1 ; RUN: llc < %s -march=arm -mcpu=cortex-a8 | FileCheck %s
2 ; Implementing vld / vst as REG_SEQUENCE eliminates the extra vmov's.
4 %struct.int16x8_t = type { <8 x i16> }
5 %struct.int32x4_t = type { <4 x i32> }
6 %struct.__neon_int8x8x2_t = type { <8 x i8>, <8 x i8> }
7 %struct.__neon_int8x8x3_t = type { <8 x i8>, <8 x i8>, <8 x i8> }
8 %struct.__neon_int16x8x2_t = type { <8 x i16>, <8 x i16> }
9 %struct.__neon_int32x4x2_t = type { <4 x i32>, <4 x i32> }
11 define void @t1(i16* %i_ptr, i16* %o_ptr, %struct.int32x4_t* nocapture %vT0ptr, %struct.int32x4_t* nocapture %vT1ptr) nounwind {
21 %0 = getelementptr inbounds %struct.int32x4_t* %vT0ptr, i32 0, i32 0 ; <<4 x i32>*> [#uses=1]
22 %1 = load <4 x i32>* %0, align 16 ; <<4 x i32>> [#uses=1]
23 %2 = getelementptr inbounds %struct.int32x4_t* %vT1ptr, i32 0, i32 0 ; <<4 x i32>*> [#uses=1]
24 %3 = load <4 x i32>* %2, align 16 ; <<4 x i32>> [#uses=1]
25 %4 = bitcast i16* %i_ptr to i8* ; <i8*> [#uses=1]
26 %5 = tail call <8 x i16> @llvm.arm.neon.vld1.v8i16(i8* %4, i32 1) ; <<8 x i16>> [#uses=1]
27 %6 = bitcast <8 x i16> %5 to <2 x double> ; <<2 x double>> [#uses=2]
28 %7 = extractelement <2 x double> %6, i32 0 ; <double> [#uses=1]
29 %8 = bitcast double %7 to <4 x i16> ; <<4 x i16>> [#uses=1]
30 %9 = sext <4 x i16> %8 to <4 x i32> ; <<4 x i32>> [#uses=1]
31 %10 = extractelement <2 x double> %6, i32 1 ; <double> [#uses=1]
32 %11 = bitcast double %10 to <4 x i16> ; <<4 x i16>> [#uses=1]
33 %12 = sext <4 x i16> %11 to <4 x i32> ; <<4 x i32>> [#uses=1]
34 %13 = mul <4 x i32> %1, %9 ; <<4 x i32>> [#uses=1]
35 %14 = mul <4 x i32> %3, %12 ; <<4 x i32>> [#uses=1]
36 %15 = tail call <4 x i16> @llvm.arm.neon.vshiftn.v4i16(<4 x i32> %13, <4 x i32> <i32 -12, i32 -12, i32 -12, i32 -12>) ; <<4 x i16>> [#uses=1]
37 %16 = tail call <4 x i16> @llvm.arm.neon.vshiftn.v4i16(<4 x i32> %14, <4 x i32> <i32 -12, i32 -12, i32 -12, i32 -12>) ; <<4 x i16>> [#uses=1]
38 %17 = shufflevector <4 x i16> %15, <4 x i16> %16, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7> ; <<8 x i16>> [#uses=1]
39 %18 = bitcast i16* %o_ptr to i8* ; <i8*> [#uses=1]
40 tail call void @llvm.arm.neon.vst1.v8i16(i8* %18, <8 x i16> %17, i32 1)
44 define void @t2(i16* %i_ptr, i16* %o_ptr, %struct.int16x8_t* nocapture %vT0ptr, %struct.int16x8_t* nocapture %vT1ptr) nounwind {
55 %0 = getelementptr inbounds %struct.int16x8_t* %vT0ptr, i32 0, i32 0 ; <<8 x i16>*> [#uses=1]
56 %1 = load <8 x i16>* %0, align 16 ; <<8 x i16>> [#uses=1]
57 %2 = getelementptr inbounds %struct.int16x8_t* %vT1ptr, i32 0, i32 0 ; <<8 x i16>*> [#uses=1]
58 %3 = load <8 x i16>* %2, align 16 ; <<8 x i16>> [#uses=1]
59 %4 = bitcast i16* %i_ptr to i8* ; <i8*> [#uses=1]
60 %5 = tail call <8 x i16> @llvm.arm.neon.vld1.v8i16(i8* %4, i32 1) ; <<8 x i16>> [#uses=1]
61 %6 = getelementptr inbounds i16* %i_ptr, i32 8 ; <i16*> [#uses=1]
62 %7 = bitcast i16* %6 to i8* ; <i8*> [#uses=1]
63 %8 = tail call <8 x i16> @llvm.arm.neon.vld1.v8i16(i8* %7, i32 1) ; <<8 x i16>> [#uses=1]
64 %9 = mul <8 x i16> %1, %5 ; <<8 x i16>> [#uses=1]
