[Merge] (DirectGLES, GLState, ShaderTranspiler): land GL43 wave5 with its two new passes inside the L2 boundary

MobileGL/MG_State/GLState/ProgramState/ProgramLinkTask.cpp

@@ -1313,6 +1313,24 @@ namespace MobileGL::MG_State::GLState {
                 // span is left; grow the table instead of leaving the uniform without
                 // a location (which would make it unsettable via glUniform*).
                 const SizeT base = artifacts.uniformIndexInTProgram.size();
+                // The growth stops at the pool GL advertises. GL 4.6 core 7.6.1 bounds every
+                // uniform location by GL_MAX_UNIFORM_LOCATIONS, and the conformance suite reads a
+                // returned location >= the advertised maximum as a failure outright
+                // (KHR-GLES31.explicit_uniform_location.uniform-loc-mix-with-implicit-max). Minting
+                // 4095, 4096, ... is strictly worse than refusing: those are locations no
+                // application may legally name and no later query can make legal, so they would
+                // only turn a link-time exhaustion into a silently unwritable uniform. Unreachable
+                // for any program that fits glslang's per-stage uniform-component limits - it takes
+                // a fragmented pool of thousands of explicitly-located slots to get here.
+                if (base + static_cast<SizeT>(locationSpan) > kMaxUniformLocations) {
+                    artifacts.infoLog = std::format(
+                        "Uniform locations exhausted: '{}' needs {} location(s) and no free span is left below "
+                        "GL_MAX_UNIFORM_LOCATIONS ({}).",
+                        uniform.name, locationSpan, kMaxUniformLocations);
+                    DeferLog(std::format("ProgramObject {}: Link failed - {}", in.externalIndex, artifacts.infoLog));
+                    ProgramObject::ResetLinkArtifacts(artifacts);
+                    return false;
+                }
                 artifacts.uniformIndexInTProgram.resize(base + locationSpan,
                                                         glslang::TQualifier::layoutLocationEnd);
                 artifacts.uniformSamplerOrImageUnitIndex.resize(base + locationSpan, -1);

MobileGL/MG_State/GLState/ProgramState/ProgramObject.h

@@ -281,6 +281,9 @@ namespace MobileGL::MG_State::GLState {
         }
 
         GLenum GetActiveUniformType(Uint index) const {
+            // The lowered counter is a plain uint inside a synthesized block; what the GL
+            // client declared - and what glGetActiveUniform must report - is an atomic_uint.
+            if (IsActiveUniformAtomicCounter(index)) return GL_UNSIGNED_INT_ATOMIC_COUNTER;
             return UniformAt(TProgramUniformIndex(index)).glDefineType;
         }
 
@@ -298,10 +301,57 @@ namespace MobileGL::MG_State::GLState {
         }
 
         Int GetActiveUniformBlockIndex(Uint index) const {
+            // An atomic counter is a DEFAULT-BLOCK uniform to GL, whatever block the
+            // transpiler lowered it onto (GL 4.6 core 7.6, table 7.6): -1.
+            if (IsActiveUniformAtomicCounter(index)) return -1;
             // Members of the synthesized global UBO are default-block uniforms to GL: -1.
             return GlBlockIndexFromTProgram(UniformAt(TProgramUniformIndex(index)).index);
         }
 
+        // The transpiler lowers every atomic_uint onto a synthesized gl_AtomicCounterBlock_N
+        // block, but GL keeps seeing an atomic counter as a default-block uniform of type
+        // GL_UNSIGNED_INT_ATOMIC_COUNTER that points at an atomic-counter BUFFER. These two
+        // answer for that GL-level declaration; without them the query surface reports the
+        // lowering instead (GL_UNSIGNED_INT, block index 0) and
+        // KHR-GL43.shader_atomic_counters.basic-program-query fails on both.
+        //
+        // The returned value is an index into the GL_ACTIVE_ATOMIC_COUNTER_BUFFERS list, i.e.
+        // the RANK of the owning counter block among the counter blocks in glslang's block
+        // order - exactly how ProgramInterface numbers the GL_ATOMIC_COUNTER_BUFFER
+        // resources glGetActiveAtomicCounterBufferiv answers from. -1 when this uniform is
+        // not an atomic counter.
+        // Answered from the OWNED reflection snapshot, never from Artifacts().program. This
+        // arrived reading the live TProgram, which is null for every program served from the
+        // translation cache's L1 - and unlike the other query-surface accessors that made the
+        // same mistake, this one DEREFERENCES it, so the second program built from a given set
+        // of sources would have taken the process down rather than answered wrongly. The
+        // snapshot carries the same three facts in the same TPROGRAM index space:
+        // getUniform(i).index -> UniformAt(i).index, getNumUniformBlocks() ->
+        // blockReflection.size(), getUniformBlock(i).name -> BlockAt(i).name.
+        Int GetActiveUniformAtomicCounterBufferIndex(Uint index) const {
+            const Int tIndex = TProgramUniformIndex(index);
+            if (tIndex < 0) return -1;
+            const Int owner = UniformAt(tIndex).index;
+            if (owner < 0) return -1;
+            const Int blockCount = static_cast<Int>(Artifacts().blockReflection.size());
+            if (owner >= blockCount) return -1;
+            const SizeT prefixLength = StringView(MG_Util::ShaderTranspiler::ATOMIC_COUNTER_BLOCK_PREFIX).size();
+            Int counterBufferIndex = 0;
+            for (Int i = 0; i < blockCount; ++i) {
+                const auto& blockName = BlockAt(i).name;
+                if (blockName.compare(0, prefixLength, MG_Util::ShaderTranspiler::ATOMIC_COUNTER_BLOCK_PREFIX) != 0) {
+                    continue;
+                }
+                if (i == owner) return counterBufferIndex;
+                ++counterBufferIndex;
+            }
+            return -1;
+        }
+
+        Bool IsActiveUniformAtomicCounter(Uint index) const {
+            return GetActiveUniformAtomicCounterBufferIndex(index) >= 0;
+        }
+
         // GL_UNIFORM_OFFSET: byte offset within the owning named block; -1 for a default-block
         // uniform. The relaxed parse gives global-UBO members real byte offsets, but GL must keep
         // seeing them as default-block uniforms, so gate on the GL-visible block index.

