| blob | 043ff4df1e4280fa1cbc992d0f2a971cb6e6b4a5 |
1 From cbe53e134d4c3a656880a906738ce19fdcd38e8b Mon Sep 17 00:00:00 2001
2 From: Max Filippov <jcmvbkbc@gmail.com>
3 Date: Fri, 1 May 2015 11:39:12 +0300
4 Subject: [PATCH] xtensa: optimize trampolines relaxation
6 Currently every fixup in the current segment is checked when relaxing
7 trampoline frag. This is very expensive. Make a searchable array of
8 fixups pointing at potentially oversized jumps at the beginning of every
9 relaxation pass and only check subset of this cache in the reach of
10 single jump from the trampoline frag currently being relaxed.
12 Original profile:
14 % time self children called name
15 -----------------------------------------
16 370.16 593.38 12283048/12283048 relax_segment
17 98.4 370.16 593.38 12283048 xtensa_relax_frag
18 58.91 269.26 2691463834/2699602236 xtensa_insnbuf_from_chars
19 68.35 68.17 811266668/813338977 S_GET_VALUE
20 36.85 29.51 2684369246/2685538060 xtensa_opcode_decode
21 28.34 8.84 2684369246/2685538060 xtensa_format_get_slot
22 12.39 5.94 2691463834/2699775044 xtensa_format_decode
23 0.03 4.60 4101109/4101109 relax_frag_for_align
24 0.18 1.76 994617/994617 relax_frag_immed
25 0.07 0.09 24556277/24851220 new_logical_line
26 0.06 0.00 12283048/14067410 as_where
27 0.04 0.00 7094588/15460506 xtensa_format_num_slots
28 0.00 0.00 1/712477 xtensa_insnbuf_alloc
29 -----------------------------------------
31 Same data, after optimization:
33 % time self children called name
34 -----------------------------------------
35 0.51 7.47 12283048/12283048 relax_segment
36 58.0 0.51 7.47 12283048 xtensa_relax_frag
37 0.02 4.08 4101109/4101109 relax_frag_for_align
38 0.18 1.39 994617/994617 relax_frag_immed
39 0.01 0.98 555/555 xtensa_cache_relaxable_fixups
40 0.21 0.25 7094588/16693271 xtensa_insnbuf_from_chars
41 0.06 0.12 24556277/24851220 new_logical_line
42 0.06 0.00 7094588/15460506 xtensa_format_num_slots
43 0.02 0.04 7094588/16866079 xtensa_format_decode
44 0.05 0.00 12283048/14067410 as_where
45 0.00 0.00 1/712477 xtensa_insnbuf_alloc
46 0.00 0.00 93808/93808 xtensa_find_first_cached_fixup
47 -----------------------------------------
49 2015-05-02 Max Filippov <jcmvbkbc@gmail.com>
50 gas/
51 * config/tc-xtensa.c (cached_fixupS, fixup_cacheS): New typedefs.
52 (struct cached_fixup, struct fixup_cache): New structures.
53 (fixup_order, xtensa_make_cached_fixup),
54 (xtensa_realloc_fixup_cache, xtensa_cache_relaxable_fixups),
55 (xtensa_find_first_cached_fixup, xtensa_delete_cached_fixup),
56 (xtensa_add_cached_fixup): New functions.
57 (xtensa_relax_frag): Cache fixups pointing at potentially
58 oversized jumps at the beginning of every relaxation pass. Only
59 check subset of this cache in the reach of single jump from the
60 trampoline frag currently being relaxed.
