From b47cb20619064771276a6798dec43aed22aa963a Mon Sep 17 00:00:00 2001
From: lioncash <mathew1800@gmail.com>
Date: Thu, 4 Nov 2021 16:25:27 -0400
Subject: [PATCH] OpcodeDispatcher: Implement handling for BLSI

Now all that remains is handling for BLSMSK and BLSR
---
 .../Interface/Core/OpcodeDispatcher.cpp       | 64 +++++++++++++++++++
 .../Source/Interface/Core/OpcodeDispatcher.h  |  1 +
 .../Interface/Core/X86Tables/VEXTables.cpp    |  2 +-
 unittests/ASM/VEX/blsi.asm                    | 34 ++++++++++
 4 files changed, 100 insertions(+), 1 deletion(-)
 create mode 100644 unittests/ASM/VEX/blsi.asm

diff --git a/External/FEXCore/Source/Interface/Core/OpcodeDispatcher.cpp b/External/FEXCore/Source/Interface/Core/OpcodeDispatcher.cpp
index 01b174599..a1aacd267 100644
--- a/External/FEXCore/Source/Interface/Core/OpcodeDispatcher.cpp
+++ b/External/FEXCore/Source/Interface/Core/OpcodeDispatcher.cpp
@@ -2204,6 +2204,63 @@ void OpDispatchBuilder::BEXTRBMIOp(OpcodeArgs) {
   SetRFLAG<X86State::RFLAG_ZF_LOC>(ZeroOp);
 }
 
+void OpDispatchBuilder::BLSIBMIOp(OpcodeArgs) {
+  // Equivalent to performing: SRC & -SRC
+
+  auto* Src = LoadSource(GPRClass, Op, Op->Src[0], Op->Flags, -1);
+  auto NegatedSrc = _Neg(Src);
+  auto Result = _And(Src, NegatedSrc);
+
+  // ...and we're done. Painless!
+  StoreResult(GPRClass, Op, Result, -1);
+
+  // Now for the flags:
+  //
+  // Only CF, SF, ZF and OF are defined as being updated
+  // CF is cleared if Src is zero, otherwise it's set.
+  // SF is set to the value of the most significant operand bit of Result.
+  // OF is always cleared
+  // ZF is set, as usual, if Result is zero or not.
+  //
+  // AF and PF are documented as being in an undefined state after
+  // a BLSI operation, however, we choose to reliably clear them.
+
+  auto Zero = _Constant(0);
+  auto One = _Constant(1);
+
+  SetRFLAG<X86State::RFLAG_OF_LOC>(Zero);
+  SetRFLAG<X86State::RFLAG_AF_LOC>(Zero);
+  if (CTX->Config.ABINoPF) {
+    _InvalidateFlags(1UL << X86State::RFLAG_PF_LOC);
+  } else {
+    SetRFLAG<X86State::RFLAG_PF_LOC>(Zero);
+  }
+
+  // ZF
+  {
+    auto ZFOp = _Select(IR::COND_EQ,
+                        Result, Zero,
+                        One, Zero);
+    SetRFLAG<X86State::RFLAG_ZF_LOC>(ZFOp);
+  }
+
+  // CF
+  {
+    auto CFOp = _Select(IR::COND_EQ,
+                        Src, Zero,
+                        Zero, One);
+    SetRFLAG<X86State::RFLAG_CF_LOC>(CFOp);
+  }
+
+  // SF
+  {
+    auto SignBit = _Constant((GetSrcSize(Op) * 8) - 1);
+    auto SFOp = _Lshr(Result, SignBit);
+
+    SetRFLAG<X86State::RFLAG_SF_LOC>(SFOp);
+  }
+}
+
 void OpDispatchBuilder::RCROp1Bit(OpcodeArgs) {
   OrderedNode *Dest = LoadSource(GPRClass, Op, Op->Dest, Op->Flags, -1);
   auto Size = GetSrcSize(Op) * 8;
@@ -5876,6 +5933,12 @@ constexpr uint16_t PF_F2 = 3;
   };
 #undef OPD
 
