174 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py2;test_i16_extend NOTE: Assertions have been autogenerated by utils/update_test_checks.py3; RUN: opt -S -mtriple=aarch64--linux-gnu -passes=slp-vectorizer -slp-threshold=-5 -pass-remarks-output=%t < %s | FileCheck %s4; RUN: cat %t | FileCheck -check-prefix=YAML %s5; RUN: opt -S -mtriple=aarch64--linux-gnu -passes=slp-vectorizer -slp-threshold=-5 -pass-remarks-output=%t < %s | FileCheck %s6; RUN: cat %t | FileCheck -check-prefix=YAML %s7 8 9@global = internal global { ptr } zeroinitializer, align 810 11; YAML-LABEL: --- !Passed12; YAML-NEXT: Pass: slp-vectorizer13; YAML-NEXT: Name: VectorizedList14; YAML-NEXT: Function: test_i16_extend15; YAML-NEXT: Args:16; YAML-NEXT: - String: 'SLP vectorized with cost '17; YAML-NEXT: - Cost: '-1618; YAML-NEXT: - String: ' and with tree size '19; YAML-NEXT: - TreeSize: '5'20; YAML-NEXT: ...21 22; Make sure we vectorize to maximize the load with when loading i16 and23; extending it for compute operations.24define void @test_i16_extend(ptr %p.1, ptr %p.2, i32 %idx.i32) {25; CHECK-LABEL: @test_i16_extend(26; CHECK-NEXT: [[P_0:%.*]] = load ptr, ptr @global, align 827; CHECK-NEXT: [[IDX_0:%.*]] = zext i32 [[IDX_I32:%.*]] to i6428; CHECK-NEXT: [[T53:%.*]] = getelementptr inbounds i16, ptr [[P_1:%.*]], i64 [[IDX_0]]29; CHECK-NEXT: [[T56:%.*]] = getelementptr inbounds i16, ptr [[P_2:%.*]], i64 [[IDX_0]]30; CHECK-NEXT: [[TMP1:%.*]] = load <8 x i16>, ptr [[T53]], align 231; CHECK-NEXT: [[TMP2:%.*]] = zext <8 x i16> [[TMP1]] to <8 x i32>32; CHECK-NEXT: [[TMP3:%.*]] = load <8 x i16>, ptr [[T56]], align 233; CHECK-NEXT: [[TMP4:%.*]] = zext <8 x i16> [[TMP3]] to <8 x i32>34; CHECK-NEXT: [[TMP5:%.*]] = sub <8 x i32> [[TMP2]], [[TMP4]]35; CHECK-NEXT: [[TMP6:%.*]] = extractelement <8 x i32> [[TMP5]], i32 036; CHECK-NEXT: [[TMP7:%.*]] = sext i32 [[TMP6]] to i6437; CHECK-NEXT: [[T60:%.*]] = getelementptr inbounds i32, ptr [[P_0]], i64 [[TMP7]]38; CHECK-NEXT: [[L_1:%.*]] = load i32, ptr [[T60]], align 439; CHECK-NEXT: [[TMP8:%.*]] = extractelement <8 x i32> [[TMP5]], i32 140; CHECK-NEXT: [[TMP9:%.*]] = sext i32 [[TMP8]] to i6441; CHECK-NEXT: [[T71:%.*]] = getelementptr inbounds i32, ptr [[P_0]], i64 [[TMP9]]42; CHECK-NEXT: [[L_2:%.*]] = load i32, ptr [[T71]], align 443; CHECK-NEXT: [[TMP10:%.*]] = extractelement <8 x i32> [[TMP5]], i32 244; CHECK-NEXT: [[TMP11:%.*]] = sext i32 [[TMP10]] to i6445; CHECK-NEXT: [[T82:%.*]] = getelementptr inbounds i32, ptr [[P_0]], i64 [[TMP11]]46; CHECK-NEXT: [[L_3:%.*]] = load i32, ptr [[T82]], align 447; CHECK-NEXT: [[TMP12:%.