x87 refactor slowpath wip

FEX-Emu · Apr 24, 2024 · 4419e74 · 4419e74
1 parent 7ebba82
commit 4419e74
Show file tree

Hide file tree

Showing 5 changed files with 482 additions and 353 deletions.
diff --git a/FEXCore/Source/Interface/Core/OpcodeDispatcher.h b/FEXCore/Source/Interface/Core/OpcodeDispatcher.h
@@ -716,7 +716,7 @@ class OpDispatchBuilder final : public IREmitter {
   void VZEROOp(OpcodeArgs);
 
   // X87 Ops
-  OrderedNode* ReconstructFSW();
+  OrderedNode* ReconstructFSW(OrderedNode* T = nullptr);
   // Returns new x87 stack top from FSW.
   OrderedNode* ReconstructX87StateFromFSW(OrderedNode* FSW);
   template<size_t width>

diff --git a/FEXCore/Source/Interface/Core/OpcodeDispatcher/X87.cpp b/FEXCore/Source/Interface/Core/OpcodeDispatcher/X87.cpp
@@ -89,24 +89,6 @@ void OpDispatchBuilder::SetX87Top(OrderedNode* Value) {
   _StoreContext(1, GPRClass, Value, offsetof(FEXCore::Core::CPUState, flags) + FEXCore::X86State::X87FLAG_TOP_LOC);
 }
 
-OrderedNode* OpDispatchBuilder::ReconstructFSW() {
-  // We must construct the FSW from our various bits
-  OrderedNode* FSW = _Constant(0);
-  auto Top = GetX87Top();
-  FSW = _Bfi(OpSize::i64Bit, 3, 11, FSW, Top);
-
-  auto C0 = GetRFLAG(FEXCore::X86State::X87FLAG_C0_LOC);
-  auto C1 = GetRFLAG(FEXCore::X86State::X87FLAG_C1_LOC);
-  auto C2 = GetRFLAG(FEXCore::X86State::X87FLAG_C2_LOC);
-  auto C3 = GetRFLAG(FEXCore::X86State::X87FLAG_C3_LOC);
-
-  FSW = _Orlshl(OpSize::i64Bit, FSW, C0, 8);
-  FSW = _Orlshl(OpSize::i64Bit, FSW, C1, 9);
-  FSW = _Orlshl(OpSize::i64Bit, FSW, C2, 10);
-  FSW = _Orlshl(OpSize::i64Bit, FSW, C3, 14);
-  return FSW;
-}
-
 OrderedNode* OpDispatchBuilder::ReconstructX87StateFromFSW(OrderedNode* FSW) {
   auto Top = _Bfe(OpSize::i32Bit, 3, 11, FSW);
   SetX87Top(Top);
@@ -343,10 +325,6 @@ void OpDispatchBuilder::X87LDSW(OpcodeArgs) {
   ReconstructX87StateFromFSW(NewFSW);
 }
 
-void OpDispatchBuilder::X87FNSTSW(OpcodeArgs) {
-  StoreResult(GPRClass, Op, ReconstructFSW(), -1);
-}
-
 void OpDispatchBuilder::X87FNSAVE(OpcodeArgs) {
   // 14 bytes for 16bit
   // 2 Bytes : FCW

diff --git a/FEXCore/Source/Interface/Core/OpcodeDispatcher/X87New.cpp b/FEXCore/Source/Interface/Core/OpcodeDispatcher/X87New.cpp
@@ -541,13 +541,10 @@ void OpDispatchBuilder::F80SCALE(OpcodeArgs) {
 template<bool Inc>
 void OpDispatchBuilder::X87ModifySTP(OpcodeArgs) {
   CurrentHeader->HasX87 = true;
-  auto orig_top = GetX87Top();
   if (Inc) {
-    auto top = _And(OpSize::i32Bit, _Add(OpSize::i32Bit, orig_top, _Constant(1)), _Constant(7));
-    SetX87Top(top);
+    _IncStackTop();
   } else {
-    auto top = _And(OpSize::i32Bit, _Sub(OpSize::i32Bit, orig_top, _Constant(1)), _Constant(7));
-    SetX87Top(top);
+    _DecStackTop();
   }
 }
 
@@ -582,4 +579,39 @@ void OpDispatchBuilder::F80F2XM1(OpcodeArgs) {
   _F80F2XM1Stack();
 }
 
