826 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc -o - %s | FileCheck %s3 4target datalayout = "e-m:o-i64:64-i128:128-n32:64-S128"5target triple = "arm64-apple-ios"6 7; CHECK-LABEL: lCPI0_0:8; CHECK-NEXT: .byte 0 ; 0x09; CHECK-NEXT: .byte 4 ; 0x410; CHECK-NEXT: .byte 8 ; 0x811; CHECK-NEXT: .byte 12 ; 0xc12; CHECK-NEXT: .byte 16 ; 0x1013; CHECK-NEXT: .byte 20 ; 0x1414; CHECK-NEXT: .byte 24 ; 0x1815; CHECK-NEXT: .byte 28 ; 0x1c16; CHECK-NEXT: .byte 255 ; 0xff17; CHECK-NEXT: .byte 255 ; 0xff18; CHECK-NEXT: .byte 255 ; 0xff19; CHECK-NEXT: .byte 255 ; 0xff20; CHECK-NEXT: .byte 255 ; 0xff21; CHECK-NEXT: .byte 255 ; 0xff22; CHECK-NEXT: .byte 255 ; 0xff23; CHECK-NEXT: .byte 255 ; 0xff24 25; It's profitable to convert the fptoui float -> i8 to first convert from26; float -> i32 and then use tbl for the truncate in a loop, so the mask can be27; materialized outside the loop.28define void @fptoui_v8f32_to_v8i8_in_loop(ptr %A, ptr %dst) {29; CHECK-LABEL: fptoui_v8f32_to_v8i8_in_loop:30; CHECK: ; %bb.0: ; %entry31; CHECK-NEXT: Lloh0:32; CHECK-NEXT: adrp x8, lCPI0_0@PAGE33; CHECK-NEXT: Lloh1:34; CHECK-NEXT: ldr q0, [x8, lCPI0_0@PAGEOFF]35; CHECK-NEXT: mov x8, xzr36; CHECK-NEXT: LBB0_1: ; %loop37; CHECK-NEXT: ; =>This Inner Loop Header: Depth=138; CHECK-NEXT: add x9, x0, x8, lsl #539; CHECK-NEXT: add x8, x8, #140; CHECK-NEXT: cmp x8, #100041; CHECK-NEXT: ldp q2, q1, [x9]42; CHECK-NEXT: fcvtzu.4s v4, v143; CHECK-NEXT: fcvtzu.4s v3, v244; CHECK-NEXT: tbl.16b v1, { v3, v4 }, v045; CHECK-NEXT: str d1, [x1], #1646; CHECK-NEXT: b.eq LBB0_147; CHECK-NEXT: ; %bb.2: ; %exit48; CHECK-NEXT: ret49; CHECK-NEXT: .loh AdrpLdr Lloh0, Lloh150entry:51 br label %loop52 53loop:54 %iv = phi i64 [ 0, %entry ], [ %iv.next, %loop ]55 %gep.A = getelementptr inbounds <8 x float>, ptr %A, i64 %iv56 %l.A = load <8 x float>, ptr %gep.A57 %c = fptoui <8 x float> %l.A to <8 x i8>58 %gep.dst = getelementptr inbounds <16 x i8>, ptr %dst, i64 %iv59 store <8 x i8> %c, ptr %gep.dst60 %iv.next = add i64 %iv, 161 %ec = icmp eq i64 %iv.next, 100062 br i1 %ec, label %loop, label %exit63 64exit:65 ret void66}67 68; Not profitable to use tbl, as materializing the masks requires more69; instructions.70define void @fptoui_v8f32_to_v8i8_no_loop(ptr %A, ptr %dst) {71; CHECK-LABEL: fptoui_v8f32_to_v8i8_no_loop:72; CHECK: ; %bb.0: ; %entry73; CHECK-NEXT: ldp q0, q1, [x0]74; CHECK-NEXT: fcvtzs.4s v1, v175; CHECK-NEXT: fcvtzs.4s v0, v076; CHECK-NEXT: uzp1.8h v0, v0, v177; CHECK-NEXT: xtn.8b v0, v078; CHECK-NEXT: str d0, [x1]79; CHECK-NEXT: ret80entry:81 %l.A = load <8 x float>, ptr %A82 %c = fptoui <8 x float> %l.A to <8 x i8>83 store <8 x i8> %c, ptr %dst84 ret void85}86 87; CHECK-LABEL: lCPI2_0:88; CHECK-NEXT: .byte 0 ; 0x089; CHECK-NEXT: .byte 4 ; 0x490; CHECK-NEXT: .byte 8 ; 0x891; CHECK-NEXT: .byte 12 ; 0xc92; CHECK-NEXT: .byte 16 ; 0x1093; CHECK-NEXT: .byte 20 ; 0x1494; CHECK-NEXT: .byte 24 ; 0x1895; CHECK-NEXT: .byte 28 ; 0x1c96; CHECK-NEXT: .byte 32 ; 0x2097; CHECK-NEXT: .byte 36 ; 0x2498; CHECK-NEXT: .byte 40 ; 0x2899; CHECK-NEXT: .byte 44 ; 0x2c100; CHECK-NEXT: .byte 48 ; 0x30101; CHECK-NEXT: .byte 52 ; 0x34102; CHECK-NEXT: .byte 56 ; 0x38103; CHECK-NEXT: .byte 60 ; 0x3c104 105; Tbl can also be used when combining multiple fptoui using a shuffle. The loop106; vectorizer may create such patterns.107define void @fptoui_2x_v8f32_to_v8i8_in_loop(ptr %A, ptr %B, ptr %dst) {108; CHECK-LABEL: fptoui_2x_v8f32_to_v8i8_in_loop:109; CHECK: ; %bb.0: ; %entry110; CHECK-NEXT: Lloh2:111; CHECK-NEXT: