AMDGPU/SI: Detect uniform branches and emit s_cbranch instructions

Reviewers: arsenm

Subscribers: mareko, MatzeB, qcolombet, arsenm, llvm-commits

Differential Revision: http://reviews.llvm.org/D16603

llvm-svn: 260765

1 files changed, 55 insertions, 0 deletions

diff --git a/llvm/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp b/llvm/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp
index c2a1beb63c5..fb9a66cca84 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp
@@ -13,6 +13,7 @@
 //===----------------------------------------------------------------------===//
 
 #include "AMDGPUInstrInfo.h"
+#include "AMDGPUIntrinsicInfo.h"
 #include "AMDGPUISelLowering.h" // For AMDGPUISD
 #include "AMDGPURegisterInfo.h"
 #include "AMDGPUSubtarget.h"
@@ -36,6 +37,20 @@ using namespace llvm;
 //===----------------------------------------------------------------------===//
 
 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 {
@@ -82,6 +97,8 @@ private:
   bool isLocalLoad(const LoadSDNode *N) const;
   bool isRegionLoad(const LoadSDNode *N) const;
 
+  bool isUniformBr(const SDNode *N) const;
+
   SDNode *glueCopyToM0(SDNode *N) const;
 
   const TargetRegisterClass *getOperandRegClass(SDNode *N, unsigned OpNo) const;
@@ -143,6 +160,7 @@ private:
                    uint32_t Offset, uint32_t Width);
   SDNode *SelectS_BFEFromShifts(SDNode *N);
   SDNode *SelectS_BFE(SDNode *N);
+  SDNode *SelectBRCOND(SDNode *N);
 
   // Include the pieces autogenerated from the target description.
 #include "AMDGPUGenDAGISel.inc"
@@ -509,6 +527,8 @@ SDNode *AMDGPUDAGToDAGISel::Select(SDNode *N) {
       break;
 
     return SelectS_BFE(N);
+  case ISD::BRCOND:
+    return SelectBRCOND(N);
   }
 
   return SelectCode(N);
@@ -623,6 +643,11 @@ bool AMDGPUDAGToDAGISel::isPrivateLoad(const LoadSDNode *N) const {
   return false;
 }
 
+bool AMDGPUDAGToDAGISel::isUniformBr(const SDNode *N) const {
+  const BasicBlock *BB = FuncInfo->MBB->getBasicBlock();
+  return BB->getTerminator()->getMetadata("amdgpu.uniform");
+}
+
 const char *AMDGPUDAGToDAGISel::getPassName() const {
   return "AMDGPU DAG->DAG Pattern Instruction Selection";
 }
@@ -1365,6 +1390,36 @@ SDNode *AMDGPUDAGToDAGISel::SelectS_BFE(SDNode *N) {
   return SelectCode(N);
 }
 
+SDNode *AMDGPUDAGToDAGISel::SelectBRCOND(SDNode *N) {
+  SDValue Cond = N->getOperand(1);
+
+  if (isCBranchSCC(N)) {
+    // This brcond will use S_CBRANCH_SCC*, so let tablegen handle it.
+    return SelectCode(N);
+  }
+
+  // The result of VOPC instructions is or'd against ~EXEC before it is
+  // written to vcc or another SGPR.  This means that the value '1' is always
+  // written to the corresponding bit for results that are masked.  In order
+  // to correctly check against vccz, we need to and VCC with the EXEC
+  // register in order to clear the value from the masked bits.
+
+  SDLoc SL(N);
+
+  SDNode *MaskedCond =
+        CurDAG->getMachineNode(AMDGPU::S_AND_B64, SL, MVT::i1,
+                               CurDAG->getRegister(AMDGPU::EXEC, MVT::i1),
+                               Cond);
+  SDValue VCC = CurDAG->getCopyToReg(N->getOperand(0), SL, AMDGPU::VCC,
+                                     SDValue(MaskedCond, 0),
+                                     SDValue()); // Passing SDValue() adds a
+                                                 // glue output.
+  return CurDAG->SelectNodeTo(N, AMDGPU::S_CBRANCH_VCCNZ, MVT::Other,
+                              N->getOperand(2), // Basic Block
+                              VCC.getValue(0),  // Chain
+                              VCC.getValue(1)); // Glue
+}
+
 bool AMDGPUDAGToDAGISel::SelectVOP3Mods(SDValue In, SDValue &Src,
                                         SDValue &SrcMods) const {