Merge pull request #7325 from Cleverdee/fix/image-table-nested-addend

fix(recompiler): fold chained immediate addends of image-table record offsets

core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp

@@ -1904,6 +1904,40 @@ void verifyPointerTableOffsetCarry() {
 #endif
 }
 
+void verifyChainedTableOffset() {
+    using namespace ShaderRecompiler;
+    namespace Abi = ImageTableAbi;
+    static TableTexture textures[2];
+    alignas(256) static std::array<std::uint32_t, 0x3000> root{};
+    static std::array<std::uint32_t, 8> keys{};
+    static std::array<std::uint32_t, 4> output{};
+    const auto keysV = TableV(keys.data(), 0u, static_cast<std::uint32_t>(sizeof(keys)));
+    const auto outputV = TableV(output.data(), 0u, static_cast<std::uint32_t>(sizeof(output)));
+    std::copy(keysV.begin(), keysV.end(), root.begin());
+    std::copy(outputV.begin(), outputV.end(), root.begin() + 4);
+    const auto at = [&](std::uint32_t offset, std::array<std::uint32_t, 8> words) {
+        std::copy(words.begin(), words.end(), root.begin() + offset / 4u);
+    };
+    at(0x60u + 3u * 32u, TableT(textures[0].bytes.data(), TableFormat8888UNorm, TableType2D));
+    at(0x60u + 5u * 32u, TableT(textures[1].bytes.data(), TableFormat32UInt, TableType2D));
+    const auto base = static_cast<std::uint64_t>(reinterpret_cast<std::uintptr_t>(root.data()));
+    const std::array<std::uint32_t, 2> userData{static_cast<std::uint32_t>(base), static_cast<std::uint32_t>(base >> 32u)};
+    const std::vector<std::uint32_t> code{0xf4080200u, 0xfa000000u, 0xf4080700u, 0xfa000010u, 0xf4080800u, 0xfa000020u, 0x7e200500u, 0x8f108210u, 0xf4200444u,
+        0x20000000u, 0x8711ff11u, 0x000000ffu, 0x8f128511u, 0x8113a012u, 0x81149013u, 0xf4080900u, 0x26000040u, 0xf4080a00u, 0x28000040u, 0xf09c0f08u,
+        0x01090000u, 0xe0700000u, 0x80070000u, 0xbf810000u};
+    auto request = TableRequest(code, userData);
+    request.shader.codeAddress = 0x37000u;
+    const auto plan = GetResourcePlan(request);
+    require(plan->info.images.size() == 1u && plan->info.images[0].table == 0u, "chained table offset: a T# whose second load offset adds to the first one's was not planned as a table");
+    const auto& column = *plan->descriptorSources.at(plan->info.images[0].source).tableColumn;
+    require(column.address && column.stride == 32u && column.addend == 0x20u && column.offset == 0x40u && column.maxKey == 0xffu && column.foldedAddend == 0x20u && column.foldedCarry == 16u, "chained table offset: the column is wrong");
+    const auto compiled = CompileTable(request);
+    const auto& snapshot = compiled.capture->snapshot.tables.tables.at(0);
+    const auto& words = compiled.capture->snapshot.tables.words;
+    require(snapshot.keys == 256u && (snapshot.codes[3] & Abi::PoisonFlag) == 0u && (snapshot.codes[5] & Abi::PoisonFlag) == 0u, "chained table offset: the records were not resolved");
+    require(words.at(snapshot.codes[3]).dwords[0] == root[(0x60u + 3u * 32u) / 4u] && words.at(snapshot.codes[3]).dwords[7] == root[(0x60u + 3u * 32u) / 4u + 7u] && words.at(snapshot.codes[5]).dwords[0] == root[(0x60u + 5u * 32u) / 4u] && snapshot.codes[3] != snapshot.codes[5], "chained table offset: a record was read from the wrong offset");
+}
+
 void verifyDescriptorPhis() {
     using namespace ShaderRecompiler;
     constexpr std::uint32_t Format8888UNorm = 56;
@@ -5307,6 +5341,7 @@ int main(int argc, char** argv) {
             verifyScalarAddressRule();
             verifyPointerImageTable();
             verifyPointerTableOffsetCarry();
+            verifyChainedTableOffset();
             std::cout << "Image table keys, maps, records and registration passed\n";
             return 0;
         }
@@ -5338,6 +5373,7 @@ int main(int argc, char** argv) {
         verifyScalarAddressRule();
         verifyPointerImageTable();
         verifyPointerTableOffsetCarry();
+        verifyChainedTableOffset();
         verifyDescriptorPhis();
         verifyUniformTableImage();
         verifyLaneSelectedDescriptors();

core/shader/recompiler/Optimization/src/ResourceTracker.cpp

@@ -542,11 +542,15 @@ private:
     static IrValue* SplitImmediateAddend(IrValue* value, std::uint32_t& addend) {
         value = value->Resolve();
         addend = 0;
-        if (value->Opcode() == IrOpcode::IAdd32 && value->ArgumentCount() == 2u) {
-            if (immediateU32(value->Argument(0), addend)) return value->Argument(1)->Resolve();
-            if (immediateU32(value->Argument(1), addend)) return value->Argument(0)->Resolve();
+        while (value->Opcode() == IrOpcode::IAdd32 && value->ArgumentCount() == 2u) {
+            std::uint32_t immediate = 0;
+            IrValue* rest = nullptr;
+            if (immediateU32(value->Argument(0), immediate)) rest = value->Argument(1)->Resolve();
+            else if (immediateU32(value->Argument(1), immediate)) rest = value->Argument(0)->Resolve();
+            else break;
+            addend += immediate;
+            value = rest;
         }
-        addend = 0;
         return value;
     }