233 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc -mtriple=aarch64-linux-gnu -o - %s | FileCheck %s3 4define <8 x i16> @mul_splat_sext_v8i16(ptr %x, ptr %y) {5; CHECK-LABEL: mul_splat_sext_v8i16:6; CHECK: // %bb.0: // %entry7; CHECK-NEXT: ldr d1, [x0]8; CHECK-NEXT: movi v0.2d, #00000000000000009; CHECK-NEXT: mov x8, xzr10; CHECK-NEXT: dup v1.8b, v1.b[3]11; CHECK-NEXT: .LBB0_1: // %l112; CHECK-NEXT: // =>This Inner Loop Header: Depth=113; CHECK-NEXT: ldr d2, [x1, x8]14; CHECK-NEXT: add x8, x8, #415; CHECK-NEXT: cmp w8, #416; CHECK-NEXT: smlal v0.8h, v2.8b, v1.8b17; CHECK-NEXT: b.eq .LBB0_118; CHECK-NEXT: // %bb.2: // %l219; CHECK-NEXT: ret20entry:21 %x.val = load <8 x i8>, ptr %x22 %x.ext = sext <8 x i8> %x.val to <8 x i16>23 %a = shufflevector <8 x i16> %x.ext, <8 x i16> undef, <8 x i32> <i32 3, i32 3, i32 3, i32 3, i32 3, i32 3, i32 3, i32 3>24 br label %l125 26l1:27 %p = phi i32 [ 0, %entry ], [ %pa, %l1 ]28 %q = phi <8 x i16> [ zeroinitializer, %entry ], [ %c, %l1 ]29 %y.idx = mul nuw nsw i32 %p, 430 %y.ptr = getelementptr i8, ptr %y, i32 %y.idx31 %y.val = load <8 x i8>, ptr %y.ptr32 %y.ext = sext <8 x i8> %y.val to <8 x i16>33 %b = mul <8 x i16> %y.ext, %a34 %c = add <8 x i16> %q, %b35 %pa = add i32 %p, 136 %c1 = icmp eq i32 %p, 037 br i1 %c1, label %l1, label %l238 39l2:40 ret <8 x i16> %c41}42 43define <4 x i32> @mul_splat_sext_v4i32(ptr %x, ptr %y) {44; CHECK-LABEL: mul_splat_sext_v4i32:45; CHECK: // %bb.0: // %entry46; CHECK-NEXT: movi v0.2d, #000000000000000047; CHECK-NEXT: ldr d1, [x0]48; CHECK-NEXT: mov x8, xzr49; CHECK-NEXT: .LBB1_1: // %l150; CHECK-NEXT: // =>This Inner Loop Header: Depth=151; CHECK-NEXT: ldr d2, [x1, x8]52; CHECK-NEXT: add x8, x8, #853; CHECK-NEXT: cmp w8, #854; CHECK-NEXT: smlal v0.4s, v2.4h, v1.h[3]55; CHECK-NEXT: b.eq .LBB1_156; CHECK-NEXT: // %bb.2: // %l257; CHECK-NEXT: ret58entry:59 %x.val = load <4 x i16>, ptr %x60 %x.ext = sext <4 x i16> %x.val to <4 x i32>61 %a = shufflevector <4 x i32> %x.ext, <4 x i32> undef, <4 x i32> <i32 3, i32 3, i32 3, i32 3>62 br label %l163 64l1:65 %p = phi i32 [ 0, %entry ], [ %pa, %l1 ]66 %q = phi <4 x i32> [ zeroinitializer, %entry ], [ %c, %l1 ]67 %y.idx = mul nuw nsw i32 %p, 468 %y.ptr = getelementptr i16, ptr %y, i32 %y.idx69 %y.val = load <4 x i16>, ptr %y.ptr70 %y.ext = sext <4 x i16> %y.val to <4 x i32>71 %b = mul <4 x i32> %y.ext, %a72 %c = add <4 x i32> %q, %b73 %pa = add i32 %p, 174 %c1 = icmp eq i32 %p, 075 br i1 %c1, label %l1, label %l276 77l2:78 ret <4 x i32> %c79}80 81define <2 x i64> @mul_splat_sext_v2i64(ptr %x, ptr %y) {82; CHECK-LABEL: mul_splat_sext_v2i64:83; CHECK: // %bb.0: // %entry84; CHECK-NEXT: movi v0.2d, #000000000000000085; CHECK-NEXT: ldr d1, [x0]86; CHECK-NEXT: mov x8, xzr87; CHECK-NEXT: .LBB2_1: // %l188; CHECK-NEXT: // =>This Inner Loop Header: Depth=189; CHECK-NEXT: ldr d2, [x1, x8]90; CHECK-NEXT: add x8, x8, #1691; CHECK-NEXT: cmp w8, #1692; CHECK-NEXT: smlal v0.2d, v2.2s, v1.s[1]93; CHECK-NEXT: b.eq .LBB2_194; CHECK-NEXT: // %bb.2: // %l295; CHECK-NEXT: ret96entry:97 %x.val = load <2 x i32>, ptr %x98 %x.ext = sext <2 x i32> %x.val to <2 x i64>99 %a = shufflevector <2 x i64> %x.ext, <2 x i64> undef, <2 x i32> <i32 1, i32 1>100 br label %l1101 102l1:103 %p = phi i32 [ 0, %entry ], [ %pa, %l1 ]104 %q = phi <2 x i64> [ zeroinitializer, %entry ], [ %c, %l1 ]105 %y.idx = mul nuw nsw i32 %p, 4106 %y.ptr = getelementptr i32, ptr %y, i32 %y.idx107 %y.val = load <2 x i32>, ptr %y.ptr108 %y.ext = sext <2 x i32> %y.val to <2 x i64>109 %b = mul <2 x i64> %y.ext, %a110 %c = add <2 x i64> %q, %b111 %pa = add i32 %p, 1112 %c1 = icmp eq i32 %p, 0113 br i1 %c1, label %l1, label %l2114 115l2:116 ret <2 x i64> %c117}118 119define <8 x i16> @mul_sext_splat_v8i16(ptr %x, ptr %y) {120; CHECK-LABEL: mul_sext_splat_v8i16:121; CHECK: // %bb.0: // %entry122; CHECK-NEXT: ldr d1, [x0]123; CHECK-NEXT: movi v0.2d, #0000000000000000124; CHECK-NEXT: mov x8, xzr125; CHECK-NEXT: dup v1.8b, v1.b[3]126; CHECK-NEXT: .LBB3_1: // %l1127; CHECK-NEXT: // =>This Inner Loop Header: Depth=1128; CHECK-NEXT: ldr d2, [x1, x8]129; CHECK-NEXT: add x8, x8, #4130; CHECK-NEXT: cmp w8, #4131; CHECK-NEXT: smlal v0.8h, v2.8b, v1.8b132; CHECK-NEXT: b.eq .LBB3_1133; CHECK-NEXT: // %bb.2: // %l2134; CHECK-NEXT: ret135entry:136 %x.val = load <8 x i8>, ptr %x137 %x.spt = shufflevector <8 x i8> %x.val, <8 x i8> undef, <8 x i32> <i32 3, i32 3, i32 3, i32 3, i32 3, i32 3, i32 3, i32 3>138 %a = sext <8 x i8> %x.spt to <8 x i16>139 br label %l1140 141l1:142 %p = phi i32 [ 0, %entry ], [ %pa, %l1 ]143 %q = phi <8 x i16> [ zeroinitializer, %entry ], [ %c, %l1 ]144 %y.idx = mul nuw nsw i32 %p, 4145 %y.ptr = getelementptr i8, ptr %y, i32 %y.idx146 %y.val = load <8 x i8>, ptr %y.ptr147 %y.ext = sext <8 x i8> %y.val to <8 x i16>148 %b = mul <8 x i16> %y.ext, %a149 %c = add <8 x i16> %q, %b150 %pa = add i32 %p, 1151 %c1 = icmp eq i32 %p, 0152 br i1 %c1, label %l1, label %l2153 154l2:155 ret <8 x i16> %c156}157 158define <4 x i32> @mul_sext_splat_v4i32(ptr %x, ptr %y) {159; CHECK-LABEL: mul_sext_splat_v4i32:160; CHECK: // %bb.0: // %entry161; CHECK-NEXT: movi v0.2d, #0000000000000000162; CHECK-NEXT: ldr d1, [x0]163; CHECK-NEXT: mov x8, xzr164; CHECK-NEXT: .LBB4_1: // %l1165; CHECK-NEXT: // =>This Inner Loop Header: Depth=1166; CHECK-NEXT: ldr d2, [x1, x8]167; CHECK-NEXT: add x8, x8, #8168; CHECK-NEXT: cmp w8, #8169; CHECK-NEXT: smlal v0.4s, v2.4h, v1.h[3]170; CHECK-NEXT: b.eq .LBB4_1171; CHECK-NEXT: // %bb.2: // %l2172; CHECK-NEXT: ret173entry:174 %x.val = load <4 x i16>, ptr %x175 %x.spt = shufflevector <4 x i16> %x.val, <4 x i16> undef, <4 x i32> <i32 3, i32 3, i32 3, i32 3>176 %a = sext <4 x i16> %x.spt to <4 x i32>177 br label %l1178 179l1:180 %p = phi i32 [ 0, %entry ], [ %pa, %l1 ]181 %q = phi <4 x i32> [ zeroinitializer, %entry ], [ %c, %l1 ]182 %y.idx = mul nuw nsw i32 %p, 4183 %y.ptr = getelementptr i16, ptr %y, i32 %y.idx184 %y.val = load <4 x i16>, ptr %y.ptr185 %y.ext = sext <4 x i16> %y.val to <4 x i32>186 %b = mul <4 x i32> %y.ext, %a187 %c = add <4 x i32> %q, %b188 %pa = add i32 %p, 1189 %c1 = icmp eq i32 %p, 0190 br i1 %c1, label %l1, label %l2191 192l2:193 ret <4 x i32> %c194}195 196define <2 x i64> @mul_sext_splat_v2i64(ptr %x, ptr %y) {197; CHECK-LABEL: mul_sext_splat_v2i64:198; CHECK: // %bb.0: // %entry199; CHECK-NEXT: movi v0.2d, #0000000000000000200; CHECK-NEXT: ldr d1, [x0]201; CHECK-NEXT: mov x8, xzr202; CHECK-NEXT: .LBB5_1: // %l1203; CHECK-NEXT: // =>This Inner Loop Header: Depth=1204; CHECK-NEXT: ldr d2, [x1, x8]205; CHECK-NEXT: add x8, x8, #16206; CHECK-NEXT: cmp w8, #16207; CHECK-NEXT: smlal v0.2d, v2.2s, v1.s[1]208; CHECK-NEXT: b.eq .LBB5_1209; CHECK-NEXT: // %bb.2: // %l2210; CHECK-NEXT: ret211entry:212 %x.val = load <2 x i32>, ptr %x213 %x.spt = shufflevector <2 x i32> %x.val, <2 x i32> undef, <2 x i32> <i32 1, i32 1>214 %a = sext <2 x i32> %x.spt to <2 x i64>215 br label %l1216 217l1:218 %p = phi i32 [ 0, %entry ], [ %pa, %l1 ]219 %q = phi <2 x i64> [ zeroinitializer, %entry ], [ %c, %l1 ]220 %y.idx = mul nuw nsw i32 %p, 4221 %y.ptr = getelementptr i32, ptr %y, i32 %y.idx222 %y.val = load <2 x i32>, ptr %y.ptr223 %y.ext = sext <2 x i32> %y.val to <2 x i64>224 %b = mul <2 x i64> %y.ext, %a225 %c = add <2 x i64> %q, %b226 %pa = add i32 %p, 1227 %c1 = icmp eq i32 %p, 0228 br i1 %c1, label %l1, label %l2229 230l2:231 ret <2 x i64> %c232}233