shader: support returning 32-bit LDS compare-exchange (DS_CMPST_RTN_B32) (#1578)

* shader: support DS_CMPST_RTN_B32

Assassin's Creed Valhalla (PPSA01490, #281) stops on DS opcode 0x30 in a
compute shader. DS compares DATA and stores DATA2, the opposite of
BUFFER_ATOMIC_CMPSWAP, so the 32-bit atomic path swaps the two for this
opcode, as the 64-bit form does.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* shader: reject unsupported GDS compare-store variant

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: nmzik <Nmzik@mail.ru>

tests/ShaderRecompilerComputeTests.cpp

@@ -21015,6 +21015,7 @@ CoverageClass ClassifyOpcode(ShaderOpcode opcode,
   case Opcode::DS_ADD_U32:
   case Opcode::DS_ADD_U64:
   case Opcode::DS_MAX_RTN_U64:
+  case Opcode::DS_CMPST_RTN_B32:
   case Opcode::DS_CMPST_RTN_B64:
   case Opcode::DS_ADD_RTN_U32:
   case Opcode::DS_SUB_U32:
@@ -34698,6 +34699,60 @@ TestCase DsAtomicReturn64() {
   return test;
 }
 
+TestCase DsCompareStoreReturn32() {
+  using O = ShaderOpcode;
+  TestCase test;
+  test.name = "DsCompareStoreReturn32";
+  auto &code = test.code;
+  const auto result = [&](u32 reg, u32 value) {
+    AppendStoreVgpr(&code, reg, static_cast<u32>(test.expected.size()));
+    test.expected.push_back(value);
+  };
+  struct Case { u32 before, compare, replacement; };
+  constexpr std::array cases{Case{0x12345678u, 0x12345678u, 0x9abcdef0u},
+                             Case{0x12345678u, 0x9abcdef0u, 0x12345678u},
+                             Case{0x80000000u, 0u, 0x7fffffffu},
+                             Case{UINT32_MAX, UINT32_MAX, 0u}};
+  AppendVMovU32(&code, 7, 0);
+  for (const auto &item : cases) {
+    AppendVMovLiteral(&code, 2, item.before);
+    code.push_back(EncodeDs0(0x0d, 8));
+    code.push_back(EncodeDs1(0, 2, 7));
+    AppendVMovLiteral(&code, 5, item.compare);
+    AppendVMovLiteral(&code, 19, item.replacement);
+    AppendVMovU32(&code, 4, 8);
+    // CS04b812daaacd532f pc 0x3e8: DATA compares, DATA2 replaces, VDst aliases ADDR.
+    code.insert(code.end(), {0xd8c00000u, 0x04130504u});
+    result(4, item.before);
+    code.push_back(EncodeDs0(0x36, 8));
+    code.push_back(EncodeDs1(8, 0, 7));
+    result(8, item.before == item.compare ? item.replacement : item.before);
+  }
+  AppendVMovU32(&code, 5, 0);
+  AppendVMovU32(&code, 4, 12); // Past the three LDS DWORDs.
+  code.insert(code.end(), {0xd8c00000u, 0x04130504u});
+  result(4, 0);
+  AppendVMovLiteral(&code, 2, 0x12345678u);
+  code.push_back(EncodeDs0(0x0d, 8));
+  code.push_back(EncodeDs1(0, 2, 7));
+  AppendVMovLiteral(&code, 5, 0x12345678u);
+  AppendVMovU32(&code, 4, 8);
+  code.push_back(EncodeSMovB32(126, InlineU32(0)));
+  code.insert(code.end(), {0xd8c00000u, 0x04130504u});
+  code.push_back(EncodeSMovB32(126, InlineU32(1)));
+  result(4, 8); // An inactive lane keeps its destination and leaves memory alone.
+  code.push_back(EncodeDs0(0x36, 8));
+  code.push_back(EncodeDs1(8, 0, 7));
+  result(8, 0x12345678u);
+  AppendEnd(&code);
+  test.compute_info.lds_size_dwords = 3;
+  test.opcodes = {O::V_MOV_B32, O::S_MOV_B32, O::DS_WRITE_B32, O::DS_CMPST_RTN_B32,
+                  O::DS_READ_B32, O::BUFFER_STORE_DWORD, O::S_ENDPGM};
+  test.required_spirv = {"OpAtomicCompareExchange %uint"};
+  test.forbidden_spirv = {"OpLoopMerge"};
+  return test;
+}
+
 TestCase DsAtomic64Contention(bool add, u32 wave_size) {
   using O = ShaderOpcode;
   TestCase test;
@@ -42276,6 +42331,7 @@ std::vector<TestCase> MakeCases() {
   }
   AddCase(DsAtomicReturnVariants);
   AddCase(DsAtomicReturn64);
+  AddCase(DsCompareStoreReturn32);
   for (bool add : {false, true}) {
     cases.push_back(DsAtomic64Bounds(add));
     for (u32 wave_size : {32, 64}) {
@@ -47800,7 +47856,8 @@ int main(int argc, char **argv) {
     RunCase(&vulkan, DsWideLdsPartialBounds());
     RunCase(&vulkan, DsAtomic64Bounds(false));
     RunCase(&vulkan, DsAtomic64Bounds(true));
-    for (auto test : {DsWideLdsPartialBounds(), DsAtomic64Bounds(false), DsAtomic64Bounds(true), DsAtomicReturn64()}) {
+    for (auto test : {DsWideLdsPartialBounds(), DsAtomic64Bounds(false), DsAtomic64Bounds(true), DsAtomicReturn64(),
+                      DsCompareStoreReturn32()}) {
       test.compute_info.lds_storage = true;
       std::erase(test.required_spirv, "WorkgroupMemoryExplicitLayoutKHR");
       test.required_spirv.push_back("StorageBuffer");

