[PATCH] D31601: Make SLPVectorizer pass the right type argument to getCmpSelInstrCost()

[Jonas Paulsson via Phabricator via llvm-commits]

jonpa created this revision.
Herald added a subscriber: mzolotukhin.

(This was also on llvm-commits last week without test case: http://lists.llvm.org/pipermail/llvm-commits/Week-of-Mon-20170327/442107.html)

I found that the SLP vectorizer does not call getCmpSelInstrCost() for a 
compare with the proper type. It currently passes i1 or a vector of i1s, 
instead of the type of the compare operand, like LoopVectorizer does.

My patch fixes this by getting the right scalar type for compares. No 
regressions after applying this patch.

The test case is for SystemZ, and depends on https://reviews.llvm.org/D29631#09b6b393.


https://reviews.llvm.org/D31601

Files:
  lib/Transforms/Vectorize/SLPVectorizer.cpp
  test/Transforms/SLPVectorizer/SystemZ/SLP-cmp-cost-query.ll


Index: test/Transforms/SLPVectorizer/SystemZ/SLP-cmp-cost-query.ll
===================================================================
--- /dev/null
+++ test/Transforms/SLPVectorizer/SystemZ/SLP-cmp-cost-query.ll
@@ -0,0 +1,35 @@
+; RUN: opt -mtriple=systemz-unknown -mcpu=z13 -slp-vectorizer -debug-only=SLP \
+; RUN:   -S -disable-output < %s |& FileCheck %s
+;
+; Check that SLP vectorizer gets the right cost difference for a compare
+; node.
+
+; Function Attrs: norecurse nounwind readonly
+define void @fun(i8* nocapture, i32 zeroext) local_unnamed_addr #0 {
+.lr.ph.preheader:
+  br label %.lr.ph
+
+.lr.ph:                                           ; preds = %.lr.ph.preheader, %.lr.ph
+  %2 = phi i32 [ %., %.lr.ph ], [ undef, %.lr.ph.preheader ]
+  %3 = phi i32 [ %.9, %.lr.ph ], [ undef, %.lr.ph.preheader ]
+  %4 = icmp ult i32 %2, %1
+  %5 = select i1 %4, i32 0, i32 %1
+  %. = sub i32 %2, %5
+  %6 = icmp ult i32 %3, %1
+  %7 = select i1 %6, i32 0, i32 %1
+  %.9 = sub i32 %3, %7
+  %8 = zext i32 %. to i64
+  %9 = getelementptr inbounds i8, i8* %0, i64 %8
+  %10 = load i8, i8* %9, align 1
+  %11 = zext i32 %.9 to i64
+  %12 = getelementptr inbounds i8, i8* %0, i64 %11
+  %13 = load i8, i8* %12, align 1
+  %14 = icmp eq i8 %10, %13
+  br i1 %14, label %.lr.ph, label %._crit_edge
+
+._crit_edge:                                      ; preds = %.lr.ph
+  ret void
+
+; CHECK: SLP: Adding cost -1 for bundle that starts with   %4 = icmp ult i32 %2, %1.
+}
+
Index: lib/Transforms/Vectorize/SLPVectorizer.cpp
===================================================================
--- lib/Transforms/Vectorize/SLPVectorizer.cpp
+++ lib/Transforms/Vectorize/SLPVectorizer.cpp
@@ -1702,6 +1702,8 @@
   Type *ScalarTy = VL[0]->getType();
   if (StoreInst *SI = dyn_cast<StoreInst>(VL[0]))
     ScalarTy = SI->getValueOperand()->getType();
+  if (CmpInst *CI = dyn_cast<CmpInst>(VL[0]))
+    ScalarTy = CI->getOperand(0)->getType();
   VectorType *VecTy = VectorType::get(ScalarTy, VL.size());
 
   // If we have computed a smaller type for the expression, update VecTy so


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D31601.93865.patch
Type: text/x-patch
Size: 2081 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20170403/869bf4f3/attachment.bin>