Automerge: [Hexagon] Improve QFP Optimizer (#166647)

This patch enhances HexagonQFPOptimizer in multiple ways:

1. Refactor the code for better readability and maintainability.

2. Optimize vabs,vneg and vilog2 converts

   The three instruction mentioned can be optimized like below:

  ```v1.sf = v0.qf32
    v2.qf = vneg v1.sf```

  to

  ```v2.qf = vneg v0.qf32```

  This optimization eliminates one conversion and is applicable
  to both qf32 and qf16 types.

3. Enable vsub fusion with mixed arguments Previously, QFPOptimizer did
not fuse partial qfloat operands with vsub. This update allows selective
use of vsub_hf_mix, vsub_sf_mix, vsub_qf16_mix, and vsub_qf32_mix when
appropriate. It also enables QFP simplifications involving vector pair
subregisters.

Example scenario in a machine basic block targeting Hexagon: ```v1.qf32
= ... // result of a vadd
      v2.sf   = v1.qf32
      v3.qf32 = vmpy(v2.sf, v2.sf)```

4. Remove redundant conversions Under certain conditions, we previously
bailed out before removing qf-to-sf/hf conversions. This patch removes
that bailout, enabling more aggressive elimination of unnecessary
conversions.

5. Don't optimize equals feeding into multiply: Removing converts
feeding into multiply loses precision. This patch avoids optimizing
multiplies along with giving the users an option to enable this by a
flag.

Patch By: Fateme Hosseini

Co-authored-by: Kaushik Kulkarni <[email protected]>
Co-authored-by: Santanu Das <[email protected]>

Author: 3 people

llvm/lib/Target/Hexagon/HexagonInstrInfo.cpp

@@ -4753,6 +4753,19 @@ bool HexagonInstrInfo::getBundleNoShuf(const MachineInstr &MIB) const {
   return (Operand.isImm() && (Operand.getImm() & memShufDisabledMask) != 0);
 }
 
+bool HexagonInstrInfo::isQFPMul(const MachineInstr *MI) const {
+  return (MI->getOpcode() == Hexagon::V6_vmpy_qf16_hf ||
+          MI->getOpcode() == Hexagon::V6_vmpy_qf16_mix_hf ||
+          MI->getOpcode() == Hexagon::V6_vmpy_qf32_hf ||
+          MI->getOpcode() == Hexagon::V6_vmpy_qf32_mix_hf ||
+          MI->getOpcode() == Hexagon::V6_vmpy_qf32_sf ||
+          MI->getOpcode() == Hexagon::V6_vmpy_qf16_mix_hf ||
+          MI->getOpcode() == Hexagon::V6_vmpy_qf16 ||
+          MI->getOpcode() == Hexagon::V6_vmpy_qf32_mix_hf ||
+          MI->getOpcode() == Hexagon::V6_vmpy_qf32_qf16 ||
+          MI->getOpcode() == Hexagon::V6_vmpy_qf32);
+}
+
 // Addressing mode relations.
 short HexagonInstrInfo::changeAddrMode_abs_io(short Opc) const {
   return Opc >= 0 ? Hexagon::changeAddrMode_abs_io(Opc) : Opc;

llvm/lib/Target/Hexagon/HexagonInstrInfo.h

@@ -532,6 +532,7 @@ class HexagonInstrInfo : public HexagonGenInstrInfo {
   }
 
   MCInst getNop() const override;
+  bool isQFPMul(const MachineInstr *MF) const;
 };
 
 /// \brief Create RegSubRegPair from a register MachineOperand

llvm/lib/Target/Hexagon/HexagonQFPOptimizer.cpp

@@ -58,7 +58,7 @@
 // are PHI inst.
 //
 //===----------------------------------------------------------------------===//
-#include <unordered_set>
+
 #define HEXAGON_QFP_OPTIMIZER "QFP optimizer pass"
 
