From 04051b5fad96e340d6de5a028356530f881b2bcc Mon Sep 17 00:00:00 2001
From: Tom Stellard <thomas.stellard@amd.com>
Date: Thu, 27 Oct 2016 23:42:29 +0000
Subject: [PATCH] AMDGPU/SI: Handle hazard with sgpr lane selects for
 v_{read,write}lane

Reviewers: arsenm

Subscribers: kzhuravl, wdng, nhaehnle, yaxunl, tony-tye, llvm-commits

Differential Revision: https://reviews.llvm.org/D25637

llvm-svn: 285367
---
 llvm/lib/Target/AMDGPU/GCNHazardRecognizer.cpp     | 34 ++++++++++-
 llvm/lib/Target/AMDGPU/GCNHazardRecognizer.h       |  1 +
 .../CodeGen/MIR/AMDGPU/inserted-wait-states.mir    | 66 ++++++++++++++++++++++
 3 files changed, 100 insertions(+), 1 deletion(-)

diff --git a/llvm/lib/Target/AMDGPU/GCNHazardRecognizer.cpp b/llvm/lib/Target/AMDGPU/GCNHazardRecognizer.cpp
index 687076a..1df397b 100644
--- a/llvm/lib/Target/AMDGPU/GCNHazardRecognizer.cpp
+++ b/llvm/lib/Target/AMDGPU/GCNHazardRecognizer.cpp
@@ -50,7 +50,11 @@ static bool isSSetReg(unsigned Opcode) {
   return Opcode == AMDGPU::S_SETREG_B32 || Opcode == AMDGPU::S_SETREG_IMM32_B32;
 }
 
-static unsigned getHWReg(const SIInstrInfo *TII, const MachineInstr &RegInstr) {
+static bool isRWLane(unsigned Opcode) {
+  return Opcode == AMDGPU::V_READLANE_B32 || Opcode == AMDGPU::V_WRITELANE_B32;
+}
+
+static bool getHWReg(const SIInstrInfo *TII, const MachineInstr &RegInstr) {
 
   const MachineOperand *RegOp = TII->getNamedOperand(RegInstr,
                                                      AMDGPU::OpName::simm16);
@@ -76,6 +80,9 @@ GCNHazardRecognizer::getHazardType(SUnit *SU, int Stalls) {
   if (isDivFMas(MI->getOpcode()) && checkDivFMasHazards(MI) > 0)
     return NoopHazard;
 
+  if (isRWLane(MI->getOpcode()) && checkRWLaneHazards(MI) > 0)
+    return NoopHazard;
+
   if (isSGetReg(MI->getOpcode()) && checkGetRegHazards(MI) > 0)
     return NoopHazard;
 
@@ -105,6 +112,9 @@ unsigned GCNHazardRecognizer::PreEmitNoops(MachineInstr *MI) {
     if (isDivFMas(MI->getOpcode()))
       WaitStates = std::max(WaitStates, checkDivFMasHazards(MI));
 
+    if (isRWLane(MI->getOpcode()))
+      WaitStates = std::max(WaitStates, checkRWLaneHazards(MI));
+
     return WaitStates;
   }
 
