1 ; RUN: llc -mtriple=x86_64-unknown-unknown -mattr=+avx < %s | FileCheck %s
2 ; RUN: llc -mtriple=x86_64-unknown-unknown -mattr=+avx -addr-sink-using-gep=1 < %s | FileCheck %s
4 %struct.A = type { i8, i8, i8, i8, i8, i8, i8, i8 }
5 %struct.B = type { i32, i32, i32, i32, i32, i32, i32, i32 }
7 ; CHECK-LABEL: merge_const_store:
8 ; save 1,2,3 ... as one big integer.
9 ; CHECK: movabsq $578437695752307201
11 define void @merge_const_store(i32 %count, %struct.A* nocapture %p) nounwind uwtable noinline ssp {
12 %1 = icmp sgt i32 %count, 0
13 br i1 %1, label %.lr.ph, label %._crit_edge
15 %i.02 = phi i32 [ %10, %.lr.ph ], [ 0, %0 ]
16 %.01 = phi %struct.A* [ %11, %.lr.ph ], [ %p, %0 ]
17 %2 = getelementptr inbounds %struct.A, %struct.A* %.01, i64 0, i32 0
18 store i8 1, i8* %2, align 1
19 %3 = getelementptr inbounds %struct.A, %struct.A* %.01, i64 0, i32 1
20 store i8 2, i8* %3, align 1
21 %4 = getelementptr inbounds %struct.A, %struct.A* %.01, i64 0, i32 2
22 store i8 3, i8* %4, align 1
23 %5 = getelementptr inbounds %struct.A, %struct.A* %.01, i64 0, i32 3
24 store i8 4, i8* %5, align 1
25 %6 = getelementptr inbounds %struct.A, %struct.A* %.01, i64 0, i32 4
26 store i8 5, i8* %6, align 1
27 %7 = getelementptr inbounds %struct.A, %struct.A* %.01, i64 0, i32 5
28 store i8 6, i8* %7, align 1
29 %8 = getelementptr inbounds %struct.A, %struct.A* %.01, i64 0, i32 6
30 store i8 7, i8* %8, align 1
31 %9 = getelementptr inbounds %struct.A, %struct.A* %.01, i64 0, i32 7
32 store i8 8, i8* %9, align 1
33 %10 = add nsw i32 %i.02, 1
34 %11 = getelementptr inbounds %struct.A, %struct.A* %.01, i64 1
35 %exitcond = icmp eq i32 %10, %count
36 br i1 %exitcond, label %._crit_edge, label %.lr.ph
41 ; No vectors because we use noimplicitfloat
42 ; CHECK-LABEL: merge_const_store_no_vec:
45 define void @merge_const_store_no_vec(i32 %count, %struct.B* nocapture %p) noimplicitfloat{
46 %1 = icmp sgt i32 %count, 0
47 br i1 %1, label %.lr.ph, label %._crit_edge
49 %i.02 = phi i32 [ %10, %.lr.ph ], [ 0, %0 ]
50 %.01 = phi %struct.B* [ %11, %.lr.ph ], [ %p, %0 ]
51 %2 = getelementptr inbounds %struct.B, %struct.B* %.01, i64 0, i32 0
52 store i32 0, i32* %2, align 4
53 %3 = getelementptr inbounds %struct.B, %struct.B* %.01, i64 0, i32 1
54 store i32 0, i32* %3, align 4
55 %4 = getelementptr inbounds %struct.B, %struct.B* %.01, i64 0, i32 2
56 store i32 0, i32* %4, align 4
57 %5 = getelementptr inbounds %struct.B, %struct.B* %.01, i64 0, i32 3
58 store i32 0, i32* %5, align 4
59 %6 = getelementptr inbounds %struct.B, %struct.B* %.01, i64 0, i32 4
60 store i32 0, i32* %6, align 4
61 %7 = getelementptr inbounds %struct.B, %struct.B* %.01, i64 0, i32 5
62 store i32 0, i32* %7, align 4
63 %8 = getelementptr inbounds %struct.B, %struct.B* %.01, i64 0, i32 6
64 store i32 0, i32* %8, align 4
65 %9 = getelementptr inbounds %struct.B, %struct.B* %.01, i64 0, i32 7
66 store i32 0, i32* %9, align 4
67 %10 = add nsw i32 %i.02, 1
68 %11 = getelementptr inbounds %struct.B, %struct.B* %.01, i64 1
69 %exitcond = icmp eq i32 %10, %count
70 br i1 %exitcond, label %._crit_edge, label %.lr.ph
75 ; Move the constants using a single vector store.
