From 02320eee6bd056c04655edd5cb83940662cc3237 Mon Sep 17 00:00:00 2001 From: Roman Tereshin Date: Mon, 27 Aug 2018 21:41:37 +0000 Subject: Revert "[SCEV][NFC] Check NoWrap flags before lexicographical comparison of SCEVs" This reverts r319889. Unfortunately, wrapping flags are not a part of SCEV's identity (they do not participate in computing a hash value or in equality comparisons) and in fact they could be assigned after the fact w/o rebuilding a SCEV. Grep for const_cast's to see quite a few of examples, apparently all for AddRec's at the moment. So, if 2 expressions get built in 2 slightly different ways: one with flags set in the beginning, the other with the flags attached later on, we may end up with 2 expressions which are exactly the same but have their operands swapped in one of the commutative N-ary expressions, and at least one of them will have "sorted by complexity" invariant broken. 2 identical SCEV's won't compare equal by pointer comparison as they are supposed to. A real-world reproducer is added as a regression test: the issue described causes 2 identical SCEV expressions to have different order of operands and therefore compare not equal, which in its turn prevents LoadStoreVectorizer from vectorizing a pair of consecutive loads. On a larger example (the source of the test attached, which is a bugpoint) I have seen even weirder behavior: adding a constant to an existing SCEV changes the order of the existing terms, for instance, getAddExpr(1, ((A * B) + (C * D))) returns (1 + (C * D) + (A * B)). Differential Revision: https://reviews.llvm.org/D40645 llvm-svn: 340777 --- .../X86/compare-scev-by-complexity.ll | 76 ++++++++++++++++++++++ 1 file changed, 76 insertions(+) create mode 100644 llvm/test/Transforms/LoadStoreVectorizer/X86/compare-scev-by-complexity.ll (limited to 'llvm/test/Transforms/LoadStoreVectorizer') diff --git a/llvm/test/Transforms/LoadStoreVectorizer/X86/compare-scev-by-complexity.ll b/llvm/test/Transforms/LoadStoreVectorizer/X86/compare-scev-by-complexity.ll new file mode 100644 index 00000000000..7f29a73bcf9 --- /dev/null +++ b/llvm/test/Transforms/LoadStoreVectorizer/X86/compare-scev-by-complexity.ll @@ -0,0 +1,76 @@ +; RUN: opt -load-store-vectorizer %s -S | FileCheck %s + +; Check that setting wrapping flags after a SCEV node is created +; does not invalidate "sorted by complexity" invariant for +; operands of commutative and associative SCEV operators. + +target triple = "x86_64--" + +@global_value0 = external constant i32 +@global_value1 = external constant i32 +@other_value = external global float +@a = external global float +@b = external global float +@c = external global float +@d = external global float +@plus1 = external global i32 +@cnd = external global i8 + +; Function Attrs: nounwind +define void @main() local_unnamed_addr #0 { +; CHECK-LABEL: @main() +; CHECK: [[PTR:%[0-9]+]] = bitcast float* %preheader.load0.address to <2 x float>* +; CHECK: = load <2 x float>, <2 x float>* [[PTR]] +; CHECK-LABEL: for.body23: +entry: + %tmp = load i32, i32* @global_value0, !range !0 + %tmp2 = load i32, i32* @global_value1 + %and.i.i = and i32 %tmp2, 2 + %add.nuw.nsw.i.i = add nuw nsw i32 %and.i.i, 0 + %mul.i.i = shl nuw nsw i32 %add.nuw.nsw.i.i, 1 + %and6.i.i = and i32 %tmp2, 3 + %and9.i.i = and i32 %tmp2, 4 + %add.nuw.nsw10.i.i = add nuw nsw i32 %and6.i.i, %and9.i.i + %conv3.i42.i = add nuw nsw i32 %mul.i.i, 1 + %reass.add346.7 = add nuw nsw i32 %add.nuw.nsw10.i.i, 56 + %reass.mul347.7 = mul nuw nsw i32 %tmp, %reass.add346.7 + %add7.i.7 = add nuw nsw i32 %reass.mul347.7, 0 + %preheader.address0.idx = add nuw nsw i32 %add7.i.7, %mul.i.i + %preheader.address0.idx.zext = zext i32 %preheader.address0.idx to i64 + %preheader.load0.address = getelementptr inbounds float, float* @other_value, i64 %preheader.address0.idx.zext + %preheader.load0. = load float, float* %preheader.load0.address, align 4, !tbaa !1 + %common.address.idx = add nuw nsw i32 %add7.i.7, %conv3.i42.i + %preheader.header.common.address.idx.zext = zext i32 %common.address.idx to i64 + %preheader.load1.address = getelementptr inbounds float, float* @other_value, i64 %preheader.header.common.address.idx.zext + %preheader.load1. = load float, float* %preheader.load1.address, align 4, !tbaa !1 + br label %for.body23 + +for.body23: ; preds = %for.body23, %entry + %loop.header.load0.address = getelementptr inbounds float, float* @other_value, i64 %preheader.header.common.address.idx.zext + %loop.header.load0. = load float, float* %loop.header.load0.address, align 4, !tbaa !1 + %reass.mul343.7 = mul nuw nsw i32 %reass.add346.7, 72 + %add7.i286.7.7 = add nuw nsw i32 %reass.mul343.7, 56 + %add9.i288.7.7 = add nuw nsw i32 %add7.i286.7.7, %mul.i.i + %loop.header.address1.idx = add nuw nsw i32 %add9.i288.7.7, 1 + %loop.header.address1.idx.zext = zext i32 %loop.header.address1.idx to i64 + %loop.header.load1.address = getelementptr inbounds float, float* @other_value, i64 %loop.header.address1.idx.zext + %loop.header.load1. = load float, float* %loop.header.load1.address, align 4, !tbaa !1 + store float %preheader.load0., float* @a, align 4, !tbaa !1 + store float %preheader.load1., float* @b, align 4, !tbaa !1 + store float %loop.header.load0., float* @c, align 4, !tbaa !1 + store float %loop.header.load1., float* @d, align 4, !tbaa !1 + %loaded.cnd = load i8, i8* @cnd + %condition = trunc i8 %loaded.cnd to i1 + br i1 %condition, label %for.body23, label %exit + +exit: + ret void +} + +attributes #0 = { nounwind } + +!0 = !{i32 0, i32 65536} +!1 = !{!2, !2, i64 0} +!2 = !{!"float", !3, i64 0} +!3 = !{!"omnipotent char", !4, i64 0} +!4 = !{!"Simple C++ TBAA"} -- cgit v1.2.3