1 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
2 ; RUN: opt -slp-vectorizer -S < %s -mtriple=x86_64-unknown-linux -mcpu=corei7-avx | FileCheck %s
4 ; This file tests the look-ahead operand reordering heuristic.
7 ; This checks that operand reordering will reorder the operands of the adds
8 ; by taking into consideration the instructions beyond the immediate
11 ; A[0] B[0] C[0] D[0] C[1] D[1] A[1] B[1]
19 define void @lookahead_basic(double* %array) {
20 ; CHECK-LABEL: @lookahead_basic(
22 ; CHECK-NEXT: [[IDX0:%.*]] = getelementptr inbounds double, double* [[ARRAY:%.*]], i64 0
23 ; CHECK-NEXT: [[IDX1:%.*]] = getelementptr inbounds double, double* [[ARRAY]], i64 1
24 ; CHECK-NEXT: [[IDX2:%.*]] = getelementptr inbounds double, double* [[ARRAY]], i64 2
25 ; CHECK-NEXT: [[IDX3:%.*]] = getelementptr inbounds double, double* [[ARRAY]], i64 3
26 ; CHECK-NEXT: [[IDX4:%.*]] = getelementptr inbounds double, double* [[ARRAY]], i64 4
27 ; CHECK-NEXT: [[IDX5:%.*]] = getelementptr inbounds double, double* [[ARRAY]], i64 5
28 ; CHECK-NEXT: [[IDX6:%.*]] = getelementptr inbounds double, double* [[ARRAY]], i64 6
29 ; CHECK-NEXT: [[IDX7:%.*]] = getelementptr inbounds double, double* [[ARRAY]], i64 7
30 ; CHECK-NEXT: [[A_0:%.*]] = load double, double* [[IDX0]], align 8
31 ; CHECK-NEXT: [[A_1:%.*]] = load double, double* [[IDX1]], align 8
32 ; CHECK-NEXT: [[B_0:%.*]] = load double, double* [[IDX2]], align 8
33 ; CHECK-NEXT: [[B_1:%.*]] = load double, double* [[IDX3]], align 8
34 ; CHECK-NEXT: [[C_0:%.*]] = load double, double* [[IDX4]], align 8
35 ; CHECK-NEXT: [[C_1:%.*]] = load double, double* [[IDX5]], align 8
36 ; CHECK-NEXT: [[D_0:%.*]] = load double, double* [[IDX6]], align 8
37 ; CHECK-NEXT: [[D_1:%.*]] = load double, double* [[IDX7]], align 8
38 ; CHECK-NEXT: [[SUBAB_0:%.*]] = fsub fast double [[A_0]], [[B_0]]
39 ; CHECK-NEXT: [[SUBCD_0:%.*]] = fsub fast double [[C_0]], [[D_0]]
40 ; CHECK-NEXT: [[SUBAB_1:%.*]] = fsub fast double [[A_1]], [[B_1]]
41 ; CHECK-NEXT: [[SUBCD_1:%.*]] = fsub fast double [[C_1]], [[D_1]]
42 ; CHECK-NEXT: [[ADDABCD_0:%.*]] = fadd fast double [[SUBAB_0]], [[SUBCD_0]]
43 ; CHECK-NEXT: [[ADDCDAB_1:%.*]] = fadd fast double [[SUBCD_1]], [[SUBAB_1]]
44 ; CHECK-NEXT: store double [[ADDABCD_0]], double* [[IDX0]], align 8
45 ; CHECK-NEXT: store double [[ADDCDAB_1]], double* [[IDX1]], align 8
46 ; CHECK-NEXT: ret void
49 %idx0 = getelementptr inbounds double, double* %array, i64 0
50 %idx1 = getelementptr inbounds double, double* %array, i64 1
51 %idx2 = getelementptr inbounds double, double* %array, i64 2
52 %idx3 = getelementptr inbounds double, double* %array, i64 3
53 %idx4 = getelementptr inbounds double, double* %array, i64 4
54 %idx5 = getelementptr inbounds double, double* %array, i64 5
55 %idx6 = getelementptr inbounds double, double* %array, i64 6