 #include "Hexagon.h"
@@ -86,6 +86,9 @@ using namespace llvm;
 cl::opt<bool>
     DisableQFOptimizer("disable-qfp-opt", cl::init(false),
                        cl::desc("Disable optimization of Qfloat operations."));
+cl::opt<bool> DisableQFOptForMul(
+    "disable-qfp-opt-mul", cl::init(true),
+    cl::desc("Disable optimization of Qfloat operations for multiply."));
 
 namespace {
 const std::map<unsigned short, unsigned short> QFPInstMap{
@@ -101,11 +104,21 @@ const std::map<unsigned short, unsigned short> QFPInstMap{
     {Hexagon::V6_vmpy_qf16_mix_hf, Hexagon::V6_vmpy_qf16},
     {Hexagon::V6_vmpy_qf32_hf, Hexagon::V6_vmpy_qf32_mix_hf},
     {Hexagon::V6_vmpy_qf32_mix_hf, Hexagon::V6_vmpy_qf32_qf16},
-    {Hexagon::V6_vmpy_qf32_sf, Hexagon::V6_vmpy_qf32}};
+    {Hexagon::V6_vmpy_qf32_sf, Hexagon::V6_vmpy_qf32},
+    {Hexagon::V6_vilog2_sf, Hexagon::V6_vilog2_qf32},
+    {Hexagon::V6_vilog2_hf, Hexagon::V6_vilog2_qf16},
+    {Hexagon::V6_vabs_qf32_sf, Hexagon::V6_vabs_qf32_qf32},
+    {Hexagon::V6_vabs_qf16_hf, Hexagon::V6_vabs_qf16_qf16},
+    {Hexagon::V6_vneg_qf32_sf, Hexagon::V6_vneg_qf32_qf32},
+    {Hexagon::V6_vneg_qf16_hf, Hexagon::V6_vneg_qf16_qf16}};
 } // namespace
 
-namespace {
+namespace llvm {
+FunctionPass *createHexagonQFPOptimizer();
+void initializeHexagonQFPOptimizerPass(PassRegistry &);
+} // namespace llvm
 
+namespace {
 struct HexagonQFPOptimizer : public MachineFunctionPass {
 public:
   static char ID;
@@ -116,6 +129,10 @@ struct HexagonQFPOptimizer : public MachineFunctionPass {
 
   bool optimizeQfp(MachineInstr *MI, MachineBasicBlock *MBB);
 
+  bool optimizeQfpTwoOp(MachineInstr *MI, MachineBasicBlock *MBB);
+
+  bool optimizeQfpOneOp(MachineInstr *MI, MachineBasicBlock *MBB);
+
   StringRef getPassName() const override { return HEXAGON_QFP_OPTIMIZER; }
 
   void getAnalysisUsage(AnalysisUsage &AU) const override {
@@ -142,19 +159,69 @@ FunctionPass *llvm::createHexagonQFPOptimizer() {
 bool HexagonQFPOptimizer::optimizeQfp(MachineInstr *MI,
                                       MachineBasicBlock *MBB) {
 
-  // Early exit:
-  // - if instruction is invalid or has too few operands (QFP ops need 2 sources
-  // + 1 dest),
-  // - or does not have a transformation mapping.
-  if (MI->getNumOperands() < 3)
+  if (MI->getNumOperands() == 2)
+    return optimizeQfpOneOp(MI, MBB);
+  else if (MI->getNumOperands() == 3)
+    return optimizeQfpTwoOp(MI, MBB);
+  else
     return false;
+}
+
+bool HexagonQFPOptimizer::optimizeQfpOneOp(MachineInstr *MI,
+                                           MachineBasicBlock *MBB) {
+
+  unsigned Op0F = 0;
   auto It = QFPInstMap.find(MI->getOpcode());
   if (It == QFPInstMap.end())
     return false;
+
   unsigned short InstTy = It->second;
+  // Get the reachind defs of MI
+  MachineInstr *DefMI = MRI->getVRegDef(MI->getOperand(1).getReg());
+  MachineOperand &Res = MI->getOperand(0);
+  if (!Res.isReg())
+    return false;
+
+  LLVM_DEBUG(dbgs() << "\n[Reaching Defs of operands]: "; DefMI->dump());
+  MachineInstr *ReachDefDef = nullptr;
+
+  // Get the reaching def of the reaching def to check for W reg def
+  if (DefMI->getNumOperands() > 1 && DefMI->getOperand(1).isReg() &&
+      DefMI->getOperand(1).getReg().isVirtual())
+    ReachDefDef = MRI->getVRegDef(DefMI->getOperand(1).getReg());
+  unsigned ReachDefOp = DefMI->getOpcode();
+  MachineInstrBuilder MIB;
+
+  // Check if the reaching def is a conversion
+  if (ReachDefOp == Hexagon::V6_vconv_sf_qf32 ||
+      ReachDefOp == Hexagon::V6_vconv_hf_qf16) {
+
+    // Return if the reaching def of reaching def is W type
+    if (ReachDefDef && MRI->getRegClass(ReachDefDef->getOperand(0).getReg()) ==
+                           &Hexagon::HvxWRRegClass)
+      return false;
+
+    // Analyze the use operands of the conversion to get their KILL status
+    MachineOperand &SrcOp = DefMI->getOperand(1);
+    Op0F = getKillRegState(SrcOp.isKill());
+    SrcOp.setIsKill(false);
+    MIB = BuildMI(*MBB, MI, MI->getDebugLoc(), HII->get(InstTy), Res.getReg())
+              .addReg(SrcOp.getReg(), Op0F, SrcOp.getSubReg());
+    LLVM_DEBUG(dbgs() << "\n[Inserting]: "; MIB.getInstr()->dump());
+    return true;
+  }
+  return false;
+}
+
+bool HexagonQFPOptimizer::optimizeQfpTwoOp(MachineInstr *MI,
+                                           MachineBasicBlock *MBB) {
 