76 ; CHECK-LABEL: merge_const_store_vec:
79 define void @merge_const_store_vec(i32 %count, %struct.B* nocapture %p) nounwind uwtable noinline ssp {
80 %1 = icmp sgt i32 %count, 0
81 br i1 %1, label %.lr.ph, label %._crit_edge
83 %i.02 = phi i32 [ %10, %.lr.ph ], [ 0, %0 ]
84 %.01 = phi %struct.B* [ %11, %.lr.ph ], [ %p, %0 ]
85 %2 = getelementptr inbounds %struct.B, %struct.B* %.01, i64 0, i32 0
86 store i32 0, i32* %2, align 4
87 %3 = getelementptr inbounds %struct.B, %struct.B* %.01, i64 0, i32 1
88 store i32 0, i32* %3, align 4
89 %4 = getelementptr inbounds %struct.B, %struct.B* %.01, i64 0, i32 2
90 store i32 0, i32* %4, align 4
91 %5 = getelementptr inbounds %struct.B, %struct.B* %.01, i64 0, i32 3
92 store i32 0, i32* %5, align 4
93 %6 = getelementptr inbounds %struct.B, %struct.B* %.01, i64 0, i32 4
94 store i32 0, i32* %6, align 4
95 %7 = getelementptr inbounds %struct.B, %struct.B* %.01, i64 0, i32 5
96 store i32 0, i32* %7, align 4
97 %8 = getelementptr inbounds %struct.B, %struct.B* %.01, i64 0, i32 6
98 store i32 0, i32* %8, align 4
99 %9 = getelementptr inbounds %struct.B, %struct.B* %.01, i64 0, i32 7
100 store i32 0, i32* %9, align 4
101 %10 = add nsw i32 %i.02, 1
102 %11 = getelementptr inbounds %struct.B, %struct.B* %.01, i64 1
103 %exitcond = icmp eq i32 %10, %count
104 br i1 %exitcond, label %._crit_edge, label %.lr.ph
109 ; Move the first 4 constants as a single vector. Move the rest as scalars.
110 ; CHECK-LABEL: merge_nonconst_store:
111 ; CHECK: movl $67305985
117 define void @merge_nonconst_store(i32 %count, i8 %zz, %struct.A* nocapture %p) nounwind uwtable noinline ssp {
118 %1 = icmp sgt i32 %count, 0
119 br i1 %1, label %.lr.ph, label %._crit_edge
121 %i.02 = phi i32 [ %10, %.lr.ph ], [ 0, %0 ]
122 %.01 = phi %struct.A* [ %11, %.lr.ph ], [ %p, %0 ]
123 %2 = getelementptr inbounds %struct.A, %struct.A* %.01, i64 0, i32 0
124 store i8 1, i8* %2, align 1
125 %3 = getelementptr inbounds %struct.A, %struct.A* %.01, i64 0, i32 1
126 store i8 2, i8* %3, align 1
127 %4 = getelementptr inbounds %struct.A, %struct.A* %.01, i64 0, i32 2
128 store i8 3, i8* %4, align 1
129 %5 = getelementptr inbounds %struct.A, %struct.A* %.01, i64 0, i32 3
130 store i8 4, i8* %5, align 1
131 %6 = getelementptr inbounds %struct.A, %struct.A* %.01, i64 0, i32 4
132 store i8 %zz, i8* %6, align 1 ; <----------- Not a const;
133 %7 = getelementptr inbounds %struct.A, %struct.A* %.01, i64 0, i32 5
134 store i8 6, i8* %7, align 1
135 %8 = getelementptr inbounds %struct.A, %struct.A* %.01, i64 0, i32 6
136 store i8 7, i8* %8, align 1
137 %9 = getelementptr inbounds %struct.A, %struct.A* %.01, i64 0, i32 7
138 store i8 8, i8* %9, align 1
139 %10 = add nsw i32 %i.02, 1
140 %11 = getelementptr inbounds %struct.A, %struct.A* %.01, i64 1
141 %exitcond = icmp eq i32 %10, %count
142 br i1 %exitcond, label %._crit_edge, label %.lr.ph
148 ; CHECK-LABEL: merge_loads_i16:
154 define void @merge_loads_i16(i32 %count, %struct.A* noalias nocapture %q, %struct.A* noalias nocapture %p) nounwind uwtable noinline ssp {
155 %1 = icmp sgt i32 %count, 0
