R600: Add support for GROUP_BARRIER instruction

8 files changed, 88 insertions, 3 deletions

diff --git a/lib/Target/R600/AMDGPUIntrinsics.td b/lib/Target/R600/AMDGPUIntrinsics.td
index eecb25b04f7..9f975bf9bf4 100644
--- a/lib/Target/R600/AMDGPUIntrinsics.td
+++ b/lib/Target/R600/AMDGPUIntrinsics.td
@@ -50,6 +50,8 @@ let TargetPrefix = "AMDGPU", isTarget = 1 in {
   def int_AMDGPU_umax : Intrinsic<[llvm_i32_ty], [llvm_i32_ty, llvm_i32_ty], [IntrNoMem]>;
   def int_AMDGPU_umin : Intrinsic<[llvm_i32_ty], [llvm_i32_ty, llvm_i32_ty], [IntrNoMem]>;
   def int_AMDGPU_cube : Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty], [IntrNoMem]>;
+
+  def int_AMDGPU_barrier_local  : Intrinsic<[], [], []>;
 }
 
 let TargetPrefix = "TGSI", isTarget = 1 in {
diff --git a/lib/Target/R600/R600EmitClauseMarkers.cpp b/lib/Target/R600/R600EmitClauseMarkers.cpp
index ff5ce5a2dd7..0aea2d7c030 100644
--- a/lib/Target/R600/R600EmitClauseMarkers.cpp
+++ b/lib/Target/R600/R600EmitClauseMarkers.cpp
@@ -177,7 +177,14 @@ private:
         AluInstCount ++;
         continue;
       }
-      if (I->getOpcode() == AMDGPU::KILLGT) {
+      // XXX: GROUP_BARRIER instructions cannot be in the same ALU clause as:
+      //
+      // * KILL or INTERP instructions
+      // * Any instruction that sets UPDATE_EXEC_MASK or UPDATE_PRED bits
+      // * Uses waterfalling (i.e. INDEX_MODE = AR.X)
+      //
+      // XXX: These checks have not been implemented yet.
+      if (TII->mustBeLastInClause(I->getOpcode())) {
         I++;
         break;
       }
diff --git a/lib/Target/R600/R600InstrInfo.cpp b/lib/Target/R600/R600InstrInfo.cpp
index adfe144a267..d0aeaa6de5c 100644
--- a/lib/Target/R600/R600InstrInfo.cpp
+++ b/lib/Target/R600/R600InstrInfo.cpp
@@ -163,6 +163,16 @@ bool R600InstrInfo::usesTextureCache(const MachineInstr *MI) const {
          usesTextureCache(MI->getOpcode());
 }
 
+bool R600InstrInfo::mustBeLastInClause(unsigned Opcode) const {
+  switch (Opcode) {
+  case AMDGPU::KILLGT:
+  case AMDGPU::GROUP_BARRIER:
+    return true;
+  default:
+    return false;
+  }
+}
+
 SmallVector<std::pair<MachineOperand *, int64_t>, 3>
 R600InstrInfo::getSrcs(MachineInstr *MI) const {
   SmallVector<std::pair<MachineOperand *, int64_t>, 3> Result;
diff --git a/lib/Target/R600/R600InstrInfo.h b/lib/Target/R600/R600InstrInfo.h
index f06abf6081c..3c2e50be662 100644
--- a/lib/Target/R600/R600InstrInfo.h
+++ b/lib/Target/R600/R600InstrInfo.h
@@ -72,6 +72,8 @@ namespace llvm {
   bool usesTextureCache(unsigned Opcode) const;
   bool usesTextureCache(const MachineInstr *MI) const;
 
+  bool mustBeLastInClause(unsigned Opcode) const;
+
   /// \returns a pair for each src of an ALU instructions.
   /// The first member of a pair is the register id.
   /// If register is ALU_CONST, second member is SEL.
diff --git a/lib/Target/R600/R600Instructions.td b/lib/Target/R600/R600Instructions.td
index 31e103d6391..5231fdeb069 100644
--- a/lib/Target/R600/R600Instructions.td
+++ b/lib/Target/R600/R600Instructions.td
@@ -1489,6 +1489,36 @@ let hasSideEffects = 1 in {
 
   def UINT_TO_FLT_eg : UINT_TO_FLT_Common<0x9C>;
 
