swung0x48/MobileGL
1
0
Fork 0
mirror of https://github.com/MobileGL-Dev/MobileGL synced 2026-09-08 04:08:32 +09:00
Code Issues Packages Projects Releases Wiki Activity

[Fix] (MG_State): keep GlslangToSpv ahead of buildReflection - reflection-first drops a vertex input's Location decoration and Adreno rejects the pipeline

Browse Source
This commit is contained in:
swung0x48
2026-08-10 08:33:47 -04:00
parent 8ad07c222c
commit 71e29f9d58
4 changed files with 126 additions and 125 deletions
Download Patch File Download Diff File Expand all files Collapse all files
MobileGL/MG_State/GLState/ProgramState/ProgramLinkTask.cpp
+84 -30
View File
@@ -376,33 +376,43 @@ namespace MobileGL::MG_State::GLState {
}
}
// ---- everything below this line up to GenerateSpirv() is the GL query surface ----
// ================== DO NOT REORDER WITHOUT A DEVICE GATE ==================
//
// ORDERING NOTE (rewritten 2026-08-10; the constraint it records was RETESTED, not
// dropped on a hunch). This block used to insist that SPIR-V be generated BEFORE
// buildReflection touches artifacts.program, on the grounds that reflection's
// live-variable analysis mutates the shared intermediates in ways that change
// subsequent GlslangToSpv output - "observed: catastrophic uniform misbinding on
// DirectVulkan for UBO-heavy content", recorded with commit 0d052719.
// SPIR-V MUST be generated here - after link + mapIO, BEFORE buildReflection touches
// artifacts.program. buildReflection's live-variable analysis perturbs the shared
// intermediates in ways that change subsequent GlslangToSpv output. The full history,
// because this constraint has now been doubted once and cost a release cycle:
//
// Re-measured on the glslang pin this tree vendors, with the same method 0d052719
// used (per-module SPIR-V hashes, both orders, byte-compared): 636 modules across
// 320 programs - the whole extracted trace corpus (BSL, Complementary Reimagined,
// IterationRP, Create/Flywheel) plus adversarial synthetics - came out BYTE-IDENTICAL
// in both orders, pre-optimize and post-optimize alike. glslang's code structure
// agrees: reflection.cpp performs no AST write (no getWritableType, no const_cast, no
// qualifier assignment) and GlslangToSpv takes a const TIntermediate&.
// * 0d052719 recorded it originally - "observed: catastrophic uniform misbinding on
// DirectVulkan for UBO-heavy content" - validated by per-module SPIR-V hashes over
// a full DirectVulkan replay.
// * 2026-08-10, it was re-measured with that same method and came back GREEN:
// 636 modules / 320 programs (BSL, Complementary Reimagined, IterationRP,
// Create/Flywheel, plus adversarial synthetics) byte-identical in both orders, pre-
// and post-optimize. On that evidence the order was inverted.
// * 2026-08-11, the device said otherwise. Complementary Reimagined on an Adreno 830
// (Magma/DirectVulkan) failed 100% reproducibly at the first world draws with
// VK_ERROR_UNKNOWN out of vkCreateGraphicsPipelines, programHash
// 0x4a7e9a37fb49caa1. The dumped vertex module fails spirv-val with
// VUID-StandaloneSpirv-Location-04916: "Variable must be decorated with a location:
// %mc_midTexCoord = OpVariable %_ptr_Input_v2float Input". Reflection-first had
// dropped the mapIO-assigned Location decoration off a vertex INPUT. Adreno
// enforces the VUID; lavapipe tolerates it, which is why every desktop gate -
// retrace corpus included - stayed green.
//
// So the order is now the other way round, and deliberately: reflection, fragment
// output validation and transform-feedback resolution are what the GL query surface
// is made of, and they are also the only remaining ways a link can FAIL, so running
// them first is what lets LINK_STATUS and every query behind it become final without
// waiting for SPIR-V (and stops a program that fails validation from paying for
// ~68 s/pack-load of SPIR-V generation it is about to throw away).
// Why the 636-module A/B missed it: the corpus replayed captured shader SOURCES, so it
// never reproduced Iris's glBindAttribLocation-before-link flow. The victim decoration
// is assigned by the io-resolver those bindings drive (TMglGlslIoResolver), so the
// triggering shape was simply not in the sample. A byte-identity result is only as
// strong as the flows the corpus contains - it is not a proof about the linker.
//
// What has NOT changed: the routing tables are sized and keyed by reflection results
// AND read the OPTIMIZED SPIR-V, so BuildGlobalUboRouting still runs strictly after
// both DoReflection and GenerateSpirv.
