Lines Matching full:tmp4
643 // CHECK: [[TMP4:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x float>
645 // CHECK: [[TMP6:%.*]] = call <2 x float> @llvm.fma.v2f32(<2 x float> [[TMP4]], <2 x float> [[TMP5]], <2 x float> [[TMP3]]) #4
656 // CHECK: [[TMP4:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x float>
658 // CHECK: [[TMP6:%.*]] = call <4 x float> @llvm.fma.v4f32(<4 x float> [[TMP4]], <4 x float> [[TMP5]], <4 x float> [[TMP3]]) #4
669 // CHECK: [[TMP4:%.*]] = bitcast <16 x i8> [[TMP1]] to <2 x double>
671 // CHECK: [[TMP6:%.*]] = call <2 x double> @llvm.fma.v2f64(<2 x double> [[TMP4]], <2 x double> [[TMP5]], <2 x double> [[TMP3]]) #4
682 // CHECK: [[TMP4:%.*]] = bitcast <8 x i8> [[TMP1]] to <2 x float>
684 // CHECK: [[TMP6:%.*]] = call <2 x float> @llvm.fma.v2f32(<2 x float> [[TMP4]], <2 x float> [[TMP5]], <2 x float> [[TMP3]]) #4
696 // CHECK: [[TMP4:%.*]] = bitcast <16 x i8> [[TMP1]] to <4 x float>
698 // CHECK: [[TMP6:%.*]] = call <4 x float> @llvm.fma.v4f32(<4 x float> [[TMP4]], <4 x float> [[TMP5]], <4 x float> [[TMP3]]) #4
710 // CHECK: [[TMP4:%.*]] = bitcast <16 x i8> [[TMP1]] to <2 x double>
712 // CHECK: [[TMP6:%.*]] = call <2 x double> @llvm.fma.v2f64(<2 x double> [[TMP4]], <2 x double> [[TMP5]], <2 x double> [[TMP3]]) #4
1038 // CHECK: [[TMP4:%.*]] = bitcast <4 x i16> [[VBSL5_I]] to <8 x i8>
1039 // CHECK: ret <8 x i8> [[TMP4]]
1143 // CHECK: [[TMP4:%.*]] = xor <2 x i32> [[VBSL_I]], <i32 -1, i32 -1>
1144 // CHECK: [[VBSL4_I:%.*]] = and <2 x i32> [[TMP4]], [[VBSL2_I]]
1163 // CHECK: [[TMP4:%.*]] = bitcast <1 x i64> [[VBSL5_I]] to <1 x double>
1164 // CHECK: ret <1 x double> [[TMP4]]
1322 // CHECK: [[TMP4:%.*]] = bitcast <4 x i32> [[VBSL5_I]] to <4 x float>
1323 // CHECK: ret <4 x float> [[TMP4]]
1365 // CHECK: [[TMP4:%.*]] = bitcast <2 x i64> [[VBSL5_I]] to <2 x double>
1366 // CHECK: ret <2 x double> [[TMP4]]
1645 // CHECK: [[TMP4:%.*]] = and <4 x i16> [[TMP2]], [[TMP3]]
1646 // CHECK: [[TMP5:%.*]] = icmp ne <4 x i16> [[TMP4]], zeroinitializer
1658 // CHECK: [[TMP4:%.*]] = and <2 x i32> [[TMP2]], [[TMP3]]
1659 // CHECK: [[TMP5:%.*]] = icmp ne <2 x i32> [[TMP4]], zeroinitializer
1680 // CHECK: [[TMP4:%.*]] = and <4 x i16> [[TMP2]], [[TMP3]]
1681 // CHECK: [[TMP5:%.*]] = icmp ne <4 x i16> [[TMP4]], zeroinitializer
1693 // CHECK: [[TMP4:%.*]] = and <2 x i32> [[TMP2]], [[TMP3]]
1694 // CHECK: [[TMP5:%.*]] = icmp ne <2 x i32> [[TMP4]], zeroinitializer
1715 // CHECK: [[TMP4:%.*]] = and <8 x i16> [[TMP2]], [[TMP3]]
1716 // CHECK: [[TMP5:%.*]] = icmp ne <8 x i16> [[TMP4]], zeroinitializer
1728 // CHECK: [[TMP4:%.*]] = and <4 x i32> [[TMP2]], [[TMP3]]
1729 // CHECK: [[TMP5:%.*]] = icmp ne <4 x i32> [[TMP4]], zeroinitializer
1750 // CHECK: [[TMP4:%.*]] = and <8 x i16> [[TMP2]], [[TMP3]]
1751 // CHECK: [[TMP5:%.*]] = icmp ne <8 x i16> [[TMP4]], zeroinitializer
1763 // CHECK: [[TMP4:%.*]] = and <4 x i32> [[TMP2]], [[TMP3]]
1764 // CHECK: [[TMP5:%.*]] = icmp ne <4 x i32> [[TMP4]], zeroinitializer
1776 // CHECK: [[TMP4:%.*]] = and <2 x i64> [[TMP2]], [[TMP3]]
1777 // CHECK: [[TMP5:%.*]] = icmp ne <2 x i64> [[TMP4]], zeroinitializer
1789 // CHECK: [[TMP4:%.*]] = and <2 x i64> [[TMP2]], [[TMP3]]
1790 // CHECK: [[TMP5:%.*]] = icmp ne <2 x i64> [[TMP4]], zeroinitializer
1811 // CHECK: [[TMP4:%.*]] = and <4 x i16> [[TMP2]], [[TMP3]]
1812 // CHECK: [[TMP5:%.*]] = icmp ne <4 x i16> [[TMP4]], zeroinitializer
1833 // CHECK: [[TMP4:%.*]] = and <8 x i16> [[TMP2]], [[TMP3]]
1834 // CHECK: [[TMP5:%.*]] = icmp ne <8 x i16> [[TMP4]], zeroinitializer
1846 // CHECK: [[TMP4:%.*]] = and <1 x i64> [[TMP2]], [[TMP3]]
1847 // CHECK: [[TMP5:%.*]] = icmp ne <1 x i64> [[TMP4]], zeroinitializer
1859 // CHECK: [[TMP4:%.*]] = and <1 x i64> [[TMP2]], [[TMP3]]
1860 // CHECK: [[TMP5:%.*]] = icmp ne <1 x i64> [[TMP4]], zeroinitializer
5507 // CHECK: [[TMP4:%.*]] = add <4 x i16> [[TMP2]], [[VSRA_N]]
5508 // CHECK: ret <4 x i16> [[TMP4]]
5519 // CHECK: [[TMP4:%.*]] = add <2 x i32> [[TMP2]], [[VSRA_N]]
5520 // CHECK: ret <2 x i32> [[TMP4]]
5539 // CHECK: [[TMP4:%.*]] = add <8 x i16> [[TMP2]], [[VSRA_N]]
5540 // CHECK: ret <8 x i16> [[TMP4]]
5551 // CHECK: [[TMP4:%.*]] = add <4 x i32> [[TMP2]], [[VSRA_N]]
5552 // CHECK: ret <4 x i32> [[TMP4]]
5563 // CHECK: [[TMP4:%.*]] = add <2 x i64> [[TMP2]], [[VSRA_N]]
5564 // CHECK: ret <2 x i64> [[TMP4]]
5583 // CHECK: [[TMP4:%.*]] = add <4 x i16> [[TMP2]], [[VSRA_N]]
5584 // CHECK: ret <4 x i16> [[TMP4]]
5595 // CHECK: [[TMP4:%.*]] = add <2 x i32> [[TMP2]], [[VSRA_N]]
5596 // CHECK: ret <2 x i32> [[TMP4]]
5615 // CHECK: [[TMP4:%.*]] = add <8 x i16> [[TMP2]], [[VSRA_N]]
5616 // CHECK: ret <8 x i16> [[TMP4]]
5627 // CHECK: [[TMP4:%.*]] = add <4 x i32> [[TMP2]], [[VSRA_N]]
5628 // CHECK: ret <4 x i32> [[TMP4]]
5639 // CHECK: [[TMP4:%.*]] = add <2 x i64> [[TMP2]], [[VSRA_N]]
5640 // CHECK: ret <2 x i64> [[TMP4]]
7324 // CHECK: [[TMP4:%.*]] = bitcast <8 x i8> [[TMP3]] to <4 x i16>
7325 // CHECK: [[TMP5:%.*]] = sext <4 x i16> [[TMP4]] to <4 x i32>
7339 // CHECK: [[TMP4:%.*]] = bitcast <8 x i8> [[TMP3]] to <2 x i32>
7340 // CHECK: [[TMP5:%.*]] = sext <2 x i32> [[TMP4]] to <2 x i64>
7365 // CHECK: [[TMP4:%.*]] = bitcast <8 x i8> [[TMP3]] to <4 x i16>
7366 // CHECK: [[TMP5:%.*]] = zext <4 x i16> [[TMP4]] to <4 x i32>
7380 // CHECK: [[TMP4:%.*]] = bitcast <8 x i8> [[TMP3]] to <2 x i32>
7381 // CHECK: [[TMP5:%.*]] = zext <2 x i32> [[TMP4]] to <2 x i64>
7594 // CHECK: [[TMP4:%.*]] = bitcast <8 x i8> [[TMP3]] to <4 x i16>
7595 // CHECK: [[TMP5:%.*]] = sext <4 x i16> [[TMP4]] to <4 x i32>
7609 // CHECK: [[TMP4:%.*]] = bitcast <8 x i8> [[TMP3]] to <2 x i32>
7610 // CHECK: [[TMP5:%.*]] = sext <2 x i32> [[TMP4]] to <2 x i64>
7635 // CHECK: [[TMP4:%.*]] = bitcast <8 x i8> [[TMP3]] to <4 x i16>
7636 // CHECK: [[TMP5:%.*]] = zext <4 x i16> [[TMP4]] to <4 x i32>
7650 // CHECK: [[TMP4:%.*]] = bitcast <8 x i8> [[TMP3]] to <2 x i32>
7651 // CHECK: [[TMP5:%.*]] = zext <2 x i32> [[TMP4]] to <2 x i64>
10126 // CHECK: [[TMP4:%.*]] = bitcast %struct.uint8x16x2_t* [[__RET]] to i8*
10127 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP3]], i8* [[TMP4]], i64 32, i32 16, i1 false)
10143 // CHECK: [[TMP4:%.*]] = bitcast %struct.uint16x8x2_t* [[RETVAL]] to i8*
10145 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 32, i32 16, i1 false)
10161 // CHECK: [[TMP4:%.*]] = bitcast %struct.uint32x4x2_t* [[RETVAL]] to i8*
10163 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 32, i32 16, i1 false)
10179 // CHECK: [[TMP4:%.*]] = bitcast %struct.uint64x2x2_t* [[RETVAL]] to i8*
10181 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 32, i32 16, i1 false)
10197 // CHECK: [[TMP4:%.*]] = bitcast %struct.int8x16x2_t* [[__RET]] to i8*
10198 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP3]], i8* [[TMP4]], i64 32, i32 16, i1 false)
10214 // CHECK: [[TMP4:%.*]] = bitcast %struct.int16x8x2_t* [[RETVAL]] to i8*
10216 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 32, i32 16, i1 false)
10232 // CHECK: [[TMP4:%.*]] = bitcast %struct.int32x4x2_t* [[RETVAL]] to i8*
