AMDGPU: Use i64 scalar compare instructions

VI added eq/ne for i64, so use them.

llvm-svn: 281800

4 files changed, 45 insertions, 12 deletions

diff --git a/llvm/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp b/llvm/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp
index 89215c10a67..951db65efbb 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp
@@ -38,18 +38,6 @@ class R600InstrInfo;
 
 namespace {
 
-static bool isCBranchSCC(const SDNode *N) {
-  assert(N->getOpcode() == ISD::BRCOND);
-  if (!N->hasOneUse())
-    return false;
-
-  SDValue Cond = N->getOperand(1);
-  if (Cond.getOpcode() == ISD::CopyToReg)
-    Cond = Cond.getOperand(2);
-  return Cond.getOpcode() == ISD::SETCC &&
-         Cond.getOperand(0).getValueType() == MVT::i32 && Cond.hasOneUse();
-}
-
 /// AMDGPU specific code to select AMDGPU machine instructions for
 /// SelectionDAG operations.
 class AMDGPUDAGToDAGISel : public SelectionDAGISel {
@@ -150,6 +138,7 @@ private:
                    uint32_t Offset, uint32_t Width);
   void SelectS_BFEFromShifts(SDNode *N);
   void SelectS_BFE(SDNode *N);
+  bool isCBranchSCC(const SDNode *N) const;
   void SelectBRCOND(SDNode *N);
   void SelectATOMIC_CMP_SWAP(SDNode *N);
 
@@ -1337,6 +1326,32 @@ void AMDGPUDAGToDAGISel::SelectS_BFE(SDNode *N) {
   SelectCode(N);
 }
 
+bool AMDGPUDAGToDAGISel::isCBranchSCC(const SDNode *N) const {
+  assert(N->getOpcode() == ISD::BRCOND);
+  if (!N->hasOneUse())
+    return false;
+
+  SDValue Cond = N->getOperand(1);
+  if (Cond.getOpcode() == ISD::CopyToReg)
+    Cond = Cond.getOperand(2);
+
+  if (Cond.getOpcode() != ISD::SETCC || !Cond.hasOneUse())
+    return false;
+
+  MVT VT = Cond.getOperand(0).getSimpleValueType();
+  if (VT == MVT::i32)
+    return true;
+
+  if (VT == MVT::i64) {
+    auto ST = static_cast<const SISubtarget *>(Subtarget);
+
+    ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
+    return (CC == ISD::SETEQ || CC == ISD::SETNE) && ST->hasScalarCompareEq64();
+  }
+
+  return false;
+}
+
 void AMDGPUDAGToDAGISel::SelectBRCOND(SDNode *N) {
   SDValue Cond = N->getOperand(1);
 
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUSubtarget.h b/llvm/lib/Target/AMDGPU/AMDGPUSubtarget.h
index 6b953eadeed..1e52e1c8063 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUSubtarget.h
+++ b/llvm/lib/Target/AMDGPU/AMDGPUSubtarget.h
@@ -488,6 +488,10 @@ public:
     return Has16BitInsts;
   }
 
+  bool hasScalarCompareEq64() const {
+    return getGeneration() >= VOLCANIC_ISLANDS;
+  }
+
   bool enableSIScheduler() const {
     return EnableSIScheduler;
   }
diff --git a/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp b/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
index 9ca44b84608..5426f7f3ced 100644
--- a/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
@@ -1930,6 +1930,8 @@ unsigned SIInstrInfo::getVALUOp(const MachineInstr &MI) {
   case AMDGPU::S_CMP_GE_U32: return AMDGPU::V_CMP_GE_U32_e32;
   case AMDGPU::S_CMP_LT_U32: return AMDGPU::V_CMP_LT_U32_e32;
   case AMDGPU::S_CMP_LE_U32: return AMDGPU::V_CMP_LE_U32_e32;
+  case AMDGPU::S_CMP_EQ_U64: return AMDGPU::V_CMP_EQ_U64_e32;
+  case AMDGPU::S_CMP_LG_U64: return AMDGPU::V_CMP_NE_U64_e32;
   case AMDGPU::S_BCNT1_I32_B32: return AMDGPU::V_BCNT_U32_B32_e64;
   case AMDGPU::S_FF1_I32_B32: return AMDGPU::V_FFBL_B32_e32;
   case AMDGPU::S_FLBIT_I32_B32: return AMDGPU::V_FFBH_U32_e32;
diff --git a/llvm/lib/Target/AMDGPU/SOPInstructions.td b/llvm/lib/Target/AMDGPU/SOPInstructions.td
index 4451d1642f5..9744cd3cd07 100644
--- a/llvm/lib/Target/AMDGPU/SOPInstructions.td
+++ b/llvm/lib/Target/AMDGPU/SOPInstructions.td
@@ -630,6 +630,14 @@ class SOPC_CMP_32<bits<7> op, string opName,
   let isCommutable = 1;
 }
 
+class SOPC_CMP_64<bits<7> op, string opName,
+                  PatLeaf cond = COND_NULL, string revOp = opName>
+  : SOPC_Helper<op, SSrc_b64, i64, opName, cond>,
+    Commutable_REV<revOp, !eq(revOp, opName)> {
+  let isCompare = 1;
+  let isCommutable = 1;
+}
+
 class SOPC_32<bits<7> op, string opName, list<dag> pattern = []>
   : SOPC_Base<op, SSrc_b32, SSrc_b32, opName, pattern>;
 
@@ -655,6 +663,10 @@ def S_BITCMP0_B64 : SOPC_64_32 <0x0e, "s_bitcmp0_b64">;
 def S_BITCMP1_B64 : SOPC_64_32 <0x0f, "s_bitcmp1_b64">;
 def S_SETVSKIP : SOPC_32 <0x10, "s_setvskip">;
 
+let SubtargetPredicate = isVI in {
+def S_CMP_EQ_U64 : SOPC_CMP_64 <0x12, "s_cmp_eq_u64", COND_EQ>;
+def S_CMP_LG_U64 : SOPC_CMP_64 <0x13, "s_cmp_lg_u64", COND_NE>;
+}
 
 //===----------------------------------------------------------------------===//
 // SOPP Instructions