1 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+sse4.1 | FileCheck %s --check-prefix=SSE
3 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx2 | FileCheck %s --check-prefix=AVX
6 define <4 x i32> @combine_vec_add_to_zero(<4 x i32> %a) {
7 ; SSE-LABEL: combine_vec_add_to_zero:
11 ; AVX-LABEL: combine_vec_add_to_zero:
14 %1 = add <4 x i32> %a, zeroinitializer
18 ; fold ((c1-A)+c2) -> (c1+c2)-A
19 define <4 x i32> @combine_vec_add_constant_sub(<4 x i32> %a) {
20 ; SSE-LABEL: combine_vec_add_constant_sub:
22 ; SSE-NEXT: movdqa {{.*#+}} xmm1 = [0,2,4,6]
23 ; SSE-NEXT: psubd %xmm0, %xmm1
24 ; SSE-NEXT: movdqa %xmm1, %xmm0
27 ; AVX-LABEL: combine_vec_add_constant_sub:
29 ; AVX-NEXT: vmovdqa {{.*#+}} xmm1 = [0,2,4,6]
30 ; AVX-NEXT: vpsubd %xmm0, %xmm1, %xmm0
32 %1 = sub <4 x i32> <i32 0, i32 1, i32 2, i32 3>, %a
33 %2 = add <4 x i32> <i32 0, i32 1, i32 2, i32 3>, %1
37 ; fold ((0-A) + B) -> B-A
38 define <4 x i32> @combine_vec_add_neg0(<4 x i32> %a, <4 x i32> %b) {
39 ; SSE-LABEL: combine_vec_add_neg0:
41 ; SSE-NEXT: psubd %xmm0, %xmm1
42 ; SSE-NEXT: movdqa %xmm1, %xmm0
45 ; AVX-LABEL: combine_vec_add_neg0:
47 ; AVX-NEXT: vpsubd %xmm0, %xmm1, %xmm0
49 %1 = sub <4 x i32> zeroinitializer, %a
50 %2 = add <4 x i32> %1, %b
54 ; fold (A + (0-B)) -> A-B
55 define <4 x i32> @combine_vec_add_neg1(<4 x i32> %a, <4 x i32> %b) {
56 ; SSE-LABEL: combine_vec_add_neg1:
58 ; SSE-NEXT: psubd %xmm1, %xmm0
61 ; AVX-LABEL: combine_vec_add_neg1:
63 ; AVX-NEXT: vpsubd %xmm1, %xmm0, %xmm0
65 %1 = sub <4 x i32> zeroinitializer, %b
66 %2 = add <4 x i32> %a, %1
71 define <4 x i32> @combine_vec_add_sub0(<4 x i32> %a, <4 x i32> %b) {
72 ; SSE-LABEL: combine_vec_add_sub0:
74 ; SSE-NEXT: movaps %xmm1, %xmm0
77 ; AVX-LABEL: combine_vec_add_sub0:
79 ; AVX-NEXT: vmovaps %xmm1, %xmm0
81 %1 = sub <4 x i32> %b, %a
82 %2 = add <4 x i32> %a, %1
87 define <4 x i32> @combine_vec_add_sub1(<4 x i32> %a, <4 x i32> %b) {
88 ; SSE-LABEL: combine_vec_add_sub1:
90 ; SSE-NEXT: movaps %xmm1, %xmm0
93 ; AVX-LABEL: combine_vec_add_sub1:
95 ; AVX-NEXT: vmovaps %xmm1, %xmm0
97 %1 = sub <4 x i32> %b, %a
98 %2 = add <4 x i32> %1, %a
102 ; fold (A+(B-(A+C))) to (B-C)
103 define <4 x i32> @combine_vec_add_sub_add0(<4 x i32> %a, <4 x i32> %b, <4 x i32> %c) {
104 ; SSE-LABEL: combine_vec_add_sub_add0:
106 ; SSE-NEXT: psubd %xmm2, %xmm1
107 ; SSE-NEXT: movdqa %xmm1, %xmm0
110 ; AVX-LABEL: combine_vec_add_sub_add0:
112 ; AVX-NEXT: vpsubd %xmm2, %xmm1, %xmm0
114 %1 = add <4 x i32> %a, %c
115 %2 = sub <4 x i32> %b, %1
116 %3 = add <4 x i32> %a, %2
120 ; fold (A+(B-(C+A))) to (B-C)