10234 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 32, i32 16, i1 false)
10250 // CHECK: [[TMP4:%.*]] = bitcast %struct.int64x2x2_t* [[RETVAL]] to i8*
10252 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 32, i32 16, i1 false)
10268 // CHECK: [[TMP4:%.*]] = bitcast %struct.float16x8x2_t* [[RETVAL]] to i8*
10270 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 32, i32 16, i1 false)
10286 // CHECK: [[TMP4:%.*]] = bitcast %struct.float32x4x2_t* [[RETVAL]] to i8*
10288 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 32, i32 16, i1 false)
10304 // CHECK: [[TMP4:%.*]] = bitcast %struct.float64x2x2_t* [[RETVAL]] to i8*
10306 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 32, i32 16, i1 false)
10322 // CHECK: [[TMP4:%.*]] = bitcast %struct.poly8x16x2_t* [[__RET]] to i8*
10323 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP3]], i8* [[TMP4]], i64 32, i32 16, i1 false)
10339 // CHECK: [[TMP4:%.*]] = bitcast %struct.poly16x8x2_t* [[RETVAL]] to i8*
10341 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 32, i32 16, i1 false)
10357 // CHECK: [[TMP4:%.*]] = bitcast %struct.uint8x8x2_t* [[__RET]] to i8*
10358 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP3]], i8* [[TMP4]], i64 16, i32 8, i1 false)
10374 // CHECK: [[TMP4:%.*]] = bitcast %struct.uint16x4x2_t* [[RETVAL]] to i8*
10376 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 16, i32 8, i1 false)
10392 // CHECK: [[TMP4:%.*]] = bitcast %struct.uint32x2x2_t* [[RETVAL]] to i8*
10394 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 16, i32 8, i1 false)
10410 // CHECK: [[TMP4:%.*]] = bitcast %struct.uint64x1x2_t* [[RETVAL]] to i8*
10412 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 16, i32 8, i1 false)
10428 // CHECK: [[TMP4:%.*]] = bitcast %struct.int8x8x2_t* [[__RET]] to i8*
10429 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP3]], i8* [[TMP4]], i64 16, i32 8, i1 false)
10445 // CHECK: [[TMP4:%.*]] = bitcast %struct.int16x4x2_t* [[RETVAL]] to i8*
10447 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 16, i32 8, i1 false)
10463 // CHECK: [[TMP4:%.*]] = bitcast %struct.int32x2x2_t* [[RETVAL]] to i8*
10465 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 16, i32 8, i1 false)
10481 // CHECK: [[TMP4:%.*]] = bitcast %struct.int64x1x2_t* [[RETVAL]] to i8*
10483 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 16, i32 8, i1 false)
10499 // CHECK: [[TMP4:%.*]] = bitcast %struct.float16x4x2_t* [[RETVAL]] to i8*
10501 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 16, i32 8, i1 false)
10517 // CHECK: [[TMP4:%.*]] = bitcast %struct.float32x2x2_t* [[RETVAL]] to i8*
10519 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 16, i32 8, i1 false)
10535 // CHECK: [[TMP4:%.*]] = bitcast %struct.float64x1x2_t* [[RETVAL]] to i8*
10537 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 16, i32 8, i1 false)
10553 // CHECK: [[TMP4:%.*]] = bitcast %struct.poly8x8x2_t* [[__RET]] to i8*
10554 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP3]], i8* [[TMP4]], i64 16, i32 8, i1 false)
10570 // CHECK: [[TMP4:%.*]] = bitcast %struct.poly16x4x2_t* [[RETVAL]] to i8*
10572 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 16, i32 8, i1 false)
10588 // CHECK: [[TMP4:%.*]] = bitcast %struct.uint8x16x3_t* [[__RET]] to i8*
10589 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP3]], i8* [[TMP4]], i64 48, i32 16, i1 false)
10605 // CHECK: [[TMP4:%.*]] = bitcast %struct.uint16x8x3_t* [[RETVAL]] to i8*
10607 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 48, i32 16, i1 false)
10623 // CHECK: [[TMP4:%.*]] = bitcast %struct.uint32x4x3_t* [[RETVAL]] to i8*
10625 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 48, i32 16, i1 false)
10641 // CHECK: [[TMP4:%.*]] = bitcast %struct.uint64x2x3_t* [[RETVAL]] to i8*
10643 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 48, i32 16, i1 false)
10659 // CHECK: [[TMP4:%.*]] = bitcast %struct.int8x16x3_t* [[__RET]] to i8*
10660 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP3]], i8* [[TMP4]], i64 48, i32 16, i1 false)
10676 // CHECK: [[TMP4:%.*]] = bitcast %struct.int16x8x3_t* [[RETVAL]] to i8*
10678 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 48, i32 16, i1 false)
10694 // CHECK: [[TMP4:%.*]] = bitcast %struct.int32x4x3_t* [[RETVAL]] to i8*
10696 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 48, i32 16, i1 false)
10712 // CHECK: [[TMP4:%.*]] = bitcast %struct.int64x2x3_t* [[RETVAL]] to i8*
10714 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 48, i32 16, i1 false)
10730 // CHECK: [[TMP4
10732 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 48, i32 16, i1 false)
10748 // CHECK: [[TMP4:%.*]] = bitcast %struct.float32x4x3_t* [[RETVAL]] to i8*
10750 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 48, i32 16, i1 false)
10766 // CHECK: [[TMP4:%.*]] = bitcast %struct.float64x2x3_t* [[RETVAL]] to i8*
10768 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 48, i32 16, i1 false)
10784 // CHECK: [[TMP4:%.*]] = bitcast %struct.poly8x16x3_t* [[__RET]] to i8*
10785 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP3]], i8* [[TMP4]], i64 48, i32 16, i1 false)
10801 // CHECK: [[TMP4:%.*]] = bitcast %struct.poly16x8x3_t* [[RETVAL]] to i8*
10803 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 48, i32 16, i1 false)
10819 // CHECK: [[TMP4:%.*]] = bitcast %struct.uint8x8x3_t* [[__RET]] to i8*
10820 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP3]], i8* [[TMP4]], i64 24, i32 8, i1 false)
10836 // CHECK: [[TMP4:%.*]] = bitcast %struct.uint16x4x3_t* [[RETVAL]] to i8*
10838 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 24, i32 8, i1 false)
10854 // CHECK: [[TMP4:%.*]] = bitcast %struct.uint32x2x3_t* [[RETVAL]] to i8*
10856 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 24, i32 8, i1 false)
10872 // CHECK: [[TMP4:%.*]] = bitcast %struct.uint64x1x3_t* [[RETVAL]] to i8*
10874 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 24, i32 8, i1 false)
10890 // CHECK: [[TMP4:%.*]] = bitcast %struct.int8x8x3_t* [[__RET]] to i8*
10891 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP3]], i8* [[TMP4]], i64 24, i32 8, i1 false)
10907 // CHECK: [[TMP4:%.*]] = bitcast %struct.int16x4x3_t* [[RETVAL]] to i8*
10909 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 24, i32 8, i1 false)
10925 // CHECK: [[TMP4:%.*]] = bitcast %struct.int32x2x3_t* [[RETVAL]] to i8*
10927 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 24, i32 8, i1 false)
10943 // CHECK: [[TMP4:%.*]] = bitcast %struct.int64x1x3_t* [[RETVAL]] to i8*
10945 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 24, i32 8, i1 false)
10961 // CHECK: [[TMP4:%.*]] = bitcast %struct.float16x4x3_t* [[RETVAL]] to i8*
10963 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 24, i32 8, i1 false)
10979 // CHECK: [[TMP4:%.*]] = bitcast %struct.float32x2x3_t* [[RETVAL]] to i8*
10981 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 24, i32 8, i1 false)
