117 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py2; RUN: opt -o - -S -passes=load-store-vectorizer,dce %s | FileCheck %s3 4; Make sure LoadStoreVectorizer vectorizes the loads below.5; In order to prove that the vectorization is safe, it tries to6; match nested adds and find an expression that adds a constant7; value to an existing index and the result doesn't overflow.8 9target triple = "x86_64--"10 11define void @ld_v4i8_add_nsw(i32 %v0, i32 %v1, ptr %src, ptr %dst) {12; CHECK-LABEL: @ld_v4i8_add_nsw(13; CHECK-NEXT: bb:14; CHECK-NEXT: [[TMP:%.*]] = add nsw i32 [[V0:%.*]], -115; CHECK-NEXT: [[TMP1:%.*]] = add nsw i32 [[V1:%.*]], [[TMP]]16; CHECK-NEXT: [[TMP2:%.*]] = sext i32 [[TMP1]] to i6417; CHECK-NEXT: [[TMP3:%.*]] = getelementptr inbounds i8, ptr [[SRC:%.*]], i64 [[TMP2]]18; CHECK-NEXT: [[TMP1:%.*]] = load <4 x i8>, ptr [[TMP3]], align 119; CHECK-NEXT: [[TMP41:%.*]] = extractelement <4 x i8> [[TMP1]], i32 020; CHECK-NEXT: [[TMP82:%.*]] = extractelement <4 x i8> [[TMP1]], i32 121; CHECK-NEXT: [[TMP133:%.*]] = extractelement <4 x i8> [[TMP1]], i32 222; CHECK-NEXT: [[TMP184:%.*]] = extractelement <4 x i8> [[TMP1]], i32 323; CHECK-NEXT: [[TMP19:%.*]] = insertelement <4 x i8> poison, i8 [[TMP41]], i32 024; CHECK-NEXT: [[TMP20:%.*]] = insertelement <4 x i8> [[TMP19]], i8 [[TMP82]], i32 125; CHECK-NEXT: [[TMP21:%.*]] = insertelement <4 x i8> [[TMP20]], i8 [[TMP133]], i32 226; CHECK-NEXT: [[TMP22:%.*]] = insertelement <4 x i8> [[TMP21]], i8 [[TMP184]], i32 327; CHECK-NEXT: store <4 x i8> [[TMP22]], ptr [[DST:%.*]], align 428; CHECK-NEXT: ret void29;30bb:31 %tmp = add nsw i32 %v0, -132 %tmp1 = add nsw i32 %v1, %tmp33 %tmp2 = sext i32 %tmp1 to i6434 %tmp3 = getelementptr inbounds i8, ptr %src, i64 %tmp235 %tmp4 = load i8, ptr %tmp3, align 136 %tmp5 = add nsw i32 %v1, %v037 %tmp6 = sext i32 %tmp5 to i6438 %tmp7 = getelementptr inbounds i8, ptr %src, i64 %tmp639 %tmp8 = load i8, ptr %tmp7, align 140 %tmp9 = add nsw i32 %v0, 141 %tmp10 = add nsw i32 %v1, %tmp942 %tmp11 = sext i32 %tmp10 to i6443 %tmp12 = getelementptr inbounds i8, ptr %src, i64 %tmp1144 %tmp13 = load i8, ptr %tmp12, align 145 %tmp14 = add nsw i32 %v0, 246 %tmp15 = add nsw i32 %v1, %tmp1447 %tmp16 = sext i32 %tmp15 to i6448 %tmp17 = getelementptr inbounds i8, ptr %src, i64 %tmp1649 %tmp18 = load i8, ptr %tmp17, align 150 %tmp19 = insertelement <4 x i8> poison, i8 %tmp4, i32 051 %tmp20 = insertelement <4 x i8> %tmp19, i8 %tmp8, i32 152 %tmp21 = insertelement <4 x i8> %tmp20, i8 %tmp13, i32 253 %tmp22 = insertelement <4 x i8> %tmp21, i8 %tmp18, i32 354 store <4 x i8> %tmp22, ptr %dst55 ret void56}57 58; Make sure we don't vectorize the loads below because the source of59; sext instructions doesn't have the nsw flag.60 61define void @ld_v4i8_add_not_safe(i32 %v0, i32 %v1, ptr %src, ptr %dst) {62; CHECK-LABEL: @ld_v4i8_add_not_safe(63; CHECK-NEXT: bb:64; CHECK-NEXT: [[TMP:%.