[PATCH] D52351: [LoopVectorizer] Fix in getScalarizationOverhead()

[Jonas Paulsson via Phabricator via llvm-commits]

jonpa updated this revision to Diff 167303.
jonpa added a reviewer: uweigand.
jonpa removed a subscriber: uweigand.
jonpa added a comment.

Thanks for review.

Updated just the test to make it more clear that the address computation is in fact scalar (the multiply).


https://reviews.llvm.org/D52351

Files:
  lib/Transforms/Vectorize/LoopVectorize.cpp
  test/Transforms/LoopVectorize/SystemZ/load-scalarization-cost-0.ll


Index: test/Transforms/LoopVectorize/SystemZ/load-scalarization-cost-0.ll
===================================================================
--- /dev/null
+++ test/Transforms/LoopVectorize/SystemZ/load-scalarization-cost-0.ll
@@ -0,0 +1,27 @@
+; RUN: opt -mtriple=s390x-unknown-linux -mcpu=z13 -loop-vectorize \
+; RUN:   -force-vector-width=2 -debug-only=loop-vectorize \
+; RUN:   -disable-output < %s 2>&1 | FileCheck %s
+; REQUIRES: asserts
+;
+; Check that a scalarized load does not get operands scalarization costs added.
+
+define void @fun(i64* %data, i64 %n, i64 %s, double* %Src) {
+entry:
+  br label %for.body
+
+for.body:
+  %iv = phi i64 [ 0, %entry ], [ %iv.next, %for.body ]
+  %mul = mul nsw i64 %iv, %s
+  %gep = getelementptr inbounds double, double* %Src, i64 %mul
+  %bct = bitcast double* %gep to i64*
+  %ld = load i64, i64* %bct
+  %iv.next = add nuw nsw i64 %iv, 1
+  %cmp110.us = icmp slt i64 %iv.next, %n
+  br i1 %cmp110.us, label %for.body, label %for.end
+
+for.end:
+  ret void
+
+; CHECK: LV: Found an estimated cost of 2 for VF 2 For instruction:   %mul = mul nsw i64 %iv, %s
+; CHECK: LV: Found an estimated cost of 2 for VF 2 For instruction:   %ld = load i64, i64* %bct
+}
Index: lib/Transforms/Vectorize/LoopVectorize.cpp
===================================================================
--- lib/Transforms/Vectorize/LoopVectorize.cpp
+++ lib/Transforms/Vectorize/LoopVectorize.cpp
@@ -2887,6 +2887,10 @@
        !TTI.supportsEfficientVectorElementLoadStore()))
     Cost += TTI.getScalarizationOverhead(RetTy, true, false);
 
+  // Some targets keep addresses scalar.
+  if (isa<LoadInst>(I) && !TTI.prefersVectorizedAddressing())
+    return Cost;
+
   if (CallInst *CI = dyn_cast<CallInst>(I)) {
     SmallVector<const Value *, 4> Operands(CI->arg_operands());
     Cost += TTI.getOperandsScalarizationOverhead(Operands, VF);


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D52351.167303.patch
Type: text/x-patch
Size: 1870 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20180927/3cb59569/attachment.bin>