56 %idx7 = getelementptr inbounds double, double* %array, i64 7
58 %A_0 = load double, double *%idx0, align 8
59 %A_1 = load double, double *%idx1, align 8
60 %B_0 = load double, double *%idx2, align 8
61 %B_1 = load double, double *%idx3, align 8
62 %C_0 = load double, double *%idx4, align 8
63 %C_1 = load double, double *%idx5, align 8
64 %D_0 = load double, double *%idx6, align 8
65 %D_1 = load double, double *%idx7, align 8
67 %subAB_0 = fsub fast double %A_0, %B_0
68 %subCD_0 = fsub fast double %C_0, %D_0
70 %subAB_1 = fsub fast double %A_1, %B_1
71 %subCD_1 = fsub fast double %C_1, %D_1
73 %addABCD_0 = fadd fast double %subAB_0, %subCD_0
74 %addCDAB_1 = fadd fast double %subCD_1, %subAB_1
76 store double %addABCD_0, double *%idx0, align 8
77 store double %addCDAB_1, double *%idx1, align 8
82 ; Check whether the look-ahead operand reordering heuristic will avoid
83 ; bundling the alt opcodes. The vectorized code should have no shuffles.
85 ; A[0] B[0] A[0] B[0] A[1] A[1] A[1] B[1]
93 define void @lookahead_alt1(double* %array) {
94 ; CHECK-LABEL: @lookahead_alt1(
96 ; CHECK-NEXT: [[IDX0:%.*]] = getelementptr inbounds double, double* [[ARRAY:%.*]], i64 0
97 ; CHECK-NEXT: [[IDX1:%.*]] = getelementptr inbounds double, double* [[ARRAY]], i64 1
98 ; CHECK-NEXT: [[IDX2:%.*]] = getelementptr inbounds double, double* [[ARRAY]], i64 2
99 ; CHECK-NEXT: [[IDX3:%.*]] = getelementptr inbounds double, double* [[ARRAY]], i64 3
100 ; CHECK-NEXT: [[IDX4:%.*]] = getelementptr inbounds double, double* [[ARRAY]], i64 4
101 ; CHECK-NEXT: [[IDX5:%.*]] = getelementptr inbounds double, double* [[ARRAY]], i64 5
102 ; CHECK-NEXT: [[IDX6:%.*]] = getelementptr inbounds double, double* [[ARRAY]], i64 6
103 ; CHECK-NEXT: [[IDX7:%.*]] = getelementptr inbounds double, double* [[ARRAY]], i64 7
104 ; CHECK-NEXT: [[TMP0:%.*]] = bitcast double* [[IDX0]] to <2 x double>*
105 ; CHECK-NEXT: [[TMP1:%.*]] = load <2 x double>, <2 x double>* [[TMP0]], align 8
106 ; CHECK-NEXT: [[TMP2:%.*]] = bitcast double* [[IDX2]] to <2 x double>*
107 ; CHECK-NEXT: [[TMP3:%.*]] = load <2 x double>, <2 x double>* [[TMP2]], align 8
108 ; CHECK-NEXT: [[TMP4:%.*]] = fsub fast <2 x double> [[TMP1]], [[TMP3]]
109 ; CHECK-NEXT: [[TMP5:%.*]] = fadd fast <2 x double> [[TMP1]], [[TMP3]]
110 ; CHECK-NEXT: [[TMP6:%.*]] = fadd fast <2 x double> [[TMP5]], [[TMP4]]
111 ; CHECK-NEXT: [[TMP7:%.*]] = bitcast double* [[IDX0]] to <2 x double>*
112 ; CHECK-NEXT: store <2 x double> [[TMP6]], <2 x double>* [[TMP7]], align 8
113 ; CHECK-NEXT: ret void
116 %idx0 = getelementptr inbounds double, double* %array, i64 0
117 %idx1 = getelementptr inbounds double, double* %array, i64 1
118 %idx2 = getelementptr inbounds double, double* %array, i64 2
119 %idx3 = getelementptr inbounds double, double* %array, i64 3
120 %idx4 = getelementptr inbounds double, double* %array, i64 4
