AMDGPU: Remove unnecessary ands when f16 is legal

Add a new node to act as a fancy bitcast from f16 operations to
i32 that implicitly zero the high 16-bits of the result.

Alternatively could try making v2f16 legal and canonicalizing
on build_vectors.

llvm-svn: 299246

6 files changed, 57 insertions, 2 deletions

diff --git a/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp b/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp
index 3f41f4e695e..54ec756bb18 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp
@@ -3491,6 +3491,7 @@ const char* AMDGPUTargetLowering::getTargetNodeName(unsigned Opcode) const {
   NODE_NAME_CASE(CVT_F32_UBYTE3)
   NODE_NAME_CASE(CVT_PKRTZ_F16_F32)
   NODE_NAME_CASE(FP_TO_FP16)
+  NODE_NAME_CASE(FP16_ZEXT)
   NODE_NAME_CASE(BUILD_VERTICAL_VECTOR)
   NODE_NAME_CASE(CONST_DATA_PTR)
   NODE_NAME_CASE(PC_ADD_REL_OFFSET)
@@ -3587,7 +3588,10 @@ void AMDGPUTargetLowering::computeKnownBitsForTargetNode(
 
     break;
   }
-  case AMDGPUISD::FP_TO_FP16: {
+  case AMDGPUISD::FP_TO_FP16:
+  case AMDGPUISD::FP16_ZEXT: {
+    unsigned BitWidth = KnownZero.getBitWidth();
+
     // High bits are zero.
     KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - 16);
     break;
@@ -3621,7 +3625,9 @@ unsigned AMDGPUTargetLowering::ComputeNumSignBitsForTargetNode(
   case AMDGPUISD::CARRY:
   case AMDGPUISD::BORROW:
     return 31;
-
+  case AMDGPUISD::FP_TO_FP16:
+  case AMDGPUISD::FP16_ZEXT:
+    return 16;
   default:
     return 1;
   }
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.h b/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.h
index 66e8464fd7f..cffed2cfeb7 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.h
+++ b/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.h
@@ -353,6 +353,9 @@ enum NodeType : unsigned {
   // are known 0.
   FP_TO_FP16,
 
+  // Wrapper around fp16 results that are known to zero the high bits.
+  FP16_ZEXT,
+
   /// This node is for VLIW targets and it is used to represent a vector
   /// that is stored in consecutive registers with the same channel.
   /// For example:
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUInstrInfo.td b/llvm/lib/Target/AMDGPU/AMDGPUInstrInfo.td
index ed2c1b970e8..56f060984f0 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUInstrInfo.td
+++ b/llvm/lib/Target/AMDGPU/AMDGPUInstrInfo.td
@@ -112,6 +112,7 @@ def AMDGPUldexp : SDNode<"AMDGPUISD::LDEXP", AMDGPULdExpOp>;
 
 def AMDGPUpkrtz_f16_f32 : SDNode<"AMDGPUISD::CVT_PKRTZ_F16_F32", AMDGPUFPPackOp>;
 def AMDGPUfp_to_f16 : SDNode<"AMDGPUISD::FP_TO_FP16" , SDTFPToIntOp>;
+def AMDGPUfp16_zext : SDNode<"AMDGPUISD::FP16_ZEXT" , SDTFPToIntOp>;
 
 
 def AMDGPUfp_class : SDNode<"AMDGPUISD::FP_CLASS", AMDGPUFPClassOp>;
diff --git a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
index 924bae31a12..f1f37f4c0f7 100644
--- a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -475,6 +475,7 @@ SITargetLowering::SITargetLowering(const TargetMachine &TM,
   setTargetDAGCombine(ISD::UINT_TO_FP);
   setTargetDAGCombine(ISD::FCANONICALIZE);
   setTargetDAGCombine(ISD::SCALAR_TO_VECTOR);
+  setTargetDAGCombine(ISD::ZERO_EXTEND);
 
