LoopVectorizer: Calculate the number of pointers to disambiguate at runtime based...

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@180593 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Transforms/Vectorize/LoopVectorize.cpp b/lib/Transforms/Vectorize/LoopVectorize.cpp
index 0faab7a3c0e160a555b18434150bd8a81a42677f..9a832f7a120b57d39b16760f482e7041352171e1 100644 (file)
--- a/lib/Transforms/Vectorize/LoopVectorize.cpp
+++ b/lib/Transforms/Vectorize/LoopVectorize.cpp
@@ -114,9 +114,9 @@ TinyTripCountVectorThreshold("vectorizer-min-trip-count", cl::init(16),
 /// We don't unroll loops with a known constant trip count below this number.
 static const unsigned TinyTripCountUnrollThreshold = 128;
 
-/// When performing a runtime memory check, do not check more than this
-/// number of pointers. Notice that the check is quadratic!
-static const unsigned RuntimeMemoryCheckThreshold = 4;
+/// When performing memory disambiguation checks at runtime do not make more
+/// than this number of comparisons.
+static const unsigned RuntimeMemoryCheckThreshold = 8;
 
 /// We use a metadata with this name  to indicate that a scalar loop was
 /// vectorized and that we don't need to re-vectorize it if we run into it
@@ -2679,6 +2679,9 @@ bool LoopVectorizationLegality::canVectorizeMemory() {
     return true;
   }
 
+  unsigned NumReadPtrs = 0;
+  unsigned NumWritePtrs = 0;
+
   // Find pointers with computable bounds. We are going to use this information
   // to place a runtime bound check.
   bool CanDoRT = true;
@@ -2687,6 +2690,7 @@ bool LoopVectorizationLegality::canVectorizeMemory() {
     Value *V = (*MI).first;
     if (hasComputableBounds(V)) {
       PtrRtCheck.insert(SE, TheLoop, V, true);
+      NumWritePtrs++;
       DEBUG(dbgs() << "LV: Found a runtime check ptr:" << *V <<"\n");
     } else {
       CanDoRT = false;
@@ -2697,6 +2701,7 @@ bool LoopVectorizationLegality::canVectorizeMemory() {
     Value *V = (*MI).first;
     if (hasComputableBounds(V)) {
       PtrRtCheck.insert(SE, TheLoop, V, false);
+      NumReadPtrs++;
       DEBUG(dbgs() << "LV: Found a runtime check ptr:" << *V <<"\n");
     } else {
       CanDoRT = false;
@@ -2706,7 +2711,9 @@ bool LoopVectorizationLegality::canVectorizeMemory() {
 
   // Check that we did not collect too many pointers or found a
   // unsizeable pointer.
-  if (!CanDoRT || PtrRtCheck.Pointers.size() > RuntimeMemoryCheckThreshold) {
+  unsigned NumComparisons = (NumWritePtrs * (NumReadPtrs + NumWritePtrs - 1));
+  DEBUG(dbgs() << "LV: We need to compare " << NumComparisons << " ptrs.\n");
+  if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) {
     PtrRtCheck.reset();
     CanDoRT = false;
   }

