114 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py2; RUN: opt -S -passes=slp-vectorizer -mtriple=aarch64--linux-gnu < %s | FileCheck %s3 4target datalayout = "e-m:e-i32:64-i128:128-n32:64-S128"5 6declare void @foo(i64, i64, i64, i64)7 8define void @test1(<4 x i16> %a, <4 x i16> %b, ptr %p) {9; Make sure types of sub and its sources are not extended.10; CHECK-LABEL: @test1(11; CHECK-NEXT: entry:12; CHECK-NEXT: [[Z0:%.*]] = zext <4 x i16> [[A:%.*]] to <4 x i32>13; CHECK-NEXT: [[Z1:%.*]] = zext <4 x i16> [[B:%.*]] to <4 x i32>14; CHECK-NEXT: [[SUB0:%.*]] = sub <4 x i32> [[Z0]], [[Z1]]15; CHECK-NEXT: [[E0:%.*]] = extractelement <4 x i32> [[SUB0]], i32 016; CHECK-NEXT: [[S0:%.*]] = sext i32 [[E0]] to i6417; CHECK-NEXT: [[GEP0:%.*]] = getelementptr inbounds i64, ptr [[P:%.*]], i64 [[S0]]18; CHECK-NEXT: [[LOAD0:%.*]] = load i64, ptr [[GEP0]], align 419; CHECK-NEXT: [[TMP1:%.*]] = extractelement <4 x i32> [[SUB0]], i32 120; CHECK-NEXT: [[TMP2:%.*]] = sext i32 [[TMP1]] to i6421; CHECK-NEXT: [[GEP1:%.*]] = getelementptr inbounds i64, ptr [[P]], i64 [[TMP2]]22; CHECK-NEXT: [[LOAD1:%.*]] = load i64, ptr [[GEP1]], align 423; CHECK-NEXT: [[TMP3:%.*]] = extractelement <4 x i32> [[SUB0]], i32 224; CHECK-NEXT: [[TMP4:%.*]] = sext i32 [[TMP3]] to i6425; CHECK-NEXT: [[GEP2:%.*]] = getelementptr inbounds i64, ptr [[P]], i64 [[TMP4]]26; CHECK-NEXT: [[LOAD2:%.*]] = load i64, ptr [[GEP2]], align 427; CHECK-NEXT: [[E3:%.*]] = extractelement <4 x i32> [[SUB0]], i32 328; CHECK-NEXT: [[S3:%.*]] = sext i32 [[E3]] to i6429; CHECK-NEXT: [[GEP3:%.*]] = getelementptr inbounds i64, ptr [[P]], i64 [[S3]]30; CHECK-NEXT: [[LOAD3:%.*]] = load i64, ptr [[GEP3]], align 431; CHECK-NEXT: call void @foo(i64 [[LOAD0]], i64 [[LOAD1]], i64 [[LOAD2]], i64 [[LOAD3]])32; CHECK-NEXT: ret void33;34entry:35 %z0 = zext <4 x i16> %a to <4 x i32>36 %z1 = zext <4 x i16> %b to <4 x i32>37 %sub0 = sub <4 x i32> %z0, %z138 %e0 = extractelement <4 x i32> %sub0, i32 039 %s0 = sext i32 %e0 to i6440 %gep0 = getelementptr inbounds i64, ptr %p, i64 %s041 %load0 = load i64, ptr %gep042 %e1 = extractelement <4 x i32> %sub0, i32 143 %s1 = sext i32 %e1 to i6444 %gep1 = getelementptr inbounds i64, ptr %p, i64 %s145 %load1 = load i64, ptr %gep146 %e2 = extractelement <4 x i32> %sub0, i32 247 %s2 = sext i32 %e2 to i6448 %gep2 = getelementptr inbounds i64, ptr %p, i64 %s249 %load2 = load i64, ptr %gep250 %e3 = extractelement <4 x i32> %sub0, i32 351 %s3 = sext i32 %e3 to i6452 %gep3 = getelementptr inbounds i64, ptr %p, i64 %s353 %load3 = load i64, ptr %gep354 call void @foo(i64 %load0, i64 %load1, i64 %load2, i64 %load3)55 ret void56}57 58define void @test2(<4 x i16> %a, <4 x i16> %b, i64 %c0, i64 %c1, i64 %c2, i64 %c3, ptr %p) {59; CHECK-LABEL: @test2(60; CHECK-NEXT: entry:61; CHECK-NEXT: [[Z0:%.