+def GROUP_BARRIER : InstR600 <
+    (outs), (ins), "  GROUP_BARRIER", [(int_AMDGPU_barrier_local)], AnyALU>,
+    R600ALU_Word0,
+    R600ALU_Word1_OP2 <0x54> {
+
+  let dst = 0;
+  let dst_rel = 0;
+  let src0 = 0;
+  let src0_rel = 0;
+  let src0_neg = 0;
+  let src0_abs = 0;
+  let src1 = 0;
+  let src1_rel = 0;
+  let src1_neg = 0;
+  let src1_abs = 0;
+  let write = 0;
+  let omod = 0;
+  let clamp = 0;
+  let last = 1;
+  let bank_swizzle = 0;
+  let pred_sel = 0;
+  let update_exec_mask = 0;
+  let update_pred = 0;
+
+  let Inst{31-0}  = Word0;
+  let Inst{63-32} = Word1;
+
+  let ALUInst = 1;
+}
+
   // TRUNC is used for the FLT_TO_INT instructions to work around a
   // perceived problem where the rounding modes are applied differently
   // depending on the instruction and the slot they are in.
diff --git a/lib/Target/R600/R600MachineScheduler.cpp b/lib/Target/R600/R600MachineScheduler.cpp
index a330d885743..acc1b4d6ee3 100644
--- a/lib/Target/R600/R600MachineScheduler.cpp
+++ b/lib/Target/R600/R600MachineScheduler.cpp
@@ -269,10 +269,14 @@ R600SchedStrategy::AluKind R600SchedStrategy::getAluKind(SUnit *SU) const {
     }
 
     // Does the instruction take a whole IG ?
+    // XXX: Is it possible to add a helper function in R600InstrInfo that can
+    // be used here and in R600PacketizerList::isSoloInstruction() ?
     if(TII->isVector(*MI) ||
         TII->isCubeOp(MI->getOpcode()) ||
-        TII->isReductionOp(MI->getOpcode()))
+        TII->isReductionOp(MI->getOpcode()) ||
+        MI->getOpcode() == AMDGPU::GROUP_BARRIER) {
       return AluT_XYZW;
+    }
 
     // Is the result already assigned to a channel ?
     unsigned DestSubReg = MI->getOperand(0).getSubReg();
diff --git a/lib/Target/R600/R600Packetizer.cpp b/lib/Target/R600/R600Packetizer.cpp
index 6024fd5c85f..4c72d229675 100644
--- a/lib/Target/R600/R600Packetizer.cpp
+++ b/lib/Target/R600/R600Packetizer.cpp
@@ -82,7 +82,11 @@ private:
       int OperandIdx = TII->getOperandIdx(BI->getOpcode(), AMDGPU::OpName::write);
       if (OperandIdx > -1 && BI->getOperand(OperandIdx).getImm() == 0)
         continue;
-      unsigned Dst = BI->getOperand(0).getReg();
+      int DstIdx = TII->getOperandIdx(BI->getOpcode(), AMDGPU::OpName::dst);
+      if (DstIdx == -1) {
+        continue;
+      }
+      unsigned Dst = BI->getOperand(DstIdx).getReg();
       if (BI->getOpcode() == AMDGPU::DOT4_r600 ||
           BI->getOpcode() == AMDGPU::DOT4_eg) {
         Result[Dst] = AMDGPU::PV_X;
@@ -154,6 +158,8 @@ public:
       return true;
     if (TII->isTransOnly(MI))
       return true;
+    if (MI->getOpcode() == AMDGPU::GROUP_BARRIER)
+      return true;
     return false;
   }
 
diff --git a/test/CodeGen/R600/llvm.AMDGPU.barrier.local.ll b/test/CodeGen/R600/llvm.AMDGPU.barrier.local.ll
new file mode 100644
index 00000000000..8d3c9ca2230
--- /dev/null
+++ b/test/CodeGen/R600/llvm.AMDGPU.barrier.local.ll
@@ -0,0 +1,24 @@
+; RUN: llc < %s -march=r600 -mcpu=redwood | FileCheck %s
+
+; CHECK: GROUP_BARRIER
+
+define void @test(i32 addrspace(1)* %out) {
+entry:
+  %0 = call i32 @llvm.r600.read.tidig.x()
+  %1 = getelementptr i32 addrspace(1)* %out, i32 %0
+  store i32 %0, i32 addrspace(1)* %1
+  call void @llvm.AMDGPU.barrier.local()
+  %2 = call i32 @llvm.r600.read.local.size.x()
+  %3 = sub i32 %2, 1
+  %4 = sub i32 %3, %0
+  %5 = getelementptr i32 addrspace(1)* %out, i32 %4
+  %6 = load i32 addrspace(1)* %5
+  store i32 %6, i32 addrspace(1)* %1
+  ret void
+}
+
+declare i32 @llvm.r600.read.tidig.x() #0
+declare void @llvm.AMDGPU.barrier.local()
+declare i32 @llvm.r600.read.local.size.x() #0
+
+attributes #0 = { readnone }