121 %idx5 = getelementptr inbounds double, double* %array, i64 5
122 %idx6 = getelementptr inbounds double, double* %array, i64 6
123 %idx7 = getelementptr inbounds double, double* %array, i64 7
125 %A_0 = load double, double *%idx0, align 8
126 %A_1 = load double, double *%idx1, align 8
127 %B_0 = load double, double *%idx2, align 8
128 %B_1 = load double, double *%idx3, align 8
130 %addAB_0_L = fadd fast double %A_0, %B_0
131 %subAB_0_R = fsub fast double %A_0, %B_0
133 %subAB_1_L = fsub fast double %A_1, %B_1
134 %addAB_1_R = fadd fast double %A_1, %B_1
136 %addABCD_0 = fadd fast double %addAB_0_L, %subAB_0_R
137 %addCDAB_1 = fadd fast double %subAB_1_L, %addAB_1_R
139 store double %addABCD_0, double *%idx0, align 8
140 store double %addCDAB_1, double *%idx1, align 8
145 ; This code should get vectorized all the way to the loads with shuffles for
148 ; A[0] B[0] C[0] D[0] C[1] D[1] A[1] B[1]
156 define void @lookahead_alt2(double* %array) {
157 ; CHECK-LABEL: @lookahead_alt2(
159 ; CHECK-NEXT: [[IDX0:%.*]] = getelementptr inbounds double, double* [[ARRAY:%.*]], i64 0
160 ; CHECK-NEXT: [[IDX1:%.*]] = getelementptr inbounds double, double* [[ARRAY]], i64 1
161 ; CHECK-NEXT: [[IDX2:%.*]] = getelementptr inbounds double, double* [[ARRAY]], i64 2
162 ; CHECK-NEXT: [[IDX3:%.*]] = getelementptr inbounds double, double* [[ARRAY]], i64 3
163 ; CHECK-NEXT: [[IDX4:%.*]] = getelementptr inbounds double, double* [[ARRAY]], i64 4
164 ; CHECK-NEXT: [[IDX5:%.*]] = getelementptr inbounds double, double* [[ARRAY]], i64 5
165 ; CHECK-NEXT: [[IDX6:%.*]] = getelementptr inbounds double, double* [[ARRAY]], i64 6
166 ; CHECK-NEXT: [[IDX7:%.*]] = getelementptr inbounds double, double* [[ARRAY]], i64 7
167 ; CHECK-NEXT: [[A_0:%.*]] = load double, double* [[IDX0]], align 8
168 ; CHECK-NEXT: [[A_1:%.*]] = load double, double* [[IDX1]], align 8
169 ; CHECK-NEXT: [[B_0:%.*]] = load double, double* [[IDX2]], align 8
170 ; CHECK-NEXT: [[B_1:%.*]] = load double, double* [[IDX3]], align 8
171 ; CHECK-NEXT: [[C_0:%.*]] = load double, double* [[IDX4]], align 8
172 ; CHECK-NEXT: [[C_1:%.*]] = load double, double* [[IDX5]], align 8
173 ; CHECK-NEXT: [[D_0:%.*]] = load double, double* [[IDX6]], align 8
174 ; CHECK-NEXT: [[D_1:%.*]] = load double, double* [[IDX7]], align 8
175 ; CHECK-NEXT: [[ADDAB_0:%.*]] = fadd fast double [[A_0]], [[B_0]]
176 ; CHECK-NEXT: [[SUBCD_0:%.*]] = fsub fast double [[C_0]], [[D_0]]
177 ; CHECK-NEXT: [[ADDCD_1:%.*]] = fadd fast double [[C_1]], [[D_1]]
178 ; CHECK-NEXT: [[SUBAB_1:%.*]] = fsub fast double [[A_1]], [[B_1]]
179 ; CHECK-NEXT: [[ADDABCD_0:%.*]] = fadd fast double [[ADDAB_0]], [[SUBCD_0]]
180 ; CHECK-NEXT: [[ADDCDAB_1:%.*]] = fadd fast double [[ADDCD_1]], [[SUBAB_1]]
181 ; CHECK-NEXT: store double [[ADDABCD_0]], double* [[IDX0]], align 8
182 ; CHECK-NEXT: store double [[ADDCDAB_1]], double* [[IDX1]], align 8