*]] = zext <4 x i16> [[A:%.*]] to <4 x i32>62; CHECK-NEXT: [[Z1:%.*]] = zext <4 x i16> [[B:%.*]] to <4 x i32>63; CHECK-NEXT: [[SUB0:%.*]] = sub <4 x i32> [[Z0]], [[Z1]]64; CHECK-NEXT: [[E0:%.*]] = extractelement <4 x i32> [[SUB0]], i32 065; CHECK-NEXT: [[S0:%.*]] = sext i32 [[E0]] to i6466; CHECK-NEXT: [[A0:%.*]] = add i64 [[S0]], [[C0:%.*]]67; CHECK-NEXT: [[GEP0:%.*]] = getelementptr inbounds i64, ptr [[P:%.*]], i64 [[A0]]68; CHECK-NEXT: [[LOAD0:%.*]] = load i64, ptr [[GEP0]], align 469; CHECK-NEXT: [[E1:%.*]] = extractelement <4 x i32> [[SUB0]], i32 170; CHECK-NEXT: [[S1:%.*]] = sext i32 [[E1]] to i6471; CHECK-NEXT: [[A1:%.*]] = add i64 [[S1]], [[C1:%.*]]72; CHECK-NEXT: [[GEP1:%.*]] = getelementptr inbounds i64, ptr [[P]], i64 [[A1]]73; CHECK-NEXT: [[LOAD1:%.*]] = load i64, ptr [[GEP1]], align 474; CHECK-NEXT: [[E2:%.*]] = extractelement <4 x i32> [[SUB0]], i32 275; CHECK-NEXT: [[S2:%.*]] = sext i32 [[E2]] to i6476; CHECK-NEXT: [[A2:%.*]] = add i64 [[S2]], [[C2:%.*]]77; CHECK-NEXT: [[GEP2:%.*]] = getelementptr inbounds i64, ptr [[P]], i64 [[A2]]78; CHECK-NEXT: [[LOAD2:%.*]] = load i64, ptr [[GEP2]], align 479; CHECK-NEXT: [[E3:%.*]] = extractelement <4 x i32> [[SUB0]], i32 380; CHECK-NEXT: [[S3:%.*]] = sext i32 [[E3]] to i6481; CHECK-NEXT: [[A3:%.*]] = add i64 [[S3]], [[C3:%.*]]82; CHECK-NEXT: [[GEP3:%.*]] = getelementptr inbounds i64, ptr [[P]], i64 [[A3]]83; CHECK-NEXT: [[LOAD3:%.*]] = load i64, ptr [[GEP3]], align 484; CHECK-NEXT: call void @foo(i64 [[LOAD0]], i64 [[LOAD1]], i64 [[LOAD2]], i64 [[LOAD3]])85; CHECK-NEXT: ret void86;87entry:88 %z0 = zext <4 x i16> %a to <4 x i32>89 %z1 = zext <4 x i16> %b to <4 x i32>90 %sub0 = sub <4 x i32> %z0, %z191 %e0 = extractelement <4 x i32> %sub0, i32 092 %s0 = sext i32 %e0 to i6493 %a0 = add i64 %s0, %c094 %gep0 = getelementptr inbounds i64, ptr %p, i64 %a095 %load0 = load i64, ptr %gep096 %e1 = extractelement <4 x i32> %sub0, i32 197 %s1 = sext i32 %e1 to i6498 %a1 = add i64 %s1, %c199 %gep1 = getelementptr inbounds i64, ptr %p, i64 %a1100 %load1 = load i64, ptr %gep1101 %e2 = extractelement <4 x i32> %sub0, i32 2102 %s2 = sext i32 %e2 to i64103 %a2 = add i64 %s2, %c2104 %gep2 = getelementptr inbounds i64, ptr %p, i64 %a2105 %load2 = load i64, ptr %gep2106 %e3 = extractelement <4 x i32> %sub0, i32 3107 %s3 = sext i32 %e3 to i64108 %a3 = add i64 %s3, %c3109 %gep3 = getelementptr inbounds i64, ptr %p, i64 %a3110 %load3 = load i64, ptr %gep3111 call void @foo(i64 %load0, i64 %load1, i64 %load2, i64 %load3)112 ret void113}114