diff options
author | Hal Finkel <hfinkel@anl.gov> | 2017-12-16 02:55:24 +0000 |
---|---|---|
committer | Hal Finkel <hfinkel@anl.gov> | 2017-12-16 02:55:24 +0000 |
commit | 3f92210a791ae96f37ae215fac95668c48a42362 (patch) | |
tree | c330e81ddb85601a774eb31174afbed828ee6169 /test/Transforms | |
parent | cd27af2a407c38c042302643f87ef50a25f31fd1 (diff) |
[LV] Extend InstWidening with CM_Widen_Recursive
Changes to the original scalar loop during LV code gen cause the return value
of Legal->isConsecutivePtr() to be inconsistent with the return value during
legal/cost phases (further analysis and information of the bug is in D39346).
This patch is an alternative fix to PR34965 following the CM_Widen approach
proposed by Ayal and Gil in D39346. It extends InstWidening enum with
CM_Widen_Reverse to properly record the widening decision for consecutive
reverse memory accesses and, consequently, get rid of the
Legal->isConsetuviePtr() call in LV code gen. I think this is a simpler/cleaner
solution to PR34965 than the one in D39346.
Fixes PR34965.
Patch by Diego Caballero, thanks!
Differential Revision: https://reviews.llvm.org/D40742
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@320913 91177308-0d34-0410-b5e6-96231b3b80d8
Diffstat (limited to 'test/Transforms')
-rw-r--r-- | test/Transforms/LoopVectorize/consecutive-ptr-cg-bug.ll | 68 |
1 files changed, 68 insertions, 0 deletions
diff --git a/test/Transforms/LoopVectorize/consecutive-ptr-cg-bug.ll b/test/Transforms/LoopVectorize/consecutive-ptr-cg-bug.ll new file mode 100644 index 00000000000..456271ea1aa --- /dev/null +++ b/test/Transforms/LoopVectorize/consecutive-ptr-cg-bug.ll @@ -0,0 +1,68 @@ +; RUN: opt -loop-vectorize -S < %s | FileCheck %s + +target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128-ni:1" +target triple = "x86_64-unknown-linux-gnu" + +; PR34965/D39346 + +; LV retains the original scalar loop intact as remainder loop. However, +; after this transformation, analysis information concerning the remainder +; loop may differ from the original scalar loop. This test is an example of +; that behaviour, where values inside the remainder loop which SCEV could +; originally analyze now require flow-sensitive analysis currently not +; supported in SCEV. In particular, during LV code generation, after turning +; the original scalar loop into the remainder loop, LV expected +; Legal->isConsecutivePtr() to be consistent and return the same output as +; during legal/cost model phases (original scalar loop). Unfortunately, that +; condition was not satisfied because of the aforementioned SCEV limitation. +; After D39346, LV code generation doesn't rely on Legal->isConsecutivePtr(), +; i.e., SCEV. This test verifies that LV is able to handle the described cases. +; +; TODO: The SCEV limitation described before may affect plans to further +; optimize the remainder loop of this particular test case. One tentative +; solution is to detect the problematic IVs in LV (%7 and %8) and perform an +; in-place IV optimization by replacing: +; %8 = phi i32 [ %.ph2, %.outer ], [ %7, %6 ] with +; with +; %8 = sub i32 %7, 1. + + +; Verify that store is vectorized as stride-1 memory access. + +; CHECK: vector.body: +; CHECK: store <4 x i32> + +; Function Attrs: uwtable +define void @test() { + br label %.outer + +; <label>:1: ; preds = %2 + ret void + +; <label>:2: ; preds = %._crit_edge.loopexit + %3 = add nsw i32 %.ph, -2 + br i1 undef, label %1, label %.outer + +.outer: ; preds = %2, %0 + %.ph = phi i32 [ %3, %2 ], [ 336, %0 ] + %.ph2 = phi i32 [ 62, %2 ], [ 110, %0 ] + %4 = and i32 %.ph, 30 + %5 = add i32 %.ph2, 1 + br label %6 + +; <label>:6: ; preds = %6, %.outer + %7 = phi i32 [ %5, %.outer ], [ %13, %6 ] + %8 = phi i32 [ %.ph2, %.outer ], [ %7, %6 ] + %9 = add i32 %8, 2 + %10 = zext i32 %9 to i64 + %11 = getelementptr inbounds i32, i32 addrspace(1)* undef, i64 %10 + %12 = ashr i32 undef, %4 + store i32 %12, i32 addrspace(1)* %11, align 4 + %13 = add i32 %7, 1 + %14 = icmp sgt i32 %13, 61 + br i1 %14, label %._crit_edge.loopexit, label %6 + +._crit_edge.loopexit: ; preds = %._crit_edge.loopexit, %6 + br i1 undef, label %2, label %._crit_edge.loopexit +} + |