156 br i1 %1, label %.lr.ph, label %._crit_edge
159 %2 = getelementptr inbounds %struct.A, %struct.A* %q, i64 0, i32 0
160 %3 = getelementptr inbounds %struct.A, %struct.A* %q, i64 0, i32 1
163 ; <label>:4 ; preds = %4, %.lr.ph
164 %i.02 = phi i32 [ 0, %.lr.ph ], [ %9, %4 ]
165 %.01 = phi %struct.A* [ %p, %.lr.ph ], [ %10, %4 ]
166 %5 = load i8, i8* %2, align 1
167 %6 = load i8, i8* %3, align 1
168 %7 = getelementptr inbounds %struct.A, %struct.A* %.01, i64 0, i32 0
169 store i8 %5, i8* %7, align 1
170 %8 = getelementptr inbounds %struct.A, %struct.A* %.01, i64 0, i32 1
171 store i8 %6, i8* %8, align 1
172 %9 = add nsw i32 %i.02, 1
173 %10 = getelementptr inbounds %struct.A, %struct.A* %.01, i64 1
174 %exitcond = icmp eq i32 %9, %count
175 br i1 %exitcond, label %._crit_edge, label %4
177 ._crit_edge: ; preds = %4, %0
181 ; The loads and the stores are interleaved. Can't merge them.
182 ; CHECK-LABEL: no_merge_loads:
188 define void @no_merge_loads(i32 %count, %struct.A* noalias nocapture %q, %struct.A* noalias nocapture %p) nounwind uwtable noinline ssp {
189 %1 = icmp sgt i32 %count, 0
190 br i1 %1, label %.lr.ph, label %._crit_edge
193 %2 = getelementptr inbounds %struct.A, %struct.A* %q, i64 0, i32 0
194 %3 = getelementptr inbounds %struct.A, %struct.A* %q, i64 0, i32 1
197 a4: ; preds = %4, %.lr.ph
198 %i.02 = phi i32 [ 0, %.lr.ph ], [ %a9, %a4 ]
199 %.01 = phi %struct.A* [ %p, %.lr.ph ], [ %a10, %a4 ]
200 %a5 = load i8, i8* %2, align 1
201 %a7 = getelementptr inbounds %struct.A, %struct.A* %.01, i64 0, i32 0
202 store i8 %a5, i8* %a7, align 1
203 %a8 = getelementptr inbounds %struct.A, %struct.A* %.01, i64 0, i32 1
204 %a6 = load i8, i8* %3, align 1
205 store i8 %a6, i8* %a8, align 1
206 %a9 = add nsw i32 %i.02, 1
207 %a10 = getelementptr inbounds %struct.A, %struct.A* %.01, i64 1
208 %exitcond = icmp eq i32 %a9, %count
209 br i1 %exitcond, label %._crit_edge, label %a4
211 ._crit_edge: ; preds = %4, %0
216 ; CHECK-LABEL: merge_loads_integer:
222 define void @merge_loads_integer(i32 %count, %struct.B* noalias nocapture %q, %struct.B* noalias nocapture %p) nounwind uwtable noinline ssp {
223 %1 = icmp sgt i32 %count, 0
224 br i1 %1, label %.lr.ph, label %._crit_edge
227 %2 = getelementptr inbounds %struct.B, %struct.B* %q, i64 0, i32 0
228 %3 = getelementptr inbounds %struct.B, %struct.B* %q, i64 0, i32 1
231 ; <label>:4 ; preds = %4, %.lr.ph
232 %i.02 = phi i32 [ 0, %.lr.ph ], [ %9, %4 ]
233 %.01 = phi %struct.B* [ %p, %.lr.ph ], [ %10, %4 ]
234 %5 = load i32, i32* %2
235 %6 = load i32, i32* %3
236 %7 = getelementptr inbounds %struct.B, %struct.B* %.01, i64 0, i32 0
237 store i32 %5, i32* %7
238 %8 = getelementptr inbounds %struct.B, %struct.B* %.01, i64 0, i32 1
239 store i32 %6, i32* %8
240 %9 = add nsw i32 %i.02, 1
241 %10 = getelementptr inbounds %struct.B, %struct.B* %.01, i64 1
242 %exitcond = icmp eq i32 %9, %count
243 br i1 %exitcond, label %._crit_edge, label %4
245 ._crit_edge: ; preds = %4, %0
250 ; CHECK-LABEL: merge_loads_vector:
256 define void @merge_loads_vector(i32 %count, %struct.B* noalias nocapture %q, %struct.B* noalias nocapture %p) nounwind uwtable noinline ssp {
257 %a1 = icmp sgt i32 %count, 0
258 br i1 %a1, label %.lr.ph, label %._crit_edge
261 %a2 = getelementptr inbounds %struct.B, %struct.B* %q, i64 0, i32 0
262 %a3 = getelementptr inbounds %struct.B, %struct.B* %q, i64 0, i32 1
263 %a4 = getelementptr inbounds %struct.B, %struct.B* %q, i64 0, i32 2
264 %a5 = getelementptr inbounds %struct.B, %struct.B* %q, i64 0, i32 3
267 block4: ; preds = %4, %.lr.ph
268 %i.02 = phi i32 [ 0, %.lr.ph ], [ %c9, %block4 ]
269 %.01 = phi %struct.B* [ %p, %.lr.ph ], [ %c10, %block4 ]
270 %a7 = getelementptr inbounds %struct.B, %struct.B* %.01, i64 0, i32 0
271 %a8 = getelementptr inbounds %struct.B, %struct.B* %.01, i64 0, i32 1
272 %a9 = getelementptr inbounds %struct.B, %struct.B* %.01, i64 0, i32 2
273 %a10 = getelementptr inbounds %struct.B, %struct.B* %.01, i64 0, i32 3
274 %b1 = load i32, i32* %a2
275 %b2 = load i32, i32* %a3
276 %b3 = load i32, i32* %a4
277 %b4 = load i32, i32* %a5
278 store i32 %b1, i32* %a7
279 store i32 %b2, i32* %a8
280 store i32 %b3, i32* %a9
281 store i32 %b4, i32* %a10
282 %c9 = add nsw i32 %i.02, 1
283 %c10 = getelementptr inbounds %struct.B, %struct.B* %.01, i64 1
284 %exitcond = icmp eq i32 %c9, %count
285 br i1 %exitcond, label %._crit_edge, label %block4
287 ._crit_edge: ; preds = %4, %0
291 ;; On x86, even unaligned copies can be merged to vector ops.
292 ; CHECK-LABEL: merge_loads_no_align:
298 define void @merge_loads_no_align(i32 %count, %struct.B* noalias nocapture %q, %struct.B* noalias nocapture %p) nounwind uwtable noinline ssp {
299 %a1 = icmp sgt i32 %count, 0
300 br i1 %a1, label %.lr.ph, label %._crit_edge
303 %a2 = getelementptr inbounds %struct.B, %struct.B* %q, i64 0, i32 0
304 %a3 = getelementptr inbounds %struct.B, %struct.B* %q, i64 0, i32 1
305 %a4 = getelementptr inbounds %struct.B, %struct.B* %q, i64 0, i32 2
306 %a5 = getelementptr inbounds %struct.B, %struct.B* %q, i64 0, i32 3
309 block4: ; preds = %4, %.lr.ph
310 %i.02 = phi i32 [ 0, %.lr.ph ], [ %c9, %block4 ]
311 %.01 = phi %struct.B* [ %p, %.lr.ph ], [ %c10, %block4 ]
312 %a7 = getelementptr inbounds %struct.B, %struct.B* %.01, i64 0, i32 0
313 %a8 = getelementptr inbounds %struct.B, %struct.B* %.01, i64 0, i32 1
314 %a9 = getelementptr inbounds %struct.B, %struct.B* %.01, i64 0, i32 2
315 %a10 = getelementptr inbounds %struct.B, %struct.B* %.01, i64 0, i32 3
316 %b1 = load i32, i32* %a2, align 1
317 %b2 = load i32, i32* %a3, align 1
318 %b3 = load i32, i32* %a4, align 1
319 %b4 = load i32, i32* %a5, align 1
320 store i32 %b1, i32* %a7, align 1
321 store i32 %b2, i32* %a8, align 1
322 store i32 %b3, i32* %a9, align 1
323 store i32 %b4, i32* %a10, align 1
324 %c9 = add nsw i32 %i.02, 1
325 %c10 = getelementptr inbounds %struct.B, %struct.B* %.01, i64 1
326 %exitcond = icmp eq i32 %c9, %count
327 br i1 %exitcond, label %._crit_edge, label %block4
329 ._crit_edge: ; preds = %4, %0
333 ; Make sure that we merge the consecutive load/store sequence below and use a
334 ; word (16 bit) instead of a byte copy.