65 %10 = mul <8 x i16> %3, %8 ; <<8 x i16>> [#uses=1]
66 %11 = bitcast i16* %o_ptr to i8* ; <i8*> [#uses=1]
67 tail call void @llvm.arm.neon.vst1.v8i16(i8* %11, <8 x i16> %9, i32 1)
68 %12 = getelementptr inbounds i16* %o_ptr, i32 8 ; <i16*> [#uses=1]
69 %13 = bitcast i16* %12 to i8* ; <i8*> [#uses=1]
70 tail call void @llvm.arm.neon.vst1.v8i16(i8* %13, <8 x i16> %10, i32 1)
74 define <8 x i8> @t3(i8* %A, i8* %B) nounwind {
80 %tmp1 = call %struct.__neon_int8x8x3_t @llvm.arm.neon.vld3.v8i8(i8* %A, i32 1) ; <%struct.__neon_int8x8x3_t> [#uses=2]
81 %tmp2 = extractvalue %struct.__neon_int8x8x3_t %tmp1, 0 ; <<8 x i8>> [#uses=1]
82 %tmp3 = extractvalue %struct.__neon_int8x8x3_t %tmp1, 2 ; <<8 x i8>> [#uses=1]
83 %tmp4 = extractvalue %struct.__neon_int8x8x3_t %tmp1, 1 ; <<8 x i8>> [#uses=1]
84 %tmp5 = sub <8 x i8> %tmp3, %tmp4
85 %tmp6 = add <8 x i8> %tmp2, %tmp3 ; <<8 x i8>> [#uses=1]
86 %tmp7 = mul <8 x i8> %tmp4, %tmp2
87 tail call void @llvm.arm.neon.vst3.v8i8(i8* %B, <8 x i8> %tmp5, <8 x i8> %tmp6, <8 x i8> %tmp7, i32 1)
91 define void @t4(i32* %in, i32* %out) nounwind {
99 %tmp1 = bitcast i32* %in to i8* ; <i8*> [#uses=1]
100 %tmp2 = tail call %struct.__neon_int32x4x2_t @llvm.arm.neon.vld2.v4i32(i8* %tmp1, i32 1) ; <%struct.__neon_int32x4x2_t> [#uses=2]
101 %tmp3 = getelementptr inbounds i32* %in, i32 8 ; <i32*> [#uses=1]
102 %tmp4 = bitcast i32* %tmp3 to i8* ; <i8*> [#uses=1]
103 %tmp5 = tail call %struct.__neon_int32x4x2_t @llvm.arm.neon.vld2.v4i32(i8* %tmp4, i32 1) ; <%struct.__neon_int32x4x2_t> [#uses=2]
104 %tmp8 = bitcast i32* %out to i8* ; <i8*> [#uses=1]
105 br i1 undef, label %return1, label %return2
110 ; CHECK-NEXT: vadd.i32
111 ; CHECK-NEXT: vadd.i32
112 ; CHECK-NEXT: vst2.32
113 %tmp52 = extractvalue %struct.__neon_int32x4x2_t %tmp2, 0 ; <<4 x i32>> [#uses=1]
114 %tmp57 = extractvalue %struct.__neon_int32x4x2_t %tmp2, 1 ; <<4 x i32>> [#uses=1]
115 %tmp = extractvalue %struct.__neon_int32x4x2_t %tmp5, 0 ; <<4 x i32>> [#uses=1]
116 %tmp39 = extractvalue %struct.__neon_int32x4x2_t %tmp5, 1 ; <<4 x i32>> [#uses=1]
117 %tmp6 = add <4 x i32> %tmp52, %tmp ; <<4 x i32>> [#uses=1]
118 %tmp7 = add <4 x i32> %tmp57, %tmp39 ; <<4 x i32>> [#uses=1]
119 tail call void @llvm.arm.neon.vst2.v4i32(i8* %tmp8, <4 x i32> %tmp6, <4 x i32> %tmp7, i32 1)
127 ; CHECK: vst2.32 {d0, d1, d2, d3}
128 %tmp100 = extractvalue %struct.__neon_int32x4x2_t %tmp2, 0 ; <<4 x i32>> [#uses=1]
129 %tmp101 = extractvalue %struct.__neon_int32x4x2_t %tmp5, 1 ; <<4 x i32>> [#uses=1]
130 %tmp102 = add <4 x i32> %tmp100, %tmp101 ; <<4 x i32>> [#uses=1]
131 tail call void @llvm.arm.neon.vst2.v4i32(i8* %tmp8, <4 x i32> %tmp102, <4 x i32> %tmp101, i32 1)
132 call void @llvm.trap()
136 define <8 x i16> @t5(i16* %A, <8 x i16>* %B) nounwind {
141 ; CHECK: vld2.16 {d0[1], d2[1]}, [r0]
144 %tmp0 = bitcast i16* %A to i8* ; <i8*> [#uses=1]
145 %tmp1 = load <8 x i16>* %B ; <<8 x i16>> [#uses=2]
