1 ; RUN: opt -mtriple armv7-linux-gnueabihf -loop-vectorize -S %s -debug-only=loop-vectorize -o /dev/null 2>&1 | FileCheck %s --check-prefix=CHECK --check-prefix=LINUX
2 ; RUN: opt -mtriple armv8-linux-gnu -loop-vectorize -S %s -debug-only=loop-vectorize -o /dev/null 2>&1 | FileCheck %s --check-prefix=CHECK --check-prefix=LINUX
3 ; RUN: opt -mtriple armv8.1.m-none-eabi -mattr=+mve.fp -loop-vectorize -S %s -debug-only=loop-vectorize -o /dev/null 2>&1 | FileCheck %s --check-prefix=CHECK --check-prefix=MVE
4 ; RUN: opt -mtriple armv7-unknwon-darwin -loop-vectorize -S %s -debug-only=loop-vectorize -o /dev/null 2>&1 | FileCheck %s --check-prefix=CHECK --check-prefix=DARWIN
7 ; Testing the ability of the loop vectorizer to tell when SIMD is safe or not
8 ; regarding IEEE 754 standard.
9 ; On Linux, we only want the vectorizer to work when -ffast-math flag is set,
10 ; because NEON is not IEEE compliant.
11 ; Darwin, on the other hand, doesn't support subnormals, and all optimizations
12 ; are allowed, even without -ffast-math.
14 ; Integer loops are always vectorizeable
15 ; CHECK: Checking a loop in "sumi"
16 ; CHECK: We can vectorize this loop!
17 define void @sumi(i32* noalias nocapture readonly %A, i32* noalias nocapture readonly %B, i32* noalias nocapture %C, i32 %N) {
19 %cmp5 = icmp eq i32 %N, 0
20 br i1 %cmp5, label %for.end, label %for.body.preheader
22 for.body.preheader: ; preds = %entry
25 for.body: ; preds = %for.body.preheader, %for.body
26 %i.06 = phi i32 [ %inc, %for.body ], [ 0, %for.body.preheader ]
27 %arrayidx = getelementptr inbounds i32, i32* %A, i32 %i.06
28 %0 = load i32, i32* %arrayidx, align 4
29 %arrayidx1 = getelementptr inbounds i32, i32* %B, i32 %i.06
30 %1 = load i32, i32* %arrayidx1, align 4
31 %mul = mul nsw i32 %1, %0
32 %arrayidx2 = getelementptr inbounds i32, i32* %C, i32 %i.06
33 store i32 %mul, i32* %arrayidx2, align 4
34 %inc = add nuw nsw i32 %i.06, 1
35 %exitcond = icmp eq i32 %inc, %N
36 br i1 %exitcond, label %for.end.loopexit, label %for.body
38 for.end.loopexit: ; preds = %for.body
41 for.end: ; preds = %for.end.loopexit, %entry
45 ; Floating-point loops need fast-math to be vectorizeable
46 ; LINUX: Checking a loop in "sumf"
47 ; LINUX: Potentially unsafe FP op prevents vectorization
48 ; MVE: Checking a loop in "sumf"
49 ; MVE: We can vectorize this loop!
50 ; DARWIN: Checking a loop in "sumf"
51 ; DARWIN: We can vectorize this loop!
52 define void @sumf(float* noalias nocapture readonly %A, float* noalias nocapture readonly %B, float* noalias nocapture %C, i32 %N) {
54 %cmp5 = icmp eq i32 %N, 0
55 br i1 %cmp5, label %for.end, label %for.body.preheader
57 for.body.preheader: ; preds = %entry
60 for.body: ; preds = %for.body.preheader, %for.body
61 %i.06 = phi i32 [ %inc, %for.body ], [ 0, %for.body.preheader ]
62 %arrayidx = getelementptr inbounds float, float* %A, i32 %i.06
63 %0 = load float, float* %arrayidx, align 4
64 %arrayidx1 = getelementptr inbounds float, float* %B, i32 %i.06
65 %1 = load float, float* %arrayidx1, align 4
66 %mul = fmul float %0, %1
67 %arrayidx2 = getelementptr inbounds float, float* %C, i32 %i.06
68 store float %mul, float* %arrayidx2, align 4
69 %inc = add nuw nsw i32 %i.06, 1
70 %exitcond = icmp eq i32 %inc, %N
71 br i1 %exitcond, label %for.end.loopexit, label %for.body
73 for.end.loopexit: ; preds = %for.body
76 for.end: ; preds = %for.end.loopexit, %entry
80 ; Integer loops are always vectorizeable
81 ; CHECK: Checking a loop in "redi"
82 ; CHECK: We can vectorize this loop!
