From aa88d94755e4621e1bcc95d97faf65118a44697c Mon Sep 17 00:00:00 2001
From: nmzik <Nmzik@mail.ru>
Date: Tue, 4 Aug 2026 23:12:47 +0200
Subject: [PATCH] shader: implement V_SUB_NC_U32

---
 .../recompiler/decompiler/VectorAluOps.cpp    |  2 +-
 tests/ShaderRecompilerComputeTests.cpp        | 44 +++++++++++++++++++
 tests/shaderCfgTests.cpp                      |  6 +++
 3 files changed, 51 insertions(+), 1 deletion(-)

diff --git a/src/graphics/shader/recompiler/decompiler/VectorAluOps.cpp b/src/graphics/shader/recompiler/decompiler/VectorAluOps.cpp
index 664dacad3..d00a81c43 100644
--- a/src/graphics/shader/recompiler/decompiler/VectorAluOps.cpp
+++ b/src/graphics/shader/recompiler/decompiler/VectorAluOps.cpp
@@ -834,7 +834,7 @@ constexpr Vop2SdwaRule VOP2_SDWA_RULES[] = {
     {Opcode::VXorB32, SdwaSelWords() | SdwaSelFull(), SdwaSelAll(), SdwaSelAll(), true, false},
     {Opcode::VXnorB32, SdwaSelWords() | SdwaSelFull(), SdwaSelAll(), SdwaSelAll(), true, false},
     {Opcode::VAddNcU32, SdwaSelFull(), SdwaSelAll(), SdwaSelAll(), false, false},
-    {Opcode::VSubNcU32, SdwaSelFull(), SdwaSelAll(), SdwaSelAll(), false, false},
+    {Opcode::VSubNcU32, SdwaSelAll(), SdwaSelAll(), SdwaSelAll(), true, false},
     {Opcode::VSubrevNcU32, SdwaSelFull(), SdwaSelAll(), SdwaSelAll(), false, false},
 };
 
diff --git a/tests/ShaderRecompilerComputeTests.cpp b/tests/ShaderRecompilerComputeTests.cpp
index 981acd8ee..8f0fb494b 100644
--- a/tests/ShaderRecompilerComputeTests.cpp
+++ b/tests/ShaderRecompilerComputeTests.cpp
@@ -10116,6 +10116,48 @@ TestCase VectorIntegerOps() {
 	         O::VCndmaskB32, O::BufferStoreDword, O::SEndpgm}};
 }
 
+TestCase Vop2SdwaSubNcExactByte2Destination() {
+	using O = ShaderOpcode;
+
+	std::vector<u32> code;
+	AppendVMovU32(&code, 11, 2);
+	code.push_back(0x4c1616f9u);
+	code.push_back(0x0686128du);
+	AppendStoreVgpr(&code, 11, 0);
+	AppendEnd(&code);
+
+	TestCase test;
+	test.name           = "Vop2SdwaSubNcExactByte2Destination";
+	test.code           = code;
+	test.expected       = {0x000b0002u};
+	test.opcodes        = {O::VMovB32, O::VSubNcU32, O::BufferStoreDword, O::SEndpgm};
+	test.required_spirv = {"OpISub", "OpShiftLeftLogical", "OpBitwiseOr"};
+	return test;
+}
+
+TestCase Vop2SdwaSubNcPreservesByteAndWordDestinations() {
+	using O = ShaderOpcode;
+
+	std::vector<u32> code;
+	AppendVMovU32(&code, 0, 3);
+	for (u32 dst = 10; dst <= 15; dst++) {
+		AppendVMovLiteral(&code, dst, 0xa1b2c3d4u);
+	}
+	for (u32 sel = 0; sel <= 5; sel++) {
+		code.push_back(EncodeVop2(0x26, 10 + sel, 249, 0));
+		code.push_back(EncodeVop2Sdwa(InlineU32(20), sel, 2, 6, 6, 0, 0, 0, 0, 0, 0, 1));
+		AppendStoreVgpr(&code, 10 + sel, sel);
+	}
+	AppendEnd(&code);
+
+	return {"Vop2SdwaSubNcPreservesByteAndWordDestinations",
+	        code,
+	        {},
+	        {0xa1b2c311u, 0xa1b211d4u, 0xa111c3d4u, 0x11b2c3d4u, 0xa1b20011u,
+	         0x0011c3d4u},
+	        {O::VMovB32, O::VSubNcU32, O::BufferStoreDword, O::SEndpgm}};
+}
+
 TestCase VectorShiftCountsMaskLowBits() {
 	using O = ShaderOpcode;
 