// The corollary for the two-phase split: only work that does NOT touch the
// intermediates may move off this critical path. spirv-opt and the global-UBO routing
// tables operate on the finished module words, so they live in ProgramSpirvTask;
// GlslangToSpv stays here, in front of reflection, where it has always belonged.
MGLOG_D("ProgramObject %u: Starting SPIR-V generation", in.externalIndex);
GenerateSpirv();
MGLOG_D("ProgramObject %u: Starting reflection", in.externalIndex);
// TEMP-STAGE-PROBE: "reflection" - buildReflection + the GL location assignment.
const bool tempStageProbeReflectionOk = [&] {
@@ -428,20 +438,22 @@ namespace MobileGL::MG_State::GLState {
}
// ---- past this point the link cannot fail any more ----
// Everything left is SPIR-V work, and it belongs to phase B. Hand it what it needs
// and stop: from the join's point of view this program is now fully linked.
// What is left is spirv-opt and the routing tables, and neither touches a glslang
// intermediate - they work on the finished module words. Hand them over and stop:
// from the join's point of view this program is now fully linked.
//
// The TShaders move rather than copy - `attrib` borrowed them into the TProgram as
// raw pointers and this node is now their owner of record, for as long as phase B
// (which holds this node) needs the intermediates hanging off them.
spirvHandoff.shaders = Move(attrib.shaders);
// Note what is NOT in the handoff any more: the TShaders. GlslangToSpv ran above, in
// this body, so the parsed ASTs die with `attrib` when this function returns, exactly
// as they did before the split. That erases both the borrowed-intermediate lifetime
// question and the peak-RSS arena backlog the earlier cut had to reason about.
// (spirvHandoff.rawSpirv was filled by GenerateSpirv() above.)
spirvHandoff.shaderTypes.resize(in.shaders.size());
for (SizeT i = 0; i < in.shaders.size(); i++) {
spirvHandoff.shaderTypes[i] = MG_Util::ConvertShaderStageToGLEnum(in.shaders[i].stage);
}
// Copied, not referenced: `artifacts` is MOVED out of this node by the join, and
// phase B runs after that. Measured at ~20 us per program, which is noise against the
// ~450 ms phase B spends on the same program.
// ~400 ms phase B spends on the same program.
spirvHandoff.reflection.program = artifacts.program;
spirvHandoff.reflection.uniformLocations = artifacts.uniformLocations;
spirvHandoff.reflection.uniformIndexInTProgram = artifacts.uniformIndexInTProgram;
@@ -452,6 +464,48 @@ namespace MobileGL::MG_State::GLState {
spirvHandoff.shaderTypes.size());
}
// Raw GlslangToSpv only - one module per attached stage, unoptimized. Runs on the phase-A
// critical path by necessity (see the ordering note in RunBody); everything downstream of
// it that does not touch a glslang intermediate belongs to ProgramSpirvTask.
void ProgramLinkTask::GenerateSpirv() {
/* As we passed first stage compilation/linking,
* we'll assume all the operations here should
* pass. We may be able to employ some optimizations
* here without the burden of error reporting.
*/
using namespace MG_Util::ShaderTranspiler;
MGLOG_D("ProgramObject %u: GenerateSpirv - start", in.externalIndex);
// The shaders were parsed once, in the link-compatible (relaxed Vulkan-rules)
// configuration, and artifacts.program linked those parses - so artifacts.program IS
// the program the backends consume. Generate SPIR-V straight from its intermediates.
Vector<GLenum> shaderTypes(in.shaders.size());
for (SizeT i = 0; i < in.shaders.size(); i++) {
shaderTypes[i] = MG_Util::ConvertShaderStageToGLEnum(in.shaders[i].stage);
}
ProgramBinaryAttrib binaryAttrib{
.shaderTypes = shaderTypes,
.program = *artifacts.program,
};
MGLOG_D("ProgramObject %u: GenerateSpirv - requesting SPIR-V binary from program", in.externalIndex);
// TEMP-STAGE-PROBE: "spirv-gen" - GlslangToSpv for every stage of this program.
