg_arithmetic_op(dnp, dlp, drp, DIF_OP_SUB);
1541		break;
1542
1543	case DT_TOK_MUL:
1544		dt_cg_arithmetic_op(dnp, dlp, drp, DIF_OP_MUL);
1545		break;
1546
1547	case DT_TOK_DIV:
1548		dt_cg_arithmetic_op(dnp, dlp, drp,
1549		    (dnp->dn_flags & DT_NF_SIGNED) ? DIF_OP_SDIV : DIF_OP_UDIV);
1550		break;
1551
1552	case DT_TOK_MOD:
1553		dt_cg_arithmetic_op(dnp, dlp, drp,
1554		    (dnp->dn_flags & DT_NF_SIGNED) ? DIF_OP_SREM : DIF_OP_UREM);
1555		break;
1556
1557	case DT_TOK_LNEG:
1558		dt_cg_logical_neg(dnp, dlp, drp);
1559		break;
1560
1561	case DT_TOK_BNEG:
1562		dt_cg_node(dnp->dn_child, dlp, drp);
1563		dnp->dn_reg = dnp->dn_child->dn_reg;
1564		instr = DIF_INSTR_NOT(dnp->dn_reg, dnp->dn_reg);
1565		dt_irlist_append(dlp, dt_cg_node_alloc(DT_LBL_NONE, instr));
1566		break;
1567
1568	case DT_TOK_PREINC:
1569		dt_cg_prearith_op(dnp, dlp, drp, DIF_OP_ADD);
1570		break;
1571
1572	case DT_TOK_POSTINC:
1573		dt_cg_postarith_op(dnp, dlp, drp, DIF_OP_ADD);
1574		break;
1575
1576	case DT_TOK_PREDEC:
1577		dt_cg_prearith_op(dnp, dlp, drp, DIF_OP_SUB);
1578		break;
1579
1580	case DT_TOK_POSTDEC:
1581		dt_cg_postarith_op(dnp, dlp, drp, DIF_OP_SUB);
1582		break;
1583
1584	case DT_TOK_IPOS:
1585		dt_cg_node(dnp->dn_child, dlp, drp);
1586		dnp->dn_reg = dnp->dn_child->dn_reg;
1587		break;
1588
1589	case DT_TOK_INEG:
1590		dt_cg_node(dnp->dn_child, dlp, drp);
1591		dnp->dn_reg = dnp->dn_child->dn_reg;
1592
1593		instr = DIF_INSTR_FMT(DIF_OP_SUB, DIF_REG_R0,
1594		    dnp->dn_reg, dnp->dn_reg);
1595
1596		dt_irlist_append(dlp, dt_cg_node_alloc(DT_LBL_NONE, instr));
1597		break;
1598
1599	case DT_TOK_DEREF:
1600		dt_cg_node(dnp->dn_child, dlp, drp);
1601		dnp->dn_reg = dnp->dn_child->dn_reg;
1602
1603		if (!(dnp->dn_flags & DT_NF_REF)) {
1604			uint_t ubit = dnp->dn_flags & DT_NF_USERLAND;
1605
1606			/*
1607			 * Save and restore DT_NF_USERLAND across dt_cg_load():
1608			 * we need the sign bit from dnp and the user bit from
1609			 * dnp->dn_child in order to get the proper opcode.
1610			 */
1611			dnp->dn_flags |=
1612			    (dnp->dn_child->dn_flags & DT_NF_USERLAND);
1613
1614			instr = DIF_INSTR_LOAD(dt_cg_load(dnp, ctfp,
1615			    dnp->dn_type), dnp->dn_reg, dnp->dn_reg);
1616
1617			dnp->dn_flags &= ~DT_NF_USERLAND;
1618			dnp->dn_flags |= ubit;
1619
1620			dt_irlist_append(dlp,
1621			    dt_cg_node_alloc(DT_LBL_NONE, instr));
1622		}
1623		break;
1624
1625	case DT_TOK_ADDROF: {
1626		uint_t rbit = dnp->dn_child->dn_flags & DT_NF_REF;
1627
1628		dnp->dn_child->dn_flags |= DT_NF_REF; /* force pass-by-ref */
1629		dt_cg_node(dnp->dn_child, dlp, drp);
1630		dnp->dn_reg = dnp->dn_child->dn_reg;
1631
1632		dnp->dn_child->dn_flags &= ~DT_NF_REF;
1633		dnp->dn_child->dn_flags |= rbit;
1634		break;
1635	}
1636
1637	case DT_TOK_SIZEOF: {
1638		size_t size = dt_node_sizeof(dnp->dn_child);
1639		dnp->dn_reg = dt_regset_alloc(drp);
1640		assert(size != 0);
1641		dt_cg_setx(dlp, dnp->dn_reg, size);
1642		break;
1643	}
1644
1645	case DT_TOK_STRINGOF:
1646		dt_cg_node(dnp->dn_child, dlp, drp);
1647		dnp->dn_reg = dnp->dn_child->dn_reg;
1648		break;
1649
1650	case DT_TOK_XLATE:
1651		/*
1652		 * An xlate operator appears in either an XLATOR, indicating a
1653		 * reference to a dynamic translator, or an OP2, indicating
1654		 * use of the xlate operator in the user's program.  For the
1655		 * dynamic case, generate an xlate opcode with a reference to
1656		 * the corresponding member, pre-computed for us in dn_members.