121 define <4 x i32> @combine_vec_add_sub_add1(<4 x i32> %a, <4 x i32> %b, <4 x i32> %c) {
122 ; SSE-LABEL: combine_vec_add_sub_add1:
124 ; SSE-NEXT: psubd %xmm2, %xmm1
125 ; SSE-NEXT: movdqa %xmm1, %xmm0
128 ; AVX-LABEL: combine_vec_add_sub_add1:
130 ; AVX-NEXT: vpsubd %xmm2, %xmm1, %xmm0
132 %1 = add <4 x i32> %c, %a
133 %2 = sub <4 x i32> %b, %1
134 %3 = add <4 x i32> %a, %2
138 ; fold (A+((B-A)+C)) to (B+C)
139 define <4 x i32> @combine_vec_add_sub_add2(<4 x i32> %a, <4 x i32> %b, <4 x i32> %c) {
140 ; SSE-LABEL: combine_vec_add_sub_add2:
142 ; SSE-NEXT: paddd %xmm2, %xmm1
143 ; SSE-NEXT: movdqa %xmm1, %xmm0
146 ; AVX-LABEL: combine_vec_add_sub_add2:
148 ; AVX-NEXT: vpaddd %xmm2, %xmm1, %xmm0
150 %1 = sub <4 x i32> %b, %a
151 %2 = add <4 x i32> %1, %c
152 %3 = add <4 x i32> %a, %2
156 ; fold (A+((B-A)-C)) to (B-C)
157 define <4 x i32> @combine_vec_add_sub_add3(<4 x i32> %a, <4 x i32> %b, <4 x i32> %c) {
158 ; SSE-LABEL: combine_vec_add_sub_add3:
160 ; SSE-NEXT: psubd %xmm2, %xmm1
161 ; SSE-NEXT: movdqa %xmm1, %xmm0
164 ; AVX-LABEL: combine_vec_add_sub_add3:
166 ; AVX-NEXT: vpsubd %xmm2, %xmm1, %xmm0
168 %1 = sub <4 x i32> %b, %a
169 %2 = sub <4 x i32> %1, %c
170 %3 = add <4 x i32> %a, %2
174 ; fold (A-B)+(C-D) to (A+C)-(B+D) when A or C is constant
175 define <4 x i32> @combine_vec_add_sub_sub(<4 x i32> %a, <4 x i32> %b, <4 x i32> %d) {
176 ; SSE-LABEL: combine_vec_add_sub_sub:
178 ; SSE-NEXT: paddd {{.*}}(%rip), %xmm0
179 ; SSE-NEXT: paddd %xmm2, %xmm1
180 ; SSE-NEXT: psubd %xmm1, %xmm0
183 ; AVX-LABEL: combine_vec_add_sub_sub:
185 ; AVX-NEXT: vpaddd {{.*}}(%rip), %xmm0, %xmm0
186 ; AVX-NEXT: vpaddd %xmm2, %xmm1, %xmm1
187 ; AVX-NEXT: vpsubd %xmm1, %xmm0, %xmm0
189 %1 = sub <4 x i32> %a, %b
190 %2 = sub <4 x i32> <i32 0, i32 1, i32 2, i32 3>, %d
191 %3 = add <4 x i32> %1, %2
195 ; fold (a+b) -> (a|b) iff a and b share no bits.
196 define <4 x i32> @combine_vec_add_uniquebits(<4 x i32> %a, <4 x i32> %b) {
197 ; SSE-LABEL: combine_vec_add_uniquebits:
199 ; SSE-NEXT: andps {{.*}}(%rip), %xmm0
200 ; SSE-NEXT: andps {{.*}}(%rip), %xmm1
201 ; SSE-NEXT: orps %xmm1, %xmm0
204 ; AVX-LABEL: combine_vec_add_uniquebits:
206 ; AVX-NEXT: vbroadcastss {{.*}}(%rip), %xmm2
207 ; AVX-NEXT: vandps %xmm2, %xmm0, %xmm0
208 ; AVX-NEXT: vbroadcastss {{.*}}(%rip), %xmm2
209 ; AVX-NEXT: vandps %xmm2, %xmm1, %xmm1
210 ; AVX-NEXT: vorps %xmm1, %xmm0, %xmm0
212 %1 = and <4 x i32> %a, <i32 61680, i32 61680, i32 61680, i32 61680>
213 %2 = and <4 x i32> %b, <i32 3855, i32 3855, i32 3855, i32 3855>
214 %3 = add <4 x i32> %1, %2
218 ; fold (add x, shl(0 - y, n)) -> sub(x, shl(y, n))
219 define <4 x i32> @combine_vec_add_shl_neg0(<4 x i32> %x, <4 x i32> %y) {