10997 // CHECK: [[TMP4:%.*]] = bitcast %struct.float64x1x3_t* [[RETVAL]] to i8*
10999 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 24, i32 8, i1 false)
11015 // CHECK: [[TMP4:%.*]] = bitcast %struct.poly8x8x3_t* [[__RET]] to i8*
11016 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP3]], i8* [[TMP4]], i64 24, i32 8, i1 false)
11032 // CHECK: [[TMP4:%.*]] = bitcast %struct.poly16x4x3_t* [[RETVAL]] to i8*
11034 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 24, i32 8, i1 false)
11050 // CHECK: [[TMP4:%.*]] = bitcast %struct.uint8x16x4_t* [[__RET]] to i8*
11051 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP3]], i8* [[TMP4]], i64 64, i32 16, i1 false)
11067 // CHECK: [[TMP4:%.*]] = bitcast %struct.uint16x8x4_t* [[RETVAL]] to i8*
11069 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 64, i32 16, i1 false)
11085 // CHECK: [[TMP4:%.*]] = bitcast %struct.uint32x4x4_t* [[RETVAL]] to i8*
11087 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 64, i32 16, i1 false)
11103 // CHECK: [[TMP4:%.*]] = bitcast %struct.uint64x2x4_t* [[RETVAL]] to i8*
11105 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 64, i32 16, i1 false)
11121 // CHECK: [[TMP4:%.*]] = bitcast %struct.int8x16x4_t* [[__RET]] to i8*
11122 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP3]], i8* [[TMP4]], i64 64, i32 16, i1 false)
11138 // CHECK: [[TMP4:%.*]] = bitcast %struct.int16x8x4_t* [[RETVAL]] to i8*
11140 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 64, i32 16, i1 false)
11156 // CHECK: [[TMP4:%.*]] = bitcast %struct.int32x4x4_t* [[RETVAL]] to i8*
11158 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 64, i32 16, i1 false)
11174 // CHECK: [[TMP4:%.*]] = bitcast %struct.int64x2x4_t* [[RETVAL]] to i8*
11176 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 64, i32 16, i1 false)
11192 // CHECK: [[TMP4:%.*]] = bitcast %struct.float16x8x4_t* [[RETVAL]] to i8*
11194 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 64, i32 16, i1 false)
11210 // CHECK: [[TMP4:%.*]] = bitcast %struct.float32x4x4_t* [[RETVAL]] to i8*
11212 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 64, i32 16, i1 false)
11228 // CHECK: [[TMP4:%.*]] = bitcast %struct.float64x2x4_t* [[RETVAL]] to i8*
11230 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 64, i32 16, i1 false)
11246 // CHECK: [[TMP4:%.*]] = bitcast %struct.poly8x16x4_t* [[__RET]] to i8*
11247 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP3]], i8* [[TMP4]], i64 64, i32 16, i1 false)
11263 // CHECK: [[TMP4:%.*]] = bitcast %struct.poly16x8x4_t* [[RETVAL]] to i8*
11265 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 64, i32 16, i1 false)
11281 // CHECK: [[TMP4:%.*]] = bitcast %struct.uint8x8x4_t* [[__RET]] to i8*
11282 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP3]], i8* [[TMP4]], i64 32, i32 8, i1 false)
11298 // CHECK: [[TMP4:%.*]] = bitcast %struct.uint16x4x4_t* [[RETVAL]] to i8*
11300 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 32, i32 8, i1 false)
11316 // CHECK: [[TMP4:%.*]] = bitcast %struct.uint32x2x4_t* [[RETVAL]] to i8*
11318 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 32, i32 8, i1 false)
11334 // CHECK: [[TMP4:%.*]] = bitcast %struct.uint64x1x4_t* [[RETVAL]] to i8*
11336 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 32, i32 8, i1 false)
11352 // CHECK: [[TMP4:%.*]] = bitcast %struct.int8x8x4_t* [[__RET]] to i8*
11353 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP3]], i8* [[TMP4]], i64 32, i32 8, i1 false)
11369 // CHECK: [[TMP4:%.*]] = bitcast %struct.int16x4x4_t* [[RETVAL]] to i8*
11371 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 32, i32 8, i1 false)
11387 // CHECK: [[TMP4:%.*]] = bitcast %struct.int32x2x4_t* [[RETVAL]] to i8*
11389 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 32, i32 8, i1 false)
11405 // CHECK: [[TMP4:%.*]] = bitcast %struct.int64x1x4_t* [[RETVAL]] to i8*
11407 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 32, i32 8, i1 false)
11423 // CHECK: [[TMP4:%.*]] = bitcast %struct.float16x4x4_t* [[RETVAL]] to i8*
11425 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 32, i32 8, i1 false)
11441 // CHECK: [[TMP4:%.*]] = bitcast %struct.float32x2x4_t* [[RETVAL]] to i8*
11443 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 32, i32 8, i1 false)
11459 // CHECK: [[TMP4:%.*]] = bitcast %struct.float64x1x4_t* [[RETVAL]] to i8*
11461 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 32, i32 8, i1 false)
11477 // CHECK: [[TMP4:%.*]] = bitcast %struct.poly8x8x4_t* [[__RET]] to i8*
11478 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP3]], i8* [[TMP4]], i64 32, i32 8, i1 false)
11494 // CHECK: [[TMP4:%.*]] = bitcast %struct.poly16x4x4_t* [[RETVAL]] to i8*
11496 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 32, i32 8, i1 false)
11803 // CHECK: [[TMP4:%.*]] = bitcast <8 x i16> [[TMP3]] to <16 x i8>
11808 // CHECK: [[TMP7:%.*]] = bitcast <16 x i8> [[TMP4]] to <8 x i16>
11828 // CHECK: [[TMP4:%.*]] = bitcast <4 x i32> [[TMP3]] to <16 x i8>
11833 // CHECK: [[TMP7:%.*]] = bitcast <16 x i8> [[TMP4]] to <4 x i32>
11853 // CHECK: [[TMP4:%.*]] = bitcast <2 x i64> [[TMP3]] to <16 x i8>
11858 // CHECK: [[TMP7:%.*]] = bitcast <16 x i8> [[TMP4]] to <2 x i64>
11898 // CHECK: [[TMP4:%.*]] = bitcast <8 x i16> [[TMP3]] to <16 x i8>
11903 // CHECK: [[TMP7:%.*]] = bitcast <16 x i8> [[TMP4]] to <8 x i16>
11923 // CHECK: [[TMP4:%.*]] = bitcast <4 x i32> [[TMP3]] to <16 x i8>
11928 // CHECK: [[TMP7:%.*]] = bitcast <16 x i8> [[TMP4]] to <4 x i32>
11948 // CHECK: [[TMP4:%.*]] = bitcast <2 x i64> [[TMP3]] to <16 x i8>
11953 // CHECK: [[TMP7:%.*]] = bitcast <16 x i8> [[TMP4]] to <2 x i64>
11973 // CHECK: [[TMP4:%.*]] = bitcast <8 x half> [[TMP3]] to <16 x i8>
11978 // CHECK: [[TMP7:%.*]] = bitcast <16 x i8> [[TMP4]] to <8 x i16>
11998 // CHECK: [[TMP4:%.*]] = bitcast <4 x float> [[TMP3]] to <16 x i8>
12003 // CHECK: [[TMP7:%.*]] = bitcast <16 x i8> [[TMP4]] to <4 x float>
12023 // CHECK: [[TMP4:%.*]] = bitcast <2 x double> [[TMP3]] to <16 x i8>
12028 // CHECK: [[TMP7:%.*]] = bitcast <16 x i8> [[TMP4]] to <2 x double>
12068 // CHECK: [[TMP4:%.*]] = bitcast <8 x i16> [[TMP3]] to <16 x i8>
12073 // CHECK: [[TMP7:%.*]] = bitcast <16 x i8> [[TMP4]] to <8 x i16>
12113 // CHECK: [[TMP4:%.*]] = bitcast <4 x i16> [[TMP3]] to <8 x i8>
12118 // CHECK: [[TMP7:%.*]] = bitcast <8 x i8> [[TMP4]] to <4 x i16>
12138 // CHECK: [[TMP4:%.*]] = bitcast <2 x i32> [[TMP3]] to <8 x i8>
12143 // CHECK: [[TMP7:%.*]] = bitcast <8 x i8> [[TMP4]] to <2 x i32>
12163 // CHECK: [[TMP4:%.*]] = bitcast <1 x i64> [[TMP3]] to <8 x i8>
12168 // CHECK: [[TMP7:%.*]] = bitcast <8 x i8> [[TMP4]] to <1 x i64>
12208 // CHECK: [[TMP4:%.*]] = bitcast <4 x i16> [[TMP3]] to <8 x i8>
12213 // CHECK: [[TMP7:%.*]] = bitcast <8 x i8> [[TMP4]] to <4 x i16>
12233 // CHECK: [[TMP4:%.*]] = bitcast <2 x i32> [[TMP3]] to <8 x i8>
12238 // CHECK: [[TMP7:%.*]] = bitcast <8 x i8> [[TMP4]] to <2 x i32>