1657		 */
1658		if (dnp->dn_kind == DT_NODE_XLATOR) {
1659			dt_xlator_t *dxp = dnp->dn_xlator;
1660
1661			assert(dxp->dx_ident->di_flags & DT_IDFLG_CGREG);
1662			assert(dxp->dx_ident->di_id != 0);
1663
1664			dnp->dn_reg = dt_regset_alloc(drp);
1665
1666			if (dxp->dx_arg == -1) {
1667				instr = DIF_INSTR_MOV(
1668				    dxp->dx_ident->di_id, dnp->dn_reg);
1669				dt_irlist_append(dlp,
1670				    dt_cg_node_alloc(DT_LBL_NONE, instr));
1671				op = DIF_OP_XLATE;
1672			} else
1673				op = DIF_OP_XLARG;
1674
1675			instr = DIF_INSTR_XLATE(op, 0, dnp->dn_reg);
1676			dt_irlist_append(dlp,
1677			    dt_cg_node_alloc(DT_LBL_NONE, instr));
1678
1679			dlp->dl_last->di_extern = dnp->dn_xmember;
1680			break;
1681		}
1682
1683		assert(dnp->dn_kind == DT_NODE_OP2);
1684		dt_cg_node(dnp->dn_right, dlp, drp);
1685		dnp->dn_reg = dnp->dn_right->dn_reg;
1686		break;
1687
1688	case DT_TOK_LPAR:
1689		dt_cg_node(dnp->dn_right, dlp, drp);
1690		dnp->dn_reg = dnp->dn_right->dn_reg;
1691		dt_cg_typecast(dnp->dn_right, dnp, dlp, drp);
1692		break;
1693
1694	case DT_TOK_PTR:
1695	case DT_TOK_DOT:
1696		assert(dnp->dn_right->dn_kind == DT_NODE_IDENT);
1697		dt_cg_node(dnp->dn_left, dlp, drp);
1698
1699		/*
1700		 * If the left-hand side of PTR or DOT is a dynamic variable,
1701		 * we expect it to be the output of a D translator.   In this
1702		 * case, we look up the parse tree corresponding to the member
1703		 * that is being accessed and run the code generator over it.
1704		 * We then cast the result as if by the assignment operator.
1705		 */
1706		if ((idp = dt_node_resolve(
1707		    dnp->dn_left, DT_IDENT_XLSOU)) != NULL ||
1708		    (idp = dt_node_resolve(
1709		    dnp->dn_left, DT_IDENT_XLPTR)) != NULL) {
1710
1711			dt_xlator_t *dxp;
1712			dt_node_t *mnp;
1713
1714			dxp = idp->di_data;
1715			mnp = dt_xlator_member(dxp, dnp->dn_right->dn_string);
1716			assert(mnp != NULL);
1717
1718			dxp->dx_ident->di_flags |= DT_IDFLG_CGREG;
1719			dxp->dx_ident->di_id = dnp->dn_left->dn_reg;
1720
1721			dt_cg_node(mnp->dn_membexpr, dlp, drp);
1722			dnp->dn_reg = mnp->dn_membexpr->dn_reg;
1723			dt_cg_typecast(mnp->dn_membexpr, dnp, dlp, drp);
1724
1725			dxp->dx_ident->di_flags &= ~DT_IDFLG_CGREG;
1726			dxp->dx_ident->di_id = 0;
1727
1728			if (dnp->dn_left->dn_reg != -1)
1729				dt_regset_free(drp, dnp->dn_left->dn_reg);
1730			break;
1731		}
1732
1733		ctfp = dnp->dn_left->dn_ctfp;
1734		type = ctf_type_resolve(ctfp, dnp->dn_left->dn_type);
1735
1736		if (dnp->dn_op == DT_TOK_PTR) {
1737			type = ctf_type_reference(ctfp, type);
1738			type = ctf_type_resolve(ctfp, type);
1739		}
1740
1741		if ((ctfp = dt_cg_membinfo(octfp = ctfp, type,
1742		    dnp->dn_right->dn_string, &m)) == NULL) {
1743			yypcb->pcb_hdl->dt_ctferr = ctf_errno(octfp);
1744			longjmp(yypcb->pcb_jmpbuf, EDT_CTF);
1745		}
1746
1747		if (m.ctm_offset != 0) {
1748			int reg;
1749
1750			reg = dt_regset_alloc(drp);
1751
1752			/*
1753			 * If the offset is not aligned on a byte boundary, it
1754			 * is a bit-field member and we will extract the value
1755			 * bits below after we generate the appropriate load.