83 define i32 @redi(i32* noalias nocapture readonly %a, i32* noalias nocapture readonly %b, i32 %N) {
85 %cmp5 = icmp eq i32 %N, 0
86 br i1 %cmp5, label %for.end, label %for.body.preheader
88 for.body.preheader: ; preds = %entry
91 for.body: ; preds = %for.body.preheader, %for.body
92 %i.07 = phi i32 [ %inc, %for.body ], [ 0, %for.body.preheader ]
93 %Red.06 = phi i32 [ %add, %for.body ], [ undef, %for.body.preheader ]
94 %arrayidx = getelementptr inbounds i32, i32* %a, i32 %i.07
95 %0 = load i32, i32* %arrayidx, align 4
96 %arrayidx1 = getelementptr inbounds i32, i32* %b, i32 %i.07
97 %1 = load i32, i32* %arrayidx1, align 4
98 %mul = mul nsw i32 %1, %0
99 %add = add nsw i32 %mul, %Red.06
100 %inc = add nuw nsw i32 %i.07, 1
101 %exitcond = icmp eq i32 %inc, %N
102 br i1 %exitcond, label %for.end.loopexit, label %for.body
104 for.end.loopexit: ; preds = %for.body
105 %add.lcssa = phi i32 [ %add, %for.body ]
108 for.end: ; preds = %for.end.loopexit, %entry
109 %Red.0.lcssa = phi i32 [ undef, %entry ], [ %add.lcssa, %for.end.loopexit ]
113 ; Floating-point loops need fast-math to be vectorizeable
114 ; LINUX: Checking a loop in "redf"
115 ; LINUX: Potentially unsafe FP op prevents vectorization
116 ; MVE: Checking a loop in "redf"
117 ; MVE: We can vectorize this loop!
118 ; DARWIN: Checking a loop in "redf"
119 ; DARWIN: We can vectorize this loop!
120 define float @redf(float* noalias nocapture readonly %a, float* noalias nocapture readonly %b, i32 %N) {
122 %cmp5 = icmp eq i32 %N, 0
123 br i1 %cmp5, label %for.end, label %for.body.preheader
125 for.body.preheader: ; preds = %entry
128 for.body: ; preds = %for.body.preheader, %for.body
129 %i.07 = phi i32 [ %inc, %for.body ], [ 0, %for.body.preheader ]
130 %Red.06 = phi float [ %add, %for.body ], [ undef, %for.body.preheader ]
131 %arrayidx = getelementptr inbounds float, float* %a, i32 %i.07
132 %0 = load float, float* %arrayidx, align 4
133 %arrayidx1 = getelementptr inbounds float, float* %b, i32 %i.07
134 %1 = load float, float* %arrayidx1, align 4
135 %mul = fmul float %0, %1
136 %add = fadd float %Red.06, %mul
137 %inc = add nuw nsw i32 %i.07, 1
138 %exitcond = icmp eq i32 %inc, %N
139 br i1 %exitcond, label %for.end.loopexit, label %for.body
141 for.end.loopexit: ; preds = %for.body
142 %add.lcssa = phi float [ %add, %for.body ]
145 for.end: ; preds = %for.end.loopexit, %entry
146 %Red.0.lcssa = phi float [ undef, %entry ], [ %add.lcssa, %for.end.loopexit ]
147 ret float %Red.0.lcssa
150 ; Make sure calls that turn into builtins are also covered
151 ; LINUX: Checking a loop in "fabs"
152 ; LINUX: Potentially unsafe FP op prevents vectorization
153 ; DARWIN: Checking a loop in "fabs"
154 ; DARWIN: We can vectorize this loop!
155 define void @fabs(float* noalias nocapture readonly %A, float* noalias nocapture readonly %B, float* noalias nocapture %C, i32 %N) {
157 %cmp10 = icmp eq i32 %N, 0
158 br i1 %cmp10, label %for.end, label %for.body
160 for.body: ; preds = %entry, %for.body
161 %i.011 = phi i32 [ %inc, %for.body ], [ 0, %entry ]
162 %arrayidx = getelementptr inbounds float, float* %A, i32 %i.011
163 %0 = load float, float* %arrayidx, align 4
164 %arrayidx1 = getelementptr inbounds float, float* %B, i32 %i.011
165 %1 = load float, float* %arrayidx1, align 4
166 %fabsf = tail call float @fabsf(float %1) #1
167 %conv3 = fmul float %0, %fabsf
168 %arrayidx4 = getelementptr inbounds float, float* %C, i32 %i.011
169 store float %conv3, float* %arrayidx4, align 4
170 %inc = add nuw nsw i32 %i.011, 1
171 %exitcond = icmp eq i32 %inc, %N
172 br i1 %exitcond, label %for.end, label %for.body
174 for.end: ; preds = %for.body, %entry
178 ; Integer loops are always vectorizeable
179 ; CHECK: Checking a loop in "sumi_fast"
180 ; CHECK: We can vectorize this loop!