12258 // CHECK: [[TMP4:%.*]] = bitcast <1 x i64> [[TMP3]] to <8 x i8>
12263 // CHECK: [[TMP7:%.*]] = bitcast <8 x i8> [[TMP4]] to <1 x i64>
12283 // CHECK: [[TMP4:%.*]] = bitcast <4 x half> [[TMP3]] to <8 x i8>
12288 // CHECK: [[TMP7:%.*]] = bitcast <8 x i8> [[TMP4]] to <4 x i16>
12308 // CHECK: [[TMP4:%.*]] = bitcast <2 x float> [[TMP3]] to <8 x i8>
12313 // CHECK: [[TMP7:%.*]] = bitcast <8 x i8> [[TMP4]] to <2 x float>
12333 // CHECK: [[TMP4:%.*]] = bitcast <1 x double> [[TMP3]] to <8 x i8>
12338 // CHECK: [[TMP7:%.*]] = bitcast <8 x i8> [[TMP4]] to <1 x double>
12378 // CHECK: [[TMP4:%.*]] = bitcast <4 x i16> [[TMP3]] to <8 x i8>
12383 // CHECK: [[TMP7:%.*]] = bitcast <8 x i8> [[TMP4]] to <4 x i16>
12407 // CHECK: [[TMP4:%.*]] = load <16 x i8>, <16 x i8>* [[ARRAYIDX4]], align 16
12408 // CHECK: call void @llvm.aarch64.neon.st3.v16i8.p0i8(<16 x i8> [[TMP2]], <16 x i8> [[TMP3]], <16 x i8> [[TMP4]], i8* %a)
12426 // CHECK: [[TMP4:%.*]] = bitcast <8 x i16> [[TMP3]] to <16 x i8>
12435 // CHECK: [[TMP9:%.*]] = bitcast <16 x i8> [[TMP4]] to <8 x i16>
12456 // CHECK: [[TMP4:%.*]] = bitcast <4 x i32> [[TMP3]] to <16 x i8>
12465 // CHECK: [[TMP9:%.*]] = bitcast <16 x i8> [[TMP4]] to <4 x i32>
12486 // CHECK: [[TMP4:%.*]] = bitcast <2 x i64> [[TMP3]] to <16 x i8>
12495 // CHECK: [[TMP9:%.*]] = bitcast <16 x i8> [[TMP4]] to <2 x i64>
12520 // CHECK: [[TMP4:%.*]] = load <16 x i8>, <16 x i8>* [[ARRAYIDX4]], align 16
12521 // CHECK: call void @llvm.aarch64.neon.st3.v16i8.p0i8(<16 x i8> [[TMP2]], <16 x i8> [[TMP3]], <16 x i8> [[TMP4]], i8* %a)
12539 // CHECK: [[TMP4:%.*]] = bitcast <8 x i16> [[TMP3]] to <16 x i8>
12548 // CHECK: [[TMP9:%.*]] = bitcast <16 x i8> [[TMP4]] to <8 x i16>
12569 // CHECK: [[TMP4:%.*]] = bitcast <4 x i32> [[TMP3]] to <16 x i8>
12578 // CHECK: [[TMP9:%.*]] = bitcast <16 x i8> [[TMP4]] to <4 x i32>
12599 // CHECK: [[TMP4:%.*]] = bitcast <2 x i64> [[TMP3]] to <16 x i8>
12608 // CHECK: [[TMP9:%.*]] = bitcast <16 x i8> [[TMP4]] to <2 x i64>
12629 // CHECK: [[TMP4:%.*]] = bitcast <8 x half> [[TMP3]] to <16 x i8>
12638 // CHECK: [[TMP9:%.*]] = bitcast <16 x i8> [[TMP4]] to <8 x i16>
12659 // CHECK: [[TMP4:%.*]] = bitcast <4 x float> [[TMP3]] to <16 x i8>
12668 // CHECK: [[TMP9:%.*]] = bitcast <16 x i8> [[TMP4]] to <4 x float>
12689 // CHECK: [[TMP4:%.*]] = bitcast <2 x double> [[TMP3]] to <16 x i8>
12698 // CHECK: [[TMP9:%.*]] = bitcast <16 x i8> [[TMP4]] to <2 x double>
12723 // CHECK: [[TMP4:%.*]] = load <16 x i8>, <16 x i8>* [[ARRAYIDX4]], align 16
12724 // CHECK: call void @llvm.aarch64.neon.st3.v16i8.p0i8(<16 x i8> [[TMP2]], <16 x i8> [[TMP3]], <16 x i8> [[TMP4]], i8* %a)
12742 // CHECK: [[TMP4:%.*]] = bitcast <8 x i16> [[TMP3]] to <16 x i8>
12751 // CHECK: [[TMP9:%.*]] = bitcast <16 x i8> [[TMP4]] to <8 x i16>
12776 // CHECK: [[TMP4:%.*]] = load <8 x i8>, <8 x i8>* [[ARRAYIDX4]], align 8
12777 // CHECK: call void @llvm.aarch64.neon.st3.v8i8.p0i8(<8 x i8> [[TMP2]], <8 x i8> [[TMP3]], <8 x i8> [[TMP4]], i8* %a)
12795 // CHECK: [[TMP4:%.*]] = bitcast <4 x i16> [[TMP3]] to <8 x i8>
12804 // CHECK: [[TMP9:%.*]] = bitcast <8 x i8> [[TMP4]] to <4 x i16>
12825 // CHECK: [[TMP4:%.*]] = bitcast <2 x i32> [[TMP3]] to <8 x i8>
12834 // CHECK: [[TMP9:%.*]] = bitcast <8 x i8> [[TMP4]] to <2 x i32>
12855 // CHECK: [[TMP4:%.*]] = bitcast <1 x i64> [[TMP3]] to <8 x i8>
12864 // CHECK: [[TMP9:%.*]] = bitcast <8 x i8> [[TMP4]] to <1 x i64>
12889 // CHECK: [[TMP4:%.*]] = load <8 x i8>, <8 x i8>* [[ARRAYIDX4]], align 8
12890 // CHECK: call void @llvm.aarch64.neon.st3.v8i8.p0i8(<8 x i8> [[TMP2]], <8 x i8> [[TMP3]], <8 x i8> [[TMP4]], i8* %a)
12908 // CHECK: [[TMP4:%.*]] = bitcast <4 x i16> [[TMP3]] to <8 x i8>
12917 // CHECK: [[TMP9:%.*]] = bitcast <8 x i8> [[TMP4]] to <4 x i16>
12938 // CHECK: [[TMP4:%.*]] = bitcast <2 x i32> [[TMP3]] to <8 x i8>
12947 // CHECK: [[TMP9:%.*]] = bitcast <8 x i8> [[TMP4]] to <2 x i32>
12968 // CHECK: [[TMP4:%.*]] = bitcast <1 x i64> [[TMP3]] to <8 x i8>
12977 // CHECK: [[TMP9:%.*]] = bitcast <8 x i8> [[TMP4]] to <1 x i64>
12998 // CHECK: [[TMP4:%.*]] = bitcast <4 x half> [[TMP3]] to <8 x i8>
13007 // CHECK: [[TMP9:%.*]] = bitcast <8 x i8> [[TMP4]] to <4 x i16>
13028 // CHECK: [[TMP4:%.*]] = bitcast <2 x float> [[TMP3]] to <8 x i8>
13037 // CHECK: [[TMP9:%.*]] = bitcast <8 x i8> [[TMP4]] to <2 x float>
13058 // CHECK: [[TMP4:%.*]] = bitcast <1 x double> [[TMP3]] to <8 x i8>
13067 // CHECK: [[TMP9:%.*]] = bitcast <8 x i8> [[TMP4]] to <1 x double>
13092 // CHECK: [[TMP4:%.*]] = load <8 x i8>, <8 x i8>* [[ARRAYIDX4]], align 8
13093 // CHECK: call void @llvm.aarch64.neon.st3.v8i8.p0i8(<8 x i8> [[TMP2]], <8 x i8> [[TMP3]], <8 x i8> [[TMP4]], i8* %a)
13111 // CHECK: [[TMP4:%.*]] = bitcast <4 x i16> [[TMP3]] to <8 x i8>
13120 // CHECK: [[TMP9:%.*]] = bitcast <8 x i8> [[TMP4]] to <4 x i16>
13145 // CHECK: [[TMP4:%.*]] = load <16 x i8>, <16 x i8>* [[ARRAYIDX4]], align 16
13149 // CHECK: call void @llvm.aarch64.neon.st4.v16i8.p0i8(<16 x i8> [[TMP2]], <16 x i8> [[TMP3]], <16 x i8> [[TMP4]], <16 x i8> [[TMP5]], i8* %a)
13167 // CHECK: [[TMP4:%.*]] = bitcast <8 x i16> [[TMP3]] to <16 x i8>
13180 // CHECK: [[TMP11:%.*]] = bitcast <16 x i8> [[TMP4]] to <8 x i16>
13202 // CHECK: [[TMP4:%.*]] = bitcast <4 x i32> [[TMP3]] to <16 x i8>
13215 // CHECK: [[TMP11:%.*]] = bitcast <16 x i8> [[TMP4]] to <4 x i32>
13237 // CHECK: [[TMP4:%.*]] = bitcast <2 x i64> [[TMP3]] to <16 x i8>
13250 // CHECK: [[TMP11:%.*]] = bitcast <16 x i8> [[TMP4]] to <2 x i64>
13276 // CHECK: [[TMP4:%.*]] = load <16 x i8>, <16 x i8>* [[ARRAYIDX4]], align 16
13280 // CHECK: call void @llvm.aarch64.neon.st4.v16i8.p0i8(<16 x i8> [[TMP2]], <16 x i8> [[TMP3]], <16 x i8> [[TMP4]], <16 x i8> [[TMP5]], i8* %a)
13298 // CHECK: [[TMP4:%.*]] = bitcast <8 x i16> [[TMP3]] to <16 x i8>
13311 // CHECK: [[TMP11:%.*]] = bitcast <16 x i8> [[TMP4]] to <8 x i16>
13333 // CHECK: [[TMP4:%.*]] = bitcast <4 x i32> [[TMP3]] to <16 x i8>
13346 // CHECK: [[TMP11:%.*]] = bitcast <16 x i8> [[TMP4]] to <4 x i32>
13368 // CHECK: [[TMP4:%.*]] = bitcast <2 x i64> [[TMP3]] to <16 x i8>
13381 // CHECK: [[TMP11:%.*]] = bitcast <16 x i8> [[TMP4]] to <2 x i64>
13403 // CHECK: [[TMP4:%.*]] = bitcast <8 x half> [[TMP3]] to <16 x i8>
13416 // CHECK: [[TMP11:%.*]] = bitcast <16 x i8> [[TMP4]] to <8 x i16>
13438 // CHECK: [[TMP4:%.*]] = bitcast <4 x float> [[TMP3]] to <16 x i8>
13451 // CHECK: [[TMP11:%.*]] = bitcast <16 x i8> [[TMP4]] to <4 x float>
13473 // CHECK: [[TMP4:%.*]] = bitcast <2 x double> [[TMP3]] to <16 x i8>
13486 // CHECK: [[TMP11:%.*]] = bitcast <16 x i8> [[TMP4]] to <2 x double>
13512 // CHECK: [[TMP4:%.*]] = load <16 x i8>, <16 x i8>* [[ARRAYIDX4]], align 16
13516 // CHECK: call void @llvm.aarch64.neon.st4.v16i8.p0i8(<16 x i8> [[TMP2]], <16 x i8> [[TMP3]], <16 x i8> [[TMP4]], <16 x i8> [[TMP5]], i8* %a)
13534 // CHECK: [[TMP4:%.*]] = bitcast <8 x i16> [[TMP3]] to <16 x i8>
13547 // CHECK: [[TMP11:%.*]] = bitcast <16 x i8> [[TMP4]] to <8 x i16>
13573 // CHECK: [[TMP4:%.*]] = load <8 x i8>, <8 x i8>* [[ARRAYIDX4]], align 8
