scriptcomp: Optimize constant->variable assignment (#80)

The nwscript construct `int n = 3;` gets compiled into the following: ``` RUNSTACK_ADD, TYPE_INTEGER CONSTANT, TYPE_INTEGER, 3 ASSIGNMENT, TYPE_VOID, -8, 4 MODIFY_STACK_POINTER, -4 ``` This ends up with just the CONSTI 3 on the top of stack, but the dance is necessary as `n = 3` is also an expression which returns 3. Then, the last MODSP discards the result of the expression. Here, we detect the pattern when it is discarded, and replace the whole set of instructions with just CONSTI 3 Partial fix for #70 ## Testing Added variables.nss test. Prior to this change, it got compiled into this mess (without asserts): ``` 0 JSR, 8 6 RET 8 RUNSTACK_ADD, TYPE_INTEGER 10 CONSTANT, TYPE_INTEGER, 1 16 ASSIGNMENT, TYPE_VOID, -8, 4 24 MODIFY_STACK_POINTER, -4 30 RUNSTACK_ADD, TYPE_INTEGER 32 CONSTANT, TYPE_INTEGER, 2 38 ASSIGNMENT, TYPE_VOID, -8, 4 46 MODIFY_STACK_POINTER, -4 52 RUNSTACK_ADD, TYPE_INTEGER 54 CONSTANT, TYPE_INTEGER, 3 60 ASSIGNMENT, TYPE_VOID, -8, 4 68 MODIFY_STACK_POINTER, -4 74 RUNSTACK_ADD, TYPE_INTEGER 76 CONSTANT, TYPE_INTEGER, 4 82 ASSIGNMENT, TYPE_VOID, -8, 4 90 MODIFY_STACK_POINTER, -4 96 RUNSTACK_ADD, TYPE_STRING 98 CONSTANT, TYPE_STRING, "A" 103 ASSIGNMENT, TYPE_VOID, -8, 4 111 MODIFY_STACK_POINTER, -4 117 RUNSTACK_ADD, TYPE_FLOAT 119 CONSTANT, TYPE_FLOAT, 1.0 125 ASSIGNMENT, TYPE_VOID, -8, 4 133 MODIFY_STACK_POINTER, -4 139 CONSTANT, TYPE_INTEGER, 1 145 JZ, 143 151 RUNSTACK_ADD, TYPE_INTEGER 153 CONSTANT, TYPE_INTEGER, 10 159 ASSIGNMENT, TYPE_VOID, -8, 4 167 MODIFY_STACK_POINTER, -4 173 RUNSTACK_ADD, TYPE_INTEGER 175 CONSTANT, TYPE_INTEGER, 20 181 ASSIGNMENT, TYPE_VOID, -8, 4 189 MODIFY_STACK_POINTER, -4 195 RUNSTACK_ADD, TYPE_INTEGER 197 CONSTANT, TYPE_INTEGER, 30 203 ASSIGNMENT, TYPE_VOID, -8, 4 211 MODIFY_STACK_POINTER, -4 217 RUNSTACK_ADD, TYPE_INTEGER 219 CONSTANT, TYPE_INTEGER, 40 225 ASSIGNMENT, TYPE_VOID, -8, 4 233 MODIFY_STACK_POINTER, -4 239 RUNSTACK_ADD, TYPE_STRING 241 CONSTANT, TYPE_STRING, "B" 246 ASSIGNMENT, TYPE_VOID, -8, 4 254 MODIFY_STACK_POINTER, -4 260 RUNSTACK_ADD, TYPE_FLOAT 262 CONSTANT, TYPE_FLOAT, 10.0 268 ASSIGNMENT, TYPE_VOID, -8, 4 276 MODIFY_STACK_POINTER, -28 282 JMP, 6 288 MODIFY_STACK_POINTER, -24 294 RET ``` after the change, it becomes: ``` 0 JSR, 8 6 RET 8 CONSTANT, TYPE_INTEGER, 1 14 CONSTANT, TYPE_INTEGER, 2 20 CONSTANT, TYPE_INTEGER, 3 26 CONSTANT, TYPE_INTEGER, 4 32 CONSTANT, TYPE_STRING, "A" 37 CONSTANT, TYPE_FLOAT, 1.0 43 CONSTANT, TYPE_INTEGER, 1 49 JZ, 53 55 CONSTANT, TYPE_INTEGER, 10 61 CONSTANT, TYPE_INTEGER, 20 67 CONSTANT, TYPE_INTEGER, 30 73 CONSTANT, TYPE_INTEGER, 40 79 CONSTANT, TYPE_STRING, "B" 84 CONSTANT, TYPE_FLOAT, 10.0 90 MODIFY_STACK_POINTER, -24 96 JMP, 6 102 MODIFY_STACK_POINTER, -24 108 RET ``` ## Changelog ### Performance Improvements - nwscript compiler will now generate more optimal code for assigning values to variables ## Licence - [x] I am licencing my change under the project's MIT licence, including all changes to GPL-3.0 licenced parts of the codebase.
niv · Jul 16, 2023 · 07a13f1 · 07a13f1
1 parent 85fc548
commit 07a13f1
Show file tree

Hide file tree

Showing 3 changed files with 82 additions and 4 deletions.
diff --git a/neverwinter/nwscript/native/scriptcomp.h b/neverwinter/nwscript/native/scriptcomp.h
@@ -606,6 +606,8 @@ class CScriptCompiler
 	int32_t         m_nOutputCodeLength;
 	std::vector<int32_t> m_aOutputCodeInstructionBoundaries;
 