62 Signed-off-by: Max Filippov <jcmvbkbc@gmail.com>
63 ---
64 Backported from: b76f99d702c3501ac320396ea06bc7f9237173c3
65 Changes to ChangeLog are dropped.
67 gas/config/tc-xtensa.c | 220 +++++++++++++++++++++++++++++++++++++++++++------
68 1 file changed, 194 insertions(+), 26 deletions(-)
70 diff --git a/gas/config/tc-xtensa.c b/gas/config/tc-xtensa.c
71 index 3e85b69..31c0b6b 100644
72 --- a/gas/config/tc-xtensa.c
73 +++ b/gas/config/tc-xtensa.c
74 @@ -8785,6 +8785,154 @@ static long relax_frag_for_align (fragS *, long);
75 static long relax_frag_immed
76 (segT, fragS *, long, int, xtensa_format, int, int *, bfd_boolean);
78 +typedef struct cached_fixup cached_fixupS;
79 +struct cached_fixup
80 +{
81 + int addr;
82 + int target;
83 + int delta;
84 + fixS *fixP;
85 +};
86 +
87 +typedef struct fixup_cache fixup_cacheS;
88 +struct fixup_cache
89 +{
90 + cached_fixupS *fixups;
91 + unsigned n_fixups;
92 + unsigned n_max;
93 +
94 + segT seg;
95 + fragS *first_frag;
96 +};
97 +
98 +static int fixup_order (const void *a, const void *b)
99 +{
100 + const cached_fixupS *pa = a;
101 + const cached_fixupS *pb = b;
102 +
103 + if (pa->addr == pb->addr)
104 + {
105 + if (pa->target == pb->target)
106 + {
107 + if (pa->fixP->fx_r_type == pb->fixP->fx_r_type)
108 + return 0;
109 + return pa->fixP->fx_r_type < pb->fixP->fx_r_type ? -1 : 1;
110 + }
111 + return pa->target - pb->target;
112 + }
113 + return pa->addr - pb->addr;
114 +}
115 +
116 +static bfd_boolean xtensa_make_cached_fixup (cached_fixupS *o, fixS *fixP)
117 +{
118 + xtensa_isa isa = xtensa_default_isa;
119 + int addr = fixP->fx_frag->fr_address;
120 + int target;
121 + int delta;
122 + symbolS *s = fixP->fx_addsy;
123 + int slot;
124 + xtensa_format fmt;
125 + xtensa_opcode opcode;
126 +
127 + if (fixP->fx_r_type < BFD_RELOC_XTENSA_SLOT0_OP ||
128 + fixP->fx_r_type > BFD_RELOC_XTENSA_SLOT14_OP)
129 + return FALSE;
130 + target = S_GET_VALUE (s);
131 + delta = target - addr;
132 +
133 + if (abs(delta) < J_RANGE / 2)
134 + return FALSE;
135 +
136 + xtensa_insnbuf_from_chars (isa, trampoline_buf,
137 + (unsigned char *) fixP->fx_frag->fr_literal +
138 + fixP->fx_where, 0);
139 + fmt = xtensa_format_decode (isa, trampoline_buf);
140 + gas_assert (fmt != XTENSA_UNDEFINED);
141 + slot = fixP->tc_fix_data.slot;
142 + xtensa_format_get_slot (isa, fmt, slot, trampoline_buf, trampoline_slotbuf);
143 + opcode = xtensa_opcode_decode (isa, fmt, slot, trampoline_slotbuf);
144 + if (opcode != xtensa_j_opcode)
145 + return FALSE;
146 +
147 + o->addr = addr;
148 + o->target = target;
149 + o->delta = delta;
150 + o->fixP = fixP;
151 +
152 + return TRUE;
153 +}
154 +
155 +static void xtensa_realloc_fixup_cache (fixup_cacheS *cache, unsigned add)
156 +{
157 + if (cache->n_fixups + add > cache->n_max)
158 + {
159 + cache->n_max = (cache->n_fixups + add) * 2;
160 + cache->fixups = xrealloc (cache->fixups,
161 + sizeof (*cache->fixups) * cache->n_max);