335 ; CHECK-LABEL: MergeLoadStoreBaseIndexOffset:
336 ; CHECK: movw (%{{.*}},%{{.*}}), [[REG:%[a-z]+]]
337 ; CHECK: movw [[REG]], (%{{.*}})
338 define void @MergeLoadStoreBaseIndexOffset(i64* %a, i8* %b, i8* %c, i32 %n) {
342 %.09 = phi i32 [ %n, %0 ], [ %11, %1 ]
343 %.08 = phi i8* [ %b, %0 ], [ %10, %1 ]
344 %.0 = phi i64* [ %a, %0 ], [ %2, %1 ]
345 %2 = getelementptr inbounds i64, i64* %.0, i64 1
346 %3 = load i64, i64* %.0, align 1
347 %4 = getelementptr inbounds i8, i8* %c, i64 %3
348 %5 = load i8, i8* %4, align 1
350 %7 = getelementptr inbounds i8, i8* %c, i64 %6
351 %8 = load i8, i8* %7, align 1
352 store i8 %5, i8* %.08, align 1
353 %9 = getelementptr inbounds i8, i8* %.08, i64 1
354 store i8 %8, i8* %9, align 1
355 %10 = getelementptr inbounds i8, i8* %.08, i64 2
356 %11 = add nsw i32 %.09, -1
357 %12 = icmp eq i32 %11, 0
358 br i1 %12, label %13, label %1
364 ; Make sure that we merge the consecutive load/store sequence below and use a
365 ; word (16 bit) instead of a byte copy even if there are intermediate sign
367 ; CHECK-LABEL: MergeLoadStoreBaseIndexOffsetSext:
368 ; CHECK: movw (%{{.*}},%{{.*}}), [[REG:%[a-z]+]]
369 ; CHECK: movw [[REG]], (%{{.*}})
370 define void @MergeLoadStoreBaseIndexOffsetSext(i8* %a, i8* %b, i8* %c, i32 %n) {
374 %.09 = phi i32 [ %n, %0 ], [ %12, %1 ]
375 %.08 = phi i8* [ %b, %0 ], [ %11, %1 ]
376 %.0 = phi i8* [ %a, %0 ], [ %2, %1 ]
377 %2 = getelementptr inbounds i8, i8* %.0, i64 1
378 %3 = load i8, i8* %.0, align 1
379 %4 = sext i8 %3 to i64
380 %5 = getelementptr inbounds i8, i8* %c, i64 %4
381 %6 = load i8, i8* %5, align 1
383 %8 = getelementptr inbounds i8, i8* %c, i64 %7
384 %9 = load i8, i8* %8, align 1
385 store i8 %6, i8* %.08, align 1
386 %10 = getelementptr inbounds i8, i8* %.08, i64 1
387 store i8 %9, i8* %10, align 1
388 %11 = getelementptr inbounds i8, i8* %.08, i64 2
389 %12 = add nsw i32 %.09, -1
390 %13 = icmp eq i32 %12, 0
391 br i1 %13, label %14, label %1
397 ; However, we can only merge ignore sign extensions when they are on all memory
399 ; CHECK-LABEL: loadStoreBaseIndexOffsetSextNoSex:
400 ; CHECK-NOT: movw (%{{.*}},%{{.*}}), [[REG:%[a-z]+]]
401 ; CHECK-NOT: movw [[REG]], (%{{.*}})
402 define void @loadStoreBaseIndexOffsetSextNoSex(i8* %a, i8* %b, i8* %c, i32 %n) {
406 %.09 = phi i32 [ %n, %0 ], [ %12, %1 ]