181 define void @sumi_fast(i32* noalias nocapture readonly %A, i32* noalias nocapture readonly %B, i32* noalias nocapture %C, i32 %N) {
183 %cmp5 = icmp eq i32 %N, 0
184 br i1 %cmp5, label %for.end, label %for.body.preheader
186 for.body.preheader: ; preds = %entry
189 for.body: ; preds = %for.body.preheader, %for.body
190 %i.06 = phi i32 [ %inc, %for.body ], [ 0, %for.body.preheader ]
191 %arrayidx = getelementptr inbounds i32, i32* %A, i32 %i.06
192 %0 = load i32, i32* %arrayidx, align 4
193 %arrayidx1 = getelementptr inbounds i32, i32* %B, i32 %i.06
194 %1 = load i32, i32* %arrayidx1, align 4
195 %mul = mul nsw i32 %1, %0
196 %arrayidx2 = getelementptr inbounds i32, i32* %C, i32 %i.06
197 store i32 %mul, i32* %arrayidx2, align 4
198 %inc = add nuw nsw i32 %i.06, 1
199 %exitcond = icmp eq i32 %inc, %N
200 br i1 %exitcond, label %for.end.loopexit, label %for.body
202 for.end.loopexit: ; preds = %for.body
205 for.end: ; preds = %for.end.loopexit, %entry
209 ; Floating-point loops can be vectorizeable with fast-math
210 ; CHECK: Checking a loop in "sumf_fast"
211 ; CHECK: We can vectorize this loop!
212 define void @sumf_fast(float* noalias nocapture readonly %A, float* noalias nocapture readonly %B, float* noalias nocapture %C, i32 %N) {
214 %cmp5 = icmp eq i32 %N, 0
215 br i1 %cmp5, label %for.end, label %for.body.preheader
217 for.body.preheader: ; preds = %entry
220 for.body: ; preds = %for.body.preheader, %for.body
221 %i.06 = phi i32 [ %inc, %for.body ], [ 0, %for.body.preheader ]
222 %arrayidx = getelementptr inbounds float, float* %A, i32 %i.06
223 %0 = load float, float* %arrayidx, align 4
224 %arrayidx1 = getelementptr inbounds float, float* %B, i32 %i.06
225 %1 = load float, float* %arrayidx1, align 4
226 %mul = fmul fast float %1, %0
227 %arrayidx2 = getelementptr inbounds float, float* %C, i32 %i.06
228 store float %mul, float* %arrayidx2, align 4
229 %inc = add nuw nsw i32 %i.06, 1
230 %exitcond = icmp eq i32 %inc, %N
231 br i1 %exitcond, label %for.end.loopexit, label %for.body
233 for.end.loopexit: ; preds = %for.body
236 for.end: ; preds = %for.end.loopexit, %entry
240 ; Integer loops are always vectorizeable
241 ; CHECK: Checking a loop in "redi_fast"
242 ; CHECK: We can vectorize this loop!
243 define i32 @redi_fast(i32* noalias nocapture readonly %a, i32* noalias nocapture readonly %b, i32 %N) {
245 %cmp5 = icmp eq i32 %N, 0
246 br i1 %cmp5, label %for.end, label %for.body.preheader
248 for.body.preheader: ; preds = %entry
251 for.body: ; preds = %for.body.preheader, %for.body
252 %i.07 = phi i32 [ %inc, %for.body ], [ 0, %for.body.preheader ]
253 %Red.06 = phi i32 [ %add, %for.body ], [ undef, %for.body.preheader ]
254 %arrayidx = getelementptr inbounds i32, i32* %a, i32 %i.07
255 %0 = load i32, i32* %arrayidx, align 4
256 %arrayidx1 = getelementptr inbounds i32, i32* %b, i32 %i.07
257 %1 = load i32, i32* %arrayidx1, align 4
258 %mul = mul nsw i32 %1, %0
259 %add = add nsw i32 %mul, %Red.06
260 %inc = add nuw nsw i32 %i.07, 1
261 %exitcond = icmp eq i32 %inc, %N
262 br i1 %exitcond, label %for.end.loopexit, label %for.body
264 for.end.loopexit: ; preds = %for.body
265 %add.lcssa = phi i32 [ %add, %for.body ]
268 for.end: ; preds = %for.end.loopexit, %entry
269 %Red.0.lcssa = phi i32 [ undef, %entry ], [ %add.lcssa, %for.end.loopexit ]
273 ; Floating-point loops can be vectorizeable with fast-math
274 ; CHECK: Checking a loop in "redf_fast"
275 ; CHECK: We can vectorize this loop!