auto binaryResult = [&] {
const MG_Util::Debug::TempStageProbeScope tempStageProbeSpirvGen(
MG_Util::Debug::kTempStageProbeSpirvGen);
return ShaderCompiler::GetSpirvBinaryFromProgram(binaryAttrib);
}();
if (!binaryResult) {
DeferLog(std::format("ProgramObject {}: GenerateSpirv - GetSpirvBinaryFromProgram failed",
in.externalIndex));
}
MOBILEGL_ASSERT(binaryResult, "GetSpirvBinaryFromProgram failed");
if (!binaryResult) return;
spirvHandoff.rawSpirv = Move(binaryResult.value());
MGLOG_D("ProgramObject %u: GenerateSpirv - generated %zu SPIR-V modules", in.externalIndex,
spirvHandoff.rawSpirv.size());
}
Bool ProgramLinkTask::ConsumeShaders(Vector<SharedPtr<glslang::TShader>>& outShaders) {
outShaders.assign(in.shaders.size(), nullptr);
MobileGL/MG_State/GLState/ProgramState/ProgramLinkTask.h
+19 -20
View File
@@ -36,9 +36,13 @@ namespace MobileGL::MG_State::GLState {
//
// Every one of the eight ways a link can fail lives here, so once this node has published
// through EnsureLinkJoined() the program's LINK_STATUS, info log and entire query surface
// are FINAL and truthful. SPIR-V generation, spirv-opt and the global-UBO routing tables
// moved to ProgramSpirvTask, which chains behind this node and is joined by only five
// getters (see ProgramObject::EnsureSpirvJoined).
// are FINAL and truthful. spirv-opt and the global-UBO routing tables moved to
// ProgramSpirvTask, which chains behind this node and is joined by only five getters (see
// ProgramObject::EnsureSpirvJoined).
//
// GlslangToSpv itself does NOT move: it has to run before buildReflection, and the reason
// is a device-confirmed correctness constraint rather than a preference. Read the ordering
// note in RunBody before touching the sequence.
//
// Same ownership rule as ShaderCompileTask: the body reads nothing but `in` (all of it
// owned or immutable) and writes nothing but `artifacts`. No GL call, no
@@ -85,24 +89,17 @@ namespace MobileGL::MG_State::GLState {
// and this node's JobState, and nothing else on it. Reading `artifacts` or
// `diagnostics` from phase B would race the publish.
struct SpirvHandoff {
// MANDATORY, and the reason this struct exists at all: TProgram::addShader stores
// a RAW TShader*, and for the one-shader-per-stage case getIntermediate() returns
// the TShader's own intermediate rather than a copy. These used to die when
// RunBody() returned, which was safe only because nothing called getIntermediate()
// afterwards. GlslangToSpv does exactly that, so phase B has to own them.
// The RAW GlslangToSpv output, one module per attached stage. Phase A generates it
// (it has to - see the ordering note in RunBody) and hands over the words; phase B
// optimizes them in place and reflects the result.
//
// MEMORY NOTE: this is the one thing the split makes live LONGER than it used to -
// a glslang arena per stage, megabytes for a shaderpack, now alive from the end of
// phase A until phase B runs instead of dying with the link body, so a deep
// phase-B backlog holds one arena per queued program. Phase B clears this vector
// as soon as GlslangToSpv returns, but read that call site's comment before
// relying on it: for the COMMON case (a shader linked into exactly one program)
// the compile node co-owns the same TShader and phase A pins that node, so the
// clear frees nothing and only the re-parsed CAS-loser shaders are actually
// released. If peak RSS ever becomes the binding constraint on a pack load, THIS
// is the field to attack - by bounding the backlog, by releasing the compile
// node's own reference at claim time, or by moving GlslangToSpv back into phase A.
Vector<SharedPtr<glslang::TShader>> shaders;
// Deliberately the module WORDS and not the parsed ASTs. An earlier cut passed the
// Vector<SharedPtr<glslang::TShader>> instead, because TProgram::addShader stores
// raw TShader* and borrows their intermediates - which meant the glslang arenas
// (megabytes per shaderpack program) had to stay alive across the whole phase-B
// backlog. Generating in phase A means they die with the link body exactly as they
// always did, and phase B holds nothing but words.
Vector<Vector<unsigned>> rawSpirv;
// GL enum per entry of `in.shaders`, in the same order (GetSpirvBinaryFromProgram
// walks it to pick the intermediates).
Vector<GLenum> shaderTypes;
@@ -145,6 +142,8 @@ namespace MobileGL::MG_State::GLState {
Bool ValidateFragmentOutputLocations();
Bool ResolveTransformFeedbackVaryings();
void ResolveGsTriangleStripCapture(const glslang::TIntermediate* captureIntermediate);
// Raw GlslangToSpv, into spirvHandoff.rawSpirv. Must run before DoReflection.
void GenerateSpirv();
// Worker-side MGLOG replacement: appended to diagnostics.logLines and replayed by the
// join, on the GL thread, where a serial implementation would have printed it.
