1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py 2; RUN: opt -S -loop-predication -loop-predication-enable-iv-truncation=true < %s 2>&1 | FileCheck %s 3; RUN: opt -S -passes='require<scalar-evolution>,loop-mssa(loop-predication)' -verify-memoryssa < %s 2>&1 | FileCheck %s 4declare void @llvm.experimental.guard(i1, ...) 5 6declare i32 @length(i8*) 7 8declare i16 @short_length(i8*) 9; Consider range check of type i16 and i32, while IV is of type i64 10; We can loop predicate this because the IV range is within i16 and within i32. 11define i64 @iv_wider_type_rc_two_narrow_types(i32 %offA, i16 %offB, i8* %arrA, i8* %arrB) { 12; CHECK-LABEL: @iv_wider_type_rc_two_narrow_types( 13; CHECK-NEXT: entry: 14; CHECK-NEXT: [[LENGTHA:%.*]] = call i32 @length(i8* [[ARRA:%.*]]) 15; CHECK-NEXT: [[LENGTHB:%.*]] = call i16 @short_length(i8* [[ARRB:%.*]]) 16; CHECK-NEXT: [[TMP0:%.*]] = sub i16 [[LENGTHB]], [[OFFB:%.*]] 17; CHECK-NEXT: [[TMP1:%.*]] = icmp ule i16 16, [[TMP0]] 18; CHECK-NEXT: [[TMP2:%.*]] = icmp ult i16 [[OFFB]], [[LENGTHB]] 19; CHECK-NEXT: [[TMP3:%.*]] = and i1 [[TMP2]], [[TMP1]] 20; CHECK-NEXT: [[TMP4:%.*]] = sub i32 [[LENGTHA]], [[OFFA:%.*]] 21; CHECK-NEXT: [[TMP5:%.*]] = icmp ule i32 16, [[TMP4]] 22; CHECK-NEXT: [[TMP6:%.*]] = icmp ult i32 [[OFFA]], [[LENGTHA]] 23; CHECK-NEXT: [[TMP7:%.*]] = and i1 [[TMP6]], [[TMP5]] 24; CHECK-NEXT: [[TMP8:%.*]] = and i1 [[TMP3]], [[TMP7]] 25; CHECK-NEXT: br label [[LOOP:%.*]] 26; CHECK: loop: 27; CHECK-NEXT: [[IV:%.*]] = phi i64 [ 0, [[ENTRY:%.*]] ], [ [[IV_NEXT:%.*]], [[LOOP]] ] 28; CHECK-NEXT: [[IV_TRUNC_32:%.*]] = trunc i64 [[IV]] to i32 29; CHECK-NEXT: [[IV_TRUNC_16:%.*]] = trunc i64 [[IV]] to i16 30; CHECK-NEXT: [[INDEXA:%.*]] = add i32 [[IV_TRUNC_32]], [[OFFA]] 31; CHECK-NEXT: [[INDEXB:%.*]] = add i16 [[IV_TRUNC_16]], [[OFFB]] 32; CHECK-NEXT: call void (i1, ...) @llvm.experimental.guard(i1 [[TMP8]], i32 9) [ "deopt"() ] 33; CHECK-NEXT: [[INDEXA_EXT:%.*]] = zext i32 [[INDEXA]] to i64 34; CHECK-NEXT: [[ADDRA:%.*]] = getelementptr inbounds i8, i8* [[ARRA]], i64 [[INDEXA_EXT]] 35; CHECK-NEXT: [[ELTA:%.*]] = load i8, i8* [[ADDRA]] 36; CHECK-NEXT: [[INDEXB_EXT:%.*]] = zext i16 [[INDEXB]] to i64 37; CHECK-NEXT: [[ADDRB:%.*]] = getelementptr inbounds i8, i8* [[ARRB]], i64 [[INDEXB_EXT]] 38; CHECK-NEXT: store i8 [[ELTA]], i8* [[ADDRB]] 39; CHECK-NEXT: [[IV_NEXT]] = add nuw nsw i64 [[IV]], 1 40; CHECK-NEXT: [[LATCH_CHECK:%.