1756			 */
1757			dt_cg_setx(dlp, reg, m.ctm_offset / NBBY);
1758
1759			instr = DIF_INSTR_FMT(DIF_OP_ADD,
1760			    dnp->dn_left->dn_reg, reg, dnp->dn_left->dn_reg);
1761
1762			dt_irlist_append(dlp,
1763			    dt_cg_node_alloc(DT_LBL_NONE, instr));
1764			dt_regset_free(drp, reg);
1765		}
1766
1767		if (!(dnp->dn_flags & DT_NF_REF)) {
1768			uint_t ubit = dnp->dn_flags & DT_NF_USERLAND;
1769
1770			/*
1771			 * Save and restore DT_NF_USERLAND across dt_cg_load():
1772			 * we need the sign bit from dnp and the user bit from
1773			 * dnp->dn_left in order to get the proper opcode.
1774			 */
1775			dnp->dn_flags |=
1776			    (dnp->dn_left->dn_flags & DT_NF_USERLAND);
1777
1778			instr = DIF_INSTR_LOAD(dt_cg_load(dnp,
1779			    ctfp, m.ctm_type), dnp->dn_left->dn_reg,
1780			    dnp->dn_left->dn_reg);
1781
1782			dnp->dn_flags &= ~DT_NF_USERLAND;
1783			dnp->dn_flags |= ubit;
1784
1785			dt_irlist_append(dlp,
1786			    dt_cg_node_alloc(DT_LBL_NONE, instr));
1787
1788			if (dnp->dn_flags & DT_NF_BITFIELD)
1789				dt_cg_field_get(dnp, dlp, drp, ctfp, &m);
1790		}
1791
1792		dnp->dn_reg = dnp->dn_left->dn_reg;
1793		break;
1794
1795	case DT_TOK_STRING:
1796		dnp->dn_reg = dt_regset_alloc(drp);
1797
1798		assert(dnp->dn_kind == DT_NODE_STRING);
1799		stroff = dt_strtab_insert(yypcb->pcb_strtab, dnp->dn_string);
1800
1801		if (stroff == -1L)
1802			longjmp(yypcb->pcb_jmpbuf, EDT_NOMEM);
1803		if (stroff > DIF_STROFF_MAX)
1804			longjmp(yypcb->pcb_jmpbuf, EDT_STR2BIG);
1805
1806		instr = DIF_INSTR_SETS((ulong_t)stroff, dnp->dn_reg);
1807		dt_irlist_append(dlp, dt_cg_node_alloc(DT_LBL_NONE, instr));
1808		break;
1809
1810	case DT_TOK_IDENT:
1811		/*
1812		 * If the specified identifier is a variable on which we have
1813		 * set the code generator register flag, then this variable
1814		 * has already had code generated for it and saved in di_id.
1815		 * Allocate a new register and copy the existing value to it.
1816		 */
1817		if (dnp->dn_kind == DT_NODE_VAR &&
1818		    (dnp->dn_ident->di_flags & DT_IDFLG_CGREG)) {
1819			dnp->dn_reg = dt_regset_alloc(drp);
1820			instr = DIF_INSTR_MOV(dnp->dn_ident->di_id,
1821			    dnp->dn_reg);
1822			dt_irlist_append(dlp,
1823			    dt_cg_node_alloc(DT_LBL_NONE, instr));
1824			break;
1825		}
1826
1827		/*
1828		 * Identifiers can represent function calls, variable refs, or
1829		 * symbols.  First we check for inlined variables, and handle
1830		 * them by generating code for the inline parse tree.