diff --git a/test/Transforms/LoopVectorize/runtime-limit.ll b/test/Transforms/LoopVectorize/runtime-limit.ll
new file mode 100644 (file)
index 0000000..d783974
--- /dev/null
+++ b/test/Transforms/LoopVectorize/runtime-limit.ll
@@ -0,0 +1,84 @@
+; RUN: opt < %s  -loop-vectorize -force-vector-unroll=1 -force-vector-width=4 -dce -instcombine -S | FileCheck %s
+
+target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64-S128"
+target triple = "x86_64-apple-macosx10.8.0"
+
+; We are vectorizing with 6 runtime checks.
+;CHECK: func1x6
+;CHECK: <4 x i32>
+;CHECK: ret
+define i32 @func1x6(i32* nocapture %out, i32* nocapture %A, i32* nocapture %B, i32* nocapture %C, i32* nocapture %D, i32* nocapture %E, i32* nocapture %F) {
+entry:
+  br label %for.body
+
+for.body:                                         ; preds = %for.body, %entry
+  %i.016 = phi i64 [ 0, %entry ], [ %inc, %for.body ]
+  %arrayidx = getelementptr inbounds i32* %A, i64 %i.016
+  %0 = load i32* %arrayidx, align 4
+  %arrayidx1 = getelementptr inbounds i32* %B, i64 %i.016
+  %1 = load i32* %arrayidx1, align 4
+  %add = add nsw i32 %1, %0
+  %arrayidx2 = getelementptr inbounds i32* %C, i64 %i.016
+  %2 = load i32* %arrayidx2, align 4
+  %add3 = add nsw i32 %add, %2
+  %arrayidx4 = getelementptr inbounds i32* %E, i64 %i.016
+  %3 = load i32* %arrayidx4, align 4
+  %add5 = add nsw i32 %add3, %3
+  %arrayidx6 = getelementptr inbounds i32* %F, i64 %i.016
+  %4 = load i32* %arrayidx6, align 4
+  %add7 = add nsw i32 %add5, %4
+  %arrayidx8 = getelementptr inbounds i32* %out, i64 %i.016
+  store i32 %add7, i32* %arrayidx8, align 4
+  %inc = add i64 %i.016, 1
+  %exitcond = icmp eq i64 %inc, 256
+  br i1 %exitcond, label %for.end, label %for.body
+
+for.end:                                          ; preds = %for.body
+  ret i32 undef
+}
+
+; We are not vectorizing with 12 runtime checks.
+;CHECK: func2x6
+;CHECK-NOT: <4 x i32>
+;CHECK: ret
+define i32 @func2x6(i32* nocapture %out, i32* nocapture %out2, i32* nocapture %A, i32* nocapture %B, i32* nocapture %C, i32* nocapture %D, i32* nocapture %E, i32* nocapture %F) {
+entry:
+  br label %for.body
+
+for.body:                                         ; preds = %for.body, %entry
+  %i.037 = phi i64 [ 0, %entry ], [ %inc, %for.body ]
+  %arrayidx = getelementptr inbounds i32* %A, i64 %i.037
+  %0 = load i32* %arrayidx, align 4
+  %arrayidx1 = getelementptr inbounds i32* %B, i64 %i.037
+  %1 = load i32* %arrayidx1, align 4
+  %add = add nsw i32 %1, %0
+  %arrayidx2 = getelementptr inbounds i32* %C, i64 %i.037
+  %2 = load i32* %arrayidx2, align 4
+  %add3 = add nsw i32 %add, %2
+  %arrayidx4 = getelementptr inbounds i32* %E, i64 %i.037
+  %3 = load i32* %arrayidx4, align 4
+  %add5 = add nsw i32 %add3, %3
+  %arrayidx6 = getelementptr inbounds i32* %F, i64 %i.037
+  %4 = load i32* %arrayidx6, align 4
+  %add7 = add nsw i32 %add5, %4
+  %arrayidx8 = getelementptr inbounds i32* %out, i64 %i.037
+  store i32 %add7, i32* %arrayidx8, align 4
+  %5 = load i32* %arrayidx, align 4
+  %6 = load i32* %arrayidx1, align 4
+  %add11 = add nsw i32 %6, %5
+  %7 = load i32* %arrayidx2, align 4
+  %add13 = add nsw i32 %add11, %7
+  %8 = load i32* %arrayidx4, align 4
+  %add15 = add nsw i32 %add13, %8
+  %9 = load i32* %arrayidx6, align 4
+  %add17 = add nsw i32 %add15, %9
+  %arrayidx18 = getelementptr inbounds i32* %out2, i64 %i.037
+  store i32 %add17, i32* %arrayidx18, align 4
+  %inc = add i64 %i.037, 1
+  %exitcond = icmp eq i64 %inc, 256
+  br i1 %exitcond, label %for.end, label %for.body
+
+for.end:                                          ; preds = %for.body
+  ret i32 undef
+}
+