13577 // CHECK: call void @llvm.aarch64.neon.st4.v8i8.p0i8(<8 x i8> [[TMP2]], <8 x i8> [[TMP3]], <8 x i8> [[TMP4]], <8 x i8> [[TMP5]], i8* %a)
13595 // CHECK: [[TMP4:%.*]] = bitcast <4 x i16> [[TMP3]] to <8 x i8>
13608 // CHECK: [[TMP11:%.*]] = bitcast <8 x i8> [[TMP4]] to <4 x i16>
13630 // CHECK: [[TMP4:%.*]] = bitcast <2 x i32> [[TMP3]] to <8 x i8>
13643 // CHECK: [[TMP11:%.*]] = bitcast <8 x i8> [[TMP4]] to <2 x i32>
13665 // CHECK: [[TMP4:%.*]] = bitcast <1 x i64> [[TMP3]] to <8 x i8>
13678 // CHECK: [[TMP11:%.*]] = bitcast <8 x i8> [[TMP4]] to <1 x i64>
13704 // CHECK: [[TMP4:%.*]] = load <8 x i8>, <8 x i8>* [[ARRAYIDX4]], align 8
13708 // CHECK: call void @llvm.aarch64.neon.st4.v8i8.p0i8(<8 x i8> [[TMP2]], <8 x i8> [[TMP3]], <8 x i8> [[TMP4]], <8 x i8> [[TMP5]], i8* %a)
13726 // CHECK: [[TMP4:%.*]] = bitcast <4 x i16> [[TMP3]] to <8 x i8>
13739 // CHECK: [[TMP11:%.*]] = bitcast <8 x i8> [[TMP4]] to <4 x i16>
13761 // CHECK: [[TMP4:%.*]] = bitcast <2 x i32> [[TMP3]] to <8 x i8>
13774 // CHECK: [[TMP11:%.*]] = bitcast <8 x i8> [[TMP4]] to <2 x i32>
13796 // CHECK: [[TMP4:%.*]] = bitcast <1 x i64> [[TMP3]] to <8 x i8>
13809 // CHECK: [[TMP11:%.*]] = bitcast <8 x i8> [[TMP4]] to <1 x i64>
13831 // CHECK: [[TMP4:%.*]] = bitcast <4 x half> [[TMP3]] to <8 x i8>
13844 // CHECK: [[TMP11:%.*]] = bitcast <8 x i8> [[TMP4]] to <4 x i16>
13866 // CHECK: [[TMP4:%.*]] = bitcast <2 x float> [[TMP3]] to <8 x i8>
13879 // CHECK: [[TMP11:%.*]] = bitcast <8 x i8> [[TMP4]] to <2 x float>
13901 // CHECK: [[TMP4:%.*]] = bitcast <1 x double> [[TMP3]] to <8 x i8>
13914 // CHECK: [[TMP11:%.*]] = bitcast <8 x i8> [[TMP4]] to <1 x double>
13940 // CHECK: [[TMP4:%.*]] = load <8 x i8>, <8 x i8>* [[ARRAYIDX4]], align 8
13944 // CHECK: call void @llvm.aarch64.neon.st4.v8i8.p0i8(<8 x i8> [[TMP2]], <8 x i8> [[TMP3]], <8 x i8> [[TMP4]], <8 x i8> [[TMP5]], i8* %a)
13962 // CHECK: [[TMP4:%.*]] = bitcast <4 x i16> [[TMP3]] to <8 x i8>
13975 // CHECK: [[TMP11:%.*]] = bitcast <8 x i8> [[TMP4]] to <4 x i16>
13995 // CHECK: [[TMP4:%.*]] = load %struct.uint8x16x2_t, %struct.uint8x16x2_t* [[RETVAL]], align 16
13996 // CHECK: ret %struct.uint8x16x2_t [[TMP4]]
14010 // CHECK: [[TMP4:%.*]] = bitcast %struct.uint16x8x2_t* [[RETVAL]] to i8*
14012 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 32, i32 16, i1 false)
14028 // CHECK: [[TMP4:%.*]] = bitcast %struct.uint32x4x2_t* [[RETVAL]] to i8*
14030 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 32, i32 16, i1 false)
14046 // CHECK: [[TMP4:%.*]] = bitcast %struct.uint64x2x2_t* [[RETVAL]] to i8*
14048 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 32, i32 16, i1 false)
14065 // CHECK: [[TMP4:%.*]] = load %struct.int8x16x2_t, %struct.int8x16x2_t* [[RETVAL]], align 16
14066 // CHECK: ret %struct.int8x16x2_t [[TMP4]]
14080 // CHECK: [[TMP4:%.*]] = bitcast %struct.int16x8x2_t* [[RETVAL]] to i8*
14082 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 32, i32 16, i1 false)
14098 // CHECK: [[TMP4:%.*]] = bitcast %struct.int32x4x2_t* [[RETVAL]] to i8*
14100 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 32, i32 16, i1 false)
14116 // CHECK: [[TMP4:%.*]] = bitcast %struct.int64x2x2_t* [[RETVAL]] to i8*
14118 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 32, i32 16, i1 false)
14134 // CHECK: [[TMP4:%.*]] = bitcast %struct.float16x8x2_t* [[RETVAL]] to i8*
14136 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 32, i32 16, i1 false)
14152 // CHECK: [[TMP4:%.*]] = bitcast %struct.float32x4x2_t* [[RETVAL]] to i8*
14154 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 32, i32 16, i1 false)
14170 // CHECK: [[TMP4:%.*]] = bitcast %struct.float64x2x2_t* [[RETVAL]] to i8*
14172 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 32, i32 16, i1 false)
14189 // CHECK: [[TMP4:%.*]] = load %struct.poly8x16x2_t, %struct.poly8x16x2_t* [[RETVAL]], align 16
14190 // CHECK: ret %struct.poly8x16x2_t [[TMP4]]
14204 // CHECK: [[TMP4:%.*]] = bitcast %struct.poly16x8x2_t* [[RETVAL]] to i8*
14206 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 32, i32 16, i1 false)
14222 // CHECK: [[TMP4:%.*]] = bitcast %struct.poly64x2x2_t* [[RETVAL]] to i8*
14224 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 32, i32 16, i1 false)
14241 // CHECK: [[TMP4:%.*]] = load %struct.uint8x8x2_t, %struct.uint8x8x2_t* [[RETVAL]], align 8
14242 // CHECK: ret %struct.uint8x8x2_t [[TMP4]]
14256 // CHECK: [[TMP4:%.*]] = bitcast %struct.uint16x4x2_t* [[RETVAL]] to i8*
14258 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 16, i32 8, i1 false)
14274 // CHECK: [[TMP4:%.*]] = bitcast %struct.uint32x2x2_t* [[RETVAL]] to i8*
14276 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 16, i32 8, i1 false)
14292 // CHECK: [[TMP4:%.*]] = bitcast %struct.uint64x1x2_t* [[RETVAL]] to i8*
14294 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 16, i32 8, i1 false)
14311 // CHECK: [[TMP4:%.*]] = load %struct.int8x8x2_t, %struct.int8x8x2_t* [[RETVAL]], align 8
14312 // CHECK: ret %struct.int8x8x2_t [[TMP4]]
14326 // CHECK: [[TMP4:%.*]] = bitcast %struct.int16x4x2_t* [[RETVAL]] to i8*
14328 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 16, i32 8, i1 false)
14344 // CHECK: [[TMP4:%.*]] = bitcast %struct.int32x2x2_t* [[RETVAL]] to i8*
14346 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 16, i32 8, i1 false)
14362 // CHECK: [[TMP4:%.*]] = bitcast %struct.int64x1x2_t* [[RETVAL]] to i8*
14364 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 16, i32 8, i1 false)
14380 // CHECK: [[TMP4:%.*]] = bitcast %struct.float16x4x2_t* [[RETVAL]] to i8*
14382 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 16, i32 8, i1 false)
14398 // CHECK: [[TMP4:%.*]] = bitcast %struct.float32x2x2_t* [[RETVAL]] to i8*
14400 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 16, i32 8, i1 false)
14416 // CHECK: [[TMP4:%.*]] = bitcast %struct.float64x1x2_t* [[RETVAL]] to i8*
14418 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 16, i32 8, i1 false)
14435 // CHECK: [[TMP4:%.*]] = load %struct.poly8x8x2_t, %struct.poly8x8x2_t* [[RETVAL]], align 8
14436 // CHECK: ret %struct.poly8x8x2_t [[TMP4]]
14450 // CHECK: [[TMP4:%.*]] = bitcast %struct.poly16x4x2_t* [[RETVAL]] to i8*
14452 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 16, i32 8, i1 false)
14468 // CHECK: [[TMP4
14470 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 16, i32 8, i1 false)
14487 // CHECK: [[TMP4:%.*]] = load %struct.uint8x16x3_t, %struct.uint8x16x3_t* [[RETVAL]], align 16
14488 // CHECK: ret %struct.uint8x16x3_t [[TMP4]]
14502 // CHECK: [[TMP4:%.*]] = bitcast %struct.uint16x8x3_t* [[RETVAL]] to i8*
14504 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 48, i32 16, i1 false)
14520 // CHECK: [[TMP4:%.*]] = bitcast %struct.uint32x4x3_t* [[RETVAL]] to i8*
14522 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 48, i32 16, i1 false)
14538 // CHECK: [[TMP4:%.*]] = bitcast %struct.uint64x2x3_t* [[RETVAL]] to i8*