162 + }
163 +}
164 +
165 +static void xtensa_cache_relaxable_fixups (fixup_cacheS *cache,
166 + segment_info_type *seginfo)
167 +{
168 + fixS *fixP;
169 +
170 + cache->n_fixups = 0;
171 +
172 + for (fixP = seginfo->fix_root; fixP ; fixP = fixP->fx_next)
173 + {
174 + xtensa_realloc_fixup_cache (cache, 1);
175 +
176 + if (xtensa_make_cached_fixup (cache->fixups + cache->n_fixups, fixP))
177 + ++cache->n_fixups;
178 + }
179 + qsort (cache->fixups, cache->n_fixups, sizeof (*cache->fixups), fixup_order);
180 +}
181 +
182 +static unsigned xtensa_find_first_cached_fixup (const fixup_cacheS *cache,
183 + int addr)
184 +{
185 + unsigned a = 0;
186 + unsigned b = cache->n_fixups;
187 +
188 + while (b - a > 1)
189 + {
190 + unsigned c = (a + b) / 2;
191 +
192 + if (cache->fixups[c].addr < addr)
193 + a = c;
194 + else
195 + b = c;
196 + }
197 + return a;
198 +}
199 +
200 +static void xtensa_delete_cached_fixup (fixup_cacheS *cache, unsigned i)
201 +{
202 + memmove (cache->fixups + i, cache->fixups + i + 1,
203 + (cache->n_fixups - i - 1) * sizeof (*cache->fixups));
204 + --cache->n_fixups;
205 +}
206 +
207 +static bfd_boolean xtensa_add_cached_fixup (fixup_cacheS *cache, fixS *fixP)
208 +{
209 + cached_fixupS o;
210 + unsigned i;
211 +
212 + if (!xtensa_make_cached_fixup (&o, fixP))
213 + return FALSE;
214 + xtensa_realloc_fixup_cache (cache, 1);
215 + i = xtensa_find_first_cached_fixup (cache, o.addr);
216 + if (i < cache->n_fixups)
217 + {
218 + ++i;
219 + memmove (cache->fixups + i + 1, cache->fixups + i,
220 + (cache->n_fixups - i) * sizeof (*cache->fixups));
221 + }
222 + cache->fixups[i] = o;
223 + ++cache->n_fixups;
224 + return TRUE;
225 +}
227 /* Return the number of bytes added to this fragment, given that the
228 input has been stretched already by "stretch". */
229 @@ -8896,35 +9044,42 @@ xtensa_relax_frag (fragS *fragP, long stretch, int *stretched_p)
230 case RELAX_TRAMPOLINE:
231 if (fragP->tc_frag_data.relax_seen)
232 {
233 - segment_info_type *seginfo = seg_info (now_seg);
234 - fragS *fP; /* The out-of-range jump. */
235 - fixS *fixP;
236 + static fixup_cacheS fixup_cache;
237 + segment_info_type *seginfo = seg_info (now_seg);
238 + int trampaddr = fragP->fr_address + fragP->fr_fix;
239 + int searchaddr = trampaddr < J_RANGE ? 0 : trampaddr - J_RANGE;
240 + unsigned i;
241 +
242 + if (now_seg != fixup_cache.seg ||
243 + fragP == fixup_cache.first_frag ||
244 + fixup_cache.first_frag == NULL)
245 + {
246 + xtensa_cache_relaxable_fixups (&fixup_cache, seginfo);
247 + fixup_cache.seg = now_seg;
248 + fixup_cache.first_frag = fragP;
249 + }
251 /* Scan for jumps that will not reach. */
252 - for (fixP = seginfo->fix_root; fixP ; fixP = fixP->fx_next)
253 + for (i = xtensa_find_first_cached_fixup (&fixup_cache, searchaddr);
254 + i < fixup_cache.n_fixups; ++i)
255 +
256 {
257 - symbolS *s = fixP->fx_addsy;
258 - xtensa_opcode opcode;
259 - int target;
260 - int addr;