146 %tmp2 = call %struct.__neon_int16x8x2_t @llvm.arm.neon.vld2lane.v8i16(i8* %tmp0, <8 x i16> %tmp1, <8 x i16> %tmp1, i32 1, i32 1) ; <%struct.__neon_int16x8x2_t> [#uses=2]
147 %tmp3 = extractvalue %struct.__neon_int16x8x2_t %tmp2, 0 ; <<8 x i16>> [#uses=1]
148 %tmp4 = extractvalue %struct.__neon_int16x8x2_t %tmp2, 1 ; <<8 x i16>> [#uses=1]
149 %tmp5 = add <8 x i16> %tmp3, %tmp4 ; <<8 x i16>> [#uses=1]
153 define <8 x i8> @t6(i8* %A, <8 x i8>* %B) nounwind {
157 ; CHECK-NEXT: vld2.8 {d0[1], d1[1]}
158 %tmp1 = load <8 x i8>* %B ; <<8 x i8>> [#uses=2]
159 %tmp2 = call %struct.__neon_int8x8x2_t @llvm.arm.neon.vld2lane.v8i8(i8* %A, <8 x i8> %tmp1, <8 x i8> %tmp1, i32 1, i32 1) ; <%struct.__neon_int8x8x2_t> [#uses=2]
160 %tmp3 = extractvalue %struct.__neon_int8x8x2_t %tmp2, 0 ; <<8 x i8>> [#uses=1]
161 %tmp4 = extractvalue %struct.__neon_int8x8x2_t %tmp2, 1 ; <<8 x i8>> [#uses=1]
162 %tmp5 = add <8 x i8> %tmp3, %tmp4 ; <<8 x i8>> [#uses=1]
166 define void @t7(i32* %iptr, i32* %optr) nounwind {
171 ; CHECK: vld1.32 {d0, d1},
174 ; CHECK: vuzp.32 q0, q1
176 %0 = bitcast i32* %iptr to i8* ; <i8*> [#uses=2]
177 %1 = tail call %struct.__neon_int32x4x2_t @llvm.arm.neon.vld2.v4i32(i8* %0, i32 1) ; <%struct.__neon_int32x4x2_t> [#uses=2]
178 %tmp57 = extractvalue %struct.__neon_int32x4x2_t %1, 0 ; <<4 x i32>> [#uses=1]
179 %tmp60 = extractvalue %struct.__neon_int32x4x2_t %1, 1 ; <<4 x i32>> [#uses=1]
180 %2 = bitcast i32* %optr to i8* ; <i8*> [#uses=2]
181 tail call void @llvm.arm.neon.vst2.v4i32(i8* %2, <4 x i32> %tmp57, <4 x i32> %tmp60, i32 1)
182 %3 = tail call <4 x i32> @llvm.arm.neon.vld1.v4i32(i8* %0, i32 1) ; <<4 x i32>> [#uses=1]
183 %4 = shufflevector <4 x i32> %3, <4 x i32> undef, <4 x i32> <i32 0, i32 2, i32 0, i32 2> ; <<4 x i32>> [#uses=1]
184 tail call void @llvm.arm.neon.vst1.v4i32(i8* %2, <4 x i32> %4, i32 1)
189 define arm_aapcs_vfpcc i32 @t8() nounwind {
191 ; CHECK: vrsqrte.f32 q0, q0
192 bb.nph55.bb.nph55.split_crit_edge:
195 bb3: ; preds = %bb3, %bb.nph55.bb.nph55.split_crit_edge
196 br i1 undef, label %bb5, label %bb3
201 bb.i25: ; preds = %bb.i25, %bb5
202 %0 = shufflevector <2 x float> undef, <2 x float> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3> ; <<4 x float>> [#uses=1]
203 %1 = call <4 x float> @llvm.arm.neon.vrsqrte.v4f32(<4 x float> %0) nounwind ; <<4 x float>> [#uses=1]
204 %2 = fmul <4 x float> %1, undef ; <<4 x float>> [#uses=1]
205 %3 = fmul <4 x float> undef, %2 ; <<4 x float>> [#uses=1]
206 %tmp26.i = bitcast <4 x float> %3 to <2 x double> ; <<2 x double>> [#uses=1]
207 %4 = extractelement <2 x double> %tmp26.i, i32 0 ; <double> [#uses=1]
208 %5 = bitcast double %4 to <2 x float> ; <<2 x float>> [#uses=1]
209 %6 = extractelement <2 x float> %5, i32 1 ; <float> [#uses=1]
210 store float %6, float* undef, align 4
211 br i1 undef, label %bb6, label %bb.i25
213 bb6: ; preds = %bb.i25
214 br i1 undef, label %bb7, label %bb14
219 bb.i49: ; preds = %bb.i49, %bb7
220 br i1 undef, label %bb.i19, label %bb.i49
222 bb.i19: ; preds = %bb.i19, %bb.i49
223 br i1 undef, label %exit, label %bb.i19
225 exit: ; preds = %bb.i19
