Дормидонтов Егор. Лабораторная работа 3. Backend. Вариант 4.

llvm/compiler-course/backend/dormidontov_egor_lab3/CMakeLists.txt

@@ -0,0 +1,17 @@
+set(Title "FmsubOptPass")
+set(Student "Dormidontov_Egor")
+set(Group "FIIT2")
+set(TARGET_NAME "${Title}_${Student}_${Group}_BACKEND")
+
+file(GLOB_RECURSE SOURCES *.cpp *.h *.hpp)
+
+add_llvm_pass_plugin(${TARGET_NAME}
+    ${SOURCES}
+    DEPENDS
+    intrinsics_gen
+    X86
+    BUILDTREE_ONLY
+)
+
+target_include_directories(${TARGET_NAME} PUBLIC ${PATH_TO_X86})
+set(LLVM_TEST_DEPENDS ${TARGET_NAME} ${LLVM_TEST_DEPENDS} PARENT_SCOPE)

llvm/compiler-course/backend/dormidontov_egor_lab3/lab3.cpp

@@ -0,0 +1,164 @@
+#include "X86.h"
+#include "X86InstrInfo.h"
+#include "X86Subtarget.h"
+#include "llvm/CodeGen/MachineFunctionPass.h"
+#include "llvm/CodeGen/MachineInstrBuilder.h"
+#include "llvm/Support/raw_ostream.h"
+
+using namespace llvm;
+
+namespace {
+
+struct FmaPattern {
+  unsigned MulOp, SubOp, AddOp, XorOp, VFMSUB213, VFNMADD213;
+  const char *Name;
+};
+
+static const FmaPattern Patterns[] = {
+    // Packed single-precision (SSE + AVX)
+    {X86::MULPSrr, X86::SUBPSrr, X86::ADDPSrr, X86::XORPSrr, X86::VFMSUB213PSr,
+     X86::VFNMADD213PSr, "PS"},
+    {X86::VMULPSrr, X86::VSUBPSrr, X86::VADDPSrr, X86::XORPSrr,
+     X86::VFMSUB213PSr, X86::VFNMADD213PSr, "VPS"},
+
+    // Packed double-precision (SSE2 + AVX)
+    {X86::MULPDrr, X86::SUBPDrr, X86::ADDPDrr, X86::XORPDrr, X86::VFMSUB213PDr,
+     X86::VFNMADD213PDr, "PD"},
+    {X86::VMULPDrr, X86::VSUBPDrr, X86::VADDPDrr, X86::XORPDrr,
+     X86::VFMSUB213PDr, X86::VFNMADD213PDr, "VPD"},
+
+    // Scalar single-precision (SSE + AVX)
+    {X86::MULSSrr, X86::SUBSSrr, X86::ADDSSrr, X86::XORPSrr, X86::VFMSUB213SSr,
+     X86::VFNMADD213SSr, "SS"},
+    {X86::VMULSSrr, X86::VSUBSSrr, X86::VADDSSrr, X86::XORPSrr,
+     X86::VFMSUB213SSr, X86::VFNMADD213SSr, "VSS"},
+
+    // Scalar double-precision (SSE2 + AVX)
+    {X86::MULSDrr, X86::SUBSDrr, X86::ADDSDrr, X86::XORPDrr, X86::VFMSUB213SDr,
+     X86::VFNMADD213SDr, "SD"},
+    {X86::VMULSDrr, X86::VSUBSDrr, X86::VADDSDrr, X86::XORPDrr,
+     X86::VFMSUB213SDr, X86::VFNMADD213SDr, "VSD"},
+};
+
+class FmsubOptPass : public MachineFunctionPass {
+public:
+  static char ID;
+  FmsubOptPass() : MachineFunctionPass(ID) {}
+  bool runOnMachineFunction(MachineFunction &MF) override;
+};
+
+char FmsubOptPass::ID = 0;
+
+bool FmsubOptPass::runOnMachineFunction(MachineFunction &MF) {
+  bool Changed = false;
+  const X86InstrInfo *TII =
+      static_cast<const X86InstrInfo *>(MF.getSubtarget().getInstrInfo());
+
+  for (auto &MBB : MF) {
+    for (auto MI = MBB.begin(); MI != MBB.end();) {
+      bool PatternMatched = false;
+
+      for (const auto &P : Patterns) {
+        // (a * b) - c → VFMSUB213
+        if (MI->getOpcode() == P.SubOp && MI != MBB.begin()) {
+          auto MulMI = std::prev(MI);
+          if (MulMI->getOpcode() == P.MulOp) {
+            auto &MulDst = MulMI->getOperand(0);
+            auto &MulA = MulMI->getOperand(1);
+            auto &MulB = MulMI->getOperand(2);
+            auto &SubDst = MI->getOperand(0);
+            auto &SubA = MI->getOperand(1);
+            auto &SubB = MI->getOperand(2);
+
+            if (SubA.isReg() && SubA.getReg() == MulDst.getReg() &&