*]] = icmp ult i64 [[IV_NEXT]], 16 41; CHECK-NEXT: br i1 [[LATCH_CHECK]], label [[LOOP]], label [[EXIT:%.*]] 42; CHECK: exit: 43; CHECK-NEXT: [[IV_LCSSA:%.*]] = phi i64 [ [[IV]], [[LOOP]] ] 44; CHECK-NEXT: ret i64 [[IV_LCSSA]] 45; 46entry: 47 %lengthA = call i32 @length(i8* %arrA) 48 %lengthB = call i16 @short_length(i8* %arrB) 49 br label %loop 50 51loop: 52 %iv = phi i64 [0, %entry ], [ %iv.next, %loop ] 53 %iv.trunc.32 = trunc i64 %iv to i32 54 %iv.trunc.16 = trunc i64 %iv to i16 55 %indexA = add i32 %iv.trunc.32, %offA 56 %indexB = add i16 %iv.trunc.16, %offB 57 %rcA = icmp ult i32 %indexA, %lengthA 58 %rcB = icmp ult i16 %indexB, %lengthB 59 %wide.chk = and i1 %rcA, %rcB 60 call void (i1, ...) @llvm.experimental.guard(i1 %wide.chk, i32 9) [ "deopt"() ] 61 %indexA.ext = zext i32 %indexA to i64 62 %addrA = getelementptr inbounds i8, i8* %arrA, i64 %indexA.ext 63 %eltA = load i8, i8* %addrA 64 %indexB.ext = zext i16 %indexB to i64 65 %addrB = getelementptr inbounds i8, i8* %arrB, i64 %indexB.ext 66 store i8 %eltA, i8* %addrB 67 %iv.next = add nuw nsw i64 %iv, 1 68 %latch.check = icmp ult i64 %iv.next, 16 69 br i1 %latch.check, label %loop, label %exit 70 71exit: 72 ret i64 %iv 73} 74 75 76; Consider an IV of type long and an array access into int array. 77; IV is of type i64 while the range check operands are of type i32 and i64. 78define i64 @iv_rc_different_types(i32 %offA, i32 %offB, i8* %arrA, i8* %arrB, i64 %max) 79; CHECK-LABEL: @iv_rc_different_types( 80; CHECK-NEXT: entry: 81; CHECK-NEXT: [[LENGTHA:%.*]] = call i32 @length(i8* [[ARRA:%.*]]) 82; CHECK-NEXT: [[LENGTHB:%.*]] = call i32 @length(i8* [[ARRB:%.*]]) 83; CHECK-NEXT: [[TMP0:%.*]] = add i32 [[LENGTHB]], -1 84; CHECK-NEXT: [[TMP1:%.*]] = sub i32 [[TMP0]], [[OFFB:%.*]] 85; CHECK-NEXT: [[TMP2:%.*]] = icmp ule i32 15, [[TMP1]] 86; CHECK-NEXT: [[TMP3:%.*]] = icmp ult i32 [[OFFB]], [[LENGTHB]] 87; CHECK-NEXT: [[TMP4:%.*]] = and i1 [[TMP3]], [[TMP2]] 88; CHECK-NEXT: [[TMP5:%.*]] = add i64 [[MAX:%.*]], -1 89; CHECK-NEXT: [[TMP6:%.*]] = icmp ule i64 15, [[TMP5]] 90; CHECK-NEXT: [[TMP7:%.*]] = icmp ult i64 0, [[MAX]] 91; CHECK-NEXT: [[TMP8:%.*]] = and i1 [[TMP7]], [[TMP6]] 92; CHECK-NEXT: [[TMP9:%.*]] = add i32 [[LENGTHA]], -1 93; CHECK-NEXT: [[TMP10:%.*]] = sub i32 [[TMP9]], [[OFFA:%.*]] 94; CHECK-NEXT: [[TMP11:%.*]] = icmp ule i32 15, [[TMP10]] 95; CHECK-NEXT: [[TMP12:%.