232 %0 = type { %1, %1, %1, %1 }
234 %2 = type { <4 x float> }
238 define arm_aapcs_vfpcc float @t9(%0* nocapture, %3* nocapture) nounwind {
241 ; CHECK-NOT: vmov d{{.*}}, d0
243 ; CHECK-NEXT: vstmia r0, {d0, d1}
244 ; CHECK-NEXT: vstmia r0, {d0, d1}
245 %3 = bitcast double 0.000000e+00 to <2 x float> ; <<2 x float>> [#uses=2]
246 %4 = shufflevector <2 x float> %3, <2 x float> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3> ; <<4 x float>> [#uses=1]
247 store <4 x float> %4, <4 x float>* undef, align 16
248 %5 = shufflevector <2 x float> %3, <2 x float> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 2, i32 3> ; <<4 x float>> [#uses=1]
249 store <4 x float> %5, <4 x float>* undef, align 16
252 ; <label>:6 ; preds = %8
255 ; <label>:7 ; preds = %6
258 ; <label>:8 ; preds = %7, %2
261 ; <label>:9 ; preds = %8
264 ; <label>:10 ; preds = %6
265 ret float 9.990000e+02
269 define arm_aapcs_vfpcc i32 @t10() nounwind {
272 ; CHECK: vmov.i32 q1, #0x3F000000
274 ; CHECK: vmla.f32 q0, q0, d0[0]
275 %0 = shufflevector <4 x float> zeroinitializer, <4 x float> undef, <4 x i32> zeroinitializer ; <<4 x float>> [#uses=1]
276 %1 = insertelement <4 x float> %0, float undef, i32 1 ; <<4 x float>> [#uses=1]
277 %2 = insertelement <4 x float> %1, float undef, i32 2 ; <<4 x float>> [#uses=1]
278 %3 = insertelement <4 x float> %2, float undef, i32 3 ; <<4 x float>> [#uses=1]
279 %tmp54.i = bitcast <4 x float> %3 to <2 x double> ; <<2 x double>> [#uses=1]
280 %4 = extractelement <2 x double> %tmp54.i, i32 1 ; <double> [#uses=1]
281 %5 = bitcast double %4 to <2 x float> ; <<2 x float>> [#uses=1]
282 %6 = shufflevector <2 x float> %5, <2 x float> undef, <4 x i32> zeroinitializer ; <<4 x float>> [#uses=1]
283 %7 = fmul <4 x float> undef, %6 ; <<4 x float>> [#uses=1]
284 %8 = fadd <4 x float> %7, undef ; <<4 x float>> [#uses=1]
285 %9 = fadd <4 x float> %8, undef ; <<4 x float>> [#uses=1]
286 %10 = shufflevector <4 x float> undef, <4 x float> %9, <4 x i32> <i32 0, i32 1, i32 2, i32 7> ; <<4 x float>> [#uses=1]
287 %11 = fmul <4 x float> %10, <float 5.000000e-01, float 5.000000e-01, float 5.000000e-01, float 5.000000e-01> ; <<4 x float>> [#uses=1]
288 %12 = shufflevector <4 x float> %11, <4 x float> undef, <4 x i32> <i32 3, i32 undef, i32 undef, i32 undef> ; <<4 x float>> [#uses=1]
289 %13 = shufflevector <4 x float> %12, <4 x float> undef, <4 x i32> zeroinitializer ; <<4 x float>> [#uses=1]
290 %14 = fmul <4 x float> %13, undef ; <<4 x float>> [#uses=1]
291 %15 = fadd <4 x float> undef, %14 ; <<4 x float>> [#uses=1]
292 %16 = shufflevector <4 x float> undef, <4 x float> %15, <4 x i32> <i32 0, i32 1, i32 6, i32 3> ; <<4 x float>> [#uses=1]
293 %17 = fmul <4 x float> %16, undef ; <<4 x float>> [#uses=1]
294 %18 = extractelement <4 x float> %17, i32 2 ; <float> [#uses=1]
295 store float %18, float* undef, align 4
296 br i1 undef, label %exit, label %bb14
298 exit: ; preds = %bb.i19
305 ; This test crashes the coalescer because live variables were not updated properly.