   // All memory operations. Some folding on the pointer operand is done to help
   // matching the constant offsets in the addressing modes.
@@ -4058,6 +4059,42 @@ SDValue SITargetLowering::performXorCombine(SDNode *N,
   return SDValue();
 }
 
+static bool fp16SrcZerosHighBits(unsigned Opc) {
+  switch (Opc) {
+  case ISD::SELECT:
+  case ISD::EXTRACT_VECTOR_ELT:
+    return false;
+  default:
+    return true;
+  }
+}
+
+SDValue SITargetLowering::performZeroExtendCombine(SDNode *N,
+                                                   DAGCombinerInfo &DCI) const {
+  if (!Subtarget->has16BitInsts() ||
+      DCI.getDAGCombineLevel() < AfterLegalizeDAG)
+    return SDValue();
+
+  EVT VT = N->getValueType(0);
+  if (VT != MVT::i32)
+    return SDValue();
+
+  SDValue Src = N->getOperand(0);
+  if (Src.getValueType() != MVT::i16)
+    return SDValue();
+
+  // (i32 zext (i16 (bitcast f16:$src))) -> fp16_zext $src
+  // FIXME: It is not universally true that the high bits are zeroed on gfx9.
+  if (Src.getOpcode() == ISD::BITCAST) {
+    SDValue BCSrc = Src.getOperand(0);
+    if (BCSrc.getValueType() == MVT::f16 &&
+        fp16SrcZerosHighBits(BCSrc.getOpcode()))
+      return DCI.DAG.getNode(AMDGPUISD::FP16_ZEXT, SDLoc(N), VT, BCSrc);
+  }
+
+  return SDValue();
+}
+
 SDValue SITargetLowering::performClassCombine(SDNode *N,
                                               DAGCombinerInfo &DCI) const {
   SelectionDAG &DAG = DCI.DAG;
@@ -4594,6 +4631,8 @@ SDValue SITargetLowering::PerformDAGCombine(SDNode *N,
     return performOrCombine(N, DCI);
   case ISD::XOR:
     return performXorCombine(N, DCI);
+  case ISD::ZERO_EXTEND:
+    return performZeroExtendCombine(N, DCI);
   case AMDGPUISD::FP_CLASS:
     return performClassCombine(N, DCI);
   case ISD::FCANONICALIZE:
diff --git a/llvm/lib/Target/AMDGPU/SIISelLowering.h b/llvm/lib/Target/AMDGPU/SIISelLowering.h
index 39f7a671749..550e8a2e38f 100644
--- a/llvm/lib/Target/AMDGPU/SIISelLowering.h
+++ b/llvm/lib/Target/AMDGPU/SIISelLowering.h
@@ -81,6 +81,7 @@ class SITargetLowering final : public AMDGPUTargetLowering {
   SDValue performAndCombine(SDNode *N, DAGCombinerInfo &DCI) const;
   SDValue performOrCombine(SDNode *N, DAGCombinerInfo &DCI) const;
   SDValue performXorCombine(SDNode *N, DAGCombinerInfo &DCI) const;
+  SDValue performZeroExtendCombine(SDNode *N, DAGCombinerInfo &DCI) const;
   SDValue performClassCombine(SDNode *N, DAGCombinerInfo &DCI) const;
   SDValue performFCanonicalizeCombine(SDNode *N, DAGCombinerInfo &DCI) const;
 
diff --git a/llvm/lib/Target/AMDGPU/SIInstructions.td b/llvm/lib/Target/AMDGPU/SIInstructions.td
index 9e343aab71e..fde14792ea6 100644
--- a/llvm/lib/Target/AMDGPU/SIInstructions.td
+++ b/llvm/lib/Target/AMDGPU/SIInstructions.td
@@ -1076,6 +1076,11 @@ def : Pat <
 //===----------------------------------------------------------------------===//
 // Miscellaneous Patterns
 //===----------------------------------------------------------------------===//
+def : Pat <
+  (i32 (AMDGPUfp16_zext f16:$src)),
+  (COPY $src)
+>;
+
 
 def : Pat <
   (i32 (trunc i64:$a)),