+                SubDst.isReg()) {
+              DebugLoc DL = MI->getDebugLoc();
+              BuildMI(MBB, MI, DL, TII->get(P.VFMSUB213), SubDst.getReg())
+                  .add(MulA)
+                  .add(MulB)
+                  .add(SubB);
+              MI = MBB.erase(MI);
+              MBB.erase(MulMI);
+              Changed = true;
+              PatternMatched = true;
+              break;
+            }
+          }
+        }
+        // c - (a * b) → VFNMADD213
+        if (MI->getOpcode() == P.SubOp && MI != MBB.begin()) {
+          auto MulMI = std::prev(MI);
+          if (MulMI->getOpcode() == P.MulOp) {
+            auto &MulDst = MulMI->getOperand(0);
+            auto &MulA = MulMI->getOperand(1);
+            auto &MulB = MulMI->getOperand(2);
+            auto &SubDst = MI->getOperand(0);
+            auto &SubA = MI->getOperand(1);
+            auto &SubB = MI->getOperand(2);
+
+            if (SubB.isReg() && SubB.getReg() == MulDst.getReg() &&
+                SubDst.isReg()) {
+              DebugLoc DL = MI->getDebugLoc();
+              BuildMI(MBB, MI, DL, TII->get(P.VFNMADD213), SubDst.getReg())
+                  .add(MulA)
+                  .add(MulB)
+                  .add(SubA);
+              MI = MBB.erase(MI);
+              MBB.erase(MulMI);
+              Changed = true;
+              PatternMatched = true;
+              break;
+            }
+          }
+        }
+        // -(a * b) + c → VFNMADD213 (через XOR/NEG)
+        if (MI->getOpcode() == P.AddOp && MI != MBB.begin()) {
+          auto XorMI = std::prev(MI);
+          if (XorMI->getOpcode() == P.XorOp && XorMI != MBB.begin()) {
+            auto MulMI = std::prev(XorMI);
+            if (MulMI->getOpcode() == P.MulOp) {
+              auto &MulDst = MulMI->getOperand(0);
+              auto &MulA = MulMI->getOperand(1);
+              auto &MulB = MulMI->getOperand(2);
+
+              auto &XorDst = XorMI->getOperand(0);
+
+              auto &AddDst = MI->getOperand(0);
+              auto &AddA = MI->getOperand(1);
+              auto &AddB = MI->getOperand(2);
+              if (XorDst.isReg() && MulDst.isReg() &&
+                  XorDst.getReg() == MulDst.getReg() &&
+                  ((AddA.isReg() && AddA.getReg() == XorDst.getReg()) ||
+                   (AddB.isReg() && AddB.getReg() == XorDst.getReg()))) {
+                Register cReg = (AddA.getReg() == XorDst.getReg())
+                                    ? AddB.getReg()
+                                    : AddA.getReg();
+                DebugLoc DL = MI->getDebugLoc();
+                BuildMI(MBB, MI, DL, TII->get(P.VFNMADD213), AddDst.getReg())
+                    .add(MulA)
+                    .add(MulB)
+                    .addReg(cReg);
+                MI = MBB.erase(MI);
+                MBB.erase(XorMI);
+                MBB.erase(MulMI);
+                Changed = true;
+                PatternMatched = true;
+                break;
+              }
+            }
+          }
+        }
+      }
+      if (!PatternMatched)
+        ++MI;
+    }
+  }
+  return Changed;
+}
+
+} // namespace
+
+static RegisterPass<FmsubOptPass>
+    X("fmsub-opt",
+      "Fused multiply-subtract/add for X86 (VFMSUB/VFNMADD) (SSE+AVX+FMA)",
+      false, false);