R600/SI: implement indirect adressing for SI

Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Tom Stellard <thomas.stellard@amd.com>

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@177277 91177308-0d34-0410-b5e6-96231b3b80d8

3 files changed, 190 insertions, 1 deletions

diff --git a/lib/Target/R600/SIISelLowering.cpp b/lib/Target/R600/SIISelLowering.cpp
index 7a1cd94..93f8c38 100644
--- a/lib/Target/R600/SIISelLowering.cpp
+++ b/lib/Target/R600/SIISelLowering.cpp
@@ -58,6 +58,11 @@ SITargetLowering::SITargetLowering(TargetMachine &TM) :
 
   computeRegisterProperties();
 
+  setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i32, Expand);
+  setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8f32, Expand);
+  setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v16i32, Expand);
+  setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v16f32, Expand);
+
   setOperationAction(ISD::ADD, MVT::i64, Legal);
   setOperationAction(ISD::ADD, MVT::i32, Legal);
 
diff --git a/lib/Target/R600/SIInstructions.td b/lib/Target/R600/SIInstructions.td
index 5a3c901..05b04a9 100644
--- a/lib/Target/R600/SIInstructions.td
+++ b/lib/Target/R600/SIInstructions.td
@@ -1149,6 +1149,31 @@ def SI_KILL : InstSI <
 } // end mayLoad = 1, mayStore = 1, hasSideEffects = 1
   // Uses = [EXEC], Defs = [EXEC]
 
+let Uses = [EXEC], Defs = [EXEC,VCC,M0] in {
+
+def SI_INDIRECT_SRC : InstSI <
+  (outs VReg_32:$dst, SReg_64:$temp),
+  (ins unknown:$src, VSrc_32:$idx, i32imm:$off),
+  "SI_INDIRECT_SRC $dst, $temp, $src, $idx, $off",
+  []
+>;
+
+class SI_INDIRECT_DST<RegisterClass rc> : InstSI <
+  (outs rc:$dst, SReg_64:$temp),
+  (ins unknown:$src, VSrc_32:$idx, i32imm:$off, VReg_32:$val),
+  "SI_INDIRECT_DST $dst, $temp, $src, $idx, $off, $val",
+  []
+> {
+  let Constraints = "$src = $dst";
+}
+
+def SI_INDIRECT_DST_V2 : SI_INDIRECT_DST<VReg_64>;
+def SI_INDIRECT_DST_V4 : SI_INDIRECT_DST<VReg_128>;
+def SI_INDIRECT_DST_V8 : SI_INDIRECT_DST<VReg_256>;
+def SI_INDIRECT_DST_V16 : SI_INDIRECT_DST<VReg_512>;
+
+} // Uses = [EXEC,VCC,M0], Defs = [EXEC,VCC,M0]
+
 } // end IsCodeGenOnly, isPseudo
 
 def : Pat<
@@ -1521,4 +1546,48 @@ defm : SMRD_Pattern <S_LOAD_DWORD_IMM, S_LOAD_DWORD_SGPR, i32>;
 defm : SMRD_Pattern <S_LOAD_DWORDX4_IMM, S_LOAD_DWORDX4_SGPR, v16i8>;
 defm : SMRD_Pattern <S_LOAD_DWORDX8_IMM, S_LOAD_DWORDX8_SGPR, v32i8>;
 