*]] = icmp ult i32 [[OFFA]], [[LENGTHA]] 96; CHECK-NEXT: [[TMP13:%.*]] = and i1 [[TMP12]], [[TMP11]] 97; CHECK-NEXT: [[TMP14:%.*]] = and i1 [[TMP4]], [[TMP8]] 98; CHECK-NEXT: [[TMP15:%.*]] = and i1 [[TMP14]], [[TMP13]] 99; CHECK-NEXT: br label [[LOOP:%.*]] 100; CHECK: loop: 101; CHECK-NEXT: [[IV:%.*]] = phi i64 [ 0, [[ENTRY:%.*]] ], [ [[IV_NEXT:%.*]], [[LOOP]] ] 102; CHECK-NEXT: [[IV_TRUNC:%.*]] = trunc i64 [[IV]] to i32 103; CHECK-NEXT: [[INDEXA:%.*]] = add i32 [[IV_TRUNC]], [[OFFA]] 104; CHECK-NEXT: [[INDEXB:%.*]] = add i32 [[IV_TRUNC]], [[OFFB]] 105; CHECK-NEXT: call void (i1, ...) @llvm.experimental.guard(i1 [[TMP15]], i32 9) [ "deopt"() ] 106; CHECK-NEXT: [[INDEXA_EXT:%.*]] = zext i32 [[INDEXA]] to i64 107; CHECK-NEXT: [[ADDRA:%.*]] = getelementptr inbounds i8, i8* [[ARRA]], i64 [[INDEXA_EXT]] 108; CHECK-NEXT: [[ELTA:%.*]] = load i8, i8* [[ADDRA]] 109; CHECK-NEXT: [[INDEXB_EXT:%.*]] = zext i32 [[INDEXB]] to i64 110; CHECK-NEXT: [[ADDRB:%.*]] = getelementptr inbounds i8, i8* [[ARRB]], i64 [[INDEXB_EXT]] 111; CHECK-NEXT: [[ELTB:%.*]] = load i8, i8* [[ADDRB]] 112; CHECK-NEXT: [[RESULT:%.*]] = xor i8 [[ELTA]], [[ELTB]] 113; CHECK-NEXT: store i8 [[RESULT]], i8* [[ADDRA]] 114; CHECK-NEXT: [[IV_NEXT]] = add nuw nsw i64 [[IV]], 1 115; CHECK-NEXT: [[LATCH_CHECK:%.*]] = icmp ult i64 [[IV]], 15 116; CHECK-NEXT: br i1 [[LATCH_CHECK]], label [[LOOP]], label [[EXIT:%.*]] 117; CHECK: exit: 118; CHECK-NEXT: [[IV_LCSSA:%.*]] = phi i64 [ [[IV]], [[LOOP]] ] 119; CHECK-NEXT: ret i64 [[IV_LCSSA]] 120; 121{ 122entry: 123 %lengthA = call i32 @length(i8* %arrA) 124 %lengthB = call i32 @length(i8* %arrB) 125 br label %loop 126 127loop: 128 %iv = phi i64 [0, %entry ], [ %iv.next, %loop ] 129 %iv.trunc = trunc i64 %iv to i32 130 %indexA = add i32 %iv.trunc, %offA 131 %indexB = add i32 %iv.trunc, %offB 132 %rcA = icmp ult i32 %indexA, %lengthA 133 %rcIV = icmp ult i64 %iv, %max 134 %wide.chk = and i1 %rcA, %rcIV 135 %rcB = icmp ult i32 %indexB, %lengthB 136 %wide.chk.final = and i1 %wide.chk, %rcB 137 call void (i1, ...) @llvm.experimental.guard(i1 %wide.chk.final, i32 9) [ "deopt"() ] 138 %indexA.ext = zext i32 %indexA to i64 139 %addrA = getelementptr inbounds i8, i8* %arrA, i64 %indexA.ext 140 %eltA = load i8, i8* %addrA 141 %indexB.ext = zext i32 %indexB to i64 142 %addrB = getelementptr inbounds i8, i8* %arrB, i64 %indexB.ext 143 %eltB = load i8, i8* %addrB 144 %result = xor i8 %eltA, %eltB 145 store i8 %result, i8* %addrA 146 %iv.next = add nuw nsw i64 %iv, 1 147 %latch.check = icmp ult i64 %iv, 15 148 br i1 %latch.check, label %loop, label %exit 149 150exit: 151 ret i64 %iv 152} 153 154; cannot narrow the IV to the range type, because we lose information. 