From 51911dfbbe15ca7aa20a8dd0afd30f4395061417 Mon Sep 17 00:00:00 2001
From: Ryan Houdek <Sonicadvance1@gmail.com>
Date: Fri, 21 Jun 2024 05:03:30 -0700
Subject: [PATCH] AVX128: Implement support for VPCLMULQDQ

This is just the 128-bit version twice.
---
 .../Interface/Core/OpcodeDispatcher.cpp       |  4 ++-
 .../Source/Interface/Core/OpcodeDispatcher.h  |  2 ++
 .../Core/OpcodeDispatcher/AVX_128.cpp         | 29 +++++++++++++++++++
 3 files changed, 34 insertions(+), 1 deletion(-)

diff --git a/FEXCore/Source/Interface/Core/OpcodeDispatcher.cpp b/FEXCore/Source/Interface/Core/OpcodeDispatcher.cpp
index 7c4b4285ea..4ee31213ac 100644
--- a/FEXCore/Source/Interface/Core/OpcodeDispatcher.cpp
+++ b/FEXCore/Source/Interface/Core/OpcodeDispatcher.cpp
@@ -5415,13 +5415,15 @@ void OpDispatchBuilder::InstallHostSpecificOpcodeHandlers() {
   if (CTX->HostFeatures.SupportsSVE256) {
     InstallToTable(FEXCore::X86Tables::VEXTableOps, AVXTable);
     InstallToTable(FEXCore::X86Tables::VEXTableGroupOps, VEXTableGroupOps);
+    if (CTX->HostFeatures.SupportsPMULL_128Bit) {
+      InstallToTable(FEXCore::X86Tables::VEXTableOps, VEX_PCLMUL);
+    }
   } else if (CTX->HostFeatures.SupportsAVX) {
     InstallAVX128Handlers();
   }
 
   if (CTX->HostFeatures.SupportsPMULL_128Bit) {
     InstallToTable(FEXCore::X86Tables::H0F3ATableOps, H0F3A_PCLMUL);
-    InstallToTable(FEXCore::X86Tables::VEXTableOps, VEX_PCLMUL);
   }
   Initialized = true;
 }
diff --git a/FEXCore/Source/Interface/Core/OpcodeDispatcher.h b/FEXCore/Source/Interface/Core/OpcodeDispatcher.h
index e862ee1d62..687aad13e5 100644
--- a/FEXCore/Source/Interface/Core/OpcodeDispatcher.h
+++ b/FEXCore/Source/Interface/Core/OpcodeDispatcher.h
@@ -1193,6 +1193,8 @@ class OpDispatchBuilder final : public IREmitter {
 
   void AVX128_VPERMD(OpcodeArgs);
 
+  void AVX128_VPCLMULQDQ(OpcodeArgs);
+
   // End of AVX 128-bit implementation
   void InvalidOp(OpcodeArgs);
 
diff --git a/FEXCore/Source/Interface/Core/OpcodeDispatcher/AVX_128.cpp b/FEXCore/Source/Interface/Core/OpcodeDispatcher/AVX_128.cpp
index e204a90781..f41d640933 100644
--- a/FEXCore/Source/Interface/Core/OpcodeDispatcher/AVX_128.cpp
+++ b/FEXCore/Source/Interface/Core/OpcodeDispatcher/AVX_128.cpp
@@ -402,6 +402,12 @@ void OpDispatchBuilder::InstallAVX128Handlers() {
   };
 #undef OPD
 
+#define OPD(map_select, pp, opcode) (((map_select - 1) << 10) | (pp << 8) | (opcode))
+  constexpr std::tuple<uint16_t, uint8_t, FEXCore::X86Tables::OpDispatchPtr> VEX128_PCLMUL[] = {
+    {OPD(3, 0b01, 0x44), 1, &OpDispatchBuilder::AVX128_VPCLMULQDQ},
+  };
+#undef OPD
+
   auto InstallToTable = [](auto& FinalTable, auto& LocalTable) {
     for (auto Op : LocalTable) {
       auto OpNum = std::get<0>(Op);
@@ -415,6 +421,10 @@ void OpDispatchBuilder::InstallAVX128Handlers() {
 
   InstallToTable(FEXCore::X86Tables::VEXTableOps, AVX128Table);
   InstallToTable(FEXCore::X86Tables::VEXTableGroupOps, VEX128TableGroupOps);
+  if (CTX->HostFeatures.SupportsPMULL_128Bit) {
+    InstallToTable(FEXCore::X86Tables::VEXTableOps, VEX128_PCLMUL);
+  }
+
   SaveAVXStateFunc = &OpDispatchBuilder::AVX128_SaveAVXState;
   RestoreAVXStateFunc = &OpDispatchBuilder::AVX128_RestoreAVXState;
   DefaultAVXStateFunc = &OpDispatchBuilder::AVX128_DefaultAVXState;
@@ -2862,4 +2872,23 @@ void OpDispatchBuilder::AVX128_VPERMD(OpcodeArgs) {
   AVX128_StoreResult_WithOpSize(Op, Op->Dest, Result);
 }
 
+void OpDispatchBuilder::AVX128_VPCLMULQDQ(OpcodeArgs) {
+  LOGMAN_THROW_A_FMT(Op->Src[2].IsLiteral(), "Selector needs to be literal here");
+
+  const auto Size = GetDstSize(Op);
+  const auto Is128Bit = Size == Core::CPUState::XMM_SSE_REG_SIZE;
+
+  auto Src1 = AVX128_LoadSource_WithOpSize(Op, Op->Src[0], Op->Flags, !Is128Bit);
+  auto Src2 = AVX128_LoadSource_WithOpSize(Op, Op->Src[1], Op->Flags, !Is128Bit);
+
+  const auto Selector = static_cast<uint8_t>(Op->Src[2].Data.Literal.Value);
+
+  RefPair Result {};
+  Result.Low = _PCLMUL(OpSize::i128Bit, Src1.Low, Src2.Low, Selector);
+  if (!Is128Bit) {
+    Result.High = _PCLMUL(OpSize::i128Bit, Src1.High, Src2.High, Selector);
+  }
+  AVX128_StoreResult_WithOpSize(Op, Op->Dest, Result);
+}
+
 } // namespace FEXCore::IR