+	char *InstructionLookback(uint32_t last=1);
+
 	// Resolving code to its proper location ... some buffers!
 	char       *m_pchResolvedOutputBuffer;
 	int32_t         m_nResolvedOutputBufferSize;

diff --git a/neverwinter/nwscript/native/scriptcompfinalcode.cpp b/neverwinter/nwscript/native/scriptcompfinalcode.cpp
@@ -6966,6 +6966,13 @@ int32_t CScriptCompiler::WriteDebuggerOutputToFile(CExoString sFileName)
 	return 0;
 }
 
+char *CScriptCompiler::InstructionLookback(uint32_t last)
+{
+	if (last == 0 || last > m_aOutputCodeInstructionBoundaries.size())
+		return NULL;
+
+	return &m_pchOutputCode[m_aOutputCodeInstructionBoundaries[m_aOutputCodeInstructionBoundaries.size() - 1 - last]];
+}
 
 void CScriptCompiler::WriteByteSwap32(char *buffer, int32_t value)
 {
@@ -6996,14 +7003,50 @@ void CScriptCompiler::EmitModifyStackPointer(int32_t nModifyBy)
 {
 	if (m_bOptimizeBinarySpace) // TODO: Rename or add separate toggle?
 	{
-		int32_t last = m_aOutputCodeInstructionBoundaries[m_aOutputCodeInstructionBoundaries.size()-2];
-		if (m_pchOutputCode[last + CVIRTUALMACHINE_OPCODE_LOCATION] == CVIRTUALMACHINE_OPCODE_MODIFY_STACK_POINTER)
+		char *last = InstructionLookback(1);
+		// Multiple MODIFY_STACK_POINTER instructions can always be merged into a single one
+		if (last[CVIRTUALMACHINE_OPCODE_LOCATION] == CVIRTUALMACHINE_OPCODE_MODIFY_STACK_POINTER)
 		{
-			int32_t mod = ReadByteSwap32(&m_pchOutputCode[last + CVIRTUALMACHINE_EXTRA_DATA_LOCATION]);
+			int32_t mod = ReadByteSwap32(&last[CVIRTUALMACHINE_EXTRA_DATA_LOCATION]);
 			mod += nModifyBy;
-			WriteByteSwap32(&m_pchOutputCode[last + CVIRTUALMACHINE_EXTRA_DATA_LOCATION], mod);
+			WriteByteSwap32(&last[CVIRTUALMACHINE_EXTRA_DATA_LOCATION], mod);
 			return;
 		}
+
+		// The nwscript construct `int n = 3;` gets compiled into the following:
+		//     RUNSTACK_ADD, TYPE_INTEGER
+		//     CONSTANT, TYPE_INTEGER, 3
+		//     ASSIGNMENT, TYPE_VOID, -8, 4
+		//     MODIFY_STACK_POINTER, -4
+		// This ends up with just the CONSTI 3 on the top of stack, but the
+		// dance is necessary as `n = 3` is also an expression which returns 3.
+		// Then, the last MODSP discards the result of the expression.
+		// Here, we detect the pattern when it is discarded, and replace the
+		// whole set of instructions with just CONSTI 3
+		if (last[CVIRTUALMACHINE_OPCODE_LOCATION] == CVIRTUALMACHINE_OPCODE_ASSIGNMENT)
+		{
+			char *instConstant    = InstructionLookback(2);
+			char *instRunstackAdd = InstructionLookback(3);
+
+			if (instConstant    && instConstant[CVIRTUALMACHINE_OPCODE_LOCATION]    == CVIRTUALMACHINE_OPCODE_CONSTANT &&
+			    instRunstackAdd && instRunstackAdd[CVIRTUALMACHINE_OPCODE_LOCATION] == CVIRTUALMACHINE_OPCODE_RUNSTACK_ADD &&
+			    instConstant[CVIRTUALMACHINE_AUXCODE_LOCATION] == instRunstackAdd[CVIRTUALMACHINE_AUXCODE_LOCATION] &&
+			    ReadByteSwap32(&last[CVIRTUALMACHINE_EXTRA_DATA_LOCATION]) == -8
+			   )
+			{
+				// Move the CONST instruction up to where runstack add is..
+				const int32_t instConstantSize = (last - instConstant);
+				memmove(instRunstackAdd, instConstant, instConstantSize);
+				// Roll back output code length to where the old instConstant started..
+				m_nOutputCodeLength = (instRunstackAdd - m_pchOutputCode) + instConstantSize;
+				// Pop the last two instruction boundaries (CONSTANT, ASSIGNMENT) since those don't exist anymore
+				m_aOutputCodeInstructionBoundaries.pop_back();
+				m_aOutputCodeInstructionBoundaries.pop_back();
+				m_aOutputCodeInstructionBoundaries.pop_back();
+				m_aOutputCodeInstructionBoundaries.push_back(m_nOutputCodeLength);
+				return;
+			}
+		}
 	}
 
 	char *buf = EmitInstruction(CVIRTUALMACHINE_OPCODE_MODIFY_STACK_POINTER, 0, 4);

diff --git a/tests/scriptcomp/corpus/variables.nss b/tests/scriptcomp/corpus/variables.nss
@@ -0,0 +1,33 @@
+void main()
+{
+    int a = 1;
+    int b = 2;
+    int c = 3;
+    int d = 4;
+    string s = "A";
+    float f = 1.0f;
+
+    if (1)
+    {
+        int a = 10;
+        int b = 20;
+        int c = 30;
+        int d = 40;
+        string s = "B";
+        float f = 10.0f;
+
+        Assert(a == 10);
+        Assert(b == 20);
+        Assert(c == 30);
+        Assert(d == 40);
+        Assert(s == "B");
+        Assert(f == 10.0f);
+    }
+
+    Assert(a == 1);
+    Assert(b == 2);
+    Assert(c == 3);
+    Assert(d == 4);
+    Assert(s == "A");
+    Assert(f == 1.0f);
+}