[PATCH] D40945: [ScalarEvolution] Improve high cost heuristic in SCEVExpander.

[Bevin Hansson via Phabricator via llvm-commits]

bevinh updated this revision to Diff 126100.
bevinh added a comment.

Enabled runtime loop unroll and added i32 as a legal integer type to prevent SCEVExpander from considering the trip count expensive due to that.


https://reviews.llvm.org/D40945

Files:
  lib/Analysis/ScalarEvolutionExpander.cpp
  test/Analysis/ScalarEvolution/expensive-expansion.ll
  test/Transforms/IndVarSimplify/no-iv-rewrite.ll


Index: test/Transforms/IndVarSimplify/no-iv-rewrite.ll
===================================================================
--- test/Transforms/IndVarSimplify/no-iv-rewrite.ll
+++ test/Transforms/IndVarSimplify/no-iv-rewrite.ll
@@ -225,17 +225,17 @@
 
 ; This test originally checked that the OR instruction was cloned. Now the
 ; ScalarEvolution is able to understand the loop evolution and that '%iv' at the
-; end of the loop is an even value. Thus '%val' is computed at the end of the
-; loop and the OR instruction is replaced by an ADD keeping the result
-; equivalent.
+; end of the loop is an even value. However, the expression is too expensive to
+; compute as the loop is not guarded, so the OR is not cloned (and changed into
+; an add) as it was before.
 ;
 ; CHECK: sext
 ; CHECK: loop:
 ; CHECK: phi i64
 ; CHECK-NOT: sext
 ; CHECK: icmp slt i64
 ; CHECK: exit:
-; CHECK: add i64
+; CHECK-NOT: add i64
 loop:
   %iv = phi i32 [ 0, %entry], [ %iv.next, %loop ]
   %t1 = sext i32 %iv to i64
Index: test/Analysis/ScalarEvolution/expensive-expansion.ll
===================================================================
--- /dev/null
+++ test/Analysis/ScalarEvolution/expensive-expansion.ll
@@ -0,0 +1,41 @@
+; RUN: opt -loop-unroll -unroll-runtime -S < %s | FileCheck %s
+
+; Add 32 as a legal integer width so that SCEVExpander won't consider the loop
+; trip count as expensive due to type width.
+target datalayout = "n32"
+
+ at g = global i32 0, align 4
+
+; Function Attrs: nounwind uwtable
+define void @fn(i32 %start) local_unnamed_addr #0 {
+; CHECK-LABEL: fn
+; CHECK-NOT: for.body.prol
+
+; LoopUnroll should not unroll this loop as the computed trip count is much
+; too expensive. The SCEV under consideration is:
+;  ((2 + (-10 smax (-1 + (-1 * %start))) + %start) /u 2)
+; SCEVExpander previously thought this was cheap as it had a legal
+; division-by-power-of-2 on the RHS, but the LHS is quite large already.
+
+entry:
+  %cmp1 = icmp sgt i32 %start, 7
+  br i1 %cmp1, label %for.body.preheader, label %for.end
+
+for.body.preheader:                               ; preds = %entry
+  br label %for.body
+
+for.body:                                         ; preds = %for.body.preheader, %for.body
+  %i.02 = phi i32 [ %sub, %for.body ], [ %start, %for.body.preheader ]
+  %0 = load volatile i32, i32* @g, align 4
+  %inc = add nsw i32 %0, 1
+  store volatile i32 %inc, i32* @g, align 4
+  %sub = add nsw i32 %i.02, -2
+  %cmp = icmp sgt i32 %i.02, 9
+  br i1 %cmp, label %for.body, label %for.end.loopexit
+
+for.end.loopexit:                                 ; preds = %for.body
+  br label %for.end
+
+for.end:                                          ; preds = %for.end.loopexit, %entry
+  ret void
+}
Index: lib/Analysis/ScalarEvolutionExpander.cpp
===================================================================
--- lib/Analysis/ScalarEvolutionExpander.cpp
+++ lib/Analysis/ScalarEvolutionExpander.cpp
@@ -2051,7 +2051,8 @@
         const DataLayout &DL =
             L->getHeader()->getParent()->getParent()->getDataLayout();
         unsigned Width = cast<IntegerType>(UDivExpr->getType())->getBitWidth();
-        return DL.isIllegalInteger(Width);
+        return DL.isIllegalInteger(Width) ||
+               isHighCostExpansionHelper(UDivExpr->getLHS(), L, At, Processed);
       }
 
     // UDivExpr is very likely a UDiv that ScalarEvolution's HowFarToZero or


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D40945.126100.patch
Type: text/x-patch
Size: 3404 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20171208/29dfa3fe/attachment.bin>