+/********** ====================== **********/
+/**********   Indirect adressing   **********/
+/********** ====================== **********/
+
+multiclass SI_INDIRECT_Pattern <RegisterClass rc, ValueType vt,
+                                SI_INDIRECT_DST IndDst> {
+  // 1. Extract with offset
+  def : Pat<
+    (vector_extract (vt rc:$vec),
+      (i64 (zext (i32 (add VReg_32:$idx, imm:$off))))
+    ),
+    (f32 (SI_INDIRECT_SRC (IMPLICIT_DEF), rc:$vec, VReg_32:$idx, imm:$off))
+  >;
+
+  // 2. Extract without offset
+  def : Pat<
+    (vector_extract (vt rc:$vec),
+      (i64 (zext (i32 VReg_32:$idx)))
+    ),
+    (f32 (SI_INDIRECT_SRC (IMPLICIT_DEF), rc:$vec, VReg_32:$idx, 0))
+  >;
+
+  // 3. Insert with offset
+  def : Pat<
+    (vector_insert (vt rc:$vec), (f32 VReg_32:$val),
+      (i64 (zext (i32 (add VReg_32:$idx, imm:$off))))
+    ),
+    (vt (IndDst (IMPLICIT_DEF), rc:$vec, VReg_32:$idx, imm:$off, VReg_32:$val))
+  >;
+
+  // 4. Insert without offset
+  def : Pat<
+    (vector_insert (vt rc:$vec), (f32 VReg_32:$val),
+      (i64 (zext (i32 VReg_32:$idx)))
+    ),
+    (vt (IndDst (IMPLICIT_DEF), rc:$vec, VReg_32:$idx, 0, VReg_32:$val))
+  >;
+}
+
+defm : SI_INDIRECT_Pattern <VReg_64, v2f32, SI_INDIRECT_DST_V2>;
+defm : SI_INDIRECT_Pattern <VReg_128, v4f32, SI_INDIRECT_DST_V4>;
+defm : SI_INDIRECT_Pattern <VReg_256, v8f32, SI_INDIRECT_DST_V8>;
+defm : SI_INDIRECT_Pattern <VReg_512, v16f32, SI_INDIRECT_DST_V16>;
+
 } // End isSI predicate
diff --git a/lib/Target/R600/SILowerControlFlow.cpp b/lib/Target/R600/SILowerControlFlow.cpp
index b215aa2..9a027e7 100644
--- a/lib/Target/R600/SILowerControlFlow.cpp
+++ b/lib/Target/R600/SILowerControlFlow.cpp
@@ -66,6 +66,7 @@ private:
   static const unsigned SkipThreshold = 12;
 
   static char ID;
+  const TargetRegisterInfo *TRI;
   const TargetInstrInfo *TII;
 
   bool shouldSkip(MachineBasicBlock *From, MachineBasicBlock *To);
@@ -84,9 +85,14 @@ private:
   void Kill(MachineInstr &MI);
   void Branch(MachineInstr &MI);
 
+  void LoadM0(MachineInstr &MI, MachineInstr *MovRel);
+  void IndirectSrc(MachineInstr &MI);
+  void IndirectDst(MachineInstr &MI);
+
 public:
   SILowerControlFlowPass(TargetMachine &tm) :
-    MachineFunctionPass(ID), TII(tm.getInstrInfo()) { }
+    MachineFunctionPass(ID), TRI(tm.getRegisterInfo()),
+    TII(tm.getInstrInfo()) { }
 
   virtual bool runOnMachineFunction(MachineFunction &MF);
 
@@ -302,6 +308,104 @@ void SILowerControlFlowPass::Kill(MachineInstr &MI) {
   MI.eraseFromParent();
 }
 