183 ; CHECK-NEXT: ret void
186 %idx0 = getelementptr inbounds double, double* %array, i64 0
187 %idx1 = getelementptr inbounds double, double* %array, i64 1
188 %idx2 = getelementptr inbounds double, double* %array, i64 2
189 %idx3 = getelementptr inbounds double, double* %array, i64 3
190 %idx4 = getelementptr inbounds double, double* %array, i64 4
191 %idx5 = getelementptr inbounds double, double* %array, i64 5
192 %idx6 = getelementptr inbounds double, double* %array, i64 6
193 %idx7 = getelementptr inbounds double, double* %array, i64 7
195 %A_0 = load double, double *%idx0, align 8
196 %A_1 = load double, double *%idx1, align 8
197 %B_0 = load double, double *%idx2, align 8
198 %B_1 = load double, double *%idx3, align 8
199 %C_0 = load double, double *%idx4, align 8
200 %C_1 = load double, double *%idx5, align 8
201 %D_0 = load double, double *%idx6, align 8
202 %D_1 = load double, double *%idx7, align 8
204 %addAB_0 = fadd fast double %A_0, %B_0
205 %subCD_0 = fsub fast double %C_0, %D_0
207 %addCD_1 = fadd fast double %C_1, %D_1
208 %subAB_1 = fsub fast double %A_1, %B_1
210 %addABCD_0 = fadd fast double %addAB_0, %subCD_0
211 %addCDAB_1 = fadd fast double %addCD_1, %subAB_1
213 store double %addABCD_0, double *%idx0, align 8
214 store double %addCDAB_1, double *%idx1, align 8
220 ; A[0] B[0] C[0] D[0] A[1] B[2] A[2] B[1]
228 ; SLP should reorder the operands of the RHS add taking into consideration the cost of external uses.
229 ; It is more profitable to reorder the operands of the RHS add, because A[1] has an external use.
231 define void @lookahead_external_uses(double* %A, double *%B, double *%C, double *%D, double *%S, double *%Ext1, double *%Ext2) {
232 ; CHECK-LABEL: @lookahead_external_uses(
234 ; CHECK-NEXT: [[IDXA0:%.*]] = getelementptr inbounds double, double* [[A:%.*]], i64 0
235 ; CHECK-NEXT: [[IDXB0:%.*]] = getelementptr inbounds double, double* [[B:%.*]], i64 0
236 ; CHECK-NEXT: [[IDXC0:%.*]] = getelementptr inbounds double, double* [[C:%.*]], i64 0
237 ; CHECK-NEXT: [[IDXD0:%.*]] = getelementptr inbounds double, double* [[D:%.*]], i64 0
238 ; CHECK-NEXT: [[IDXA1:%.*]] = getelementptr inbounds double, double* [[A]], i64 1
239 ; CHECK-NEXT: [[IDXB2:%.*]] = getelementptr inbounds double, double* [[B]], i64 2
240 ; CHECK-NEXT: [[IDXA2:%.*]] = getelementptr inbounds double, double* [[A]], i64 2
241 ; CHECK-NEXT: [[IDXB1:%.*]] = getelementptr inbounds double, double* [[B]], i64 1
242 ; CHECK-NEXT: [[B0:%.*]] = load double, double* [[IDXB0]], align 8
243 ; CHECK-NEXT: [[C0:%.*]] = load double, double* [[IDXC0]], align 8
244 ; CHECK-NEXT: [[D0:%.*]] = load double, double* [[IDXD0]], align 8
245 ; CHECK-NEXT: [[TMP0:%.*]] = bitcast double* [[IDXA0]] to <2 x double>*
246 ; CHECK-NEXT: [[TMP1:%.*]] = load <2 x double>, <2 x double>* [[TMP0]], align 8
247 ; CHECK-NEXT: [[B2:%.*]] = load double, double* [[IDXB2]], align 8