1831		 */
1832		if (dnp->dn_kind == DT_NODE_VAR &&
1833		    (dnp->dn_ident->di_flags & DT_IDFLG_INLINE)) {
1834			dt_cg_inline(dnp, dlp, drp);
1835			break;
1836		}
1837
1838		switch (dnp->dn_kind) {
1839		case DT_NODE_FUNC: {
1840			dtrace_hdl_t *dtp = yypcb->pcb_hdl;
1841
1842			if ((idp = dnp->dn_ident)->di_kind != DT_IDENT_FUNC) {
1843				dnerror(dnp, D_CG_EXPR, "%s %s( ) may not be "
1844				    "called from a D expression (D program "
1845				    "context required)\n",
1846				    dt_idkind_name(idp->di_kind), idp->di_name);
1847			}
1848
1849			switch (idp->di_id) {
1850			case DIF_SUBR_TYPEREF:
1851				dt_cg_func_typeref(dtp, dnp);
1852				break;
1853
1854			default:
1855				break;
1856			}
1857
1858			dt_cg_arglist(dnp->dn_ident, dnp->dn_args, dlp, drp);
1859
1860			dnp->dn_reg = dt_regset_alloc(drp);
1861			instr = DIF_INSTR_CALL(dnp->dn_ident->di_id,
1862			    dnp->dn_reg);
1863
1864			dt_irlist_append(dlp,
1865			    dt_cg_node_alloc(DT_LBL_NONE, instr));
1866
1867			break;
1868		}
1869
1870		case DT_NODE_VAR:
1871			if (dnp->dn_ident->di_kind == DT_IDENT_XLSOU ||
1872			    dnp->dn_ident->di_kind == DT_IDENT_XLPTR) {
1873				/*
1874				 * This can only happen if we have translated
1875				 * args[].  See dt_idcook_args() for details.
1876				 */
1877				assert(dnp->dn_ident->di_id == DIF_VAR_ARGS);
1878				dt_cg_array_op(dnp, dlp, drp);
1879				break;
1880			}
1881
1882			if (dnp->dn_ident->di_kind == DT_IDENT_ARRAY) {
1883				if (dnp->dn_ident->di_id > DIF_VAR_ARRAY_MAX)
1884					dt_cg_assoc_op(dnp, dlp, drp);
1885				else
1886					dt_cg_array_op(dnp, dlp, drp);
1887				break;
1888			}
1889
1890			dnp->dn_reg = dt_regset_alloc(drp);
1891
1892			if (dnp->dn_ident->di_flags & DT_IDFLG_LOCAL)
1893				op = DIF_OP_LDLS;
1894			else if (dnp->dn_ident->di_flags & DT_IDFLG_TLS)
1895				op = DIF_OP_LDTS;
1896			else
1897				op = DIF_OP_LDGS;
1898
1899			dnp->dn_ident->di_flags |= DT_IDFLG_DIFR;
1900
1901			instr = DIF_INSTR_LDV(op,
1902			    dnp->dn_ident->di_id, dnp->dn_reg);
1903
1904			dt_irlist_append(dlp,
1905			    dt_cg_node_alloc(DT_LBL_NONE, instr));
1906			break;
1907
1908		case DT_NODE_SYM: {
1909			dtrace_hdl_t *dtp = yypcb->pcb_hdl;
1910			dtrace_syminfo_t *sip = dnp->dn_ident->di_data;
1911			GElf_Sym sym;
1912
1913			if (dtrace_lookup_by_name(dtp,
1914			    sip->dts_object, sip->dts_name, &sym, NULL) == -1) {
1915				xyerror(D_UNKNOWN, "cg failed for symbol %s`%s:"
1916				    " %s\n", sip->dts_object, sip->dts_name,
1917				    dtrace_errmsg(dtp, dtrace_errno(dtp)));
1918			}
1919
1920			dnp->dn_reg = dt_regset_alloc(drp);
1921			dt_cg_xsetx(dlp, dnp->dn_ident,
1922			    DT_LBL_NONE, dnp->dn_reg, sym.st_value);
1923
1924			if (!(dnp->dn_flags & DT_NF_REF)) {
1925				instr = DIF_INSTR_LOAD(dt_cg_load(dnp, ctfp,
1926				    dnp->dn_type), dnp->dn_reg, dnp->dn_reg);