@@ -438,3 +448,25 @@ int GCNHazardRecognizer::checkVALUHazards(MachineInstr *VALU) {
   }
   return WaitStatesNeeded;
 }
+
+int GCNHazardRecognizer::checkRWLaneHazards(MachineInstr *RWLane) {
+  const SIInstrInfo *TII = ST.getInstrInfo();
+  const SIRegisterInfo *TRI = ST.getRegisterInfo();
+  const MachineRegisterInfo &MRI =
+      RWLane->getParent()->getParent()->getRegInfo();
+
+  const MachineOperand *LaneSelectOp =
+      TII->getNamedOperand(*RWLane, AMDGPU::OpName::src1);
+
+  if (!LaneSelectOp->isReg() || !TRI->isSGPRReg(MRI, LaneSelectOp->getReg()))
+    return 0;
+
+  unsigned LaneSelectReg = LaneSelectOp->getReg();
+  auto IsHazardFn = [TII] (MachineInstr *MI) {
+    return TII->isVALU(*MI);
+  };
+
+  const int RWLaneWaitStates = 4;
+  int WaitStatesSince = getWaitStatesSinceDef(LaneSelectReg, IsHazardFn);
+  return RWLaneWaitStates - WaitStatesSince;
+}
diff --git a/llvm/lib/Target/AMDGPU/GCNHazardRecognizer.h b/llvm/lib/Target/AMDGPU/GCNHazardRecognizer.h
index f0882d0..8b9ea6a 100644
--- a/llvm/lib/Target/AMDGPU/GCNHazardRecognizer.h
+++ b/llvm/lib/Target/AMDGPU/GCNHazardRecognizer.h
@@ -50,6 +50,7 @@ class GCNHazardRecognizer final : public ScheduleHazardRecognizer {
   int checkSetRegHazards(MachineInstr *SetRegInstr);
   int createsVALUHazard(const MachineInstr &MI);
   int checkVALUHazards(MachineInstr *VALU);
+  int checkRWLaneHazards(MachineInstr *RWLane);
 public:
   GCNHazardRecognizer(const MachineFunction &MF);
   // We can only issue one instruction per cycle.
diff --git a/llvm/test/CodeGen/MIR/AMDGPU/inserted-wait-states.mir b/llvm/test/CodeGen/MIR/AMDGPU/inserted-wait-states.mir
index 8936e6f..4d466d4 100644
--- a/llvm/test/CodeGen/MIR/AMDGPU/inserted-wait-states.mir
+++ b/llvm/test/CodeGen/MIR/AMDGPU/inserted-wait-states.mir
@@ -7,6 +7,7 @@
   define void @s_getreg() { ret void }
   define void @s_setreg() { ret void }
   define void @vmem_gt_8dw_store() { ret void }
+  define void @readwrite_lane() { ret void }
 ...
 ---
 # GCN-LABEL: name: div_fmas
@@ -234,3 +235,68 @@ body: |
     S_ENDPGM
 
 ...
+
+...
+---
+
+# GCN-LABEL: name: readwrite_lane
+
+# GCN-LABEL: bb.0:
+# GCN: V_ADD_I32
+# GCN: S_NOP
+# GCN: S_NOP
+# GCN: S_NOP
+# GCN: S_NOP
+# GCN: V_READLANE_B32
+
+# GCN-LABEL: bb.1:
+# GCN: V_ADD_I32
+# GCN: S_NOP
+# GCN: S_NOP
+# GCN: S_NOP
+# GCN: S_NOP
+# GCN: V_WRITELANE_B32
+
+# GCN-LABEL: bb.2:
+# GCN: V_ADD_I32
+# GCN: S_NOP
+# GCN: S_NOP
+# GCN: S_NOP
+# GCN: S_NOP
+# GCN: V_READLANE_B32
+
+# GCN-LABEL: bb.3:
+# GCN: V_ADD_I32
+# GCN: S_NOP
+# GCN: S_NOP
+# GCN: S_NOP
+# GCN: S_NOP
+# GCN: V_WRITELANE_B32
+
+name: readwrite_lane
+
+body: |
+  bb.0:
+    successors: %bb.1
+    %vgpr0,%sgpr0_sgpr1 = V_ADD_I32_e64 %vgpr1, %vgpr2, implicit %vcc, implicit %exec
+    %sgpr4 = V_READLANE_B32 %vgpr4, %sgpr0
+    S_BRANCH %bb.1
+
+  bb.1:
+    successors: %bb.2
+    %vgpr0,%sgpr0_sgpr1 = V_ADD_I32_e64 %vgpr1, %vgpr2, implicit %vcc, implicit %exec
+    %vgpr4 = V_WRITELANE_B32 %sgpr0, %sgpr0
+    S_BRANCH %bb.2
+
+  bb.2:
+    successors: %bb.3
+    %vgpr0,implicit %vcc = V_ADD_I32_e32 %vgpr1, %vgpr2, implicit %vcc, implicit %exec
+    %sgpr4 = V_READLANE_B32 %vgpr4, %vcc_lo
+    S_BRANCH %bb.3
+
+  bb.3:
+    %vgpr0,implicit %vcc = V_ADD_I32_e32 %vgpr1, %vgpr2, implicit %vcc, implicit %exec
+    %vgpr4 = V_WRITELANE_B32 %sgpr4, %vcc_lo
+    S_ENDPGM
+
+...
-- 
2.7.4