14540 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 48, i32 16, i1 false)
14557 // CHECK: [[TMP4:%.*]] = load %struct.int8x16x3_t, %struct.int8x16x3_t* [[RETVAL]], align 16
14558 // CHECK: ret %struct.int8x16x3_t [[TMP4]]
14572 // CHECK: [[TMP4:%.*]] = bitcast %struct.int16x8x3_t* [[RETVAL]] to i8*
14574 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 48, i32 16, i1 false)
14590 // CHECK: [[TMP4:%.*]] = bitcast %struct.int32x4x3_t* [[RETVAL]] to i8*
14592 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 48, i32 16, i1 false)
14608 // CHECK: [[TMP4:%.*]] = bitcast %struct.int64x2x3_t* [[RETVAL]] to i8*
14610 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 48, i32 16, i1 false)
14626 // CHECK: [[TMP4:%.*]] = bitcast %struct.float16x8x3_t* [[RETVAL]] to i8*
14628 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 48, i32 16, i1 false)
14644 // CHECK: [[TMP4:%.*]] = bitcast %struct.float32x4x3_t* [[RETVAL]] to i8*
14646 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 48, i32 16, i1 false)
14662 // CHECK: [[TMP4:%.*]] = bitcast %struct.float64x2x3_t* [[RETVAL]] to i8*
14664 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 48, i32 16, i1 false)
14681 // CHECK: [[TMP4:%.*]] = load %struct.poly8x16x3_t, %struct.poly8x16x3_t* [[RETVAL]], align 16
14682 // CHECK: ret %struct.poly8x16x3_t [[TMP4]]
14696 // CHECK: [[TMP4:%.*]] = bitcast %struct.poly16x8x3_t* [[RETVAL]] to i8*
14698 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 48, i32 16, i1 false)
14714 // CHECK: [[TMP4
14716 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 48, i32 16, i1 false)
14733 // CHECK: [[TMP4:%.*]] = load %struct.uint8x8x3_t, %struct.uint8x8x3_t* [[RETVAL]], align 8
14734 // CHECK: ret %struct.uint8x8x3_t [[TMP4]]
14748 // CHECK: [[TMP4:%.*]] = bitcast %struct.uint16x4x3_t* [[RETVAL]] to i8*
14750 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 24, i32 8, i1 false)
14766 // CHECK: [[TMP4:%.*]] = bitcast %struct.uint32x2x3_t* [[RETVAL]] to i8*
14768 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 24, i32 8, i1 false)
14784 // CHECK: [[TMP4:%.*]] = bitcast %struct.uint64x1x3_t* [[RETVAL]] to i8*
14786 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 24, i32 8, i1 false)
14803 // CHECK: [[TMP4:%.*]] = load %struct.int8x8x3_t, %struct.int8x8x3_t* [[RETVAL]], align 8
14804 // CHECK: ret %struct.int8x8x3_t [[TMP4]]
14818 // CHECK: [[TMP4:%.*]] = bitcast %struct.int16x4x3_t* [[RETVAL]] to i8*
14820 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 24, i32 8, i1 false)
14836 // CHECK: [[TMP4:%.*]] = bitcast %struct.int32x2x3_t* [[RETVAL]] to i8*
14838 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 24, i32 8, i1 false)
14854 // CHECK: [[TMP4:%.*]] = bitcast %struct.int64x1x3_t* [[RETVAL]] to i8*
14856 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 24, i32 8, i1 false)
14872 // CHECK: [[TMP4:%.*]] = bitcast %struct.float16x4x3_t* [[RETVAL]] to i8*
14874 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 24, i32 8, i1 false)
14890 // CHECK: [[TMP4:%.*]] = bitcast %struct.float32x2x3_t* [[RETVAL]] to i8*
14892 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 24, i32 8, i1 false)
14908 // CHECK: [[TMP4:%.*]] = bitcast %struct.float64x1x3_t* [[RETVAL]] to i8*
14910 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 24, i32 8, i1 false)
14927 // CHECK: [[TMP4:%.*]] = load %struct.poly8x8x3_t, %struct.poly8x8x3_t* [[RETVAL]], align 8
14928 // CHECK: ret %struct.poly8x8x3_t [[TMP4]]
14942 // CHECK: [[TMP4:%.*]] = bitcast %struct.poly16x4x3_t* [[RETVAL]] to i8*
14944 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 24, i32 8, i1 false)
14960 // CHECK: [[TMP4:%.*]] = bitcast %struct.poly64x1x3_t* [[RETVAL]] to i8*
14962 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 24, i32 8, i1 false)
14979 // CHECK: [[TMP4:%.*]] = load %struct.uint8x16x4_t, %struct.uint8x16x4_t* [[RETVAL]], align 16
14980 // CHECK: ret %struct.uint8x16x4_t [[TMP4]]
14994 // CHECK: [[TMP4:%.*]] = bitcast %struct.uint16x8x4_t* [[RETVAL]] to i8*
14996 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 64, i32 16, i1 false)
15012 // CHECK: [[TMP4:%.*]] = bitcast %struct.uint32x4x4_t* [[RETVAL]] to i8*
15014 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 64, i32 16, i1 false)
15030 // CHECK: [[TMP4:%.*]] = bitcast %struct.uint64x2x4_t* [[RETVAL]] to i8*
15032 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 64, i32 16, i1 false)
15049 // CHECK: [[TMP4:%.*]] = load %struct.int8x16x4_t, %struct.int8x16x4_t* [[RETVAL]], align 16
15050 // CHECK: ret %struct.int8x16x4_t [[TMP4]]
15064 // CHECK: [[TMP4:%.*]] = bitcast %struct.int16x8x4_t* [[RETVAL]] to i8*
15066 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 64, i32 16, i1 false)
15082 // CHECK: [[TMP4:%.*]] = bitcast %struct.int32x4x4_t* [[RETVAL]] to i8*
15084 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 64, i32 16, i1 false)
15100 // CHECK: [[TMP4:%.*]] = bitcast %struct.int64x2x4_t* [[RETVAL]] to i8*
15102 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 64, i32 16, i1 false)
15118 // CHECK: [[TMP4:%.*]] = bitcast %struct.float16x8x4_t* [[RETVAL]] to i8*
15120 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 64, i32 16, i1 false)
15136 // CHECK: [[TMP4:%.*]] = bitcast %struct.float32x4x4_t* [[RETVAL]] to i8*
15138 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 64, i32 16, i1 false)
15154 // CHECK: [[TMP4:%.*]] = bitcast %struct.float64x2x4_t* [[RETVAL]] to i8*
15156 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 64, i32 16, i1 false)
15173 // CHECK: [[TMP4:%.*]] = load %struct.poly8x16x4_t, %struct.poly8x16x4_t* [[RETVAL]], align 16
15174 // CHECK: ret %struct.poly8x16x4_t [[TMP4]]
15188 // CHECK: [[TMP4:%.*]] = bitcast %struct.poly16x8x4_t* [[RETVAL]] to i8*
15190 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 64, i32 16, i1 false)
15206 // CHECK: [[TMP4:%.*]] = bitcast %struct.poly64x2x4_t* [[RETVAL]] to i8*
15208 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 64, i32 16, i1 false)
15225 // CHECK: [[TMP4:%.*]] = load %struct.uint8x8x4_t, %struct.uint8x8x4_t* [[RETVAL]], align 8
15226 // CHECK: ret %struct.uint8x8x4_t [[TMP4]]
15240 // CHECK: [[TMP4:%.*]] = bitcast %struct.uint16x4x4_t* [[RETVAL]] to i8*
15242 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 32, i32 8, i1 false)
15258 // CHECK: [[TMP4:%.*]] = bitcast %struct.uint32x2x4_t* [[RETVAL]] to i8*
15260 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 32, i32 8, i1 false)
15276 // CHECK: [[TMP4:%.*]] = bitcast %struct.uint64x1x4_t* [[RETVAL]] to i8*
15278 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 32, i32 8, i1 false)
15295 // CHECK: [[TMP4:%.*]] = load %struct.int8x8x4_t, %struct.int8x8x4_t* [[RETVAL]], align 8
15296 // CHECK: ret %struct.int8x8x4_t [[TMP4]]
15310 // CHECK: [[TMP4:%.*]] = bitcast %struct.int16x4x4_t* [[RETVAL]] to i8*
15312 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 32, i32 8, i1 false)