adrp x8, lCPI2_0@PAGE112; CHECK-NEXT: Lloh3:113; CHECK-NEXT: ldr q0, [x8, lCPI2_0@PAGEOFF]114; CHECK-NEXT: mov x8, xzr115; CHECK-NEXT: LBB2_1: ; %loop116; CHECK-NEXT: ; =>This Inner Loop Header: Depth=1117; CHECK-NEXT: lsl x9, x8, #5118; CHECK-NEXT: add x10, x0, x9119; CHECK-NEXT: add x9, x1, x9120; CHECK-NEXT: ldp q2, q1, [x10]121; CHECK-NEXT: fcvtzu.4s v4, v1122; CHECK-NEXT: ldp q7, q1, [x9]123; CHECK-NEXT: fcvtzu.4s v3, v2124; CHECK-NEXT: fcvtzu.4s v6, v1125; CHECK-NEXT: fcvtzu.4s v5, v7126; CHECK-NEXT: tbl.16b v1, { v3, v4, v5, v6 }, v0127; CHECK-NEXT: str q1, [x2, x8, lsl #4]128; CHECK-NEXT: add x8, x8, #1129; CHECK-NEXT: cmp x8, #1000130; CHECK-NEXT: b.eq LBB2_1131; CHECK-NEXT: ; %bb.2: ; %exit132; CHECK-NEXT: ret133; CHECK-NEXT: .loh AdrpLdr Lloh2, Lloh3134entry:135 br label %loop136 137loop:138 %iv = phi i64 [ 0, %entry ], [ %iv.next, %loop ]139 %gep.A = getelementptr inbounds <8 x float>, ptr %A, i64 %iv140 %gep.B = getelementptr inbounds <8x float>, ptr %B, i64 %iv141 %l.A = load <8 x float>, ptr %gep.A142 %l.B = load <8 x float>, ptr %gep.B143 %c1 = fptoui <8 x float> %l.A to <8 x i8>144 %c2 = fptoui <8 x float> %l.B to <8 x i8>145 %s = shufflevector <8 x i8> %c1, <8 x i8> %c2, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>146 %gep.dst = getelementptr inbounds <16 x i8>, ptr %dst, i64 %iv147 store <16 x i8> %s, ptr %gep.dst148 %iv.next = add i64 %iv, 1149 %ec = icmp eq i64 %iv.next, 1000150 br i1 %ec, label %loop, label %exit151 152exit:153 ret void154}155 156; CHECK-LABEL: lCPI3_0:157; CHECK-NEXT: .byte 0 ; 0x0158; CHECK-NEXT: .byte 36 ; 0x24159; CHECK-NEXT: .byte 8 ; 0x8160; CHECK-NEXT: .byte 12 ; 0xc161; CHECK-NEXT: .byte 16 ; 0x10162; CHECK-NEXT: .byte 20 ; 0x14163; CHECK-NEXT: .byte 24 ; 0x18164; CHECK-NEXT: .byte 44 ; 0x2c165; CHECK-NEXT: .byte 32 ; 0x20166; CHECK-NEXT: .byte 36 ; 0x24167; CHECK-NEXT: .byte 40 ; 0x28168; CHECK-NEXT: .byte 44 ; 0x2c169; CHECK-NEXT: .byte 48 ; 0x30170; CHECK-NEXT: .byte 12 ; 0xc171; CHECK-NEXT: .byte 56 ; 0x38172; CHECK-NEXT: .byte 60 ; 0x3c173 174define void @fptoui_2x_v8f32_to_v8i8_in_loop_no_concat_shuffle(ptr %A, ptr %B, ptr %dst) {175; CHECK-LABEL: fptoui_2x_v8f32_to_v8i8_in_loop_no_concat_shuffle:176; CHECK: ; %bb.0: ; %entry177; CHECK-NEXT: Lloh4:178; CHECK-NEXT: adrp x8, lCPI3_0@PAGE179; CHECK-NEXT: Lloh5:180; CHECK-NEXT: ldr q0, [x8, lCPI3_0@PAGEOFF]181; CHECK-NEXT: mov x8, xzr182; CHECK-NEXT: LBB3_1: ; %loop183; CHECK-NEXT: ; =>This Inner Loop Header: Depth=1184; CHECK-NEXT: lsl x9, x8, #5185; CHECK-NEXT: add x10, x0, x9186; CHECK-NEXT: add x9, x1, x9187; CHECK-NEXT: ldp q2, q1, [x10]188; CHECK-NEXT: fcvtzu.4s v4, v1189; CHECK-NEXT: ldp q7, q1, [x9]190; CHECK-NEXT: fcvtzu.4s v3, v2191; CHECK-NEXT: fcvtzu.4s v6, v1192; CHECK-NEXT: fcvtzu.4s v5, v7193; CHECK-NEXT: tbl.16b v1, { v3, v4, v5, v6 }, v0194; CHECK-NEXT: str q1, [x2, x8, lsl #4]195; CHECK-NEXT: add x8, x8, #1196; CHECK-NEXT: cmp x8, #1000197; CHECK-NEXT: b.eq LBB3_1198; CHECK-NEXT: ; %bb.2: ; %exit199; CHECK-NEXT: ret200; CHECK-NEXT: .loh AdrpLdr Lloh4, Lloh5201entry:202 br label %loop203 204loop:205 %iv = phi i64 [ 0, %entry ], [ %iv.next, %loop ]206 %gep.A = getelementptr inbounds <8 x float>, ptr %A, i64 %iv207 %gep.B = getelementptr inbounds <8x float>, ptr %B, i64 %iv208 %l.A = load <8 x float>, ptr %gep.A209 %l.B = load <8 x float>, ptr %gep.B210 %c1 = fptoui <8 x float> %l.A to <8 x i8>211 %c2 = fptoui <8 x float> %l.B to <8 x i8>212 %s = shufflevector <8 x i8> %c1, <8 x i8> %c2, <16 x i32> <i32 0, i32 9, i32 