+// Operations dealing with loading and storing environment pieces
+
+
+// Reconstruct as a constant the Status Word of the FPU.
+// We only track stack top and each of the code conditions (C flags)
+// Top is 3 bits at bit 11.
+// C0 is 1 bit at bit 8.
+// C1 is 1 bit at bit 9.
+// C2 is 1 bit at bit 10.
+// C3 is 1 bit at bit 14.
+// Optionally we can pass a pre calculated value for Top, otherwise we calculate it
+// during the function runtime.
+OrderedNode* OpDispatchBuilder::ReconstructFSW(OrderedNode* T) {
+  // We must construct the FSW from our various bits
+  OrderedNode* FSW = _Constant(0);
+  auto* Top = T ? T : GetX87Top();
+  FSW = _Bfi(OpSize::i64Bit, 3, 11, FSW, Top);
+
+  auto C0 = GetRFLAG(FEXCore::X86State::X87FLAG_C0_LOC);
+  auto C1 = GetRFLAG(FEXCore::X86State::X87FLAG_C1_LOC);
+  auto C2 = GetRFLAG(FEXCore::X86State::X87FLAG_C2_LOC);
+  auto C3 = GetRFLAG(FEXCore::X86State::X87FLAG_C3_LOC);
+
+  FSW = _Orlshl(OpSize::i64Bit, FSW, C0, 8);
+  FSW = _Orlshl(OpSize::i64Bit, FSW, C1, 9);
+  FSW = _Orlshl(OpSize::i64Bit, FSW, C2, 10);
+  FSW = _Orlshl(OpSize::i64Bit, FSW, C3, 14);
+  return FSW;
+}
+
+void OpDispatchBuilder::X87FNSTSW(OpcodeArgs) {
+  CurrentHeader->HasX87 = true;
+  StoreResult(GPRClass, Op, ReconstructFSW(_SyncStack()), -1);
+}
+
 } // namespace FEXCore::IR
diff --git a/FEXCore/Source/Interface/IR/IR.json b/FEXCore/Source/Interface/IR/IR.json
@@ -2368,6 +2368,29 @@
       }
     },
     "F80": {
+      "GPR = SyncStack": {
+        "Desc": [
+          "Synchronizes the stack environment with the MMX registers.",
+          "Returns the current stack top."
+        ],
+        "DestSize": "8",
+        "JITDispatch": false,
+        "HasSideEffects": true
+      },
+      "IncStackTop": {
+        "Desc": [
+          "Increase stack top-pointer."
+        ],
+        "JITDispatch": false,
+        "HasSideEffects": true
+      },
+      "DecStackTop": {
+        "Desc": [
+          "Decrease stack top-pointer."
+        ],
+        "JITDispatch": false,
+        "HasSideEffects": true
+      },
       "FPR = PushStack FPR:$X80Src, OpSize:$OpSize, i1:$Float, u8:$LoadSize": {
         "Desc": [
           "Pushes the provided source on to the x87 stack.",
@@ -2415,51 +2438,46 @@
         "DestSize": "16",
         "JITDispatch": false
       },
-      "FPR = F80AddStack u8:$SrcStack1, u8:$SrcStack2": {
+      "F80AddStack u8:$SrcStack1, u8:$SrcStack2": {
         "Desc": [
           "Adds two stack locations together, storing the result in to the first stack location"
         ],
         "HasSideEffects": true,
-        "DestSize": "16",
         "JITDispatch": false
       },
-      "FPR = F80AddValue u8:$SrcStack, FPR:$X80Src": {
+      "F80AddValue u8:$SrcStack, FPR:$X80Src": {
         "Desc": [
           "Adds a operand value to a stack location. The result stored in to the stack location provided."
         ],
         "HasSideEffects": true,
-        "DestSize": "16",
         "JITDispatch": false
       },
       "FPR = F80Add FPR:$X80Src1, FPR:$X80Src2": {
         "DestSize": "16",
         "JITDispatch": false
       },
-      "FPR = F80SubStack u8:$DstStack, u8:$SrcStack1, u8:$SrcStack2": {
+      "F80SubStack u8:$DstStack, u8:$SrcStack1, u8:$SrcStack2": {
         "Desc": [
           "Subtracts the value in stack location TOP+$SrcStack2 from the value in stack location TOP+$SrcStack1.",
-          "The result is stored in stack location TOP+$DstStack and returned."
+          "The result is stored in stack location TOP+$DstStack."
         ],
         "HasSideEffects": true,