15328 // CHECK: [[TMP4:%.*]] = bitcast %struct.int32x2x4_t* [[RETVAL]] to i8*
15330 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 32, i32 8, i1 false)
15346 // CHECK: [[TMP4:%.*]] = bitcast %struct.int64x1x4_t* [[RETVAL]] to i8*
15348 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 32, i32 8, i1 false)
15364 // CHECK: [[TMP4:%.*]] = bitcast %struct.float16x4x4_t* [[RETVAL]] to i8*
15366 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 32, i32 8, i1 false)
15382 // CHECK: [[TMP4:%.*]] = bitcast %struct.float32x2x4_t* [[RETVAL]] to i8*
15384 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 32, i32 8, i1 false)
15400 // CHECK: [[TMP4:%.*]] = bitcast %struct.float64x1x4_t* [[RETVAL]] to i8*
15402 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 32, i32 8, i1 false)
15419 // CHECK: [[TMP4:%.*]] = load %struct.poly8x8x4_t, %struct.poly8x8x4_t* [[RETVAL]], align 8
15420 // CHECK: ret %struct.poly8x8x4_t [[TMP4]]
15434 // CHECK: [[TMP4:%.*]] = bitcast %struct.poly16x4x4_t* [[RETVAL]] to i8*
15436 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 32, i32 8, i1 false)
15452 // CHECK: [[TMP4:%.*]] = bitcast %struct.poly64x1x4_t* [[RETVAL]] to i8*
15454 // CHECK: call void @llvm.memcpy.p0i8.p0i8.i64(i8* [[TMP4]], i8* [[TMP5]], i64 32, i32 8, i1 false)
15493 // CHECK: [[TMP4:%.*]] = bitcast <8 x i16> [[TMP3]] to <16 x i8>
15498 // CHECK: [[TMP7:%.*]] = bitcast <16 x i8> [[TMP4]] to <8 x i16>
15519 // CHECK: [[TMP4:%.*]] = bitcast <4 x i32> [[TMP3]] to <16 x i8>
15524 // CHECK: [[TMP7:%.*]] = bitcast <16 x i8> [[TMP4]] to <4 x i32>
15545 // CHECK: [[TMP4:%.*]] = bitcast <2 x i64> [[TMP3]] to <16 x i8>
15550 // CHECK: [[TMP7:%.*]] = bitcast <16 x i8> [[TMP4]] to <2 x i64>
15591 // CHECK: [[TMP4:%.*]] = bitcast <8 x i16> [[TMP3]] to <16 x i8>
15596 // CHECK: [[TMP7:%.*]] = bitcast <16 x i8> [[TMP4]] to <8 x i16>
15617 // CHECK: [[TMP4:%.*]] = bitcast <4 x i32> [[TMP3]] to <16 x i8>
15622 // CHECK: [[TMP7:%.*]] = bitcast <16 x i8> [[TMP4]] to <4 x i32>
15643 // CHECK: [[TMP4:%.*]] = bitcast <2 x i64> [[TMP3]] to <16 x i8>
15648 // CHECK: [[TMP7:%.*]] = bitcast <16 x i8> [[TMP4]] to <2 x i64>
15669 // CHECK: [[TMP4:%.*]] = bitcast <8 x half> [[TMP3]] to <16 x i8>
15674 // CHECK: [[TMP7:%.*]] = bitcast <16 x i8> [[TMP4]] to <8 x i16>
15695 // CHECK: [[TMP4:%.*]] = bitcast <4 x float> [[TMP3]] to <16 x i8>
15700 // CHECK: [[TMP7:%.*]] = bitcast <16 x i8> [[TMP4]] to <4 x float>
15721 // CHECK: [[TMP4:%.*]] = bitcast <2 x double> [[TMP3]] to <16 x i8>
15726 // CHECK: [[TMP7:%.*]] = bitcast <16 x i8> [[TMP4]] to <2 x double>
15767 // CHECK: [[TMP4:%.*]] = bitcast <8 x i16> [[TMP3]] to <16 x i8>
15772 // CHECK: [[TMP7:%.*]] = bitcast <16 x i8> [[TMP4]] to <8 x i16>
15793 // CHECK: [[TMP4:%.*]] = bitcast <2 x i64> [[TMP3]] to <16 x i8>
15798 // CHECK: [[TMP7:%.*]] = bitcast <16 x i8> [[TMP4]] to <2 x i64>
15839 // CHECK: [[TMP4:%.*]] = bitcast <4 x i16> [[TMP3]] to <8 x i8>
15844 // CHECK: [[TMP7:%.*]] = bitcast <8 x i8> [[TMP4]] to <4 x i16>
15865 // CHECK: [[TMP4:%.*]] = bitcast <2 x i32> [[TMP3]] to <8 x i8>
15870 // CHECK: [[TMP7:%.*]] = bitcast <8 x i8> [[TMP4]] to <2 x i32>
15891 // CHECK: [[TMP4:%.*]] = bitcast <1 x i64> [[TMP3]] to <8 x i8>
15896 // CHECK: [[TMP7:%.*]] = bitcast <8 x i8> [[TMP4]] to <1 x i64>
15937 // CHECK: [[TMP4:%.*]] = bitcast <4 x i16> [[TMP3]] to <8 x i8>
15942 // CHECK: [[TMP7:%.*]] = bitcast <8 x i8> [[TMP4]] to <4 x i16>
15963 // CHECK: [[TMP4:%.*]] = bitcast <2 x i32> [[TMP3]] to <8 x i8>
15968 // CHECK: [[TMP7:%.*]] = bitcast <8 x i8> [[TMP4]] to <2 x i32>
15989 // CHECK: [[TMP4:%.*]] = bitcast <1 x i64> [[TMP3]] to <8 x i8>
15994 // CHECK: [[TMP7:%.*]] = bitcast <8 x i8> [[TMP4]] to <1 x i64>
16015 // CHECK: [[TMP4:%.*]] = bitcast <4 x half> [[TMP3]] to <8 x i8>
16020 // CHECK: [[TMP7:%.*]] = bitcast <8 x i8> [[TMP4]] to <4 x i16>
16041 // CHECK: [[TMP4:%.*]] = bitcast <2 x float> [[TMP3]] to <8 x i8>
16046 // CHECK: [[TMP7:%.*]] = bitcast <8 x i8> [[TMP4]] to <2 x float>
16067 // CHECK: [[TMP4:%.*]] = bitcast <1 x double> [[TMP3]] to <8 x i8>
16072 // CHECK: [[TMP7:%.*]] = bitcast <8 x i8> [[TMP4]] to <1 x double>
16113 // CHECK: [[TMP4:%.*]] = bitcast <4 x i16> [[TMP3]] to <8 x i8>
16118 // CHECK: [[TMP7:%.*]] = bitcast <8 x i8> [[TMP4]] to <4 x i16>
16139 // CHECK: [[TMP4:%.*]] = bitcast <1 x i64> [[TMP3]] to <8 x i8>
16144 // CHECK: [[TMP7:%.*]] = bitcast <8 x i8> [[TMP4]] to <1 x i64>
16169 // CHECK: [[TMP4:%.*]] = load <16 x i8>, <16 x i8>* [[ARRAYIDX4]], align 16
16170 // CHECK: call void @llvm.aarch64.neon.st1x3.v16i8.p0i8(<16 x i8> [[TMP2]], <16 x i8> [[TMP3]], <16 x i8> [[TMP4]], i8* %a)
16188 // CHECK: [[TMP4:%.*]] = bitcast <8 x i16> [[TMP3]] to <16 x i8>
16197 // CHECK: [[TMP9:%.*]] = bitcast <16 x i8> [[TMP4]] to <8 x i16>
16219 // CHECK: [[TMP4:%.*]] = bitcast <4 x i32> [[TMP3]] to <16 x i8>
16228 // CHECK: [[TMP9:%.*]] = bitcast <16 x i8> [[TMP4]] to <4 x i32>
16250 // CHECK: [[TMP4:%.*]] = bitcast <2 x i64> [[TMP3]] to <16 x i8>
16259 // CHECK: [[TMP9:%.*]] = bitcast <16 x i8> [[TMP4]] to <2 x i64>
16285 // CHECK: [[TMP4:%.*]] = load <16 x i8>, <16 x i8>* [[ARRAYIDX4]], align 16
16286 // CHECK: call void @llvm.aarch64.neon.st1x3.v16i8.p0i8(<16 x i8> [[TMP2]], <16 x i8> [[TMP3]], <16 x i8> [[TMP4]], i8* %a)
16304 // CHECK: [[TMP4:%.*]] = bitcast <8 x i16> [[TMP3]] to <16 x i8>
16313 // CHECK: [[TMP9:%.*]] = bitcast <16 x i8> [[TMP4]] to <8 x i16>
16335 // CHECK: [[TMP4:%.*]] = bitcast <4 x i32> [[TMP3]] to <16 x i8>
16344 // CHECK: [[TMP9:%.*]] = bitcast <16 x i8> [[TMP4]] to <4 x i32>
16366 // CHECK: [[TMP4:%.*]] = bitcast <2 x i64> [[TMP3]] to <16 x i8>
16375 // CHECK: [[TMP9:%.*]] = bitcast <16 x i8> [[TMP4]] to <2 x i64>
16397 // CHECK: [[TMP4:%.*]] = bitcast <8 x half> [[TMP3]] to <16 x i8>
16406 // CHECK: [[TMP9:%.*]] = bitcast <16 x i8> [[TMP4]] to <8 x i16>
16428 // CHECK: [[TMP4:%.*]] = bitcast <4 x float> [[TMP3]] to <16 x i8>
16437 // CHECK: [[TMP9:%.*]] = bitcast <16 x i8> [[TMP4]] to <4 x float>
16459 // CHECK: [[TMP4:%.*]] = bitcast <2 x double> [[TMP3]] to <16 x i8>
16468 // CHECK: [[TMP9:%.*]] = bitcast <16 x i8> [[TMP4]] to <2 x double>
16494 // CHECK: [[TMP4:%.*]] = load <16 x i8>, <16 x i8>* [[ARRAYIDX4]], align 16
16495 // CHECK: call void @llvm.aarch64.neon.st1x3.v16i8.p0i8(<16 x i8> [[TMP2]], <16 x i8> [[TMP3]], <16 x i8> [[TMP4]], i8* %a)
16513 // CHECK: [[TMP4:%.*]] = bitcast <8 x i16> [[TMP3]] to <16 x i8>
16522 // CHECK: [[TMP9:%.*]] = bitcast <16 x i8> [[TMP4]] to <8 x i16>
16544 // CHECK: [[TMP4:%.*]] = bitcast <2 x i64> [[TMP3]] to <16 x i8>
16553 // CHECK: [[TMP9:%.*]] = bitcast <16 x i8> [[TMP4]] to <2 x i64>
16579 // CHECK: [[TMP4:%.*]] = load <8 x i8>, <8 x i8>* [[ARRAYIDX4]], align 8
16580 // CHECK: call void @llvm.aarch64.neon.st1x3.v8i8.p0i8(<8 x i8> [[TMP2]], <8 x i8> [[TMP3]], <8 x i8> [[TMP4]], i8* %a)
16598 // CHECK: [[TMP4:%.*]] = bitcast <4 x i16> [[TMP3]] to <8 x i8>
16607 // CHECK: [[TMP9:%.*]] = bitcast <8 x i8> [[TMP4]] to <4 x i16>
