From: Nemanja Ivanovic <nemanja.i.ibm@gmail.com>
Date: Tue, 29 Nov 2016 23:36:03 +0000 (+0000)
Subject: [PowerPC] Improvements for BUILD_VECTOR Vol. 2
X-Git-Url: http://review.tizen.org/git/?a=commitdiff_plain;h=8c11e79b1773ba484db0a710aa90ebffbfea550c;p=platform%2Fupstream%2Fllvm.git

[PowerPC] Improvements for BUILD_VECTOR Vol. 2

This patch corresponds to review:
https://reviews.llvm.org/D25980

This is the 2nd patch in a series of 4 that improve the lowering and combining
for BUILD_VECTOR nodes on PowerPC. This particular patch combines a build vector
of fp-to-int conversions into an fp-to-int conversion of a build vector of fp
values. For example:
Converts (build_vector (fp_to_[su]i $A), (fp_to_[su]i $B), ...)
Into (fp_to_[su]i (build_vector $A, $B, ...))).
Which is a natural match for much cleaner code.

llvm-svn: 288218
---

diff --git a/llvm/lib/Target/PowerPC/PPCISelLowering.cpp b/llvm/lib/Target/PowerPC/PPCISelLowering.cpp
index 4797fe3..73572e9 100644
--- a/llvm/lib/Target/PowerPC/PPCISelLowering.cpp
+++ b/llvm/lib/Target/PowerPC/PPCISelLowering.cpp
@@ -10670,6 +10670,86 @@ SDValue PPCTargetLowering::DAGCombineExtBoolTrunc(SDNode *N,
       ShiftCst);
 }
 
+/// \brief Reduces the number of fp-to-int conversion when building a vector.
+///
+/// If this vector is built out of floating to integer conversions,
+/// transform it to a vector built out of floating point values followed by a
+/// single floating to integer conversion of the vector.
+/// Namely  (build_vector (fptosi $A), (fptosi $B), ...)
+/// becomes (fptosi (build_vector ($A, $B, ...)))
+SDValue PPCTargetLowering::
+combineElementTruncationToVectorTruncation(SDNode *N,
+                                           DAGCombinerInfo &DCI) const {
+  assert(N->getOpcode() == ISD::BUILD_VECTOR &&
+         "Should be called with a BUILD_VECTOR node");
+
+  SelectionDAG &DAG = DCI.DAG;
+  SDLoc dl(N);
+
+  SDValue FirstInput = N->getOperand(0);
+  assert(FirstInput.getOpcode() == PPCISD::MFVSR &&
+         "The input operand must be an fp-to-int conversion.");
+
+  // This combine happens after legalization so the fp_to_[su]i nodes are
+  // already converted to PPCSISD nodes.
+  unsigned FirstConversion = FirstInput.getOperand(0).getOpcode();
+  if (FirstConversion == PPCISD::FCTIDZ ||
+      FirstConversion == PPCISD::FCTIDUZ ||
+      FirstConversion == PPCISD::FCTIWZ ||
+      FirstConversion == PPCISD::FCTIWUZ) {
+    bool IsSplat = true;
+    bool Is32Bit = FirstConversion == PPCISD::FCTIWZ ||
+      FirstConversion == PPCISD::FCTIWUZ;
+    EVT SrcVT = FirstInput.getOperand(0).getValueType();
+    SmallVector<SDValue, 4> Ops;
+    EVT TargetVT = N->getValueType(0);
+    for (int i = 0, e = N->getNumOperands(); i < e; ++i) {
+      if (N->getOperand(i).getOpcode() != PPCISD::MFVSR)
+        return SDValue();
+      unsigned NextConversion = N->getOperand(i).getOperand(0).getOpcode();
+      if (NextConversion != FirstConversion)
+        return SDValue();
+      if (N->getOperand(i) != FirstInput)
+        IsSplat = false;
+    }
+
+    // If this is a splat, we leave it as-is since there will be only a single
+    // fp-to-int conversion followed by a splat of the integer. This is better
+    // for 32-bit and smaller ints and neutral for 64-bit ints.
+    if (IsSplat)
+      return SDValue();
+
+    // Now that we know we have the right type of node, get its operands
+    for (int i = 0, e = N->getNumOperands(); i < e; ++i) {
+      SDValue In = N->getOperand(i).getOperand(0);
+      // For 32-bit values, we need to add an FP_ROUND node.
+      if (Is32Bit) {
+        if (In.isUndef())
+          Ops.push_back(DAG.getUNDEF(SrcVT));
+        else {
+          SDValue Trunc = DAG.getNode(ISD::FP_ROUND, dl,
+                                      MVT::f32, In.getOperand(0),
+                                      DAG.getIntPtrConstant(1, dl));
+          Ops.push_back(Trunc);
+        }
+      } else
+        Ops.push_back(In.isUndef() ? DAG.getUNDEF(SrcVT) : In.getOperand(0));
+    }
+
+    unsigned Opcode;
+    if (FirstConversion == PPCISD::FCTIDZ ||
+        FirstConversion == PPCISD::FCTIWZ)
+      Opcode = ISD::FP_TO_SINT;
+    else
+      Opcode = ISD::FP_TO_UINT;
+
+    EVT NewVT = TargetVT == MVT::v2i64 ? MVT::v2f64 : MVT::v4f32;
+    SDValue BV = DAG.getBuildVector(NewVT, dl, Ops);
+    return DAG.getNode(Opcode, dl, TargetVT, BV);
+  }
+  return SDValue();
+}
+
 SDValue PPCTargetLowering::DAGCombineBuildVector(SDNode *N,
                                                  DAGCombinerInfo &DCI) const {
   assert(N->getOpcode() == ISD::BUILD_VECTOR &&
@@ -10677,7 +10757,20 @@ SDValue PPCTargetLowering::DAGCombineBuildVector(SDNode *N,
 
   SelectionDAG &DAG = DCI.DAG;
   SDLoc dl(N);
-  if (N->getValueType(0) != MVT::v2f64 || !Subtarget.hasVSX())
+
+  if (!Subtarget.hasVSX())
+    return SDValue();
+
+  // The target independent DAG combiner will leave a build_vector of
+  // float-to-int conversions intact. We can generate MUCH better code for
+  // a float-to-int conversion of a vector of floats.
+  SDValue FirstInput = N->getOperand(0);
+  if (FirstInput.getOpcode() == PPCISD::MFVSR) {
+    SDValue Reduced = combineElementTruncationToVectorTruncation(N, DCI);
+    if (Reduced)
+      return Reduced;
+  }
+  if (N->getValueType(0) != MVT::v2f64)
     return SDValue();
 
   // Looking for:
diff --git a/llvm/lib/Target/PowerPC/PPCISelLowering.h b/llvm/lib/Target/PowerPC/PPCISelLowering.h
index 1723d12..95f32f5 100644
--- a/llvm/lib/Target/PowerPC/PPCISelLowering.h
+++ b/llvm/lib/Target/PowerPC/PPCISelLowering.h
@@ -990,6 +990,10 @@ namespace llvm {
     unsigned combineRepeatedFPDivisors() const override;
 
     CCAssignFn *useFastISelCCs(unsigned Flag) const;
+
+    SDValue
+      combineElementTruncationToVectorTruncation(SDNode *N,
+                                                 DAGCombinerInfo &DCI) const;
   };
 
   namespace PPC {