tests/shaderCfgTests.cpp

@@ -8119,6 +8119,24 @@ void TestDsReturningAtomic64Decoder() {
         "captured DS_CMPST_RTN_B64 lost its comparison/replacement pairs or aliases");
 }
 
+void TestDsCompareStore32Decoder() {
+  using namespace ShaderRecompiler::Decoder;
+  // CS04b812daaacd532f pc 0x3e8 (AC Valhalla, #281).
+  const uint32_t words[] = {0xd8c00000u, 0x04130504u};
+  Instruction decoded;
+  DecodeInstruction(words, 0u, decoded);
+  Check(decoded.opcode == Opcode::DS_CMPST_RTN_B32 && decoded.word_count == 2u &&
+            decoded.data_dwords == 1u && decoded.src_count == 3u && decoded.src0.reg == 4u &&
+            decoded.src1.reg == 5u && decoded.src2.reg == 19u && decoded.dst.reg == 4u &&
+            decoded.offset == 0u && !decoded.gds,
+        "captured DS_CMPST_RTN_B32 lost its comparison/replacement order or aliases");
+
+  const uint32_t gds_words[] = {words[0] | (1u << 17u), words[1]};
+  DecodeInstruction(gds_words, 0u, decoded);
+  Check(decoded.opcode == Opcode::UNSUPPORTED && decoded.word_count == 2u && decoded.gds,
+        "DS_CMPST_RTN_B32 must reject the unimplemented GDS variant");
+}
+
 void TestDsFloatMinMaxDecoder() {
   using namespace ShaderRecompiler::Decoder;
   for (bool max_value : {false, true}) {
@@ -16480,6 +16498,7 @@ int main() {
   TestScalarAshrI64Decoder();
   TestImageAtomicWidthDecoder();
   TestDsReturningAtomic64Decoder();
+  TestDsCompareStore32Decoder();
   TestDsFloatMinMaxDecoder();
   TestNewShaderDecoderArchitecture();
   TestImageAddressOperands();