Merge pull request #4169 from Sonicadvance1/x87_loadstore_tests

unittests/ASM: Fixes x87 80-bit loads on the edge of page boundaries.

FEXCore/Source/Interface/Core/OpcodeDispatcher.cpp

@@ -4307,6 +4307,14 @@ Ref OpDispatchBuilder::LoadSource_WithOpSize(RegisterClassType Class, const X86T
   }
 
   if ((IsOperandMem(Operand, true) && LoadData) || ForceLoad) {
+    if (OpSize == OpSize::f80Bit) {
+      Ref MemSrc = LoadEffectiveAddress(A, true);
+
+      // For X87 extended doubles, Split the load.
+      auto Res = _LoadMem(Class, OpSize::i64Bit, MemSrc, Align == OpSize::iInvalid ? OpSize : Align);
+      return _VLoadVectorElement(OpSize::i128Bit, OpSize::i16Bit, Res, 4, _Add(OpSize::i64Bit, MemSrc, _InlineConstant(8)));
+    }
+
     return _LoadMemAutoTSO(Class, OpSize, A, Align == OpSize::iInvalid ? OpSize : Align);
   } else {
     return LoadEffectiveAddress(A, false, AllowUpperGarbage);

FEXCore/Source/Interface/Core/OpcodeDispatcher/X87.cpp

@@ -63,7 +63,7 @@ void OpDispatchBuilder::SetX87Top(Ref Value) {
 void OpDispatchBuilder::FLD(OpcodeArgs, IR::OpSize Width) {
   const auto ReadWidth = (Width == OpSize::f80Bit) ? OpSize::i128Bit : Width;
 
-  Ref Data = LoadSource_WithOpSize(FPRClass, Op, Op->Src[0], ReadWidth, Op->Flags);
+  Ref Data = LoadSource_WithOpSize(FPRClass, Op, Op->Src[0], Width, Op->Flags);
   Ref ConvertedData = Data;
   // Convert to 80bit float
   if (Width == OpSize::i32Bit || Width == OpSize::i64Bit) {
@@ -79,7 +79,7 @@ void OpDispatchBuilder::FLDFromStack(OpcodeArgs) {
 
 void OpDispatchBuilder::FBLD(OpcodeArgs) {
   // Read from memory
-  Ref Data = LoadSource_WithOpSize(FPRClass, Op, Op->Src[0], OpSize::i128Bit, Op->Flags);
+  Ref Data = LoadSource_WithOpSize(FPRClass, Op, Op->Src[0], OpSize::f80Bit, Op->Flags);
   Ref ConvertedData = _F80BCDLoad(Data);
   _PushStack(ConvertedData, Data, OpSize::i128Bit, true);
 }

unittests/ASM/X87/LoadAtBoundary.asm

@@ -0,0 +1,32 @@
+%ifdef CONFIG
+{
+  "RegData": {
+    "MM7":  ["0x5354555657584142", "0x0000000000005152"],
+    "MM6":  ["0xe94de5eae34fc1c0", "0x0000000000004039"]
+  },
+  "MemoryRegions": {
+    "0x100000000": "4096"
+  }
+}
+%endif
+
+finit ; enters x87 state
+
+mov rax, 0x100000000
+mov rbx, 0x4142434445464748
+mov rcx, 0x5152535455565758
+mov rdx, (0x100000000 + 0x1000 - 16)
+
+mov [rdx], rbx
+mov [rdx + 8], rcx
+
+mov rdx, 0x100000000 + 0x1000
+
+; Do an 80-bit load at the edge of a page.
+; Ensuring tword loads don't extend past the end of a page.
+fld tword [rdx - 10]
+
+; Do an 80-bit BCD load at the edge of a page.
+fbld [rdx - 10]
+
+hlt

unittests/ASM/X87/StoreAtBoundary.asm

@@ -0,0 +1,33 @@
+%ifdef CONFIG
+{
+  "MemoryRegions": {
+    "0x100000000": "4096"
+  }
+}
+%endif
+
+finit ; enters x87 state
+
+mov rax, 0x100000000
+mov rbx, 0x4142434445464748
+mov rcx, 0x5152535455565758
+mov rdx, (0x100000000 + 0x1000 - 16)
+
+mov [rdx], rbx
+mov [rdx + 8], rcx
+
+mov rdx, 0x100000000 + 0x1000
+
+; Load the data in to an x87 register for storing.
+fld tword [rdx - 16]
+fld tword [rdx - 16]
+
+; Do an 80-bit store at the edge of a page.
+; Ensuring tword stores don't extend past the end of a page.
+; If storing past the end of the page, then an unhandled SIGSEGV will occur.
+fstp tword [rdx - 10]
+
+; Do an 80-bit bcd store at the edge of a page.
+fbstp [rdx - 10]
+
+hlt

unittests/InstructionCountCI/FlagM/x87.json

@@ -6869,12 +6869,14 @@
       ]
     },
     "fld tword [rax]": {
-      "ExpectedInstructionCount": 12,
+      "ExpectedInstructionCount": 14,
       "Comment": [
         "0xdb !11b /5"
       ],
       "ExpectedArm64ASM": [
-        "ldr q2, [x4]",
+        "ldr d2, [x4]",
+        "add x20, x4, #0x8 (8)",
+        "ld1 {v2.h}[4], [x20]",
         "ldrb w20, [x28, #1019]",
         "mov w21, #0x1",
         "sub w20, w20, #0x1 (1)",
@@ -16690,12 +16692,14 @@
       ]
     },
     "fbld tword [rax]": {
-      "ExpectedInstructionCount": 40,
+      "ExpectedInstructionCount": 42,
       "Comment": [
         "0xdf !11b /4"
       ],
       "ExpectedArm64ASM": [
-        "ldr q2, [x4]",
+        "ldr d2, [x4]",
+        "add x20, x4, #0x8 (8)",
+        "ld1 {v2.h}[4], [x20]",
         "mrs x0, nzcv",
         "str w0, [x28, #1000]",
         "stp x4, x7, [x28, #280]",

unittests/InstructionCountCI/x87.json

@@ -6868,12 +6868,14 @@
       ]
     },
     "fld tword [rax]": {
-      "ExpectedInstructionCount": 12,
+      "ExpectedInstructionCount": 14,
       "Comment": [
         "0xdb !11b /5"
       ],
       "ExpectedArm64ASM": [
-        "ldr q2, [x4]",
+        "ldr d2, [x4]",
+        "add x20, x4, #0x8 (8)",
+        "ld1 {v2.h}[4], [x20]",
         "ldrb w20, [x28, #1019]",
         "mov w21, #0x1",
         "sub w20, w20, #0x1 (1)",
@@ -16548,12 +16550,14 @@
       ]
     },
     "fbld tword [rax]": {
-      "ExpectedInstructionCount": 40,
+      "ExpectedInstructionCount": 42,
       "Comment": [
         "0xdf !11b /4"
       ],
       "ExpectedArm64ASM": [
-        "ldr q2, [x4]",
+        "ldr d2, [x4]",
+        "add x20, x4, #0x8 (8)",
+        "ld1 {v2.h}[4], [x20]",
         "mrs x0, nzcv",
         "str w0, [x28, #1000]",
         "stp x4, x7, [x28, #280]",