*]] = extractelement <8 x i32> [[TMP5]], i32 348; CHECK-NEXT: [[TMP13:%.*]] = sext i32 [[TMP12]] to i6449; CHECK-NEXT: [[T93:%.*]] = getelementptr inbounds i32, ptr [[P_0]], i64 [[TMP13]]50; CHECK-NEXT: [[L_4:%.*]] = load i32, ptr [[T93]], align 451; CHECK-NEXT: [[TMP14:%.*]] = extractelement <8 x i32> [[TMP5]], i32 452; CHECK-NEXT: [[TMP15:%.*]] = sext i32 [[TMP14]] to i6453; CHECK-NEXT: [[T104:%.*]] = getelementptr inbounds i32, ptr [[P_0]], i64 [[TMP15]]54; CHECK-NEXT: [[L_5:%.*]] = load i32, ptr [[T104]], align 455; CHECK-NEXT: [[TMP16:%.*]] = extractelement <8 x i32> [[TMP5]], i32 556; CHECK-NEXT: [[TMP17:%.*]] = sext i32 [[TMP16]] to i6457; CHECK-NEXT: [[T115:%.*]] = getelementptr inbounds i32, ptr [[P_0]], i64 [[TMP17]]58; CHECK-NEXT: [[L_6:%.*]] = load i32, ptr [[T115]], align 459; CHECK-NEXT: [[TMP18:%.*]] = extractelement <8 x i32> [[TMP5]], i32 660; CHECK-NEXT: [[TMP19:%.*]] = sext i32 [[TMP18]] to i6461; CHECK-NEXT: [[T126:%.*]] = getelementptr inbounds i32, ptr [[P_0]], i64 [[TMP19]]62; CHECK-NEXT: [[L_7:%.*]] = load i32, ptr [[T126]], align 463; CHECK-NEXT: [[TMP20:%.*]] = extractelement <8 x i32> [[TMP5]], i32 764; CHECK-NEXT: [[TMP21:%.*]] = sext i32 [[TMP20]] to i6465; CHECK-NEXT: [[T137:%.*]] = getelementptr inbounds i32, ptr [[P_0]], i64 [[TMP21]]66; CHECK-NEXT: [[L_8:%.*]] = load i32, ptr [[T137]], align 467; CHECK-NEXT: call void @use(i32 [[L_1]], i32 [[L_2]], i32 [[L_3]], i32 [[L_4]], i32 [[L_5]], i32 [[L_6]], i32 [[L_7]], i32 [[L_8]])68; CHECK-NEXT: ret void69;70 %p.0 = load ptr, ptr @global, align 871 72 %idx.0 = zext i32 %idx.i32 to i6473 %idx.1 = add nsw i64 %idx.0, 174 %idx.2 = add nsw i64 %idx.0, 275 %idx.3 = add nsw i64 %idx.0, 376 %idx.4 = add nsw i64 %idx.0, 477 %idx.5 = add nsw i64 %idx.0, 578 %idx.6 = add nsw i64 %idx.0, 679 %idx.7 = add nsw i64 %idx.0, 780 81 %t53 = getelementptr inbounds i16, ptr %p.1, i64 %idx.082 %op1.l = load i16, ptr %t53, align 283 %op1.ext = zext i16 %op1.l to i6484 %t56 = getelementptr inbounds i16, ptr %p.2, i64 %idx.085 %op2.l = load i16, ptr %t56, align 286 %op2.ext = zext i16 %op2.l to i6487 %sub.1 = sub nsw i64 %op1.ext, %op2.ext88 89 %t60 = getelementptr inbounds i32, ptr %p.0, i64 %sub.190 %l.1 = load i32, ptr %t60, align 491 92 %t64 = getelementptr inbounds i16, ptr %p.1, i64 %idx.193 %t65 = load i16, ptr %t64, align 294 %t66 = zext i16 %t65 to i6495 %t67 = getelementptr inbounds i16, ptr %p.2, i64 %idx.196 %t68 = load i16, ptr %t67, align 297 %t69 = zext i16 %t68 to i6498 %sub.2 = sub nsw i64 %t66, %t6999 100 %t71 = getelementptr inbounds i32, ptr %p.0, i64 %sub.2101 %l.2 = load i32, ptr %t71, align 4102 103 %t75 = getelementptr inbounds i16, ptr %p.1, i64 %idx.2104 %t76 = load i16, ptr %t75, align 2105 %t77 = zext i16 %t76 to i64106 %t78 = getelementptr inbounds i16, ptr %p.2, i64 %idx.2107 %t79 = load i16, ptr %t78, align 2108 %t80 = zext i16 %t79 to i64109 %sub.3 = sub nsw i64 %t77, %t80110 111 %t82 = getelementptr inbounds i32, ptr %p.0, i64 %sub.3112 %l.3 = load i32, ptr %t82, align 4113 114 %t86 = getelementptr inbounds i16, ptr %p.1, i64 %idx.3115 %t87 = load i16, ptr %t86, align 2116 %t88 = zext i16 %t87 to i64117 %t89 = getelementptr inbounds i16, ptr %p.2, i64 %idx.3118 %t90 = load i16, ptr %t89, align 2119 %t91 = zext i16 %t90 to i64120 %sub.4 = sub nsw i64 %t88, %t91121 122 %t93 = getelementptr inbounds i32, ptr %p.0, i64 %sub.4123 %l.4 = load i32, ptr %t93, align 4124 125 %t97 = getelementptr inbounds i16, ptr %p.1, i64 %idx.4126 %t98 = load i16, ptr %t97, align 2127 %t99 = zext i16 %t98 to i64128 %t100 = getelementptr inbounds i16, ptr %p.2, i64 %idx.4129 %t101 = load i16, ptr %t100, align 2130 %t102 = zext i16 %t101 to i64131 %sub.5 = sub nsw i64 %t99, %t102132 133 %t104 = getelementptr inbounds i32, ptr %p.0, i64 %sub.5134 %l.5 = load i32, ptr %t104, align 4135 136 %t108 = getelementptr inbounds i16, ptr %p.1, i64 %idx.5137 %t109 = load i16, ptr %t108, align 2138 %t110 = zext i16 %t109 to i64139 %t111 = getelementptr inbounds i16, ptr %p.2, i64 %idx.5140 %t112 = load i16, ptr %t111, align 2141 %t113 = zext i16 %t112 to i64142 %sub.6 = sub nsw i64 %t110, %t113143 144 %t115 = getelementptr inbounds i32, ptr %p.0, i64 %sub.6145 %l.6 = load i32, ptr %t115, align 4146 147 %t119 = getelementptr inbounds i16, ptr %p.1, i64 %idx.6148 %t120 = load i16, ptr %t119, align 2149 %t121 = zext i16 %t120 to i64150 %t122 = getelementptr inbounds i16, ptr %p.2, i64 %idx.6151 %t123 = load i16, ptr %t122, align 2152 %t124 = zext i16 %t123 to i64153 %sub.7 = sub nsw i64 %t121, %t124154 155 %t126 = getelementptr inbounds i32, ptr %p.0, i64 %sub.7156 %l.7 = load i32, ptr %t126, align 4157 158 %t130 = getelementptr inbounds i16, ptr %p.1, i64 %idx.7159 %t131 = load i16, ptr %t130, align 2160 %t132 = zext i16 %t131 to i64161 %t133 = getelementptr inbounds i16, ptr %p.2, i64 %idx.7162 %t134 = load i16, ptr %t133, align 2163 %t135 = zext i16 %t134 to i64164 %sub.8 = sub nsw i64 %t132, %t135165 166 %t137 = getelementptr inbounds i32, ptr %p.0, i64 %sub.8167 %l.8 = load i32, ptr %t137, align 4168 169 call void @use(i32 %l.1, i32 %l.2, i32 %l.3, i32 %l.4, i32 %l.5, i32 %l.6, i32 %l.7, i32 %l.8)170 ret void171}172 173declare void @use(i32, i32, i32, i32, i32, i32, i32, i32)174