   unsigned Op0F = 0;
   unsigned Op1F = 0;
+  auto It = QFPInstMap.find(MI->getOpcode());
+  if (It == QFPInstMap.end())
+    return false;
+  unsigned short InstTy = It->second;
   // Get the reaching defs of MI, DefMI1 and DefMI2
   MachineInstr *DefMI1 = nullptr;
   MachineInstr *DefMI2 = nullptr;
@@ -167,6 +234,9 @@ bool HexagonQFPOptimizer::optimizeQfp(MachineInstr *MI,
     return false;
 
   MachineOperand &Res = MI->getOperand(0);
+  if (!Res.isReg())
+    return false;
+
   MachineInstr *Inst1 = nullptr;
   MachineInstr *Inst2 = nullptr;
   LLVM_DEBUG(dbgs() << "\n[Reaching Defs of operands]: "; DefMI1->dump();
@@ -185,7 +255,8 @@ bool HexagonQFPOptimizer::optimizeQfp(MachineInstr *MI,
   unsigned Def2OP = DefMI2->getOpcode();
 
   MachineInstrBuilder MIB;
-  // Case 1: Both reaching defs of MI are qf to sf/hf conversions
+
+  // Check if the both the reaching defs of MI are qf to sf/hf conversions
   if ((Def1OP == Hexagon::V6_vconv_sf_qf32 &&
        Def2OP == Hexagon::V6_vconv_sf_qf32) ||
       (Def1OP == Hexagon::V6_vconv_hf_qf16 &&
@@ -226,7 +297,7 @@ bool HexagonQFPOptimizer::optimizeQfp(MachineInstr *MI,
     LLVM_DEBUG(dbgs() << "\n[Inserting]: "; MIB.getInstr()->dump());
     return true;
 
-    // Case 2: Left operand is conversion to sf/hf
+    // Check if left operand's reaching def is a conversion to sf/hf
   } else if (((Def1OP == Hexagon::V6_vconv_sf_qf32 &&
                Def2OP != Hexagon::V6_vconv_sf_qf32) ||
               (Def1OP == Hexagon::V6_vconv_hf_qf16 &&
@@ -250,21 +321,14 @@ bool HexagonQFPOptimizer::optimizeQfp(MachineInstr *MI,
     LLVM_DEBUG(dbgs() << "\n[Inserting]: "; MIB.getInstr()->dump());
     return true;
 
