Merge pull request FEX-Emu#3397 from pmatos/XCHGOp

Improve XCHG operations
Sonicadvance1 · Feb 2, 2024 · 920a8db · 920a8db
2 parents 9c37c0f + 4623544
commit 920a8db
Show file tree

Hide file tree

Showing 3 changed files with 38 additions and 40 deletions.
diff --git a/FEXCore/Source/Interface/Core/OpcodeDispatcher.cpp b/FEXCore/Source/Interface/Core/OpcodeDispatcher.cpp
@@ -1367,14 +1367,17 @@ void OpDispatchBuilder::XCHGOp(OpcodeArgs) {
     // So x86-64 spec mandates this special case that even though it is a 32bit instruction and
     // is supposed to zext the result, it is a true no-op
     if (Op->Flags & FEXCore::X86Tables::DecodeFlags::FLAG_REP_PREFIX) {
-      // If this instruction has a REP prefix then this is architectually defined to be a `PAUSE` instruction.
-      // On older processors this ends up being a true `REP NOP` which is why they stuck this here.
+      // If this instruction has a REP prefix then this is architecturally
+      // defined to be a `PAUSE` instruction. On older processors this ends up
+      // being a true `REP NOP` which is why they stuck this here.
       _Yield();
     }
     return;
   }
 
-  OrderedNode *Src = LoadSource(GPRClass, Op, Op->Src[0], Op->Flags);
+  // AllowUpperGarbage: OK to allow as it will be overwritten by StoreResult.
+  OrderedNode *Src = LoadSource(GPRClass, Op, Op->Src[0], Op->Flags,
+                                {.AllowUpperGarbage = true});
   if (DestIsMem(Op)) {
     HandledLock = Op->Flags & FEXCore::X86Tables::DecodeFlags::FLAG_LOCK;
     OrderedNode *Dest = LoadSource(GPRClass, Op, Op->Dest, Op->Flags, {.LoadData = false});
@@ -1385,7 +1388,8 @@ void OpDispatchBuilder::XCHGOp(OpcodeArgs) {
     StoreResult(GPRClass, Op, Op->Src[0], Result, -1);
   }
   else {
-    OrderedNode *Dest = LoadSource(GPRClass, Op, Op->Dest, Op->Flags);
+    // AllowUpperGarbage: OK to allow as it will be overwritten by StoreResult.
+    OrderedNode *Dest = LoadSource(GPRClass, Op, Op->Dest, Op->Flags, {.AllowUpperGarbage = true});
 
     // Swap the contents
     // Order matters here since we don't want to swap context contents for one that effects the other

diff --git a/unittests/InstructionCountCI/Atomics.json b/unittests/InstructionCountCI/Atomics.json
@@ -358,31 +358,28 @@
       ]
     },
     "xchg byte [rax], cl": {
-      "ExpectedInstructionCount": 4,
+      "ExpectedInstructionCount": 3,
       "Comment": "0x86",
       "ExpectedArm64ASM": [
-        "uxtb w20, w5",
-        "mov w1, w20",
+        "mov w1, w5",
         "swpalb w1, w20, [x4]",
         "bfxil x5, x20, #0, #8"
       ]
     },
     "xchg word [rax], cx": {
-      "ExpectedInstructionCount": 4,
+      "ExpectedInstructionCount": 3,
       "Comment": "0x87",
       "ExpectedArm64ASM": [
-        "uxth w20, w5",
-        "mov w1, w20",
+        "mov w1, w5",
         "swpalh w1, w20, [x4]",
         "bfxil x5, x20, #0, #16"
       ]
     },
     "xchg dword [rax], ecx": {
-      "ExpectedInstructionCount": 3,
+      "ExpectedInstructionCount": 2,
       "Comment": "0x87",
       "ExpectedArm64ASM": [
-        "mov w20, w5",
-        "mov w1, w20",
+        "mov w1, w5",
         "swpal w1, w5, [x4]"
       ]
     },

diff --git a/unittests/InstructionCountCI/Primary.json b/unittests/InstructionCountCI/Primary.json
@@ -2210,18 +2210,17 @@
       "ExpectedInstructionCount": 4,
       "Comment": "0x86",
       "ExpectedArm64ASM": [
-        "uxtb w20, w7",
-        "uxtb w21, w5",
-        "bfxil x5, x20, #0, #8",
-        "bfxil x7, x21, #0, #8"
+        "mov x20, x5",
+        "mov x5, x20",
+        "bfxil x5, x7, #0, #8",
+        "bfxil x7, x20, #0, #8"
       ]
     },
     "xchg [rax], cl": {
-      "ExpectedInstructionCount": 4,
+      "ExpectedInstructionCount": 3,
       "Comment": "0x86",
       "ExpectedArm64ASM": [
-        "uxtb w20, w5",
-        "mov w1, w20",
+        "mov w1, w5",
         "swpalb w1, w20, [x4]",
         "bfxil x5, x20, #0, #8"
       ]
@@ -2230,18 +2229,17 @@
       "ExpectedInstructionCount": 4,
       "Comment": "0x87",
       "ExpectedArm64ASM": [
-        "uxth w20, w7",
-        "uxth w21, w5",
-        "bfxil x5, x20, #0, #16",
-        "bfxil x7, x21, #0, #16"
+        "mov x20, x5",
+        "mov x5, x20",
+        "bfxil x5, x7, #0, #16",
+        "bfxil x7, x20, #0, #16"
       ]
     },
     "xchg [rax], cx": {
-      "ExpectedInstructionCount": 4,
+      "ExpectedInstructionCount": 3,
       "Comment": "0x87",
       "ExpectedArm64ASM": [
-        "uxth w20, w5",
-        "mov w1, w20",
+        "mov w1, w5",
         "swpalh w1, w20, [x4]",
         "bfxil x5, x20, #0, #16"
       ]
@@ -2250,17 +2248,16 @@
       "ExpectedInstructionCount": 3,
       "Comment": "0x87",
       "ExpectedArm64ASM": [
-        "mov w20, w7",
-        "mov w7, w5",
-        "mov x5, x20"
+        "mov x20, x5",
+        "mov w5, w7",
+        "mov w7, w20"
       ]
     },
     "xchg [rax], ecx": {
-      "ExpectedInstructionCount": 3,
+      "ExpectedInstructionCount": 2,
       "Comment": "0x87",
       "ExpectedArm64ASM": [
-        "mov w20, w5",
-        "mov w1, w20",
+        "mov w1, w5",
         "swpal w1, w5, [x4]"
       ]
     },
@@ -2753,19 +2750,19 @@
       "ExpectedInstructionCount": 4,
       "Comment": "0x90",
       "ExpectedArm64ASM": [
-        "uxth w20, w4",
-        "uxth w21, w7",
-        "bfxil x7, x20, #0, #16",
-        "bfxil x4, x21, #0, #16"
+        "mov x20, x7",
+        "mov x7, x20",
+        "bfxil x7, x4, #0, #16",
+        "bfxil x4, x20, #0, #16"
       ]
     },
     "xchg eax, ebx": {
       "ExpectedInstructionCount": 3,
       "Comment": "0x90",
       "ExpectedArm64ASM": [
-        "mov w20, w4",
-        "mov w4, w7",
-        "mov x7, x20"
+        "mov x20, x7",
+        "mov w7, w4",
+        "mov w4, w20"
       ]
     },
     "xchg rax, rbx": {