1927				dt_irlist_append(dlp,
1928				    dt_cg_node_alloc(DT_LBL_NONE, instr));
1929			}
1930			break;
1931		}
1932
1933		default:
1934			xyerror(D_UNKNOWN, "internal error -- node type %u is "
1935			    "not valid for an identifier\n", dnp->dn_kind);
1936		}
1937		break;
1938
1939	case DT_TOK_INT:
1940		dnp->dn_reg = dt_regset_alloc(drp);
1941		dt_cg_setx(dlp, dnp->dn_reg, dnp->dn_value);
1942		break;
1943
1944	default:
1945		xyerror(D_UNKNOWN, "internal error -- token type %u is not a "
1946		    "valid D compilation token\n", dnp->dn_op);
1947	}
1948}
1949
1950void
1951dt_cg(dt_pcb_t *pcb, dt_node_t *dnp)
1952{
1953	dif_instr_t instr;
1954	dt_xlator_t *dxp;
1955	dt_ident_t *idp;
1956
1957	if (pcb->pcb_regs == NULL && (pcb->pcb_regs =
1958	    dt_regset_create(pcb->pcb_hdl->dt_conf.dtc_difintregs)) == NULL)
1959		longjmp(pcb->pcb_jmpbuf, EDT_NOMEM);
1960
1961	dt_regset_reset(pcb->pcb_regs);
1962	(void) dt_regset_alloc(pcb->pcb_regs); /* allocate %r0 */
1963
1964	if (pcb->pcb_inttab != NULL)
1965		dt_inttab_destroy(pcb->pcb_inttab);
1966
1967	if ((pcb->pcb_inttab = dt_inttab_create(yypcb->pcb_hdl)) == NULL)
1968		longjmp(pcb->pcb_jmpbuf, EDT_NOMEM);
1969
1970	if (pcb->pcb_strtab != NULL)
1971		dt_strtab_destroy(pcb->pcb_strtab);
1972
1973	if ((pcb->pcb_strtab = dt_strtab_create(BUFSIZ)) == NULL)
1974		longjmp(pcb->pcb_jmpbuf, EDT_NOMEM);
1975
1976	dt_irlist_destroy(&pcb->pcb_ir);
1977	dt_irlist_create(&pcb->pcb_ir);
1978
1979	assert(pcb->pcb_dret == NULL);
1980	pcb->pcb_dret = dnp;
1981
1982	if (dt_node_resolve(dnp, DT_IDENT_XLPTR) != NULL) {
1983		dnerror(dnp, D_CG_DYN, "expression cannot evaluate to result "
1984		    "of a translated pointer\n");
1985	}
1986
1987	/*
1988	 * If we're generating code for a translator body, assign the input
1989	 * parameter to the first available register (i.e. caller passes %r1).
1990	 */
1991	if (dnp->dn_kind == DT_NODE_MEMBER) {
1992		dxp = dnp->dn_membxlator;
1993		dnp = dnp->dn_membexpr;
1994
1995		dxp->dx_ident->di_flags |= DT_IDFLG_CGREG;
1996		dxp->dx_ident->di_id = dt_regset_alloc(pcb->pcb_regs);
1997	}
1998
1999	dt_cg_node(dnp, &pcb->pcb_ir, pcb->pcb_regs);
2000
2001	if ((idp = dt_node_resolve(dnp, DT_IDENT_XLSOU)) != NULL) {
2002		int reg = dt_cg_xlate_expand(dnp, idp,
2003		    &pcb->pcb_ir, pcb->pcb_regs);
2004		dt_regset_free(pcb->pcb_regs, dnp->dn_reg);
2005		dnp->dn_reg = reg;
2006	}
2007
2008	instr = DIF_INSTR_RET(dnp->dn_reg);
2009	dt_regset_free(pcb->pcb_regs, dnp->dn_reg);
2010	dt_irlist_append(&pcb->pcb_ir, dt_cg_node_alloc(DT_LBL_NONE, instr));
2011
2012	if (dnp->dn_kind == DT_NODE_MEMBER) {
2013		dt_regset_free(pcb->pcb_regs, dxp->dx_ident->di_id);
2014		dxp->dx_ident->di_id = 0;
2015		dxp->dx_ident->di_flags &= ~DT_IDFLG_CGREG;
2016	}
2017
2018	dt_regset_free(pcb->pcb_regs, 0);
2019	dt_regset_assert_free(pcb->pcb_regs);
2020}
2021