-    // Case 2: Left operand is conversion to sf/hf
+    // Check if right operand's reaching def is a conversion to sf/hf
   } else if (((Def1OP != Hexagon::V6_vconv_sf_qf32 &&
                Def2OP == Hexagon::V6_vconv_sf_qf32) ||
               (Def1OP != Hexagon::V6_vconv_hf_qf16 &&
                Def2OP == Hexagon::V6_vconv_hf_qf16)) &&
              !DefMI1->isPHI() &&
              (MI->getOpcode() != Hexagon::V6_vmpy_qf32_sf)) {
     // The second operand of original instruction is converted.
-    // In "mix" instructions, "qf" operand is always the first operand.
-
-    // Caveat: vsub is not commutative w.r.t operands.
-    if (InstTy == Hexagon::V6_vsub_qf16_mix ||
-        InstTy == Hexagon::V6_vsub_qf32_mix)
-      return false;
-
     if (Inst2 && MRI->getRegClass(Inst2->getOperand(0).getReg()) ==
                      &Hexagon::HvxWRRegClass)
       return false;
@@ -275,10 +339,26 @@ bool HexagonQFPOptimizer::optimizeQfp(MachineInstr *MI,
     Op1F = getKillRegState(Src2.isKill());
     Src2.setIsKill(false);
     Op0F = getKillRegState(Src1.isKill());
-    MIB = BuildMI(*MBB, MI, MI->getDebugLoc(), HII->get(InstTy), Res.getReg())
-              .addReg(Src2.getReg(), Op1F,
-                      Src2.getSubReg()) // Notice the operands are flipped.
-              .addReg(Src1.getReg(), Op0F, Src1.getSubReg());
+    if (InstTy == Hexagon::V6_vsub_qf16_mix ||
+        InstTy == Hexagon::V6_vsub_qf32_mix) {
+      if (!HST->useHVXV81Ops())
+        // vsub_(hf|sf)_mix insts are only avlbl on hvx81+
+        return false;
+      // vsub is not commutative w.r.t. operands -> treat it as a special case
+      // to choose the correct mix instruction.
+      if (Def2OP == Hexagon::V6_vconv_sf_qf32)
+        InstTy = Hexagon::V6_vsub_sf_mix;
+      else if (Def2OP == Hexagon::V6_vconv_hf_qf16)
+        InstTy = Hexagon::V6_vsub_hf_mix;
+      MIB = BuildMI(*MBB, MI, MI->getDebugLoc(), HII->get(InstTy), Res.getReg())
+                .addReg(Src1.getReg(), Op0F, Src1.getSubReg())
+                .addReg(Src2.getReg(), Op1F, Src2.getSubReg());
+    } else {
+      MIB = BuildMI(*MBB, MI, MI->getDebugLoc(), HII->get(InstTy), Res.getReg())
+                .addReg(Src2.getReg(), Op1F,
+                        Src2.getSubReg()) // Notice the operands are flipped.
+                .addReg(Src1.getReg(), Op0F, Src1.getSubReg());
+    }
     LLVM_DEBUG(dbgs() << "\n[Inserting]: "; MIB.getInstr()->dump());
     return true;
   }
@@ -309,15 +389,18 @@ bool HexagonQFPOptimizer::runOnMachineFunction(MachineFunction &MF) {
     while (MII != MBBI->instr_end()) {
       MachineInstr *MI = &*MII;
       ++MII; // As MI might be removed.
-
-      if (QFPInstMap.count(MI->getOpcode()) &&
-          MI->getOpcode() != Hexagon::V6_vconv_sf_qf32 &&
-          MI->getOpcode() != Hexagon::V6_vconv_hf_qf16) {
-        LLVM_DEBUG(dbgs() << "\n###Analyzing for removal: "; MI->dump());
-        if (optimizeQfp(MI, MBB)) {
-          MI->eraseFromParent();
-          LLVM_DEBUG(dbgs() << "\t....Removing....");
-          Changed = true;
+      if (QFPInstMap.count(MI->getOpcode())) {
+        auto OpC = MI->getOpcode();
+        if (DisableQFOptForMul && HII->isQFPMul(MI))
+          continue;
+        if (OpC != Hexagon::V6_vconv_sf_qf32 &&
+            OpC != Hexagon::V6_vconv_hf_qf16) {
+          LLVM_DEBUG(dbgs() << "\n###Analyzing for removal: "; MI->dump());
+          if (optimizeQfp(MI, MBB)) {
+            MI->eraseFromParent();
+            LLVM_DEBUG(dbgs() << "\t....Removing....");
+            Changed = true;
+          }
         }
       }
     }