Prevent a DAGCombine from firing where there are two uses of

a combined-away node and the result of the combine isn't substantially
smaller than the input, it's just canonicalized. This is the first part
of a significant (7%) performance gain for Snappy's hot decompression
loop.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@147604 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
index 8a8f0cf2bc75fce8659121771b585f69a87dc75c..5b2d86f511ddc13e9658934d2c6e371d9aa073c4 100644 (file)
--- a/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -3331,7 +3331,9 @@ SDValue DAGCombiner::visitSHL(SDNode *N) {
 
   // fold (shl (srl x, c1), c2) -> (and (shl x, (sub c2, c1), MASK) or
   //                               (and (srl x, (sub c1, c2), MASK)
-  if (N1C && N0.getOpcode() == ISD::SRL &&
+  // Only fold this if the inner shift has no other uses -- if it does, folding
+  // this will increase the total number of instructions.
+  if (N1C && N0.getOpcode() == ISD::SRL && N0.hasOneUse() &&
       N0.getOperand(1).getOpcode() == ISD::Constant) {
     uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue();
     if (c1 < VT.getSizeInBits()) {

diff --git a/test/CodeGen/X86/shift-folding.ll b/test/CodeGen/X86/shift-folding.ll
index 7eac116e020ef2701c185e51eaf4a02ea63b638e..3ea601147bb091cbfe7ad80175554d0a7b7e47b8 100644 (file)
--- a/test/CodeGen/X86/shift-folding.ll
+++ b/test/CodeGen/X86/shift-folding.ll
@@ -48,3 +48,23 @@ entry:
   %tmp512 = lshr i32 %tmp4, 24
   ret i32 %tmp512
 }
+
+define i64 @test5(i16 %i, i32* %arr) {
+; Ensure that we don't fold away shifts which have multiple uses, as they are
+; just re-introduced for the second use.
+; CHECK: test5:
+; CHECK-NOT: shrl
+; CHECK: shrl $11
+; CHECK-NOT: shrl
+; CHECK: ret
+
+entry:
+  %i.zext = zext i16 %i to i32
+  %index = lshr i32 %i.zext, 11
+  %index.zext = zext i32 %index to i64
+  %val.ptr = getelementptr inbounds i32* %arr, i64 %index.zext
+  %val = load i32* %val.ptr
+  %val.zext = zext i32 %val to i64
+  %sum = add i64 %val.zext, %index.zext
+  ret i64 %sum
+}