llvm.org GIT mirror llvm / 673968a
When forming an ICmpZero LSRUse, normalize the non-IV operand of the comparison, so that the resulting expression is fully normalized. This fixes PR9939. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@131576 91177308-0d34-0410-b5e6-96231b3b80d8 Dan Gohman 8 years ago
2 changed file(s) with 95 addition(s) and 0 deletion(s). Raw diff Collapse all Expand all
20722072 // x == y --> x - y == 0
20732073 const SCEV *N = SE.getSCEV(NV);
20742074 if (SE.isLoopInvariant(N, L)) {
2075 // S is normalized, so normalize N before folding it into S
2076 // to keep the result normalized.
2077 N = TransformForPostIncUse(Normalize, N, CI, 0,
2078 LF.PostIncLoops, SE, DT);
20752079 Kind = LSRUse::ICmpZero;
20762080 S = SE.getMinusSCEV(N, S);
20772081 }
0 ; RUN: opt -loop-reduce -S < %s | FileCheck %s
1 ; PR9939
2
3 ; LSR should property handle the post-inc offset when folding the
4 ; non-IV operand of an icmp into the IV.
5
6 ; CHECK: %tmp2 = sub i64 %sub.ptr.lhs.cast, %sub.ptr.rhs.cast
7 ; CHECK: %tmp3 = lshr i64 %tmp2, 1
8 ; CHECK: %tmp4 = mul i64 %tmp3, 2
9 ; CHECK: br label %for.body
10 ; CHECK: for.body:
11 ; CHECK: %lsr.iv5 = phi i64 [ %lsr.iv.next, %for.body ], [ %tmp4, %for.body.lr.ph ]
12 ; CHECK: %lsr.iv.next = add i64 %lsr.iv5, -2
13 ; CHECK: %lsr.iv.next6 = inttoptr i64 %lsr.iv.next to i16*
14 ; CHECK: %cmp27 = icmp eq i16* %lsr.iv.next6, null
15
16 target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64"
17 target triple = "x86_64-unknown-linux-gnu"
18
19 %struct.Vector2 = type { i16*, [64 x i16], i32 }
20
21 @.str = private unnamed_addr constant [37 x i8] c"0123456789abcdefghijklmnopqrstuvwxyz\00"
22
23 define void @_Z15IntegerToStringjjR7Vector2(i32 %i, i32 %radix, %struct.Vector2* nocapture %result) nounwind noinline {
24 entry:
25 %buffer = alloca [33 x i16], align 16
26 %add.ptr = getelementptr inbounds [33 x i16]* %buffer, i64 0, i64 33
27 br label %do.body
28
29 do.body: ; preds = %do.body, %entry
30 %0 = phi i64 [ %indvar.next44, %do.body ], [ 0, %entry ]
31 %i.addr.0 = phi i32 [ %div, %do.body ], [ %i, %entry ]
32 %tmp51 = sub i64 32, %0
33 %incdec.ptr = getelementptr [33 x i16]* %buffer, i64 0, i64 %tmp51
34 %rem = urem i32 %i.addr.0, 10
35 %div = udiv i32 %i.addr.0, 10
36 %idxprom = zext i32 %rem to i64
37 %arrayidx = getelementptr inbounds [37 x i8]* @.str, i64 0, i64 %idxprom
38 %tmp5 = load i8* %arrayidx, align 1
39 %conv = sext i8 %tmp5 to i16
40 store i16 %conv, i16* %incdec.ptr, align 2
41 %1 = icmp ugt i32 %i.addr.0, 9
42 %indvar.next44 = add i64 %0, 1
43 br i1 %1, label %do.body, label %do.end
44
45 do.end: ; preds = %do.body
46 %xap.0 = inttoptr i64 %0 to i1*
47 %cap.0 = ptrtoint i1* %xap.0 to i64
48 %sub.ptr.lhs.cast = ptrtoint i16* %add.ptr to i64
49 %sub.ptr.rhs.cast = ptrtoint i16* %incdec.ptr to i64
50 %sub.ptr.sub = sub i64 %sub.ptr.lhs.cast, %sub.ptr.rhs.cast
51 %sub.ptr.div39 = lshr exact i64 %sub.ptr.sub, 1
52 %conv11 = trunc i64 %sub.ptr.div39 to i32
53 %mLength = getelementptr inbounds %struct.Vector2* %result, i64 0, i32 2
54 %idx.ext21 = bitcast i64 %sub.ptr.div39 to i64
55 %incdec.ptr.sum = add i64 %idx.ext21, -1
56 %cp.0.sum = sub i64 %incdec.ptr.sum, %0
57 %add.ptr22 = getelementptr [33 x i16]* %buffer, i64 1, i64 %cp.0.sum
58 %cmp2740 = icmp eq i64 %idx.ext21, 0
59 br i1 %cmp2740, label %for.end, label %for.body.lr.ph
60
61 for.body.lr.ph: ; preds = %do.end
62 %tmp16 = load i32* %mLength, align 4
63 %mBegin = getelementptr inbounds %struct.Vector2* %result, i64 0, i32 0
64 %tmp14 = load i16** %mBegin, align 8
65 %tmp48 = zext i32 %tmp16 to i64
66 br label %for.body
67
68 for.body: ; preds = %for.body, %for.body.lr.ph
69 %indvar = phi i64 [ 0, %for.body.lr.ph ], [ %indvar.next, %for.body ]
70 %tmp46 = add i64 %tmp51, %indvar
71 %p.042 = getelementptr [33 x i16]* %buffer, i64 0, i64 %tmp46
72 %tmp47 = sub i64 %indvar, %0
73 %incdec.ptr32 = getelementptr [33 x i16]* %buffer, i64 1, i64 %tmp47
74 %tmp49 = add i64 %tmp48, %indvar
75 %dst.041 = getelementptr i16* %tmp14, i64 %tmp49
76 %tmp29 = load i16* %p.042, align 2
77 store i16 %tmp29, i16* %dst.041, align 2
78 %cmp27 = icmp eq i16* %incdec.ptr32, %add.ptr22
79 %indvar.next = add i64 %indvar, 1
80 br i1 %cmp27, label %for.end.loopexit, label %for.body
81
82 for.end.loopexit: ; preds = %for.body
83 br label %for.end
84
85 for.end: ; preds = %for.end.loopexit, %do.end
86 %tmp38 = load i32* %mLength, align 4
87 %add = add i32 %tmp38, %conv11
88 store i32 %add, i32* %mLength, align 4
89 ret void
90 }