2, i32 3, i32 4, i32 5, i32 6, i32 11, i32 8, i32 9, i32 10, i32 11, i32 12, i32 3, i32 14, i32 15>213 %gep.dst = getelementptr inbounds <16 x i8>, ptr %dst, i64 %iv214 store <16 x i8> %s, ptr %gep.dst215 %iv.next = add i64 %iv, 1216 %ec = icmp eq i64 %iv.next, 1000217 br i1 %ec, label %loop, label %exit218 219exit:220 ret void221}222 223; CHECK-LABEL: lCPI4_0:224; CHECK-NEXT: .byte 0 ; 0x0225; CHECK-NEXT: .byte 4 ; 0x4226; CHECK-NEXT: .byte 8 ; 0x8227; CHECK-NEXT: .byte 12 ; 0xc228; CHECK-NEXT: .byte 16 ; 0x10229; CHECK-NEXT: .byte 20 ; 0x14230; CHECK-NEXT: .byte 24 ; 0x18231; CHECK-NEXT: .byte 28 ; 0x1c232; CHECK-NEXT: .byte 32 ; 0x20233; CHECK-NEXT: .byte 36 ; 0x24234; CHECK-NEXT: .byte 40 ; 0x28235; CHECK-NEXT: .byte 44 ; 0x2c236; CHECK-NEXT: .byte 48 ; 0x30237; CHECK-NEXT: .byte 52 ; 0x34238; CHECK-NEXT: .byte 56 ; 0x38239; CHECK-NEXT: .byte 60 ; 0x3c240 241define void @fptoui_v16f32_to_v16i8_in_loop(ptr %A, ptr %dst) {242; CHECK-LABEL: fptoui_v16f32_to_v16i8_in_loop:243; CHECK: ; %bb.0: ; %entry244; CHECK-NEXT: Lloh6:245; CHECK-NEXT: adrp x8, lCPI4_0@PAGE246; CHECK-NEXT: Lloh7:247; CHECK-NEXT: ldr q0, [x8, lCPI4_0@PAGEOFF]248; CHECK-NEXT: mov x8, xzr249; CHECK-NEXT: LBB4_1: ; %loop250; CHECK-NEXT: ; =>This Inner Loop Header: Depth=1251; CHECK-NEXT: add x9, x0, x8, lsl #6252; CHECK-NEXT: add x8, x8, #1253; CHECK-NEXT: cmp x8, #1000254; CHECK-NEXT: ldp q2, q1, [x9, #32]255; CHECK-NEXT: fcvtzu.4s v6, v1256; CHECK-NEXT: ldp q7, q1, [x9]257; CHECK-NEXT: fcvtzu.4s v5, v2258; CHECK-NEXT: fcvtzu.4s v4, v1259; CHECK-NEXT: fcvtzu.4s v3, v7260; CHECK-NEXT: tbl.16b v1, { v3, v4, v5, v6 }, v0261; CHECK-NEXT: str q1, [x1], #32262; CHECK-NEXT: b.eq LBB4_1263; CHECK-NEXT: ; %bb.2: ; %exit264; CHECK-NEXT: ret265; CHECK-NEXT: .loh AdrpLdr Lloh6, Lloh7266entry:267 br label %loop268 269loop:270 %iv = phi i64 [ 0, %entry ], [ %iv.next, %loop ]271 %gep.A = getelementptr inbounds <16 x float>, ptr %A, i64 %iv272 %l.A = load <16 x float>, ptr %gep.A273 %c = fptoui <16 x float> %l.A to <16 x i8>274 %gep.dst = getelementptr inbounds <32 x i8>, ptr %dst, i64 %iv275 store <16 x i8> %c, ptr %gep.dst276 %iv.next = add i64 %iv, 1277 %ec = icmp eq i64 %iv.next, 1000278 br i1 %ec, label %loop, label %exit279 280exit:281 ret void282}283 284; CHECK-LABEL: lCPI5_0:285; CHECK-NEXT: .byte 0 ; 0x0286; CHECK-NEXT: .byte 4 ; 0x4287; CHECK-NEXT: .byte 8 ; 0x8288; CHECK-NEXT: .byte 12 ; 0xc289; CHECK-NEXT: .byte 16 ; 0x10290; CHECK-NEXT: .byte 20 ; 0x14291; CHECK-NEXT: .byte 24 ; 0x18292; CHECK-NEXT: .byte 28 ; 0x1c293; CHECK-NEXT: .byte 32 ; 0x20294; CHECK-NEXT: .byte 36 ; 0x24295; CHECK-NEXT: .byte 40 ; 0x28296; CHECK-NEXT: .byte 44 ; 0x2c297; CHECK-NEXT: .byte 48 ; 0x30298; CHECK-NEXT: .byte 52 ; 0x34299; CHECK-NEXT: .byte 56 ; 0x38300; CHECK-NEXT: .byte 60 ; 0x3c301 302define void @fptoui_2x_v16f32_to_v16i8_in_loop(ptr %A, ptr %B, ptr %dst) {303; CHECK-LABEL: fptoui_2x_v16f32_to_v16i8_in_loop:304; CHECK: ; %bb.0: ; %entry305; CHECK-NEXT: Lloh8:306; CHECK-NEXT: adrp x8, lCPI5_0@PAGE307; CHECK-NEXT: Lloh9:308; CHECK-NEXT: ldr q0, [x8, lCPI5_0@PAGEOFF]309; CHECK-NEXT: mov x8, xzr310; CHECK-NEXT: LBB5_1: ; %loop311; CHECK-NEXT: ; =>This Inner Loop Header: Depth=1312; CHECK-NEXT: lsl x9, x8, #6313; CHECK-NEXT: add x10, x1, x9314; CHECK-NEXT: add x9, x0, x9315; CHECK-NEXT: ldp q2, q1, [x10, #32]316; CHECK-NEXT: ldp q3, q4, [x9, #32]317; CHECK-NEXT: ldp q5, q6, [x10]318; CHECK-NEXT: fcvtzu.4s v19, v1319; CHECK-NEXT: fcvtzu.4s v18, v2320; CHECK-NEXT: ldp q2, q1, [x9]321; CHECK-NEXT: fcvtzu.4s v23, v4322; CHECK-NEXT: fcvtzu.4s v17, v6323; CHECK-NEXT: add x9, x2, x8, lsl #5324; CHECK-NEXT: fcvtzu.4s v22, v3325; CHECK-NEXT: fcvtzu.4s v16, v5326; CHECK-NEXT: add x8, x8, #1327; CHECK-NEXT: fcvtzu.4s v21, v1328; CHECK-NEXT: cmp x8, #1000329; CHECK-NEXT: fcvtzu.4s v20, v2330; CHECK-NEXT: tbl.16b v1, { v16, v17, v18, v19 }, v0331; CHECK-NEXT: tbl.16b v2, { v20, v21, v22, v23 }, v0332; CHECK-NEXT: stp q2, q1, [x9]333; CHECK-NEXT: b.eq LBB5_1334; CHECK-NEXT: ; %bb.2: ; %exit335; CHECK-NEXT: ret336; CHECK-NEXT: .loh AdrpLdr Lloh8, Lloh9337entry:338 br label %loop339 340loop:341 %iv = phi i64 [ 0, %entry ], [ %iv.next, %loop ]342 %gep.A = getelementptr inbounds <16 x float>, ptr %A, i64 %iv343 %gep.B = getelementptr inbounds <16 x float>, ptr %B, i64 %iv344 %l.A = load <16 x float>, ptr %gep.A345 %l.B = load <16 x float>, ptr %gep.B346 %c1 = fptoui <16 x float> %l.A to <16 x i8>347 %c2 = fptoui <16 x float> %l.B to <16 x i8>348 %s = shufflevector <16 x i8> %c1, <16 x i8> %c2, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>349 %gep.dst = getelementptr inbounds <32 x i8>, ptr %dst, i64 %iv350 store <32 x i8> %s, ptr %gep.dst351 %iv.next = add i64 %iv, 1352 %ec = icmp eq i64 %iv.next, 1000353 br i1 %ec, label %loop, label %exit354 355exit:356 ret void357}358 359define void @fptoui_v8f32_to_v8i16_in_loop(ptr %A, ptr %dst) {360; CHECK-LABEL: fptoui_v8f32_to_v8i16_in_loop:361; CHECK: ; %bb.0: ; %entry362; CHECK-NEXT: mov x8, xzr363; CHECK-NEXT: LBB6_1: ; %loop364; CHECK-NEXT: ; =>This Inner Loop Header: Depth=1365; CHECK-NEXT: add x9, x0, x8, lsl #5366; CHECK-NEXT: ldp q0, q1, [x9]367; CHECK-NEXT: fcvtzu.4s v1, v1368; CHECK-NEXT: fcvtzu.4s v0, v0369; CHECK-NEXT: uzp1.8h v0, v0, v1370; CHECK-NEXT: str q0, [x1, x8, lsl #4]371; CHECK-NEXT: add x8, x8, #1372; CHECK-NEXT: cmp x8, #1000373; CHECK-NEXT: b.eq LBB6_1374; CHECK-NEXT: ; %bb.2: ; %exit375; CHECK-NEXT: ret376entry:377 br label %loop378 379loop:380 %iv = phi i64 [ 0, %entry ], [ %iv.next, %loop ]381 %gep.A = getelementptr inbounds <8 x float>, ptr %A, i64 %iv382 %l.A = load <8 x float>, ptr %gep.A383 %c = fptoui <8 x float> %l.A to <8 x i16>384 %gep.dst = getelementptr inbounds <8 x i16>, ptr %dst, i64 %iv385 store <8 x i16> %c, ptr %gep.dst386 %iv.next = add i64 %iv, 1387 %ec = icmp eq i64 %iv.next, 1000388 br i1 %ec, label %loop, label %exit389 390exit:391 ret void392}393 394define void @fptoui_2x_v8f32_to_v8i16_in_loop(ptr %A, ptr %B, ptr %dst) {395; CHECK-LABEL: fptoui_2x_v8f32_to_v8i16_in_loop:396; CHECK: ; %bb.0: ; %entry397; CHECK-NEXT: mov x8, xzr398; CHECK-NEXT: LBB7_1: ; %loop399; CHECK-NEXT: ; =>This Inner Loop Header: Depth=1400; CHECK-NEXT: lsl x9, x8, #5401; CHECK-NEXT: add x8, x8, #1402; CHECK-NEXT: cmp x8, #1000403; CHECK-NEXT: add x10, x0, x9404; CHECK-NEXT: add x11, x1, x9405; CHECK-NEXT: add x9, x2, x9406; CHECK-NEXT: ldp q0, q1, [x10]407; CHECK-NEXT: ldp q2, q3, [x11]408; CHECK-NEXT: fcvtzu.4s v1, v1409; CHECK-NEXT: fcvtzu.4s v0, v0410; CHECK-NEXT: fcvtzu.4s v3, v3411; CHECK-NEXT: fcvtzu.4s v2, v2412; CHECK-NEXT: uzp1.8h v0, v0, v1413; CHECK-NEXT: uzp1.8h v1, v2, v3414; CHECK-NEXT: stp q0, q1, [x9]415; CHECK-NEXT: b.eq LBB7_1416; CHECK-NEXT: ; %bb.2: ; %exit417; CHECK-NEXT: ret418entry:419 br label %loop420 421loop:422 %iv = phi i64 [ 0, %entry ], [ %iv.next, %loop ]423 %gep.A = getelementptr inbounds <8 x float>, ptr %A, i64 %iv424 %gep.B = getelementptr inbounds <8 x float>, ptr %B, i64 %iv425 %l.A = load <8 x float>, ptr %gep.A426 %l.B = load <8 x