Add a transform to DAG Combiner. This improves the

code for the case where 32-bit divide by constant is
turned into 64-bit multiply by constant.  8771012.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@122090 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
index be5a5bc1a5e5a215203e5ea2f47190e0a8aabed8..522ebee73cf32ce3d6005265923ee5d6679683d8 100644 (file)
--- a/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -3171,6 +3171,26 @@ SDValue DAGCombiner::visitSRL(SDNode *N) {
                        DAG.getConstant(c1 + c2, N1.getValueType()));
   }
 
+  // fold (srl (trunc (srl x, c1)), c2) -> 0 or (trunc (srl x, (add c1, c2)))
+  // This is only valid if the OpSizeInBits + c1 = size of inner shift
+  if (N1C && N0.getOpcode() == ISD::TRUNCATE &&
+      N0.getOperand(0).getOpcode() == ISD::SRL &&
+      N0.getOperand(0)->getOperand(1).getOpcode() == ISD::Constant) {
+    uint64_t c1 = 
+      cast<ConstantSDNode>(N0.getOperand(0)->getOperand(1))->getZExtValue();
+    uint64_t c2 = N1C->getZExtValue();
+    EVT InnerShiftVT = N0.getOperand(0)->getOperand(1).getValueType();
+    uint64_t InnerShiftSize = InnerShiftVT.getScalarType().getSizeInBits();
+    if (c1 + OpSizeInBits == InnerShiftSize) {
+      if (c1 + c2 >= InnerShiftSize)
+        return DAG.getConstant(0, VT);
+      return DAG.getNode(ISD::TRUNCATE, N0->getDebugLoc(), VT,
+                         DAG.getNode(ISD::SRL, N0->getDebugLoc(), InnerShiftVT, 
+                                     N0.getOperand(0)->getOperand(0),
+                                     DAG.getConstant(c1 + c2, InnerShiftVT)));
+    }
+  }
+
   // fold (srl (shl x, c), c) -> (and x, cst2)
   if (N1C && N0.getOpcode() == ISD::SHL && N0.getOperand(1) == N1 &&
       N0.getValueSizeInBits() <= 64) {

diff --git a/test/CodeGen/X86/x86_64-mul-by-const.ll b/test/CodeGen/X86/x86_64-mul-by-const.ll
new file mode 100644 (file)
index 0000000..9cda9ab
--- /dev/null
+++ b/test/CodeGen/X86/x86_64-mul-by-const.ll
@@ -0,0 +1,9 @@
+; RUN: llc < %s -mtriple=x86_64-apple-darwin | FileCheck %s
+; Formerly there were two shifts.  8771012.
+
+define i32 @f9188_mul365384439_shift27(i32 %A) nounwind {
+; CHECK:  imulq $365384439,
+; CHECK:  shrq  $59, %rax
+        %tmp1 = udiv i32 %A, 1577682821         ; <i32> [#uses=1]
+        ret i32 %tmp1
+}