*]] = add nsw i32 [[V0:%.*]], -165; CHECK-NEXT: [[TMP1:%.*]] = add i32 [[V1:%.*]], [[TMP]]66; CHECK-NEXT: [[TMP2:%.*]] = sext i32 [[TMP1]] to i6467; CHECK-NEXT: [[TMP3:%.*]] = getelementptr inbounds i8, ptr [[SRC:%.*]], i64 [[TMP2]]68; CHECK-NEXT: [[TMP4:%.*]] = load i8, ptr [[TMP3]], align 169; CHECK-NEXT: [[TMP5:%.*]] = add i32 [[V1]], [[V0]]70; CHECK-NEXT: [[TMP6:%.*]] = sext i32 [[TMP5]] to i6471; CHECK-NEXT: [[TMP7:%.*]] = getelementptr inbounds i8, ptr [[SRC]], i64 [[TMP6]]72; CHECK-NEXT: [[TMP8:%.*]] = load i8, ptr [[TMP7]], align 173; CHECK-NEXT: [[TMP9:%.*]] = add nsw i32 [[V0]], 174; CHECK-NEXT: [[TMP10:%.*]] = add i32 [[V1]], [[TMP9]]75; CHECK-NEXT: [[TMP11:%.*]] = sext i32 [[TMP10]] to i6476; CHECK-NEXT: [[TMP12:%.*]] = getelementptr inbounds i8, ptr [[SRC]], i64 [[TMP11]]77; CHECK-NEXT: [[TMP13:%.*]] = load i8, ptr [[TMP12]], align 178; CHECK-NEXT: [[TMP14:%.*]] = add nsw i32 [[V0]], 279; CHECK-NEXT: [[TMP15:%.*]] = add i32 [[V1]], [[TMP14]]80; CHECK-NEXT: [[TMP16:%.*]] = sext i32 [[TMP15]] to i6481; CHECK-NEXT: [[TMP17:%.*]] = getelementptr inbounds i8, ptr [[SRC]], i64 [[TMP16]]82; CHECK-NEXT: [[TMP18:%.*]] = load i8, ptr [[TMP17]], align 183; CHECK-NEXT: [[TMP19:%.*]] = insertelement <4 x i8> poison, i8 [[TMP4]], i32 084; CHECK-NEXT: [[TMP20:%.*]] = insertelement <4 x i8> [[TMP19]], i8 [[TMP8]], i32 185; CHECK-NEXT: [[TMP21:%.*]] = insertelement <4 x i8> [[TMP20]], i8 [[TMP13]], i32 286; CHECK-NEXT: [[TMP22:%.*]] = insertelement <4 x i8> [[TMP21]], i8 [[TMP18]], i32 387; CHECK-NEXT: store <4 x i8> [[TMP22]], ptr [[DST:%.*]], align 488; CHECK-NEXT: ret void89;90bb:91 %tmp = add nsw i32 %v0, -192 %tmp1 = add i32 %v1, %tmp93 %tmp2 = sext i32 %tmp1 to i6494 %tmp3 = getelementptr inbounds i8, ptr %src, i64 %tmp295 %tmp4 = load i8, ptr %tmp3, align 196 %tmp5 = add i32 %v1, %v097 %tmp6 = sext i32 %tmp5 to i6498 %tmp7 = getelementptr inbounds i8, ptr %src, i64 %tmp699 %tmp8 = load i8, ptr %tmp7, align 1100 %tmp9 = add nsw i32 %v0, 1101 %tmp10 = add i32 %v1, %tmp9102 %tmp11 = sext i32 %tmp10 to i64103 %tmp12 = getelementptr inbounds i8, ptr %src, i64 %tmp11104 %tmp13 = load i8, ptr %tmp12, align 1105 %tmp14 = add nsw i32 %v0, 2106 %tmp15 = add i32 %v1, %tmp14107 %tmp16 = sext i32 %tmp15 to i64108 %tmp17 = getelementptr inbounds i8, ptr %src, i64 %tmp16109 %tmp18 = load i8, ptr %tmp17, align 1110 %tmp19 = insertelement <4 x i8> poison, i8 %tmp4, i32 0111 %tmp20 = insertelement <4 x i8> %tmp19, i8 %tmp8, i32 1112 %tmp21 = insertelement <4 x i8> %tmp20, i8 %tmp13, i32 2113 %tmp22 = insertelement <4 x i8> %tmp21, i8 %tmp18, i32 3114 store <4 x i8> %tmp22, ptr %dst115 ret void116}117