-        "DestSize": "16",
         "JITDispatch": false
       },
-      "FPR = F80SubValue u8:$SrcStack, FPR:$X80Src": {
+      "F80SubValue u8:$SrcStack, FPR:$X80Src": {
         "Desc": [
           "Subtracts the value $X80Src from the value in stack location TOP+$SrcStack.",
-          "The result is stored in stack location TOP and returned."
+          "The result is stored in stack location TOP."
         ],
         "HasSideEffects": true,
-        "DestSize": "16",
         "JITDispatch": false
       },
-      "FPR = F80SubRValue FPR:$X80Src, u8:$SrcStack": {
+      "F80SubRValue FPR:$X80Src, u8:$SrcStack": {
         "Desc": [
           "Subtracts the value in stack location TOP+$SrcStack from the value $X80Src.",
-          "The result is stored in stack location TOP and returned."
+          "The result is stored in stack location TOP."
         ],
         "HasSideEffects": true,
-        "DestSize": "16",
         "JITDispatch": false
       },
       "FPR = F80Sub FPR:$X80Src1, FPR:$X80Src2": {
@@ -2471,54 +2489,49 @@
         "DestSize": "16",
         "JITDispatch": false
       },
-      "FPR = F80MulStack u8:$SrcStack1, u8:$SrcStack2": {
+      "F80MulStack u8:$SrcStack1, u8:$SrcStack2": {
         "Desc": [
           "Multiplies two stack locations together, storing the result in to the first stack location"
         ],
         "HasSideEffects": true,
-        "DestSize": "16",
         "JITDispatch": false
       },
-      "FPR = F80MulValue u8:$SrcStack, FPR:$X80Src": {
+      "F80MulValue u8:$SrcStack, FPR:$X80Src": {
         "Desc": [
           "Multiplies a operand value to a stack location. The result stored in to the stack location provided."
         ],
         "HasSideEffects": true,
-        "DestSize": "16",
         "JITDispatch": false
       },
       "FPR = F80Mul FPR:$X80Src1, FPR:$X80Src2": {
         "DestSize": "16",
         "JITDispatch": false
       },
-      "FPR = F80DivStack u8:$DstStack, u8:$SrcStack1, u8:$SrcStack2": {
+      "F80DivStack u8:$DstStack, u8:$SrcStack1, u8:$SrcStack2": {
         "Desc": [
           "Divides the value in stack location TOP+$SrcStack1 by the value in stack location TOP+$SrcStack2.",
-          "The result is stored in stack location TOP+$DstStack and returned.",
+          "The result is stored in stack location TOP+$DstStack.",
           "`FPR|Stack[TOP+DstStack] = Stack[TOP+SrcStack1] / Stack[TOP+SrcStack2]`"
         ],
         "HasSideEffects": true,
-        "DestSize": "16",
         "JITDispatch": false
       },
-      "FPR = F80DivValue u8:$SrcStack, FPR:$X80Src": {
+      "F80DivValue u8:$SrcStack, FPR:$X80Src": {
         "Desc": [
           "Divides the value in stack location TOP+$SrcStack by the value $X80Src.",
           "The result is stored in stack location TOP and returned.",
           "`FPR|Stack[TOP] = Stack[TOP+SrcStack] / X80Src`"
         ],
         "HasSideEffects": true,
-        "DestSize": "16",
         "JITDispatch": false
       },
-      "FPR = F80DivRValue FPR:$X80Src, u8:$SrcStack": {
+      "F80DivRValue FPR:$X80Src, u8:$SrcStack": {
         "Desc": [
           "Divides the value X80Src by the value in stack location TOP+$SrcStack.",
-          "The result is stored in stack location TOP and returned.",
+          "The result is stored in stack location TOP.",
           "`FPR|Stack[TOP] = X80Src / Stack[TOP+SrcStack]`"
         ],
         "HasSideEffects": true,
-        "DestSize": "16",
         "JITDispatch": false
       },
       "FPR = F80Div FPR:$X80Src1, FPR:$X80Src2": {