306 define <8 x i8> @t11(i8* %A1, i8* %A2, i8* %A3, i8* %A4, i8* %A5, i8* %A6, i8* %A7, i8* %A8, i8* %B) nounwind {
307 %tmp1d = call %struct.__neon_int8x8x3_t @llvm.arm.neon.vld3.v8i8(i8* %A4, i32 1) ; <%struct.__neon_int8x8x3_t> [#uses=1]
308 %tmp2d = extractvalue %struct.__neon_int8x8x3_t %tmp1d, 0 ; <<8 x i8>> [#uses=1]
309 %tmp1f = call %struct.__neon_int8x8x3_t @llvm.arm.neon.vld3.v8i8(i8* %A6, i32 1) ; <%struct.__neon_int8x8x3_t> [#uses=1]
310 %tmp2f = extractvalue %struct.__neon_int8x8x3_t %tmp1f, 0 ; <<8 x i8>> [#uses=1]
311 %tmp2bd = add <8 x i8> zeroinitializer, %tmp2d ; <<8 x i8>> [#uses=1]
312 %tmp2abcd = mul <8 x i8> zeroinitializer, %tmp2bd ; <<8 x i8>> [#uses=1]
313 %tmp2ef = sub <8 x i8> zeroinitializer, %tmp2f ; <<8 x i8>> [#uses=1]
314 %tmp2efgh = mul <8 x i8> %tmp2ef, undef ; <<8 x i8>> [#uses=2]
315 call void @llvm.arm.neon.vst3.v8i8(i8* %A2, <8 x i8> undef, <8 x i8> undef, <8 x i8> %tmp2efgh, i32 1)
316 %tmp2 = sub <8 x i8> %tmp2efgh, %tmp2abcd ; <<8 x i8>> [#uses=1]
317 %tmp7 = mul <8 x i8> undef, %tmp2 ; <<8 x i8>> [#uses=1]
318 tail call void @llvm.arm.neon.vst3.v8i8(i8* %B, <8 x i8> undef, <8 x i8> undef, <8 x i8> %tmp7, i32 1)
322 declare <4 x i32> @llvm.arm.neon.vld1.v4i32(i8*, i32) nounwind readonly
324 declare <8 x i16> @llvm.arm.neon.vld1.v8i16(i8*, i32) nounwind readonly
326 declare <4 x i16> @llvm.arm.neon.vshiftn.v4i16(<4 x i32>, <4 x i32>) nounwind readnone
328 declare void @llvm.arm.neon.vst1.v4i32(i8*, <4 x i32>, i32) nounwind
330 declare void @llvm.arm.neon.vst1.v8i16(i8*, <8 x i16>, i32) nounwind
332 declare void @llvm.arm.neon.vst3.v8i8(i8*, <8 x i8>, <8 x i8>, <8 x i8>, i32)
335 declare %struct.__neon_int8x8x3_t @llvm.arm.neon.vld3.v8i8(i8*, i32) nounwind readonly
337 declare %struct.__neon_int32x4x2_t @llvm.arm.neon.vld2.v4i32(i8*, i32) nounwind readonly
339 declare %struct.__neon_int8x8x2_t @llvm.arm.neon.vld2lane.v8i8(i8*, <8 x i8>, <8 x i8>, i32, i32) nounwind readonly
341 declare %struct.__neon_int16x8x2_t @llvm.arm.neon.vld2lane.v8i16(i8*, <8 x i16>, <8 x i16>, i32, i32) nounwind readonly
343 declare void @llvm.arm.neon.vst2.v4i32(i8*, <4 x i32>, <4 x i32>, i32) nounwind
345 declare <4 x float> @llvm.arm.neon.vrsqrte.v4f32(<4 x float>) nounwind readnone
347 declare void @llvm.trap() nounwind