248 ; CHECK-NEXT: [[A2:%.*]] = load double, double* [[IDXA2]], align 8
249 ; CHECK-NEXT: [[B1:%.*]] = load double, double* [[IDXB1]], align 8
250 ; CHECK-NEXT: [[TMP2:%.*]] = insertelement <2 x double> undef, double [[B0]], i32 0
251 ; CHECK-NEXT: [[TMP3:%.*]] = insertelement <2 x double> [[TMP2]], double [[B2]], i32 1
252 ; CHECK-NEXT: [[TMP4:%.*]] = fsub fast <2 x double> [[TMP1]], [[TMP3]]
253 ; CHECK-NEXT: [[TMP5:%.*]] = insertelement <2 x double> undef, double [[C0]], i32 0
254 ; CHECK-NEXT: [[TMP6:%.*]] = insertelement <2 x double> [[TMP5]], double [[A2]], i32 1
255 ; CHECK-NEXT: [[TMP7:%.*]] = insertelement <2 x double> undef, double [[D0]], i32 0
256 ; CHECK-NEXT: [[TMP8:%.*]] = insertelement <2 x double> [[TMP7]], double [[B1]], i32 1
257 ; CHECK-NEXT: [[TMP9:%.*]] = fsub fast <2 x double> [[TMP6]], [[TMP8]]
258 ; CHECK-NEXT: [[TMP10:%.*]] = fadd fast <2 x double> [[TMP4]], [[TMP9]]
259 ; CHECK-NEXT: [[IDXS0:%.*]] = getelementptr inbounds double, double* [[S:%.*]], i64 0
260 ; CHECK-NEXT: [[IDXS1:%.*]] = getelementptr inbounds double, double* [[S]], i64 1
261 ; CHECK-NEXT: [[TMP11:%.*]] = bitcast double* [[IDXS0]] to <2 x double>*
262 ; CHECK-NEXT: store <2 x double> [[TMP10]], <2 x double>* [[TMP11]], align 8
263 ; CHECK-NEXT: [[TMP12:%.*]] = extractelement <2 x double> [[TMP1]], i32 1
264 ; CHECK-NEXT: store double [[TMP12]], double* [[EXT1:%.*]], align 8
265 ; CHECK-NEXT: ret void
268 %IdxA0 = getelementptr inbounds double, double* %A, i64 0
269 %IdxB0 = getelementptr inbounds double, double* %B, i64 0
270 %IdxC0 = getelementptr inbounds double, double* %C, i64 0
271 %IdxD0 = getelementptr inbounds double, double* %D, i64 0
273 %IdxA1 = getelementptr inbounds double, double* %A, i64 1
274 %IdxB2 = getelementptr inbounds double, double* %B, i64 2
275 %IdxA2 = getelementptr inbounds double, double* %A, i64 2
276 %IdxB1 = getelementptr inbounds double, double* %B, i64 1
278 %A0 = load double, double *%IdxA0, align 8
279 %B0 = load double, double *%IdxB0, align 8
280 %C0 = load double, double *%IdxC0, align 8
281 %D0 = load double, double *%IdxD0, align 8
283 %A1 = load double, double *%IdxA1, align 8
284 %B2 = load double, double *%IdxB2, align 8
285 %A2 = load double, double *%IdxA2, align 8
286 %B1 = load double, double *%IdxB1, align 8
288 %subA0B0 = fsub fast double %A0, %B0
289 %subC0D0 = fsub fast double %C0, %D0
291 %subA1B2 = fsub fast double %A1, %B2
292 %subA2B1 = fsub fast double %A2, %B1
294 %add0 = fadd fast double %subA0B0, %subC0D0
295 %add1 = fadd fast double %subA1B2, %subA2B1
297 %IdxS0 = getelementptr inbounds double, double* %S, i64 0
298 %IdxS1 = getelementptr inbounds double, double* %S, i64 1
300 store double %add0, double *%IdxS0, align 8
301 store double %add1, double *%IdxS1, align 8
304 store double %A1, double *%Ext1, align 8