Merge pull request #3263 from alyssarosenzweig/opt/not-garbage

OpcodeDispatcher: Make "not" not garbage
FEX-Emu · Nov 9, 2023 · 3767f36 · 3767f36
2 parents af32539 + 03087a5
commit 3767f36
Show file tree

Hide file tree

Showing 3 changed files with 53 additions and 26 deletions.
diff --git a/FEXCore/Source/Interface/Core/OpcodeDispatcher.cpp b/FEXCore/Source/Interface/Core/OpcodeDispatcher.cpp
@@ -3343,11 +3343,39 @@ void OpDispatchBuilder::NOTOp(OpcodeArgs) {
     OrderedNode *DestMem = LoadSource(GPRClass, Op, Op->Dest, Op->Flags, {.LoadData = false});
     DestMem = AppendSegmentOffset(DestMem, Op->Flags);
     _AtomicXor(IR::SizeToOpSize(Size), MaskConst, DestMem);
-  }
-  else {
+  } else if (!Op->Dest.IsGPR()) {
+    // GPR version plays fast and loose with sizes, be safe for memory tho.
     OrderedNode *Src = LoadSource(GPRClass, Op, Op->Dest, Op->Flags);
     Src = _Xor(OpSize::i64Bit, Src, MaskConst);
     StoreResult(GPRClass, Op, Src, -1);
+  } else {
+    // Specially handle high bits so we can invert in place with the correct
+    // mask and a larger type.
+    auto Dest = Op->Dest;
+    if (Dest.Data.GPR.HighBits) {
+      LOGMAN_THROW_A_FMT(Size == 1, "Only 8-bit GPRs get high bits");
+      MaskConst = _Constant(0xFF00);
+      Dest.Data.GPR.HighBits = false;
+    }
+
+    // Always load full size, we explicitly want the upper bits to get the
+    // insert behaviour for free/implicitly.
+    const uint8_t GPRSize = CTX->GetGPRSize();
+    OrderedNode *Src = LoadSource_WithOpSize(GPRClass, Op, Dest, GPRSize, Op->Flags);
+
+    // For 8/16-bit, use 64-bit invert so we invert in place, while getting
+    // insert behaviour. For 32-bit, use 32-bit invert to zero the upper bits.
+    unsigned EffectiveSize = Size == 4 ? 4 : GPRSize;
+
+    // If we're inverting the whole thing, use Not instead of Xor to save a constant.
+    if (Size >= 4)
+      Src = _Not(IR::SizeToOpSize(EffectiveSize), Src);
+    else
+      Src = _Xor(IR::SizeToOpSize(EffectiveSize), Src, MaskConst);
+
+    // Always store 64-bit, the Not/Xor correctly handle the upper bits and this
+    // way we can delete the store.
+    StoreResult_WithOpSize(GPRClass, Op, Dest, Src, GPRSize, -1);
   }
 }
 

diff --git a/unittests/InstructionCountCI/FlagM/PrimaryGroup.json b/unittests/InstructionCountCI/FlagM/PrimaryGroup.json
@@ -3081,13 +3081,11 @@
       ]
     },
     "not bl": {
-      "ExpectedInstructionCount": 3,
-      "Optimal": "No",
+      "ExpectedInstructionCount": 1,
+      "Optimal": "Yes",
       "Comment": "GROUP2 0xf6 /2",
       "ExpectedArm64ASM": [
-        "uxtb w20, w7",
-        "eor x20, x20, #0xff",
-        "bfxil x7, x20, #0, #8"
+        "eor x7, x7, #0xff"
       ]
     },
     "neg bl": {

diff --git a/unittests/InstructionCountCI/PrimaryGroup.json b/unittests/InstructionCountCI/PrimaryGroup.json
@@ -3233,13 +3233,19 @@
       ]
     },
     "not bl": {
-      "ExpectedInstructionCount": 3,
-      "Optimal": "No",
+      "ExpectedInstructionCount": 1,
+      "Optimal": "Yes",
       "Comment": "GROUP2 0xf6 /2",
       "ExpectedArm64ASM": [
-        "uxtb w20, w7",
-        "eor x20, x20, #0xff",
-        "bfxil x7, x20, #0, #8"
+        "eor x7, x7, #0xff"
+      ]
+    },
+    "not bh": {
+      "ExpectedInstructionCount": 1,
+      "Optimal": "Yes",
+      "Comment": "GROUP2 0xf6 /2",
+      "ExpectedArm64ASM": [
+        "eor x7, x7, #0xff00"
       ]
     },
     "neg bl": {
@@ -3400,32 +3406,27 @@
       ]
     },
     "not bx": {
-      "ExpectedInstructionCount": 3,
-      "Optimal": "No",
+      "ExpectedInstructionCount": 1,
+      "Optimal": "Yes",
       "Comment": "GROUP2 0xf7 /1",
       "ExpectedArm64ASM": [
-        "uxth w20, w7",
-        "eor x20, x20, #0xffff",
-        "bfxil x7, x20, #0, #16"
+        "eor x7, x7, #0xffff"
       ]
     },
     "not ebx": {
-      "ExpectedInstructionCount": 3,
-      "Optimal": "No",
+      "ExpectedInstructionCount": 1,
+      "Optimal": "Yes",
       "Comment": "GROUP2 0xf7 /1",
       "ExpectedArm64ASM": [
-        "mov w20, w7",
-        "eor x20, x20, #0xffffffff",
-        "mov w7, w20"
+        "mvn w7, w7"
       ]
     },
     "not rbx": {
-      "ExpectedInstructionCount": 2,
-      "Optimal": "No",
+      "ExpectedInstructionCount": 1,
+      "Optimal": "Yes",
       "Comment": "GROUP2 0xf7 /1",
       "ExpectedArm64ASM": [
-        "mov x20, #0xffffffffffffffff",
-        "eor x7, x7, x20"
+        "mvn x7, x7"
       ]
     },
     "neg bx": {