16629 // CHECK: [[TMP4:%.*]] = bitcast <2 x i32> [[TMP3]] to <8 x i8>
16638 // CHECK: [[TMP9:%.*]] = bitcast <8 x i8> [[TMP4]] to <2 x i32>
16660 // CHECK: [[TMP4:%.*]] = bitcast <1 x i64> [[TMP3]] to <8 x i8>
16669 // CHECK: [[TMP9:%.*]] = bitcast <8 x i8> [[TMP4]] to <1 x i64>
16695 // CHECK: [[TMP4:%.*]] = load <8 x i8>, <8 x i8>* [[ARRAYIDX4]], align 8
16696 // CHECK: call void @llvm.aarch64.neon.st1x3.v8i8.p0i8(<8 x i8> [[TMP2]], <8 x i8> [[TMP3]], <8 x i8> [[TMP4]], i8* %a)
16714 // CHECK: [[TMP4:%.*]] = bitcast <4 x i16> [[TMP3]] to <8 x i8>
16723 // CHECK: [[TMP9:%.*]] = bitcast <8 x i8> [[TMP4]] to <4 x i16>
16745 // CHECK: [[TMP4:%.*]] = bitcast <2 x i32> [[TMP3]] to <8 x i8>
16754 // CHECK: [[TMP9:%.*]] = bitcast <8 x i8> [[TMP4]] to <2 x i32>
16776 // CHECK: [[TMP4:%.*]] = bitcast <1 x i64> [[TMP3]] to <8 x i8>
16785 // CHECK: [[TMP9:%.*]] = bitcast <8 x i8> [[TMP4]] to <1 x i64>
16807 // CHECK: [[TMP4:%.*]] = bitcast <4 x half> [[TMP3]] to <8 x i8>
16816 // CHECK: [[TMP9:%.*]] = bitcast <8 x i8> [[TMP4]] to <4 x i16>
16838 // CHECK: [[TMP4:%.*]] = bitcast <2 x float> [[TMP3]] to <8 x i8>
16847 // CHECK: [[TMP9:%.*]] = bitcast <8 x i8> [[TMP4]] to <2 x float>
16869 // CHECK: [[TMP4:%.*]] = bitcast <1 x double> [[TMP3]] to <8 x i8>
16878 // CHECK: [[TMP9:%.*]] = bitcast <8 x i8> [[TMP4]] to <1 x double>
16904 // CHECK: [[TMP4:%.*]] = load <8 x i8>, <8 x i8>* [[ARRAYIDX4]], align 8
16905 // CHECK: call void @llvm.aarch64.neon.st1x3.v8i8.p0i8(<8 x i8> [[TMP2]], <8 x i8> [[TMP3]], <8 x i8> [[TMP4]], i8* %a)
16923 // CHECK: [[TMP4:%.*]] = bitcast <4 x i16> [[TMP3]] to <8 x i8>
16932 // CHECK: [[TMP9:%.*]] = bitcast <8 x i8> [[TMP4]] to <4 x i16>
16954 // CHECK: [[TMP4:%.*]] = bitcast <1 x i64> [[TMP3]] to <8 x i8>
16963 // CHECK: [[TMP9:%.*]] = bitcast <8 x i8> [[TMP4]] to <1 x i64>
16989 // CHECK: [[TMP4:%.*]] = load <16 x i8>, <16 x i8>* [[ARRAYIDX4]], align 16
16993 // CHECK: call void @llvm.aarch64.neon.st1x4.v16i8.p0i8(<16 x i8> [[TMP2]], <16 x i8> [[TMP3]], <16 x i8> [[TMP4]], <16 x i8> [[TMP5]], i8* %a)
17011 // CHECK: [[TMP4:%.*]] = bitcast <8 x i16> [[TMP3]] to <16 x i8>
17024 // CHECK: [[TMP11:%.*]] = bitcast <16 x i8> [[TMP4]] to <8 x i16>
17047 // CHECK: [[TMP4:%.*]] = bitcast <4 x i32> [[TMP3]] to <16 x i8>
17060 // CHECK: [[TMP11:%.*]] = bitcast <16 x i8> [[TMP4]] to <4 x i32>
17083 // CHECK: [[TMP4:%.*]] = bitcast <2 x i64> [[TMP3]] to <16 x i8>
17096 // CHECK: [[TMP11:%.*]] = bitcast <16 x i8> [[TMP4]] to <2 x i64>
17123 // CHECK: [[TMP4:%.*]] = load <16 x i8>, <16 x i8>* [[ARRAYIDX4]], align 16
17127 // CHECK: call void @llvm.aarch64.neon.st1x4.v16i8.p0i8(<16 x i8> [[TMP2]], <16 x i8> [[TMP3]], <16 x i8> [[TMP4]], <16 x i8> [[TMP5]], i8* %a)
17145 // CHECK: [[TMP4:%.*]] = bitcast <8 x i16> [[TMP3]] to <16 x i8>
17158 // CHECK: [[TMP11:%.*]] = bitcast <16 x i8> [[TMP4]] to <8 x i16>
17181 // CHECK: [[TMP4:%.*]] = bitcast <4 x i32> [[TMP3]] to <16 x i8>
17194 // CHECK: [[TMP11:%.*]] = bitcast <16 x i8> [[TMP4]] to <4 x i32>
17217 // CHECK: [[TMP4:%.*]] = bitcast <2 x i64> [[TMP3]] to <16 x i8>
17230 // CHECK: [[TMP11:%.*]] = bitcast <16 x i8> [[TMP4]] to <2 x i64>
17253 // CHECK: [[TMP4:%.*]] = bitcast <8 x half> [[TMP3]] to <16 x i8>
17266 // CHECK: [[TMP11:%.*]] = bitcast <16 x i8> [[TMP4]] to <8 x i16>
17289 // CHECK: [[TMP4:%.*]] = bitcast <4 x float> [[TMP3]] to <16 x i8>
17302 // CHECK: [[TMP11:%.*]] = bitcast <16 x i8> [[TMP4]] to <4 x float>
17325 // CHECK: [[TMP4:%.*]] = bitcast <2 x double> [[TMP3]] to <16 x i8>
17338 // CHECK: [[TMP11:%.*]] = bitcast <16 x i8> [[TMP4]] to <2 x double>
17365 // CHECK: [[TMP4:%.*]] = load <16 x i8>, <16 x i8>* [[ARRAYIDX4]], align 16
17369 // CHECK: call void @llvm.aarch64.neon.st1x4.v16i8.p0i8(<16 x i8> [[TMP2]], <16 x i8> [[TMP3]], <16 x i8> [[TMP4]], <16 x i8> [[TMP5]], i8* %a)
17387 // CHECK: [[TMP4:%.*]] = bitcast <8 x i16> [[TMP3]] to <16 x i8>
17400 // CHECK: [[TMP11:%.*]] = bitcast <16 x i8> [[TMP4]] to <8 x i16>
17423 // CHECK: [[TMP4:%.*]] = bitcast <2 x i64> [[TMP3]] to <16 x i8>
17436 // CHECK: [[TMP11:%.*]] = bitcast <16 x i8> [[TMP4]] to <2 x i64>
17463 // CHECK: [[TMP4:%.*]] = load <8 x i8>, <8 x i8>* [[ARRAYIDX4]], align 8
17467 // CHECK: call void @llvm.aarch64.neon.st1x4.v8i8.p0i8(<8 x i8> [[TMP2]], <8 x i8> [[TMP3]], <8 x i8> [[TMP4]], <8 x i8> [[TMP5]], i8* %a)
17485 // CHECK: [[TMP4:%.*]] = bitcast <4 x i16> [[TMP3]] to <8 x i8>
17498 // CHECK: [[TMP11:%.*]] = bitcast <8 x i8> [[TMP4]] to <4 x i16>
17521 // CHECK: [[TMP4:%.*]] = bitcast <2 x i32> [[TMP3]] to <8 x i8>
17534 // CHECK: [[TMP11:%.*]] = bitcast <8 x i8> [[TMP4]] to <2 x i32>
17557 // CHECK: [[TMP4:%.*]] = bitcast <1 x i64> [[TMP3]] to <8 x i8>
17570 // CHECK: [[TMP11:%.*]] = bitcast <8 x i8> [[TMP4]] to <1 x i64>
17597 // CHECK: [[TMP4:%.*]] = load <8 x i8>, <8 x i8>* [[ARRAYIDX4]], align 8
17601 // CHECK: call void @llvm.aarch64.neon.st1x4.v8i8.p0i8(<8 x i8> [[TMP2]], <8 x i8> [[TMP3]], <8 x i8> [[TMP4]], <8 x i8> [[TMP5]], i8* %a)
17619 // CHECK: [[TMP4:%.*]] = bitcast <4 x i16> [[TMP3]] to <8 x i8>
17632 // CHECK: [[TMP11:%.*]] = bitcast <8 x i8> [[TMP4]] to <4 x i16>
17655 // CHECK: [[TMP4:%.*]] = bitcast <2 x i32> [[TMP3]] to <8 x i8>
17668 // CHECK: [[TMP11:%.*]] = bitcast <8 x i8> [[TMP4]] to <2 x i32>
17691 // CHECK: [[TMP4:%.*]] = bitcast <1 x i64> [[TMP3]] to <8 x i8>
17704 // CHECK: [[TMP11:%.*]] = bitcast <8 x i8> [[TMP4]] to <1 x i64>
17727 // CHECK: [[TMP4:%.*]] = bitcast <4 x half> [[TMP3]] to <8 x i8>
17740 // CHECK: [[TMP11:%.*]] = bitcast <8 x i8> [[TMP4]] to <4 x i16>
17763 // CHECK: [[TMP4:%.*]] = bitcast <2 x float> [[TMP3]] to <8 x i8>
17776 // CHECK: [[TMP11:%.*]] = bitcast <8 x i8> [[TMP4]] to <2 x float>
17799 // CHECK: [[TMP4:%.*]] = bitcast <1 x double> [[TMP3]] to <8 x i8>
17812 // CHECK: [[TMP11:%.*]] = bitcast <8 x i8> [[TMP4]] to <1 x double>
17839 // CHECK: [[TMP4:%.*]] = load <8 x i8>, <8 x i8>* [[ARRAYIDX4]], align 8
17843 // CHECK: call void @llvm.aarch64.neon.st1x4.v8i8.p0i8(<8 x i8> [[TMP2]], <8 x i8> [[TMP3]], <8 x i8> [[TMP4]], <8 x i8> [[TMP5]], i8* %a)
17861 // CHECK: [[TMP4:%.*]] = bitcast <4 x i16> [[TMP3]] to <8 x i8>
17874 // CHECK: [[TMP11:%.*]] = bitcast <8 x i8> [[TMP4]] to <4 x i16>
17897 // CHECK: [[TMP4:%.*]] = bitcast <1 x i64> [[TMP3]] to <8 x i8>
17910 // CHECK: [[TMP11:%.*]] = bitcast <8 x i8> [[TMP4]] to <1 x i64>
18647 // CHECK: [[TMP4:%.*]] = add <1 x i64> [[TMP2]], [[VSRA_N]]
18648 // CHECK: ret <1 x i64> [[TMP4]]
18674 // CHECK: [[TMP4:%.*]] = add <1 x i64> [[TMP2]], [[VSRA_N]]
18675 // CHECK: ret <1 x i64> [[TMP4]]
22127 // CHECK: [[TMP4:%.*]] = bitcast <8 x i8> [[TMP1]] to <1 x double>
22129 // CHECK: [[TMP6:%.*]] = call <1 x double> @llvm.fma.v1f64(<1 x double> [[TMP4]], <1 x double> [[TMP5]], <1 x double> [[TMP3]]) #4
22141 // CHECK: [[TMP4:%.*]] = bitcast <8 x i8> [[TMP1]] to <1 x double>
22143 // CHECK: [[TMP6:%.*]] = call <1 x double> @llvm.fma.v1f64(<1 x double> [[TMP4]], <1 x double> [[TMP5]], <1 x double> [[TMP3]]) #4