261 - int delta;
262 -
263 - if (fixP->fx_r_type < BFD_RELOC_XTENSA_SLOT0_OP ||
264 - fixP->fx_r_type > BFD_RELOC_XTENSA_SLOT14_OP)
265 - continue;
266 - xtensa_insnbuf_from_chars (isa, trampoline_buf,
267 - (unsigned char *) fixP->fx_frag->fr_literal + fixP->fx_where,
268 - 0);
269 - fmt = xtensa_format_decode (isa, trampoline_buf);
270 - gas_assert (fmt != XTENSA_UNDEFINED);
271 - slot = fixP->tc_fix_data.slot;
272 - xtensa_format_get_slot (isa, fmt, slot, trampoline_buf, trampoline_slotbuf);
273 - opcode = xtensa_opcode_decode (isa, fmt, slot, trampoline_slotbuf);
274 - if (opcode != xtensa_j_opcode)
275 + fixS *fixP = fixup_cache.fixups[i].fixP;
276 + int target = fixup_cache.fixups[i].target;
277 + int addr = fixup_cache.fixups[i].addr;
278 + int delta = fixup_cache.fixups[i].delta + stretch;
279 +
280 + trampaddr = fragP->fr_address + fragP->fr_fix;
281 +
282 + if (addr + J_RANGE < trampaddr)
283 continue;
284 - target = S_GET_VALUE (s);
285 - addr = fixP->fx_frag->fr_address;
286 - delta = target - addr + stretch;
287 + if (addr > trampaddr + J_RANGE)
288 + break;
289 + if (abs (delta) < J_RANGE)
290 + continue;
291 +
292 + slot = fixP->tc_fix_data.slot;
293 +
294 if (delta > J_RANGE || delta < -1 * J_RANGE)
295 { /* Found an out-of-range jump; scan the list of trampolines for the best match. */
296 struct trampoline_seg *ts = find_trampoline_seg (now_seg);
297 @@ -8978,14 +9133,13 @@ xtensa_relax_frag (fragS *fragP, long stretch, int *stretched_p)
298 }
299 if (tf->fragP == fragP)
300 {
301 - int trampaddr = fragP->fr_address + fragP->fr_fix;
302 -
303 if (abs (addr - trampaddr) < J_RANGE)
304 { /* The trampoline is in range of original; fix it! */
305 fixS *newfixP;
306 int offset;
307 TInsn insn;
308 symbolS *lsym;
309 + fragS *fP; /* The out-of-range jump. */
311 new_stretch += init_trampoline_frag (tf);
312 offset = fragP->fr_fix; /* Where to assemble the j insn. */
313 @@ -9009,10 +9163,20 @@ xtensa_relax_frag (fragS *fragP, long stretch, int *stretched_p)
314 newfixP->tc_fix_data.X_add_symbol = lsym;
315 newfixP->tc_fix_data.X_add_number = offset;
316 newfixP->tc_fix_data.slot = slot;
317 +
318 + xtensa_delete_cached_fixup (&fixup_cache, i);
319 + xtensa_add_cached_fixup (&fixup_cache, newfixP);
320 +
321 /* Move the fix-up from the original j insn to this one. */
322 fixP->fx_frag = fragP;
323 fixP->fx_where = fragP->fr_fix - 3;
324 fixP->tc_fix_data.slot = 0;
325 +
326 + xtensa_add_cached_fixup (&fixup_cache, fixP);
327 +
328 + /* re-do current fixup */
329 + --i;
330 +
331 /* Adjust the jump around this trampoline (if present). */
332 if (tf->fixP != NULL)
333 {
334 @@ -9027,6 +9191,8 @@ xtensa_relax_frag (fragS *fragP, long stretch, int *stretched_p)
335 fragP->fr_subtype = 0;
336 /* Remove from the trampoline_list. */
337 prev->next = tf->next;
338 + if (fragP == fixup_cache.first_frag)
339 + fixup_cache.first_frag = NULL;
340 break;
341 }
342 }
343 --
344 1.8.1.4