+  #define OPD(group, pp, opcode) (((group - X86Tables::InstType::TYPE_VEX_GROUP_12) << 4) | (pp << 3) | (opcode))
+    const std::vector<std::tuple<uint8_t, uint8_t, X86Tables::OpDispatchPtr>> VEXGroupTable = {
+      {OPD(X86Tables::InstType::TYPE_VEX_GROUP_17, 0, 0b011), 1, &OpDispatchBuilder::BLSIBMIOp},
+    };
+  #undef OPD
+
   const std::vector<std::tuple<uint8_t, uint8_t, FEXCore::X86Tables::OpDispatchPtr>> EVEXTable = {
     {0x10, 2, &OpDispatchBuilder::UnimplementedOp},
     {0x59, 1, &OpDispatchBuilder::UnimplementedOp},
@@ -5936,6 +5999,7 @@ constexpr uint16_t PF_F2 = 3;
   InstallToTable(FEXCore::X86Tables::H0F38TableOps, H0F38Table);
   InstallToTable(FEXCore::X86Tables::H0F3ATableOps, H0F3ATable);
   InstallToTable(FEXCore::X86Tables::VEXTableOps, VEXTable);
+  InstallToTable(FEXCore::X86Tables::VEXTableGroupOps, VEXGroupTable);
   InstallToTable(FEXCore::X86Tables::EVEXTableOps, EVEXTable);
 }
 
diff --git a/External/FEXCore/Source/Interface/Core/OpcodeDispatcher.h b/External/FEXCore/Source/Interface/Core/OpcodeDispatcher.h
index a440b707e..331d46917 100644
--- a/External/FEXCore/Source/Interface/Core/OpcodeDispatcher.h
+++ b/External/FEXCore/Source/Interface/Core/OpcodeDispatcher.h
@@ -327,6 +327,7 @@ public:
   // BMI Ops
   void ANDNBMIOp(OpcodeArgs);
   void BEXTRBMIOp(OpcodeArgs);
+  void BLSIBMIOp(OpcodeArgs);
 
   // X87 Ops
   template<size_t width>
diff --git a/External/FEXCore/Source/Interface/Core/X86Tables/VEXTables.cpp b/External/FEXCore/Source/Interface/Core/X86Tables/VEXTables.cpp
index 8b5de202a..e9ae5b6a1 100644
--- a/External/FEXCore/Source/Interface/Core/X86Tables/VEXTables.cpp
+++ b/External/FEXCore/Source/Interface/Core/X86Tables/VEXTables.cpp
@@ -505,7 +505,7 @@ void InitializeVEXTables() {
 
     {OPD(TYPE_VEX_GROUP_17, 0, 0b001), 1, X86InstInfo{"BLSR",     TYPE_UNDEC, FLAGS_MODRM, 0, nullptr}},
     {OPD(TYPE_VEX_GROUP_17, 0, 0b010), 1, X86InstInfo{"BLSMSK",   TYPE_UNDEC, FLAGS_MODRM, 0, nullptr}},
-    {OPD(TYPE_VEX_GROUP_17, 0, 0b011), 1, X86InstInfo{"BLSI",     TYPE_UNDEC, FLAGS_MODRM, 0, nullptr}},
+    {OPD(TYPE_VEX_GROUP_17, 0, 0b011), 1, X86InstInfo{"BLSI",     TYPE_INST, FLAGS_MODRM | FLAGS_VEX_DST, 0, nullptr}},
   };
 #undef OPD
 
diff --git a/unittests/ASM/VEX/blsi.asm b/unittests/ASM/VEX/blsi.asm
new file mode 100644
index 000000000..ccddd8802
--- /dev/null
+++ b/unittests/ASM/VEX/blsi.asm
@@ -0,0 +1,34 @@
+%ifdef CONFIG
+{
+  "RegData": {
+      "RAX": "1",
+      "RBX": "0xFF00000000000000",
+      "RCX": "0x0100000000000000",
+      "RDX": "1",
+      "RSI": "0xFF000000",
+      "RDI": "0x01000000"
+  }
+}
+%endif
+
+; Trivial test, this should result in 1.
+mov rax, 11
+blsi rax, rax
+
+; Results in the lowest set bit (bit 56) being extracted
+mov rbx, 0xFF00000000000000
+mov rcx, 0
+blsi rcx, rbx
+
+; Same tests but with 32-bit registers
+
+; Trivial test, this should result in 1.
+mov edx, 11
+blsi edx, edx
+
+; Results in the lowest set bit (bit 24) being extracted
+mov rsi, 0xFF000000
+mov rdi, 0
+blsi edi, esi
+
+hlt
\ No newline at end of file