155; for (i64 i= 5; i>= 2; i++) 156; this loop wraps around after reaching 2^64. 157define i64 @iv_rc_different_type(i32 %offA, i8* %arrA) { 158; CHECK-LABEL: @iv_rc_different_type( 159; CHECK-NEXT: entry: 160; CHECK-NEXT: [[LENGTHA:%.*]] = call i32 @length(i8* [[ARRA:%.*]]) 161; CHECK-NEXT: br label [[LOOP:%.*]] 162; CHECK: loop: 163; CHECK-NEXT: [[IV:%.*]] = phi i64 [ 5, [[ENTRY:%.*]] ], [ [[IV_NEXT:%.*]], [[LOOP]] ] 164; CHECK-NEXT: [[IV_TRUNC_32:%.*]] = trunc i64 [[IV]] to i32 165; CHECK-NEXT: [[INDEXA:%.*]] = add i32 [[IV_TRUNC_32]], [[OFFA:%.*]] 166; CHECK-NEXT: [[RCA:%.*]] = icmp ult i32 [[INDEXA]], [[LENGTHA]] 167; CHECK-NEXT: call void (i1, ...) @llvm.experimental.guard(i1 [[RCA]], i32 9) [ "deopt"() ] 168; CHECK-NEXT: [[INDEXA_EXT:%.*]] = zext i32 [[INDEXA]] to i64 169; CHECK-NEXT: [[ADDRA:%.*]] = getelementptr inbounds i8, i8* [[ARRA]], i64 [[INDEXA_EXT]] 170; CHECK-NEXT: [[ELTA:%.*]] = load i8, i8* [[ADDRA]] 171; CHECK-NEXT: [[RES:%.*]] = add i8 [[ELTA]], 2 172; CHECK-NEXT: store i8 [[ELTA]], i8* [[ADDRA]] 173; CHECK-NEXT: [[IV_NEXT]] = add i64 [[IV]], 1 174; CHECK-NEXT: [[LATCH_CHECK:%.*]] = icmp sge i64 [[IV_NEXT]], 2 175; CHECK-NEXT: br i1 [[LATCH_CHECK]], label [[LOOP]], label [[EXIT:%.*]] 176; CHECK: exit: 177; CHECK-NEXT: [[IV_LCSSA:%.*]] = phi i64 [ [[IV]], [[LOOP]] ] 178; CHECK-NEXT: ret i64 [[IV_LCSSA]] 179; 180entry: 181 %lengthA = call i32 @length(i8* %arrA) 182 br label %loop 183 184loop: 185 %iv = phi i64 [ 5, %entry ], [ %iv.next, %loop ] 186 %iv.trunc.32 = trunc i64 %iv to i32 187 %indexA = add i32 %iv.trunc.32, %offA 188 %rcA = icmp ult i32 %indexA, %lengthA 189 call void (i1, ...) @llvm.experimental.guard(i1 %rcA, i32 9) [ "deopt"() ] 190 %indexA.ext = zext i32 %indexA to i64 191 %addrA = getelementptr inbounds i8, i8* %arrA, i64 %indexA.ext 192 %eltA = load i8, i8* %addrA 193 %res = add i8 %eltA, 2 194 store i8 %eltA, i8* %addrA 195 %iv.next = add i64 %iv, 1 196 %latch.check = icmp sge i64 %iv.next, 2 197 br i1 %latch.check, label %loop, label %exit 198 199exit: 200 ret i64 %iv 201} 202