[PATCH] D26169: [Reassociate] Skip reassociation for IV if its def instruction contains NSW/NUW flags to enable better IndVar and LSR

[Wei Mi via llvm-commits]

wmi created this revision.
wmi added reviewers: sanjoy, dberlin, davidxl.
wmi added a subscriber: llvm-commits.
wmi set the repository for this revision to rL LLVM.

Reassociation will remove nsw/nuw flags after change the order of computation. It may block IndVar from doing IV use widening and also block the optimization of LSR.

For %sub1 = %i.0 + %size - 1 in the following testcase, reassociation will convert it to %sub1 = %size - 1 + %i.0 and remove all the nsw flags. Then IndVar cannot do widening for %idxprom and cannot remove the sext. In addition, LSR cannot regard %arrayidx as an induction variable and it may produce suboptimal result.

for.body:

  %add = add nsw i32 %i.0, %size
  %sub1 = add nsw i32 %add, -1
  %idxprom = sext i32 %sub1 to i64
  %arrayidx = getelementptr inbounds [1000 x i32], [1000 x i32]* @maxarray, i64 0, i64 %idxprom
  %tmp0 = load i32, i32* %arrayidx, align 4
  %tmp1 = load i32, i32* @total, align 4
  %add2 = add nsw i32 %tmp1, %tmp0
  store i32 %add2, i32* @total, align 4
  %inc = add nsw i32 %i.0, 1
  br label %for.cond

The patch skips reassociation for IV use when it contains NSW/NUW flags so as to let IndVar and LSR do better jobs. Because LSR will also do reassociation work in its Formula canonicalization,  we won't lose performance opportunity mostly.


Repository:
  rL LLVM

https://reviews.llvm.org/D26169

Files:
  include/llvm/Transforms/Scalar/Reassociate.h
  lib/Transforms/Scalar/Reassociate.cpp
  test/Transforms/Reassociate/ivuse.ll

-------------- next part --------------
A non-text attachment was scrubbed...
Name: D26169.76476.patch
Type: text/x-patch
Size: 5655 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20161031/8e659460/attachment.bin>