276 define float @redf_fast(float* noalias nocapture readonly %a, float* noalias nocapture readonly %b, i32 %N) {
278 %cmp5 = icmp eq i32 %N, 0
279 br i1 %cmp5, label %for.end, label %for.body.preheader
281 for.body.preheader: ; preds = %entry
284 for.body: ; preds = %for.body.preheader, %for.body
285 %i.07 = phi i32 [ %inc, %for.body ], [ 0, %for.body.preheader ]
286 %Red.06 = phi float [ %add, %for.body ], [ undef, %for.body.preheader ]
287 %arrayidx = getelementptr inbounds float, float* %a, i32 %i.07
288 %0 = load float, float* %arrayidx, align 4
289 %arrayidx1 = getelementptr inbounds float, float* %b, i32 %i.07
290 %1 = load float, float* %arrayidx1, align 4
291 %mul = fmul fast float %1, %0
292 %add = fadd fast float %mul, %Red.06
293 %inc = add nuw nsw i32 %i.07, 1
294 %exitcond = icmp eq i32 %inc, %N
295 br i1 %exitcond, label %for.end.loopexit, label %for.body
297 for.end.loopexit: ; preds = %for.body
298 %add.lcssa = phi float [ %add, %for.body ]
301 for.end: ; preds = %for.end.loopexit, %entry
302 %Red.0.lcssa = phi float [ undef, %entry ], [ %add.lcssa, %for.end.loopexit ]
303 ret float %Red.0.lcssa
306 ; Make sure calls that turn into builtins are also covered
307 ; CHECK: Checking a loop in "fabs_fast"
308 ; CHECK: We can vectorize this loop!
309 define void @fabs_fast(float* noalias nocapture readonly %A, float* noalias nocapture readonly %B, float* noalias nocapture %C, i32 %N) {
311 %cmp10 = icmp eq i32 %N, 0
312 br i1 %cmp10, label %for.end, label %for.body
314 for.body: ; preds = %entry, %for.body
315 %i.011 = phi i32 [ %inc, %for.body ], [ 0, %entry ]
316 %arrayidx = getelementptr inbounds float, float* %A, i32 %i.011
317 %0 = load float, float* %arrayidx, align 4
318 %arrayidx1 = getelementptr inbounds float, float* %B, i32 %i.011
319 %1 = load float, float* %arrayidx1, align 4
320 %fabsf = tail call fast float @fabsf(float %1) #2
321 %conv3 = fmul fast float %fabsf, %0
322 %arrayidx4 = getelementptr inbounds float, float* %C, i32 %i.011
323 store float %conv3, float* %arrayidx4, align 4
324 %inc = add nuw nsw i32 %i.011, 1
325 %exitcond = icmp eq i32 %inc, %N
326 br i1 %exitcond, label %for.end, label %for.body
328 for.end: ; preds = %for.body, %entry
332 declare float @fabsf(float)
334 attributes #1 = { nounwind readnone "disable-tail-calls"="false" "less-precise-fpmad"="false" "no-frame-pointer-elim"="true" "no-frame-pointer-elim-non-leaf" "no-infs-fp-math"="false" "no-nans-fp-math"="false" "stack-protector-buffer-size"="8" "target-cpu"="cortex-a8" "target-features"="+dsp,+neon,+vfp3" "unsafe-fp-math"="false" "use-soft-float"="false" }
335 attributes #2 = { nounwind readnone "disable-tail-calls"="false" "less-precise-fpmad"="false" "no-frame-pointer-elim"="true" "no-frame-pointer-elim-non-leaf" "no-infs-fp-math"="true" "no-nans-fp-math"="true" "stack-protector-buffer-size"="8" "target-cpu"="cortex-a8" "target-features"="+dsp,+neon,+vfp3" "unsafe-fp-math"="true" "use-soft-float"="false" }