407 %.08 = phi i8* [ %b, %0 ], [ %11, %1 ]
408 %.0 = phi i8* [ %a, %0 ], [ %2, %1 ]
409 %2 = getelementptr inbounds i8, i8* %.0, i64 1
410 %3 = load i8, i8* %.0, align 1
411 %4 = sext i8 %3 to i64
412 %5 = getelementptr inbounds i8, i8* %c, i64 %4
413 %6 = load i8, i8* %5, align 1
415 %wrap.4 = sext i8 %7 to i64
416 %8 = getelementptr inbounds i8, i8* %c, i64 %wrap.4
417 %9 = load i8, i8* %8, align 1
418 store i8 %6, i8* %.08, align 1
419 %10 = getelementptr inbounds i8, i8* %.08, i64 1
420 store i8 %9, i8* %10, align 1
421 %11 = getelementptr inbounds i8, i8* %.08, i64 2
422 %12 = add nsw i32 %.09, -1
423 %13 = icmp eq i32 %12, 0
424 br i1 %13, label %14, label %1
430 ; PR21711 ( http://llvm.org/bugs/show_bug.cgi?id=21711 )
431 define void @merge_vec_element_store(<8 x float> %v, float* %ptr) {
432 %vecext0 = extractelement <8 x float> %v, i32 0
433 %vecext1 = extractelement <8 x float> %v, i32 1
434 %vecext2 = extractelement <8 x float> %v, i32 2
435 %vecext3 = extractelement <8 x float> %v, i32 3
436 %vecext4 = extractelement <8 x float> %v, i32 4
437 %vecext5 = extractelement <8 x float> %v, i32 5
438 %vecext6 = extractelement <8 x float> %v, i32 6
439 %vecext7 = extractelement <8 x float> %v, i32 7
440 %arrayidx1 = getelementptr inbounds float, float* %ptr, i64 1
441 %arrayidx2 = getelementptr inbounds float, float* %ptr, i64 2
442 %arrayidx3 = getelementptr inbounds float, float* %ptr, i64 3
443 %arrayidx4 = getelementptr inbounds float, float* %ptr, i64 4
444 %arrayidx5 = getelementptr inbounds float, float* %ptr, i64 5
445 %arrayidx6 = getelementptr inbounds float, float* %ptr, i64 6
446 %arrayidx7 = getelementptr inbounds float, float* %ptr, i64 7
447 store float %vecext0, float* %ptr, align 4
448 store float %vecext1, float* %arrayidx1, align 4
449 store float %vecext2, float* %arrayidx2, align 4
450 store float %vecext3, float* %arrayidx3, align 4
451 store float %vecext4, float* %arrayidx4, align 4
452 store float %vecext5, float* %arrayidx5, align 4
453 store float %vecext6, float* %arrayidx6, align 4
454 store float %vecext7, float* %arrayidx7, align 4
457 ; CHECK-LABEL: merge_vec_element_store
459 ; CHECK-NEXT: vzeroupper
463 ; PR21711 - Merge vector stores into wider vector stores.
464 ; These should be merged into 32-byte stores.