@@ -15050,6 +15092,8 @@ std::vector<TestCase> MakeCases() {
 	AddCase(VectorMoves);
 	AddCase(VectorVop3MoveAppliesFloatSourceModifiers);
 	AddCase(VectorIntegerOps);
+	AddCase(Vop2SdwaSubNcExactByte2Destination);
+	AddCase(Vop2SdwaSubNcPreservesByteAndWordDestinations);
 	AddCase(VectorShiftCountsMaskLowBits);
 	AddCase(VectorVop3IntegerOps);
 	AddCase(VectorBfeI32ArithmeticShiftMasksField);
diff --git a/tests/shaderCfgTests.cpp b/tests/shaderCfgTests.cpp
index 5b58f733b..15eb520e1 100644
--- a/tests/shaderCfgTests.cpp
+++ b/tests/shaderCfgTests.cpp
@@ -2417,6 +2417,8 @@ void TestNewShaderRecompilerBootB16PackedAndSdwaOpcodes() {
 	    0x0c860688u, // v_add_nc_u32 v5, 8, sign-extended v4.lo
 	    0x4a0c0cf9u,
 	    0x0d860688u, // v_add_nc_u32 v6, 8, sign-extended v6.hi
+	    0x4c1616f9u,
+	    0x0686128du, // v_sub_nc_u32 v11.byte2, 13, v11; preserve other destination bytes
 	    0xbf810000u,
 	};
 
@@ -2465,6 +2467,8 @@ void TestNewShaderRecompilerBootB16PackedAndSdwaOpcodes() {
 	Check(Common::ContainsStr(result.decoded_dump, "v_add_nc_u32 v6") &&
 	          Common::ContainsStr(result.decoded_dump, "v6.sdwa(sel=5,sext=1"),
 	      "new decoder did not decode V_ADD_NC_U32 SDWA sign-extended high word");
+	Check(Common::ContainsStr(result.decoded_dump, "v_sub_nc_u32 v11.sdwa(sel=2"),
+	      "new decoder did not decode V_SUB_NC_U32 SDWA byte-2 destination");
 	Check(!Common::ContainsStr(result.decoded_dump, "unsupported family=VOP2 opcode=0x00"),
 	      "literal/SDWA extension words were decoded as phantom VOP2 instructions");
 	Check(!Common::ContainsStr(result.decoded_dump,
@@ -2498,6 +2502,8 @@ void TestNewShaderRecompilerBootB16PackedAndSdwaOpcodes() {
 	Check(Common::ContainsStr(result.ir_dump, "IAddU32 v6") &&
 	          Common::ContainsStr(result.ir_dump, "v6.sdwa(sel=5,sext=1"),
 	      "V_ADD_NC_U32 SDWA sign-extended high word did not lower to IR");
+	Check(Common::ContainsStr(result.ir_dump, "ISubU32 v11.sdwa(sel=2"),
+	      "V_SUB_NC_U32 SDWA byte-2 destination did not lower to IR");
 	Check(SpirvContainsOpcode(result.spirv, 128),
 	      "SPIR-V binary does not contain OpIAdd for U16 operations");
 	Check(SpirvContainsOpcode(result.spirv, 202),