MobileGL/MG_State/GLState/ProgramState/ShaderCompileTask.cpp

@@ -195,6 +195,13 @@ namespace {
             return result;
         }
 
+        if (const std::optional<String> counterOffsetError =
+                FindAtomicCounterOffsetViolation(result.preprocessedSource)) {
+            result.outcome = ShaderPreprocessOutcome::AtomicCounterOffsetRejected;
+            result.infoLog = *counterOffsetError;
+            return result;
+        }
+
         // The parse this feeds runs in the link-compatible configuration (Vulkan-client
         // env with relaxed rules): the TShader it produces is what glLinkProgram links and
         // what the backends' SPIR-V is generated from - there is no second, GL-client

MobileGL/MG_State/GLState/ProgramState/ShaderPreprocessCache.h

@@ -29,6 +29,9 @@ namespace MobileGL::MG_State::GLState {
         // FindShaderStorageBindingViolation rejected it: a storage block declared a binding at or
         // past GL_MAX_SHADER_STORAGE_BUFFER_BINDINGS.
         ResourceBindingRejected,
+        // FindAtomicCounterOffsetViolation rejected it: an atomic counter declared a
+        // layout(offset =) that is misaligned or reaches past GL_MAX_ATOMIC_COUNTER_BUFFER_SIZE.
+        AtomicCounterOffsetRejected,
         // The source-only half was clean but glslang rejected the preprocessed source.
         // Memoizing this saves the parse itself on every later object with that source.
         ParseFailed,

MobileGL/MG_State/GLState/RenderState/RenderState.cpp

@@ -843,6 +843,21 @@ namespace MobileGL {
                     stored = box;
                     stateChanged = true;
                 }
+                // "The application has written this rectangle" is a DIFFERENT predicate from "the
+                // value moved", and the backends need the first one: glScissor(0, 0, 0, 0) as the
+                // very first scissor call leaves every stored box byte-identical to its
+                // never-written default, and that call is precisely the one whose meaning a
+                // backend must stop guessing at (see ScissorBoxWrittenMask).
+                //
+                // The transition has to count as a state change for the version too. DirectGLES'
+                // SyncRenderState early-outs on an unchanged render-state version BEFORE it
+                // reaches the span memcmp that would otherwise notice the mask, so a version-less
+                // flag flip would sit in the parameter block and never be pushed. It is a
+                // once-per-index transition, so the steady state still costs nothing.
+                if (m_parameters.ScissorBoxWrittenMask != kAllViewportsMask) {
+                    m_parameters.ScissorBoxWrittenMask = kAllViewportsMask;
+                    stateChanged = true;
+                }
                 if (stateChanged) ++m_version;
             }
 
@@ -855,9 +870,15 @@ namespace MobileGL {
                     MOBILEGL_ASSERT(false, "Scissor box index out of range: %u", index);
                     return;
                 }
-                if (m_parameters.ScissorBoxes[index] == box) return;
+                // See SetScissorBox: a first write is state even when it does not move the value,
+                // so the unchanged-value early-out may only fire once this index is already
+                // marked written.
+                const Uint32 writtenBit = 1u << index;
+                const Bool alreadyWritten = (m_parameters.ScissorBoxWrittenMask & writtenBit) != 0;
+                if (alreadyWritten && m_parameters.ScissorBoxes[index] == box) return;
 
                 m_parameters.ScissorBoxes[index] = box;
+                m_parameters.ScissorBoxWrittenMask |= writtenBit;
                 ++m_version;
             }
 

MobileGL/MG_State/GLState/RenderState/RenderState.h

@@ -328,6 +328,18 @@ namespace MobileGL {
         // turns it into a real glEnable/glDisable.
         Uint32 ScissorTestEnabledMask = 0;
         Array<IntVec4, MAX_VIEWPORTS> ScissorBoxes{}; // x, y, width, height
+        // One bit per viewport, set the first time the application writes that index's scissor
+        // rectangle - glScissor broadcasts and sets all 16, glScissorIndexed/glScissorArrayv set
+        // the indices they name. It exists because the RECTANGLE cannot answer "has the
+        // application spoken?": ScissorBoxes starts all-zero (its spec initial value is the size
+        // of a window the frontend does not know yet, see the RenderState constructor), and
+        // glScissor(0, 0, 0, 0) is a legal GL state meaning "the scissor test rejects every
+        // fragment". A backend that reads an empty rectangle as the never-written sentinel
+        // therefore INVERTS that request into "accept every fragment"; DirectGLES did exactly
+        // that and KHR-GL43.viewport_array.scissor_zero_dimension caught it. Deliberately beside
+        // ScissorBoxes so it shares their tail span (after LogicOp) and DirectGLES' span memcmp
+        // picks a transition up like any other state.
+        Uint32 ScissorBoxWrittenMask = 0;
         // glEnable(GL_CLIP_DISTANCE0 + i) for i in [0, 8), one bit each. A bitmask rather than
         // eight bools because every consumer wants the set, not an individual flag, and because
         // the SYNC_CAPABILITY/SET_CAPABILITY macros key off a "<Name>Enabled" field name that