220 ; SSE-LABEL: combine_vec_add_shl_neg0:
222 ; SSE-NEXT: pslld $5, %xmm1
223 ; SSE-NEXT: psubd %xmm1, %xmm0
226 ; AVX-LABEL: combine_vec_add_shl_neg0:
228 ; AVX-NEXT: vpslld $5, %xmm1, %xmm1
229 ; AVX-NEXT: vpsubd %xmm1, %xmm0, %xmm0
231 %1 = sub <4 x i32> zeroinitializer, %y
232 %2 = shl <4 x i32> %1, <i32 5, i32 5, i32 5, i32 5>
233 %3 = add <4 x i32> %x, %2
237 ; fold (add shl(0 - y, n), x) -> sub(x, shl(y, n))
238 define <4 x i32> @combine_vec_add_shl_neg1(<4 x i32> %x, <4 x i32> %y) {
239 ; SSE-LABEL: combine_vec_add_shl_neg1:
241 ; SSE-NEXT: pslld $5, %xmm1
242 ; SSE-NEXT: psubd %xmm1, %xmm0
245 ; AVX-LABEL: combine_vec_add_shl_neg1:
247 ; AVX-NEXT: vpslld $5, %xmm1, %xmm1
248 ; AVX-NEXT: vpsubd %xmm1, %xmm0, %xmm0
250 %1 = sub <4 x i32> zeroinitializer, %y
251 %2 = shl <4 x i32> %1, <i32 5, i32 5, i32 5, i32 5>
252 %3 = add <4 x i32> %2, %x
256 ; (add z, (and (sbbl x, x), 1)) -> (sub z, (sbbl x, x))
257 ; and similar xforms where the inner op is either ~0 or 0.
258 define <4 x i32> @combine_vec_add_and_compare(<4 x i32> %a0, <4 x i32> %a1, <4 x i32> %a2) {
259 ; SSE-LABEL: combine_vec_add_and_compare:
261 ; SSE-NEXT: pcmpeqd %xmm2, %xmm1
262 ; SSE-NEXT: psubd %xmm1, %xmm0
265 ; AVX-LABEL: combine_vec_add_and_compare:
267 ; AVX-NEXT: vpcmpeqd %xmm2, %xmm1, %xmm1
268 ; AVX-NEXT: vpsubd %xmm1, %xmm0, %xmm0
270 %1 = icmp eq <4 x i32> %a1, %a2
271 %2 = sext <4 x i1> %1 to <4 x i32>
272 %3 = and <4 x i32> %2, <i32 1, i32 1, i32 1, i32 1>
273 %4 = add <4 x i32> %a0, %3
277 ; add (sext i1), X -> sub X, (zext i1)
278 define <4 x i32> @combine_vec_add_sext(<4 x i1> %a0, <4 x i32> %a1) {
279 ; SSE-LABEL: combine_vec_add_sext:
281 ; SSE-NEXT: pslld $31, %xmm0
282 ; SSE-NEXT: psrad $31, %xmm0
283 ; SSE-NEXT: paddd %xmm1, %xmm0
286 ; AVX-LABEL: combine_vec_add_sext:
288 ; AVX-NEXT: vpslld $31, %xmm0, %xmm0
289 ; AVX-NEXT: vpsrad $31, %xmm0, %xmm0
290 ; AVX-NEXT: vpaddd %xmm1, %xmm0, %xmm0
292 %1 = sext <4 x i1> %a0 to <4 x i32>
293 %2 = add <4 x i32> %1, %a1
297 ; add (sext i1), X -> sub X, (zext i1)
298 define <4 x i32> @combine_vec_add_sextinreg(<4 x i32> %a0, <4 x i32> %a1) {
299 ; SSE-LABEL: combine_vec_add_sextinreg:
301 ; SSE-NEXT: pslld $31, %xmm0
302 ; SSE-NEXT: psrad $31, %xmm0
303 ; SSE-NEXT: paddd %xmm1, %xmm0
306 ; AVX-LABEL: combine_vec_add_sextinreg:
308 ; AVX-NEXT: vpslld $31, %xmm0, %xmm0
309 ; AVX-NEXT: vpsrad $31, %xmm0, %xmm0
310 ; AVX-NEXT: vpaddd %xmm1, %xmm0, %xmm0
312 %1 = shl <4 x i32> %a0, <i32 31, i32 31, i32 31, i32 31>
313 %2 = ashr <4 x i32> %1, <i32 31, i32 31, i32 31, i32 31>
314 %3 = add <4 x i32> %2, %a1