MobileGL/MG_State/GLState/ProgramState/ProgramSpirvTask.cpp
+12 -67
View File
@@ -74,16 +74,14 @@ namespace MobileGL::MG_State::GLState {
// ProgramLinkTask::RunBody - no GL/EGL call, no pActiveBackendObject read, no
// pGLContext->RecordError().
void ProgramSpirvTask::RunBody() {
// glslang leaves this worker's TLS pool allocator pointing at the last arena it
// touched; reset it on the way out so an unrelated later job cannot allocate out of a
// pool that has since been freed. Declared FIRST so it is destroyed LAST - the phase-A
// release below drops the TShaders (and their pools) and must happen inside it.
// spirv-tools and SPIRV-Cross do not use glslang's pool allocator, but the guard is
// kept: it is cheap, and it keeps the "a body never leaves this thread's allocator
// pointing at someone else's arena" rule uniform across both task bodies.
const GlslangThreadAllocatorGuard glslangGuard;
using namespace MG_Util::ShaderTranspiler;
// Drop phase A - and with it the TShaders, the TProgram reference and phase A's whole
// input snapshot - the moment this body is done, rather than at some later join. For a
// pack load that is the difference between W glslang arenas alive and all of them.
// Drop phase A - the module words are moved out below, and phase A's input snapshot
// (its compile nodes, its sources) has no reader here - the moment this body is done.
struct PhaseAReleaser {
SharedPtr<ProgramLinkTask>& node;
~PhaseAReleaser() { node.reset(); }
@@ -107,35 +105,12 @@ namespace MobileGL::MG_State::GLState {
const MG_Util::Debug::TempStageProbeScope tempStageProbeSpirvTask(
MG_Util::Debug::kTempStageProbeSpirvTaskTotal);
MGLOG_D("ProgramObject %u: Starting SPIR-V generation", externalIndex);
GenerateSpirv(handoff, externalIndex);
// GlslangToSpv was the only consumer of the parsed ASTs; everything after this point
// works on the SPIR-V and on the TProgram's own self-contained reflection pool. Drop
// them here rather than at the end of the body, which is ~87% of this node's runtime
// earlier (spirv-opt plus routing).
//
// WHAT THIS ACTUALLY FREES, precisely - it is LESS than "the glslang arenas", and the
// difference matters for the peak-RSS story:
// * CAS-LOSER shaders (the re-parse in ShaderCompileTask::ClaimParsedShader, i.e.
// the 2nd..Nth link of a shared shader): freed here in full. The handoff is their
// ONLY owner.
// * CAS-WINNER shaders (the common case - one shader object linked into one
// program, which is every program of an Iris pack load): NOT freed here. The
// winner branch returns a COPY of ShaderCompileTask::artifacts.shader
// (ShaderCompileTask.cpp:320) and the node never releases its own reference, while
// phase A holds that node through in.shaders[i].compiled for its whole life - and
// phase A lives until PhaseAReleaser fires at the end of this body. So the
// refcount goes 2 -> 1 here and the arena dies where it would have died anyway.
//
// Making it free the winner's arena too means releasing whatever pins the TShader
// inside the compile node, and neither obvious route is safe as a drive-by: moving out
// of artifacts.shader at claim time races ShaderObject::GetCompiledShader() on the GL
// thread and breaks JobNode's "a terminal node is immutable" invariant, and dropping
// phase A's in.shaders[i].compiled reference only helps when nothing else holds the
// node (the adoption map is a WeakPtr index, so it would also change which nodes stay
// adoptable). Both belong in a change that can be reviewed against the consume-once
// and adoption semantics on their own terms.
handoff.shaders.clear();
// Phase A already produced these (it must - GlslangToSpv has to run ahead of
// buildReflection; see ProgramLinkTask::RunBody's ordering note). Take the words.
artifacts.generatedSpirv = Move(handoff.rawSpirv);
MGLOG_D("ProgramObject %u: optimizing %zu SPIR-V module(s)", externalIndex,
artifacts.generatedSpirv.size());
OptimizeSpirv(externalIndex);
MGLOG_D("ProgramObject %u: Building global-UBO routing tables", externalIndex);
{
@@ -148,38 +123,8 @@ namespace MobileGL::MG_State::GLState {
artifacts.generatedSpirv.size());
}
void ProgramSpirvTask::GenerateSpirv(const ProgramLinkTask::SpirvHandoff& handoff, const Uint externalIndex) {
/* As we passed first stage compilation/linking,
* we'll assume all the operations here should
* pass. We may be able to employ some optimizations
* here without the burden of error reporting.
