1; RUN: opt -basic-aa -loop-accesses -analyze -enable-new-pm=0 < %s | FileCheck %s 2; RUN: opt -passes=print-access-info %s -disable-output 2>&1 | FileCheck %s 3 4target datalayout = "e-m:e-i8:8:32-i16:16:32-i64:64-i128:128-n32:64-S128" 5 6; i and i + 1 can overflow in the following kernel: 7; void test1(unsigned long long x, int *a, int *b) { 8; for (unsigned i = 0; i < x; ++i) 9; b[i] = a[i+1] + 1; 10; } 11; 12; If accesses to a and b can alias, we need to emit a run-time alias check 13; between accesses to a and b. However, when i and i + 1 can wrap, their 14; SCEV expression is not an AddRec. We need to create SCEV predicates and 15; coerce the expressions to AddRecs in order to be able to emit the run-time 16; alias check. 17; 18; The accesses at b[i] and a[i+1] correspond to the addresses %arrayidx and 19; %arrayidx4 in the test. The SCEV expressions for these are: 20; ((4 * (zext i32 {1,+,1}<%for.body> to i64))<nuw><nsw> + %a)<nsw> 21; ((4 * (zext i32 {0,+,1}<%for.body> to i64))<nuw><nsw> + %b)<nsw> 22; 23; The transformed expressions are: 24; i64 {(4 + %a),+,4}<%for.body> 25; i64 {(4 + %b),+,4}<%for.body> 26 27; CHECK-LABEL: test1 28; CHECK: Memory dependences are safe with run-time checks 29; CHECK-NEXT: Dependences: 30; CHECK-NEXT: Run-time memory checks: 31; CHECK-NEXT: Check 0: 32; CHECK-NEXT: Comparing group 33; CHECK-NEXT: %arrayidx = getelementptr inbounds i32, i32* %a, i64 %idxprom 34; CHECK-NEXT: Against group 35; CHECK-NEXT: %arrayidx4 = getelementptr inbounds i32, i32* %b, i64 %conv11 36; CHECK-NEXT: Grouped accesses: 37; CHECK-NEXT: Group 38; CHECK-NEXT: (Low: (4 + %a) High: (4 + (4 * (1 umax %x)) + %a)) 39; CHECK-NEXT: Member: {(4 + %a),+,4}<%for.body> 40; CHECK-NEXT: Group 41; CHECK-NEXT: (Low: %b High: ((4 * (1 umax %x)) + %b)) 42; CHECK-NEXT: Member: {%b,+,4}<%for.body> 43; CHECK: Non vectorizable stores to invariant address were not found in loop. 44; CHECK-NEXT: SCEV assumptions: 45; CHECK-NEXT: {1,+,1}<%for.body> Added Flags: <nusw> 46; CHECK-NEXT: {0,+,1}<%for.body> Added Flags: <nusw> 47; CHECK: Expressions re-written: 48; CHECK-NEXT: [PSE] %arrayidx = getelementptr inbounds i32, i32* %a, i64 %idxprom: 49; CHECK-NEXT: ((4 * (zext i32 {1,+,1}<%for.body> to i64))<nuw><nsw> + %a)<nuw> 50; CHECK-NEXT: --> {(4 + %a),+,4}<%for.body> 51; CHECK-NEXT: [PSE] %arrayidx4 = getelementptr inbounds i32, i32* %b, i64 %conv11: 52; CHECK-NEXT: ((4 * (zext i32 {0,+,1}<%for.body> to i64))<nuw><nsw> + %b)<nuw> 53; CHECK-NEXT: --> {%b,+,4}<%for.body> 54define void @test1(i64 %x, i32* %a, i32* %b) { 55entry: 56 br label %for.body 57 58for.body: ; preds = %for.body.preheader, %for.body 59 %conv11 = phi i64 [ %conv, %for.body ], [ 0, %entry ] 60 %i.010 = phi i32 [ %add, %for.body ], [ 0, %entry ] 61 %add = add i32 %i.010, 1 62 %idxprom = zext i32 %add to i64 63 %arrayidx = getelementptr inbounds i32, i32* %a, i64 %idxprom 64 %ld = load i32, i32* %arrayidx, align 4 65 %add2 = add nsw i32 %ld, 1 66 %arrayidx4 = getelementptr inbounds i32, i32* %b, i64 %conv11 67 store i32 %add2, i32* %arrayidx4, align 4 68 %conv = zext i32 %add to i64 69 %cmp = icmp ult i64 %conv, %x 70 br i1 %cmp, label %for.body, label %exit 71 72exit: 73 ret void 74} 75 76; i can overflow in the following kernel: 77; void test2(unsigned long long x, int *a) { 78; for (unsigned i = 0; i < x; ++i) 79; a[i] = a[i] + 1; 80; } 81; 82; We need to check that i doesn't wrap, but we don't need a run-time alias 83; check. We also need an extra no-wrap check to get the backedge taken count. 84 85; CHECK-LABEL: test2 86; CHECK: Memory dependences are safe 87; CHECK: SCEV assumptions: 88; CHECK-NEXT: {1,+,1}<%for.body> Added Flags: <nusw> 89; CHECK-NEXT: {0,+,1}<%for.body> Added Flags: <nusw> 90 define void @test2(i64 %x, i32* %a) { 91entry: 92 br label %for.body 93 94for.body: 95 %conv11 = phi i64 [ %conv, %for.body ], [ 0, %entry ] 96 %i.010 = phi i32 [ %inc, %for.body ], [ 0, %entry ] 97 %arrayidx = getelementptr inbounds i32, i32* %a, i64 %conv11 98 %ld = load i32, i32* %arrayidx, align 4 99 %add = add nsw i32 %ld, 1 100 store i32 %add, i32* %arrayidx, align 4 101 %inc = add i32 %i.010, 1 102 %conv = zext i32 %inc to i64 103 %cmp = icmp ult i64 %conv, %x 104 br i1 %cmp, label %for.body, label %exit 105 106exit: 107 ret void 108} 109