465 define void @merge_vec_extract_stores(<8 x float> %v1, <8 x float> %v2, <4 x float>* %ptr) {
466 %idx0 = getelementptr inbounds <4 x float>, <4 x float>* %ptr, i64 3
467 %idx1 = getelementptr inbounds <4 x float>, <4 x float>* %ptr, i64 4
468 %idx2 = getelementptr inbounds <4 x float>, <4 x float>* %ptr, i64 5
469 %idx3 = getelementptr inbounds <4 x float>, <4 x float>* %ptr, i64 6
470 %shuffle0 = shufflevector <8 x float> %v1, <8 x float> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
471 %shuffle1 = shufflevector <8 x float> %v1, <8 x float> undef, <4 x i32> <i32 4, i32 5, i32 6, i32 7>
472 %shuffle2 = shufflevector <8 x float> %v2, <8 x float> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
473 %shuffle3 = shufflevector <8 x float> %v2, <8 x float> undef, <4 x i32> <i32 4, i32 5, i32 6, i32 7>
474 store <4 x float> %shuffle0, <4 x float>* %idx0, align 16
475 store <4 x float> %shuffle1, <4 x float>* %idx1, align 16
476 store <4 x float> %shuffle2, <4 x float>* %idx2, align 16
477 store <4 x float> %shuffle3, <4 x float>* %idx3, align 16
480 ; CHECK-LABEL: merge_vec_extract_stores
481 ; CHECK: vmovaps %xmm0, 48(%rdi)
482 ; CHECK-NEXT: vextractf128 $1, %ymm0, 64(%rdi)
483 ; CHECK-NEXT: vmovaps %xmm1, 80(%rdi)
484 ; CHECK-NEXT: vextractf128 $1, %ymm1, 96(%rdi)
485 ; CHECK-NEXT: vzeroupper
489 ; Merging vector stores when sourced from vector loads is not currently handled.
490 define void @merge_vec_stores_from_loads(<4 x float>* %v, <4 x float>* %ptr) {
491 %load_idx0 = getelementptr inbounds <4 x float>, <4 x float>* %v, i64 0
492 %load_idx1 = getelementptr inbounds <4 x float>, <4 x float>* %v, i64 1
493 %v0 = load <4 x float>, <4 x float>* %load_idx0
494 %v1 = load <4 x float>, <4 x float>* %load_idx1
495 %store_idx0 = getelementptr inbounds <4 x float>, <4 x float>* %ptr, i64 0
496 %store_idx1 = getelementptr inbounds <4 x float>, <4 x float>* %ptr, i64 1
497 store <4 x float> %v0, <4 x float>* %store_idx0, align 16
498 store <4 x float> %v1, <4 x float>* %store_idx1, align 16
501 ; CHECK-LABEL: merge_vec_stores_from_loads
503 ; CHECK-NEXT: vmovaps
504 ; CHECK-NEXT: vmovaps
505 ; CHECK-NEXT: vmovaps
509 ; Merging vector stores when sourced from a constant vector is not currently handled.
510 define void @merge_vec_stores_of_constants(<4 x i32>* %ptr) {
511 %idx0 = getelementptr inbounds <4 x i32>, <4 x i32>* %ptr, i64 3
512 %idx1 = getelementptr inbounds <4 x i32>, <4 x i32>* %ptr, i64 4
513 store <4 x i32> <i32 0, i32 0, i32 0, i32 0>, <4 x i32>* %idx0, align 16
514 store <4 x i32> <i32 0, i32 0, i32 0, i32 0>, <4 x i32>* %idx1, align 16
517 ; CHECK-LABEL: merge_vec_stores_of_constants
519 ; CHECK-NEXT: vmovaps
520 ; CHECK-NEXT: vmovaps
524 ; This is a minimized test based on real code that was failing.
525 ; We could merge stores (and loads) like this...
527 define void @merge_vec_element_and_scalar_load([6 x i64]* %array) {
528 %idx0 = getelementptr inbounds [6 x i64], [6 x i64]* %array, i64 0, i64 0
529 %idx1 = getelementptr inbounds [6 x i64], [6 x i64]* %array, i64 0, i64 1
530 %idx4 = getelementptr inbounds [6 x i64], [6 x i64]* %array, i64 0, i64 4
531 %idx5 = getelementptr inbounds [6 x i64], [6 x i64]* %array, i64 0, i64 5
533 %a0 = load i64, i64* %idx0, align 8
534 store i64 %a0, i64* %idx4, align 8
536 %b = bitcast i64* %idx1 to <2 x i64>*
537 %v = load <2 x i64>, <2 x i64>* %b, align 8
538 %a1 = extractelement <2 x i64> %v, i32 0
539 store i64 %a1, i64* %idx5, align 8
542 ; CHECK-LABEL: merge_vec_element_and_scalar_load
543 ; CHECK: movq (%rdi), %rax
544 ; CHECK-NEXT: movq %rax, 32(%rdi)
545 ; CHECK-NEXT: movq 8(%rdi), %rax
546 ; CHECK-NEXT: movq %rax, 40(%rdi)