+void SILowerControlFlowPass::LoadM0(MachineInstr &MI, MachineInstr *MovRel) {
+
+  MachineBasicBlock &MBB = *MI.getParent();
+  DebugLoc DL = MI.getDebugLoc();
+  MachineBasicBlock::iterator I = MI;
+
+  unsigned Save = MI.getOperand(1).getReg();
+  unsigned Idx = MI.getOperand(3).getReg();
+
+  if (AMDGPU::SReg_32RegClass.contains(Idx)) {
+    BuildMI(MBB, &MI, DL, TII->get(AMDGPU::S_MOV_B32), AMDGPU::M0)
+            .addReg(Idx);
+    MBB.insert(I, MovRel);
+    MI.eraseFromParent();
+    return;
+  }
+
+  assert(AMDGPU::SReg_64RegClass.contains(Save));
+  assert(AMDGPU::VReg_32RegClass.contains(Idx));
+
+  // Save the EXEC mask
+  BuildMI(MBB, &MI, DL, TII->get(AMDGPU::S_MOV_B64), Save)
+          .addReg(AMDGPU::EXEC);
+
+  // Read the next variant into VCC (lower 32 bits) <- also loop target
+  BuildMI(MBB, &MI, DL, TII->get(AMDGPU::V_READFIRSTLANE_B32_e32), AMDGPU::VCC)
+          .addReg(Idx);
+
+  // Move index from VCC into M0
+  BuildMI(MBB, &MI, DL, TII->get(AMDGPU::S_MOV_B32), AMDGPU::M0)
+          .addReg(AMDGPU::VCC);
+
+  // Compare the just read M0 value to all possible Idx values
+  BuildMI(MBB, &MI, DL, TII->get(AMDGPU::V_CMP_EQ_U32_e32), AMDGPU::VCC)
+          .addReg(AMDGPU::M0)
+          .addReg(Idx);
+
+  // Update EXEC, save the original EXEC value to VCC
+  BuildMI(MBB, &MI, DL, TII->get(AMDGPU::S_AND_SAVEEXEC_B64), AMDGPU::VCC)
+          .addReg(AMDGPU::VCC);
+
+  // Do the actual move
+  MBB.insert(I, MovRel);
+
+  // Update EXEC, switch all done bits to 0 and all todo bits to 1
+  BuildMI(MBB, &MI, DL, TII->get(AMDGPU::S_XOR_B64), AMDGPU::EXEC)
+          .addReg(AMDGPU::EXEC)
+          .addReg(AMDGPU::VCC);
+
+  // Loop back to V_READFIRSTLANE_B32 if there are still variants to cover
+  BuildMI(MBB, &MI, DL, TII->get(AMDGPU::S_CBRANCH_EXECNZ))
+          .addImm(-7)
+          .addReg(AMDGPU::EXEC);
+
+  // Restore EXEC
+  BuildMI(MBB, &MI, DL, TII->get(AMDGPU::S_MOV_B64), AMDGPU::EXEC)
+          .addReg(Save);
+
+  MI.eraseFromParent();
+}
+
+void SILowerControlFlowPass::IndirectSrc(MachineInstr &MI) {
+
+  MachineBasicBlock &MBB = *MI.getParent();
+  DebugLoc DL = MI.getDebugLoc();
+
+  unsigned Dst = MI.getOperand(0).getReg();
+  unsigned Vec = MI.getOperand(2).getReg();
+  unsigned Off = MI.getOperand(4).getImm();
+
+  MachineInstr *MovRel = 
+    BuildMI(*MBB.getParent(), DL, TII->get(AMDGPU::V_MOVRELS_B32_e32), Dst)
+            .addReg(TRI->getSubReg(Vec, AMDGPU::sub0) + Off)
+            .addReg(AMDGPU::M0, RegState::Implicit)
+            .addReg(Vec, RegState::Implicit);
+
+  LoadM0(MI, MovRel);
+}
+
+void SILowerControlFlowPass::IndirectDst(MachineInstr &MI) {
+
+  MachineBasicBlock &MBB = *MI.getParent();
+  DebugLoc DL = MI.getDebugLoc();
+
+  unsigned Dst = MI.getOperand(0).getReg();
+  unsigned Off = MI.getOperand(4).getImm();
+  unsigned Val = MI.getOperand(5).getReg();
+
+  MachineInstr *MovRel = 
+    BuildMI(*MBB.getParent(), DL, TII->get(AMDGPU::V_MOVRELD_B32_e32))
+            .addReg(TRI->getSubReg(Dst, AMDGPU::sub0) + Off, RegState::Define)
+            .addReg(Val)
+            .addReg(AMDGPU::M0, RegState::Implicit)
+            .addReg(Dst, RegState::Implicit);
+
+  LoadM0(MI, MovRel);
+}
+
 bool SILowerControlFlowPass::runOnMachineFunction(MachineFunction &MF) {
 
   bool HaveKill = false;
@@ -363,6 +467,17 @@ bool SILowerControlFlowPass::runOnMachineFunction(MachineFunction &MF) {
         case AMDGPU::S_BRANCH:
           Branch(MI);
           break;
+
+        case AMDGPU::SI_INDIRECT_SRC:
+          IndirectSrc(MI);
+          break;
+
+        case AMDGPU::SI_INDIRECT_DST_V2:
+        case AMDGPU::SI_INDIRECT_DST_V4:
+        case AMDGPU::SI_INDIRECT_DST_V8:
+        case AMDGPU::SI_INDIRECT_DST_V16:
+          IndirectDst(MI);
+          break;
       }
     }
   }