*/
void ProgramSpirvTask::OptimizeSpirv(const Uint externalIndex) {
using namespace MG_Util::ShaderTranspiler;
MGLOG_D("ProgramObject %u: GenerateSpirv - start", externalIndex);
// The shaders were parsed once, in the link-compatible (relaxed Vulkan-rules)
// configuration, and the handoff's program linked those parses - so it IS the program
// the backends consume. Generate SPIR-V straight from its intermediates, which the
// handoff's TShaders keep alive.
ProgramBinaryAttrib binaryAttrib{
.shaderTypes = handoff.shaderTypes,
.program = *handoff.reflection.program,
};
MGLOG_D("ProgramObject %u: GenerateSpirv - requesting SPIR-V binary from program", externalIndex);
// TEMP-STAGE-PROBE: "spirv-gen" - GlslangToSpv for every stage of this program.
auto binaryResult = [&] {
const MG_Util::Debug::TempStageProbeScope tempStageProbeSpirvGen(
MG_Util::Debug::kTempStageProbeSpirvGen);
return ShaderCompiler::GetSpirvBinaryFromProgram(binaryAttrib);
}();
if (!binaryResult) {
DeferLog(std::format("ProgramObject {}: GenerateSpirv - GetSpirvBinaryFromProgram failed", externalIndex));
MOBILEGL_ASSERT(binaryResult, "GetSpirvBinaryFromProgram failed");
return; // spirvStatus stays false: linked, but not drawable.
}
artifacts.generatedSpirv = Move(binaryResult.value());
MGLOG_D("ProgramObject %u: GenerateSpirv - generated %zu SPIR-V modules", externalIndex,
artifacts.generatedSpirv.size());
// Linked SPIR-V generated, sanitize and optimize it
{
MobileGL/MG_State/GLState/ProgramState/ProgramSpirvTask.h
+11 -8
View File
@@ -12,8 +12,8 @@
#include <MG_Util/Async/JobNode.h>
namespace MobileGL::MG_State::GLState {
// PHASE B of one glLinkProgram: GlslangToSpv, spirv-opt, and the SPIRV-Cross pass that
// builds the glUniform*-to-scratch routing tables. Chained behind exactly one
// PHASE B of one glLinkProgram: spirv-opt and the SPIRV-Cross pass that builds the
// glUniform*-to-scratch routing tables. Chained behind exactly one
// ProgramLinkTask and joined by exactly five ProgramObject getters (GetGeneratedSpirv,
// GetUniformOffset, MapUBO, GetUBOData, GetUBOSize), so ~120 other getters and the whole
// GL query surface stay on the phase-A gate and answer without waiting for any of this.
@@ -28,11 +28,14 @@ namespace MobileGL::MG_State::GLState {
// synchronization - covers everything except the two members the join consumes.)
//
// ---- lifetime ----
// The handoff owns the Vector<SharedPtr<glslang::TShader>>, and that is mandatory rather
// than tidy: glslang::TProgram stores raw TShader* and, for the one-shader-per-stage case,
// BORROWS each stage's TIntermediate from its TShader. GlslangToSpv reads exactly those
// intermediates. Before the split the shaders died when ProgramLinkTask::RunBody returned,
// which was safe only because nothing called getIntermediate() afterwards.
// This node touches no glslang intermediate at all. GlslangToSpv runs in PHASE A, because
// it has to precede buildReflection (a device-confirmed constraint - see the ordering note
// in ProgramLinkTask::RunBody), so the handoff carries finished module WORDS and the parsed
// ASTs die with the link body exactly as they always did. The only glslang object reachable
// from here is the TProgram, held by SharedPtr for the reflection the routing pass reads -
// and that reflection is a deep clone in the TProgram's own pool (TObjectReflection's ctor
// does type(pType.clone()) with the allocator pointed at that pool), so it is independent
// of the TShaders' lifetime.
//
// ---- failure ----
// A cancel (relink, teardown, program destruction) or an optimizer failure publishes
@@ -65,7 +68,7 @@ namespace MobileGL::MG_State::GLState {
private:
void RunBody() override;
void GenerateSpirv(const ProgramLinkTask::SpirvHandoff& handoff, Uint externalIndex);
void OptimizeSpirv(Uint externalIndex);
void BuildGlobalUboRouting(const ProgramLinkTask::SpirvHandoff& handoff, Uint externalIndex);
// Worker-side MGLOG replacement, replayed by the join on the GL thread. Same reason as