From bed19e5d8ce641663626cdbb73d43d64c250f67b Mon Sep 17 00:00:00 2001
From: nmzik <Nmzik@mail.ru>
Date: Sun, 23 Aug 2026 03:45:57 +0200
Subject: [PATCH] shader: preserve concurrent subword buffer stores

Co-authored-by: Senaxx <Senaxx@users.noreply.github.com>
---
 .../backend/spirv/spirvEmitterMemory.cpp      |  6 ++--
 tests/ShaderRecompilerComputeTests.cpp        | 30 +++++++++++++++++++
 2 files changed, 33 insertions(+), 3 deletions(-)

diff --git a/src/graphics/shader/recompiler/backend/spirv/spirvEmitterMemory.cpp b/src/graphics/shader/recompiler/backend/spirv/spirvEmitterMemory.cpp
index 22216af91..10baff4e1 100644
--- a/src/graphics/shader/recompiler/backend/spirv/spirvEmitterMemory.cpp
+++ b/src/graphics/shader/recompiler/backend/spirv/spirvEmitterMemory.cpp
@@ -436,12 +436,12 @@ void StoreSubwordInBounds(ValueEmitContext& ctx, const IR::MemoryInfo& mem,
 		                     Unary(ctx.state, OpNot, TypeU32(ctx.state), mask)),
 		              value);
 	};
-	if (mem.kind == IR::ResourceKind::Lds || mem.kind == IR::ResourceKind::Gds) {
-		AtomicUpdate(ctx.state, pointer, mem.kind, merge);
-	} else {
+	if (mem.kind == IR::ResourceKind::Scratch) {
 		const auto old = ctx.state.builder.AllocateId();
 		ctx.state.builder.AddFunction({OpLoad, TypeU32(ctx.state), old, pointer});
 		ctx.state.builder.AddFunction({OpStore, pointer, merge(old)});
+	} else {
+		AtomicUpdate(ctx.state, pointer, mem.kind, merge);
 	}
 }
 
diff --git a/tests/ShaderRecompilerComputeTests.cpp b/tests/ShaderRecompilerComputeTests.cpp
index ab08761df..67ef05856 100644
--- a/tests/ShaderRecompilerComputeTests.cpp
+++ b/tests/ShaderRecompilerComputeTests.cpp
@@ -16112,6 +16112,35 @@ TestCase BufferStoreFormatXResource16UintWritesHalfword() {
   return test;
 }
 
+TestCase BufferStoreFormatXResource16UintPreservesAdjacentLanes() {
+  using O = ShaderOpcode;
+
+  std::vector<u32> code;
+  code.push_back(EncodeVop2(0x25, 1, InlineU32(1), 0));
+  code.push_back(EncodeMubuf0(0x04u, 0, true, false));
+  code.push_back(EncodeMubuf1(1, 0, 0));
+  AppendEnd(&code);
+
+  TestCase test;
+  test.name = "BufferStoreFormatXResource16UintPreservesAdjacentLanes";
+  test.code = code;
+  test.initial = std::vector<u32>(32, 0);
+  for (u32 index = 0; index < 32; index++) {
+    test.expected.push_back(((index * 2 + 2) << 16) | (index * 2 + 1));
+  }
+  test.opcodes = {O::V_ADD_NC_U32, O::BUFFER_STORE_FORMAT_X, O::S_ENDPGM};
+  test.compute_info.threads_num[0] = 64;
+  test.compute_info.threads_num[1] = 1;
+  test.compute_info.threads_num[2] = 1;
+  test.compute_info.thread_ids_num = 1;
+  test.compute_info.wave_size = 64;
+  test.has_compute_info = true;
+  test.user_data = MakeStructuredStorageBufferData(2, 64, false, 11);
+  test.has_user_data = true;
+  test.required_spirv = {"OpAtomicLoad", "OpAtomicCompareExchange"};
+  return test;
+}
+
 TestCase BufferLoadFormatXyResource88UintExtractsBytes() {
   using O = ShaderOpcode;
 
@@ -19742,6 +19771,7 @@ std::vector<TestCase> MakeCases() {
   AddCase(BufferLoadFormatXyzwInactiveExecPreservesOverlappingAddress);
   AddCase(BufferFormatStoreVariants);
   AddCase(BufferStoreFormatXResource16UintWritesHalfword);
+  AddCase(BufferStoreFormatXResource16UintPreservesAdjacentLanes);
   AddCase(BufferLoadFormatXResource8UintZeroExtendsByte);
   AddCase(BufferLoadFormatXyResource88UintExtractsBytes);
   AddCase(BufferLoadFormatXyResource8888UnormConvertsFirstTwoComponents);