float>, ptr %gep.B427 %c1 = fptoui <8 x float> %l.A to <8 x i16>428 %c2 = fptoui <8 x float> %l.B to <8 x i16>429 %s = shufflevector <8 x i16> %c1, <8 x i16> %c2, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>430 %gep.dst = getelementptr inbounds <16 x i16>, ptr %dst, i64 %iv431 store <16 x i16> %s, ptr %gep.dst432 %iv.next = add i64 %iv, 1433 %ec = icmp eq i64 %iv.next, 1000434 br i1 %ec, label %loop, label %exit435 436exit:437 ret void438}439 440; CHECK-LABEL: lCPI8_0:441; CHECK-NEXT: .byte 4 ; 0x4442; CHECK-NEXT: .byte 255 ; 0xff443; CHECK-NEXT: .byte 255 ; 0xff444; CHECK-NEXT: .byte 255 ; 0xff445; CHECK-NEXT: .byte 5 ; 0x5446; CHECK-NEXT: .byte 255 ; 0xff447; CHECK-NEXT: .byte 255 ; 0xff448; CHECK-NEXT: .byte 255 ; 0xff449; CHECK-NEXT: .byte 6 ; 0x6450; CHECK-NEXT: .byte 255 ; 0xff451; CHECK-NEXT: .byte 255 ; 0xff452; CHECK-NEXT: .byte 255 ; 0xff453; CHECK-NEXT: .byte 7 ; 0x7454; CHECK-NEXT: .byte 255 ; 0xff455; CHECK-NEXT: .byte 255 ; 0xff456; CHECK-NEXT: .byte 255 ; 0xff457; CHECK-NEXT: lCPI8_1:458; CHECK-NEXT: .byte 0 ; 0x0459; CHECK-NEXT: .byte 255 ; 0xff460; CHECK-NEXT: .byte 255 ; 0xff461; CHECK-NEXT: .byte 255 ; 0xff462; CHECK-NEXT: .byte 1 ; 0x1463; CHECK-NEXT: .byte 255 ; 0xff464; CHECK-NEXT: .byte 255 ; 0xff465; CHECK-NEXT: .byte 255 ; 0xff466; CHECK-NEXT: .byte 2 ; 0x2467; CHECK-NEXT: .byte 255 ; 0xff468; CHECK-NEXT: .byte 255 ; 0xff469; CHECK-NEXT: .byte 255 ; 0xff470; CHECK-NEXT: .byte 3 ; 0x3471; CHECK-NEXT: .byte 255 ; 0xff472; CHECK-NEXT: .byte 255 ; 0xff473; CHECK-NEXT: .byte 255 ; 0xff474 475define void @uitofp_v8i8_to_v8f32(ptr %src, ptr %dst) {476; CHECK-LABEL: uitofp_v8i8_to_v8f32:477; CHECK: ; %bb.0: ; %entry478; CHECK-NEXT: Lloh10:479; CHECK-NEXT: adrp x8, lCPI8_0@PAGE480; CHECK-NEXT: Lloh11:481; CHECK-NEXT: adrp x9, lCPI8_1@PAGE482; CHECK-NEXT: Lloh12:483; CHECK-NEXT: ldr q0, [x8, lCPI8_0@PAGEOFF]484; CHECK-NEXT: Lloh13:485; CHECK-NEXT: ldr q1, [x9, lCPI8_1@PAGEOFF]486; CHECK-NEXT: mov x8, xzr487; CHECK-NEXT: LBB8_1: ; %loop488; CHECK-NEXT: ; =>This Inner Loop Header: Depth=1489; CHECK-NEXT: ldr d2, [x0, x8, lsl #3]490; CHECK-NEXT: add x9, x1, x8, lsl #5491; CHECK-NEXT: add x8, x8, #1492; CHECK-NEXT: cmp x8, #1000493; CHECK-NEXT: tbl.16b v3, { v2 }, v0494; CHECK-NEXT: tbl.16b v2, { v2 }, v1495; CHECK-NEXT: ucvtf.4s v3, v3496; CHECK-NEXT: ucvtf.4s v2, v2497; CHECK-NEXT: stp q2, q3, [x9]498; CHECK-NEXT: b.eq LBB8_1499; CHECK-NEXT: ; %bb.2: ; %exit500; CHECK-NEXT: ret501; CHECK-NEXT: .loh AdrpLdr Lloh11, Lloh13502; CHECK-NEXT: .loh AdrpLdr Lloh10, Lloh12503entry:504 br label %loop505 506loop:507 %iv = phi i64 [ 0, %entry ], [ %iv.next, %loop ]508 %gep.src = getelementptr inbounds <8 x i8>, ptr %src, i64 %iv509 %l = load <8 x i8>, ptr %gep.src510 %conv = uitofp <8 x i8> %l to <8 x float>511 %gep.dst = getelementptr inbounds <8 x float>, ptr %dst, i64 %iv512 store <8 x float> %conv, ptr %gep.dst513 %iv.next = add i64 %iv, 1514 %ec = icmp eq i64 %iv.next, 1000515 br i1 %ec, label %loop, label %exit516 517exit:518 ret void519}520 521; CHECK-LABEL: lCPI9_0:522; CHECK-NEXT: .byte 12 ; 0xc523; CHECK-NEXT: .byte 255 ; 0xff524; CHECK-NEXT: .byte 255 ; 0xff525; CHECK-NEXT: .byte 255 ; 0xff526; CHECK-NEXT: .byte 13 ; 0xd527; CHECK-NEXT: .byte 255 ; 0xff528; CHECK-NEXT: .byte 255 ; 0xff529; CHECK-NEXT: .byte 255 ; 0xff530; CHECK-NEXT: .byte 14 ; 0xe531; CHECK-NEXT: .byte 255 ; 0xff532; CHECK-NEXT: .byte 255 ; 0xff533; CHECK-NEXT: .byte 255 ; 0xff534; CHECK-NEXT: .byte 15 ; 0xf535; CHECK-NEXT: .byte 255 ; 0xff536; CHECK-NEXT: .byte 255 ; 0xff537; CHECK-NEXT: .byte 255 ; 0xff538; CHECK-NEXT: lCPI9_1:539; CHECK-NEXT: .byte 8 ; 0x8540; CHECK-NEXT: .byte 255 ; 0xff541; CHECK-NEXT: .byte 255 ; 0xff542; CHECK-NEXT: .byte 255 ; 0xff543; CHECK-NEXT: .byte 9 ; 0x9544; CHECK-NEXT: .byte 255 ; 0xff545; CHECK-NEXT: .byte 255 ; 0xff546; CHECK-NEXT: .byte 255 ; 0xff547; CHECK-NEXT: .byte 10 ; 0xa548; CHECK-NEXT: .byte 255 ; 0xff549; CHECK-NEXT: .byte 255 ; 0xff550; CHECK-NEXT: .byte 255 ; 0xff551; CHECK-NEXT: .byte 11 ; 0xb552; CHECK-NEXT: .byte 255 ; 0xff553; CHECK-NEXT: .byte 255 ; 0xff554; CHECK-NEXT: .byte 255 ; 0xff555; CHECK-NEXT: lCPI9_2:556; CHECK-NEXT: .byte 4 ; 0x4557; CHECK-NEXT: .byte 255 ; 0xff558; CHECK-NEXT: .byte 255 ; 0xff559; CHECK-NEXT: .byte 255 ; 0xff560; CHECK-NEXT: .byte 5 ; 0x5561; CHECK-NEXT: .byte 255 ; 0xff562; CHECK-NEXT: .byte 255 ; 0xff563; CHECK-NEXT: .byte 255 ; 0xff564; CHECK-NEXT: .byte 6 ; 0x6565; CHECK-NEXT: .byte 255 ; 0xff566; CHECK-NEXT: .byte 255 ; 0xff567; CHECK-NEXT: .byte 255 ; 0xff568; CHECK-NEXT: .byte 7 ; 0x7569; CHECK-NEXT: .byte 255 ; 0xff570; CHECK-NEXT: .byte 255 ; 0xff571; CHECK-NEXT: .byte 255 ; 0xff572; CHECK-NEXT: lCPI9_3:573; CHECK-NEXT: .byte 0 ; 0x0574; CHECK-NEXT: .byte 255 ; 0xff575; CHECK-NEXT: .byte 255 ; 0xff576; CHECK-NEXT: .byte 255 ; 0xff577; CHECK-NEXT: .byte 1 ; 0x1578; CHECK-NEXT: .byte 255 ; 0xff579; CHECK-NEXT: .byte 255 ; 0xff580; CHECK-NEXT: .byte 255 ; 0xff581; CHECK-NEXT: .byte 2 ; 0x2582; CHECK-NEXT: .byte 255 ; 0xff583; CHECK-NEXT: .byte 255 ; 0xff584; CHECK-NEXT: .byte 255 ; 0xff585; CHECK-NEXT: .byte 3 ; 0x3586; CHECK-NEXT: .byte 255 ; 0xff587; CHECK-NEXT: .byte 255 ; 0xff588; CHECK-NEXT: .byte 255 ; 0xff589 590define void @uitofp_v16i8_to_v16f32(ptr %src, ptr %dst) {591; CHECK-LABEL: uitofp_v16i8_to_v16f32:592; CHECK: ; %bb.0: ; %entry593; CHECK-NEXT: Lloh14:594; CHECK-NEXT: adrp x8, lCPI9_0@PAGE595; CHECK-NEXT: Lloh15:596; CHECK-NEXT: adrp x9, lCPI9_1@PAGE597; CHECK-NEXT: Lloh16:598; CHECK-NEXT: adrp x10, lCPI9_2@PAGE599; CHECK-NEXT: Lloh17:600; CHECK-NEXT: ldr q0, [x8, lCPI9_0@PAGEOFF]601; CHECK-NEXT: Lloh18:602; CHECK-NEXT: adrp x8, lCPI9_3@PAGE603; CHECK-NEXT: Lloh19:604; CHECK-NEXT: ldr q1, [x9, lCPI9_1@PAGEOFF]605; CHECK-NEXT: Lloh20:606; CHECK-NEXT: ldr q2, [x10, lCPI9_2@PAGEOFF]607; CHECK-NEXT: Lloh21:608; CHECK-NEXT: ldr q3, [x8, lCPI9_3@PAGEOFF]609; CHECK-NEXT: mov x8, xzr610; CHECK-NEXT: LBB9_1: ; %loop611; CHECK-NEXT: ; =>This Inner Loop Header: Depth=1612; CHECK-NEXT: ldr q4, [x0, x8, lsl #4]613; CHECK-NEXT: add x9, x1, x8, lsl #6614; CHECK-NEXT: add x8, x8, #1615; CHECK-NEXT: cmp x8, #1000616; CHECK-NEXT: tbl.16b v5, { v4 }, v0617; CHECK-NEXT: tbl.16b v6, { v4 }, v1618; CHECK-NEXT: tbl.16b v7, { v4 }, v2619; CHECK-NEXT: tbl.16b v4, { v4 }, v3620; CHECK-NEXT: ucvtf.4s v5, v5621; CHECK-NEXT: ucvtf.4s v6, v6622; CHECK-NEXT: ucvtf.4s v7, v7623; CHECK-NEXT: ucvtf.4s v4, v4624; CHECK-NEXT: stp q6, q5, [x9, #32]625; CHECK-NEXT: stp q4, q7, [x9]626; CHECK-NEXT: b.eq LBB9_1627; CHECK-NEXT: ; %bb.2: ; %exit628; CHECK-NEXT: ret629; CHECK-NEXT: .loh AdrpLdr Lloh18, Lloh21630; CHECK-NEXT: .loh AdrpLdr Lloh16, Lloh20631; CHECK-NEXT: .loh AdrpLdr Lloh15, Lloh19632; CHECK-NEXT: .loh AdrpAdrp Lloh14, Lloh18633; CHECK-NEXT: .loh AdrpLdr Lloh14, Lloh17634entry:635 br label %loop636 637loop:638 %iv = phi i64 [ 0, %entry ], [ %iv.next, %loop ]639 %gep.src = getelementptr inbounds <16 x i8>, ptr %src, i64 %iv640 %l = load <16 x i8>, ptr %gep.src641 %conv = uitofp <16 x i8> %l to <16 x float>642 %gep.dst = getelementptr inbounds <16 x float>, ptr %dst, i64 %iv643 store <16 x float> %conv, ptr %gep.dst644 %iv.next = add i64 %iv, 1645 %ec = icmp eq i64 %iv.next, 1000646 br i1 %ec, label %loop, label %exit647 648exit:649 ret void650}651 652define void @uitofp_v8i16_to_v8f64(ptr nocapture noundef readonly %x, ptr nocapture noundef writeonly %y, i32 noundef %n) {653; CHECK-LABEL: uitofp_v8i16_to_v8f64:654; CHECK: ; %bb.0: ; %entry655; CHECK-NEXT: Lloh22:656; CHECK-NEXT: adrp x8, lCPI10_0@PAGE657; CHECK-NEXT: Lloh23:658; CHECK-NEXT: adrp x9, lCPI10_1@PAGE659; CHECK-NEXT: Lloh24:660; CHECK-NEXT: adrp x10, lCPI10_2@PAGE661; CHECK-NEXT: Lloh25:662; CHECK-NEXT: ldr q0, [x8, lCPI10_0@PAGEOFF]663; CHECK-NEXT: Lloh26:664; CHECK-NEXT: adrp x8, lCPI10_3@PAGE665; CHECK-NEXT: Lloh27:666; CHECK-NEXT: ldr q1, [x9, lCPI10_1@PAGEOFF]667; CHECK-NEXT: Lloh28:668; CHECK-NEXT: ldr q2, [x10, lCPI10_2@PAGEOFF]669; CHECK-NEXT: Lloh29:670; CHECK-NEXT: ldr q3, [x8, lCPI10_3@PAGEOFF]671; CHECK-NEXT: mov x8, xzr672; CHECK-NEXT: LBB10_1: ; %vector.body673; CHECK-NEXT: ; =>This Inner Loop Header: Depth=1674; CHECK-NEXT: ldr q4, [x0, x8]675; CHECK-NEXT: add x9, x1, x8676; CHECK-NEXT: add x8, x8, #64677; CHECK-NEXT: cmp x8, #2, lsl #12 ; =8192678; CHECK-NEXT: tbl.16b v5, { v4 }, v0679; CHECK-NEXT: tbl.16b v6, { v4 }, v1680; CHECK-NEXT: tbl.16b v7, { v4 }, v2681; CHECK-NEXT: tbl.16b v4, { v4 }, v3682; CHECK-NEXT: ucvtf.2d v5, v5683; CHECK-NEXT: ucvtf.2d v6, v6684; CHECK-NEXT: ucvtf.2d v7, v7685; CHECK-NEXT: ucvtf.2d v4, v4686; CHECK-NEXT: stp q6, q5, [x9, #32]687; CHECK-NEXT: stp q4, q7, [x9]688; CHECK-NEXT: b.ne LBB10_1689; CHECK-NEXT: ; %bb.2: ; %for.cond.cleanup690; CHECK-NEXT: ret691; CHECK-NEXT: .loh AdrpLdr Lloh26, Lloh29692; CHECK-NEXT: .loh AdrpLdr Lloh24, Lloh28693; CHECK-NEXT: .loh AdrpLdr Lloh23, Lloh27694; CHECK-NEXT: .loh AdrpAdrp Lloh22, Lloh26695; CHECK-NEXT: .loh AdrpLdr Lloh22, Lloh25696entry:697 br label %vector.body698 699vector.body:700 %index = phi i64 [ 0, %entry ], [ %index.next, %vector.body ]701 %.idx = shl nsw i64 %index, 3702 %g = getelementptr inbounds i8, ptr %x, i64 %.idx703 %wide.vec = load <8 x i16>, ptr %g, align 2704 %u = uitofp <8 x i16> %wide.vec to <8 x double>705 %h = getelementptr inbounds double, ptr %y, i64 %index706 store <8 x double> %u, ptr %h, align 8707 %index.next = add nuw i64 %index, 8708 %c = icmp eq i64 %index.next, 1024709 br i1 %c, label %for.cond.cleanup, label %vector.body710 711for.cond.cleanup:712 ret void713}714 715define void @uitofp_ld4_v32i16_to_v8f64(ptr nocapture noundef readonly %x, ptr nocapture noundef writeonly %y, i32 noundef %n) {716; CHECK-LABEL: uitofp_ld4_v32i16_to_v8f64:717; CHECK: ; %bb.0: ; %entry718; CHECK-NEXT: Lloh30:719; CHECK-NEXT: adrp x8, lCPI11_0@PAGE720; CHECK-NEXT: Lloh31:721; CHECK-NEXT: adrp x9, lCPI11_1@PAGE722; CHECK-NEXT: Lloh32:723; CHECK-NEXT: adrp x10, lCPI11_2@PAGE724; CHECK-NEXT: Lloh33:725; CHECK-NEXT: ldr q0, [x8, lCPI11_0@PAGEOFF]726; CHECK-NEXT: Lloh34:727; CHECK-NEXT: adrp x8, lCPI11_3@PAGE728; CHECK-NEXT: Lloh35:729; CHECK-NEXT: ldr q1, [x9, lCPI11_1@PAGEOFF]730; CHECK-NEXT: Lloh36:731; CHECK-NEXT: ldr q2, [x10, lCPI11_2@PAGEOFF]732; CHECK-NEXT: Lloh37:733; CHECK-NEXT: ldr q3, [x8, lCPI11_3@PAGEOFF]734; CHECK-NEXT: mov x8, xzr735; CHECK-NEXT: LBB11_1: ; %vector.body736; CHECK-NEXT: ; =>This Inner Loop Header: Depth=1737; CHECK-NEXT: add x9, x0, x8738; CHECK-NEXT: ldp q5, q4, [x9, #32]739; CHECK-NEXT: ldp q7, q6, [x9]740; CHECK-NEXT: add x9, x1, x8741; CHECK-NEXT: add