[FIX] Teach the ScopExpander about parallel subfunctions

llvm-svn: 265824

2 files changed, 58 insertions, 5 deletions

diff --git a/polly/lib/Support/ScopHelper.cpp b/polly/lib/Support/ScopHelper.cpp
index e73946f5042..469c00d52d3 100644
--- a/polly/lib/Support/ScopHelper.cpp
+++ b/polly/lib/Support/ScopHelper.cpp
@@ -225,9 +225,9 @@ void polly::splitEntryBlockForAlloca(BasicBlock *EntryBlock, Pass *P) {
 struct ScopExpander : SCEVVisitor<ScopExpander, const SCEV *> {
   friend struct SCEVVisitor<ScopExpander, const SCEV *>;
 
-  explicit ScopExpander(const Region &R, ScalarEvolution &SE,
+  explicit ScopExpander(const Region &R, Function &F, ScalarEvolution &SE,
                         const DataLayout &DL, const char *Name, ValueMapT *VMap)
-      : Expander(SCEVExpander(SE, DL, Name)), SE(SE), Name(Name), R(R),
+      : Expander(SCEVExpander(SE, DL, Name)), F(F), SE(SE), Name(Name), R(R),
         VMap(VMap) {}
 
   Value *expandCodeFor(const SCEV *E, Type *Ty, Instruction *I) {
@@ -241,6 +241,10 @@ struct ScopExpander : SCEVVisitor<ScopExpander, const SCEV *> {
 
 private:
   SCEVExpander Expander;
+
+  /// @brief The function in which the code is placed.
+  Function &F;
+
   ScalarEvolution &SE;
   const char *Name;
   const Region &R;
@@ -264,10 +268,15 @@ private:
                   Inst->getOpcode() != Instruction::SDiv))
       return E;
 
-    if (!R.contains(Inst))
+    // If the instruction is outside the SCoP we can just use it without the
+    // need to recompute it. However, if it is in another function we need to
+    // recompute it as the definition does not dominate the use.
+    bool SameFunction = (&F == R.getEntry()->getParent());
+    if (!R.contains(Inst) && SameFunction)
       return E;
 
-    Instruction *StartIP = R.getEnteringBlock()->getTerminator();
+    Instruction *StartIP = SameFunction ? R.getEnteringBlock()->getTerminator()
+                                        : F.getEntryBlock().getTerminator();
 
     const SCEV *LHSScev = visit(SE.getSCEV(Inst->getOperand(0)));
     const SCEV *RHSScev = visit(SE.getSCEV(Inst->getOperand(1)));
@@ -333,7 +342,7 @@ private:
 Value *polly::expandCodeFor(Scop &S, ScalarEvolution &SE, const DataLayout &DL,
                             const char *Name, const SCEV *E, Type *Ty,
                             Instruction *IP, ValueMapT *VMap) {
-  ScopExpander Expander(S.getRegion(), SE, DL, Name, VMap);
+  ScopExpander Expander(S.getRegion(), *IP->getFunction(), SE, DL, Name, VMap);
   return Expander.expandCodeFor(E, Ty, IP);
 }
 
diff --git a/polly/test/Isl/CodeGen/OpenMP/recomputed-srem.ll b/polly/test/Isl/CodeGen/OpenMP/recomputed-srem.ll
new file mode 100644
index 00000000000..bfbfce176f8
--- /dev/null
+++ b/polly/test/Isl/CodeGen/OpenMP/recomputed-srem.ll
@@ -0,0 +1,44 @@
+; RUN: opt %loadPolly -polly-codegen -polly-parallel \
+; RUN: -polly-parallel-force -S < %s | FileCheck %s
+;
+; Test to verify that we recompute %rem96 in the parallel subfunction.
+;
+; CHECK:        %rem96polly = srem i32 %polly.subfunc.arg.n2, 16
+; CHECK-NEXT:   br label %polly.par.checkNext
+;
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+
+; Function Attrs: nounwind uwtable
+define void @dmxpy(i32 %n1, float* %y, i32 %n2, float* %x) #0 {
+entry:
+  %rem96 = srem i32 %n2, 16
+  %0 = sext i32 %rem96 to i64
+  %1 = add i64 %0, 15
+  br label %for.cond195.preheader
+
+for.cond195.preheader:                            ; preds = %for.inc363, %entry
+  %indvars.iv262 = phi i64 [ %1, %entry ], [ %indvars.iv.next263, %for.inc363 ]
+  %j.0236 = phi i32 [ 0, %entry ], [ %add364, %for.inc363 ]
+  br label %for.body197
+
+for.body197:                                      ; preds = %for.body197, %for.cond195.preheader
+  %indvars.iv = phi i64 [ %indvars.iv.next, %for.body197 ], [ 0, %for.cond195.preheader ]
+  %arrayidx199 = getelementptr inbounds float, float* %y, i64 %indvars.iv
+  %2 = add nsw i64 %indvars.iv262, -6
+  %arrayidx292 = getelementptr inbounds float, float* %x, i64 %2
+  %3 = load float, float* %arrayidx292, align 4
+  store float undef, float* %arrayidx199, align 4
+  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
+  %lftr.wideiv = trunc i64 %indvars.iv.next to i32
+  %exitcond = icmp ne i32 %lftr.wideiv, %n1
+  br i1 %exitcond, label %for.body197, label %for.inc363
+
+for.inc363:                                       ; preds = %for.body197
+  %add364 = add nsw i32 %j.0236, 16
+  %cmp193 = icmp slt i32 %add364, %n2
+  %indvars.iv.next263 = add i64 %indvars.iv262, 16
+  br i1 %cmp193, label %for.cond195.preheader, label %for.end365
+
+for.end365:                                       ; preds = %for.inc363
+  ret void
+}