x8, x8, #64742; CHECK-NEXT: tbl.16b v16, { v4 }, v0743; CHECK-NEXT: tbl.16b v17, { v5 }, v0744; CHECK-NEXT: tbl.16b v21, { v4 }, v1745; CHECK-NEXT: tbl.16b v18, { v6 }, v0746; CHECK-NEXT: tbl.16b v19, { v7 }, v0747; CHECK-NEXT: tbl.16b v20, { v7 }, v1748; CHECK-NEXT: tbl.16b v22, { v5 }, v1749; CHECK-NEXT: tbl.16b v23, { v5 }, v2750; CHECK-NEXT: tbl.16b v24, { v4 }, v2751; CHECK-NEXT: tbl.16b v25, { v7 }, v2752; CHECK-NEXT: tbl.16b v5, { v5 }, v3753; CHECK-NEXT: tbl.16b v4, { v4 }, v3754; CHECK-NEXT: tbl.16b v7, { v7 }, v3755; CHECK-NEXT: tbl.16b v26, { v6 }, v1756; CHECK-NEXT: tbl.16b v27, { v6 }, v2757; CHECK-NEXT: tbl.16b v6, { v6 }, v3758; CHECK-NEXT: ucvtf.2d v17, v17759; CHECK-NEXT: ucvtf.2d v16, v16760; CHECK-NEXT: ucvtf.2d v19, v19761; CHECK-NEXT: ucvtf.2d v18, v18762; CHECK-NEXT: ucvtf.2d v22, v22763; CHECK-NEXT: ucvtf.2d v23, v23764; CHECK-NEXT: ucvtf.2d v5, v5765; CHECK-NEXT: ucvtf.2d v21, v21766; CHECK-NEXT: ucvtf.2d v24, v24767; CHECK-NEXT: ucvtf.2d v4, v4768; CHECK-NEXT: cmp x8, #2, lsl #12 ; =8192769; CHECK-NEXT: ucvtf.2d v20, v20770; CHECK-NEXT: ucvtf.2d v25, v25771; CHECK-NEXT: ucvtf.2d v7, v7772; CHECK-NEXT: ucvtf.2d v26, v26773; CHECK-NEXT: ucvtf.2d v27, v27774; CHECK-NEXT: ucvtf.2d v6, v6775; CHECK-NEXT: fadd.2d v17, v22, v17776; CHECK-NEXT: fadd.2d v5, v23, v5777; CHECK-NEXT: fadd.2d v16, v21, v16778; CHECK-NEXT: fadd.2d v4, v24, v4779; CHECK-NEXT: fadd.2d v19, v20, v19780; CHECK-NEXT: fadd.2d v7, v25, v7781; CHECK-NEXT: fadd.2d v18, v26, v18782; CHECK-NEXT: fadd.2d v6, v27, v6783; CHECK-NEXT: fadd.2d v5, v17, v5784; CHECK-NEXT: fadd.2d v4, v16, v4785; CHECK-NEXT: fadd.2d v7, v19, v7786; CHECK-NEXT: fadd.2d v6, v18, v6787; CHECK-NEXT: stp q5, q4, [x9, #32]788; CHECK-NEXT: stp q7, q6, [x9]789; CHECK-NEXT: b.ne LBB11_1790; CHECK-NEXT: ; %bb.2: ; %for.cond.cleanup791; CHECK-NEXT: ret792; CHECK-NEXT: .loh AdrpLdr Lloh34, Lloh37793; CHECK-NEXT: .loh AdrpLdr Lloh32, Lloh36794; CHECK-NEXT: .loh AdrpLdr Lloh31, Lloh35795; CHECK-NEXT: .loh AdrpAdrp Lloh30, Lloh34796; CHECK-NEXT: .loh AdrpLdr Lloh30, Lloh33797entry:798 br label %vector.body799 800vector.body:801 %index = phi i64 [ 0, %entry ], [ %index.next, %vector.body ]802 %.idx = shl nsw i64 %index, 3803 %0 = getelementptr inbounds i8, ptr %x, i64 %.idx804 %wide.vec = load <32 x i16>, ptr %0, align 2805 %strided.vec = shufflevector <32 x i16> %wide.vec, <32 x i16> poison, <8 x i32> <i32 0, i32 4, i32 8, i32 12, i32 16, i32 20, i32 24, i32 28>806 %strided.vec36 = shufflevector <32 x i16> %wide.vec, <32 x i16> poison, <8 x i32> <i32 1, i32 5, i32 9, i32 13, i32 17, i32 21, i32 25, i32 29>807 %strided.vec37 = shufflevector <32 x i16> %wide.vec, <32 x i16> poison, <8 x i32> <i32 2, i32 6, i32 10, i32 14, i32 18, i32 22, i32 26, i32 30>808 %strided.vec38 = shufflevector <32 x i16> %wide.vec, <32 x i16> poison, <8 x i32> <i32 3, i32 7, i32 11, i32 15, i32 19, i32 23, i32 27, i32 31>809 %1 = uitofp <8 x i16> %strided.vec to <8 x double>810 %2 = uitofp <8 x i16> %strided.vec36 to <8 x double>811 %3 = fadd fast <8 x double> %2, %1812 %4 = uitofp <8 x i16> %strided.vec37 to <8 x double>813 %5 = fadd fast <8 x double> %3, %4814 %6 = uitofp <8 x i16> %strided.vec38 to <8 x double>815 %7 = fadd fast <8 x double> %5, %6816 %8 = getelementptr inbounds double, ptr %y, i64 %index817 store <8 x double> %7, ptr %8, align 8818 %index.next = add nuw i64 %index, 8819 %9 = icmp eq i64 %index.next, 1024820 br i1 %9, label %for.cond.cleanup, label %vector.body821 822for.cond.cleanup:823 ret void824}825 826