Draft
Conversation
yuri91
force-pushed
the
as/codegen
branch
2 times, most recently
from
00178ad to
b669cdd
Compare
yuri91
force-pushed
the
as/codegen
branch
2 times, most recently
from
036ff2f to
fbcc4b5
Compare
yuri91
force-pushed
the
as/codegen
branch
2 times, most recently
from
a12703a to
c89941f
Compare
Introduce a small helper that maps a data address space to the
corresponding function address space:
Default -> Default
GenericJS -> Client (functions live in the Client AS on the
JS side, not in GenericJS)
Wasm -> Wasm
Available in both a strongly-typed CheerpAS overload and an
`unsigned` overload. Used by the coroutine and later by
ItaniumCXXABI / LowerGlobalDtors / CodeGenModule / the asan
path to pick the correct AS for function pointers.
The helper wraps Function::Create with an explicit Cheerp AS, since Module::getOrInsertFunction picks the module's default AS.
CheerpNativeRewriterPass rewrites C++ `new ClientType(...)` patterns in IR. For client-transparent JS host types it turns the allocation into a stack slot (alloca of pointer-to-T) that the constructor writes into and the caller loads from; for non-transparent cases it emits a synthetic `cheerpCreate<Name>` wrapper that invokes the constructor and returns `this`. Thread Cheerp address spaces through both paths: - rewriteNativeAllocationUsers: the alloca lives in GenericJS (the slot is JS-side stack), and its element type is `T*` in the Client AS (the pointer it holds references a client-side object). - rewriteIfNativeConstructorCall: the bitcast feeding the store into that slot takes the GenericJS AS to match the alloca. - getReturningConstructor: create the synthetic `cheerpCreate<Name>` wrapper in the Client AS via cheerp::getOrCreateFunction, rather than the default AS from M.getOrInsertFunction. TODO (ordering, Phase 2): this commit calls cheerp::getOrCreateFunction, whose definition currently lives in a later commit (FreeAndDeleteRemoval: make __genericjs__free trampoline AS-aware). Reorder so the helper is introduced before its first use.
Clang was returning LangAS::Default for string/constant literals in Cheerp mode, and the matching section-based "asmjs" detection in LLVM's IRBuilder fell back to the default AS whenever a caller didn't have a section set. Thread Cheerp address spaces through every site that materializes a constant string: - Clang CodeGenModule::GetGlobalConstantAddressSpace: return LangAS::cheerp_wasm for cheerp-wasm targets, LangAS::cheerp_genericjs otherwise. - Clang CodeGenModule::castStringLiteralToDefaultAddressSpace: skip the default-AS cast in Cheerp mode; the AS is already correct. - Clang CodeGenModule::GenerateStringLiteral: override AddrSpace to CheerpAS::Wasm on cheerp-wasm so the produced GlobalVariable lands in the wasm AS even on paths where the numeric AS derived from the LangAS lookup differs. - LLVM IRBuilderBase::CreateGlobalString: detect "this string belongs in asmjs" from the AS argument (CheerpAS::Wasm + cheerp-wasm triple) instead of inspecting the basic block's parent function section — drives section from AS, not the reverse. - LLVM OMPIRBuilder::getOrCreateSrcLocStr: pick the Cheerp AS by target triple when creating OpenMP source-location strings (previously always AS 0). - LLVM SimplifyLibCalls: thread the source string's AS into CreateGlobalString/CreateGlobalStringPtr in optimizeStringNCpy and optimizePrintFString so generated scratch strings inherit the caller's AS.
Prepare CGExpr/CGExprScalar for typed-ptrcast on AS boundaries:
- EmitStoreOfScalar: if the value's AS doesn't match the address
element type's AS, insert a cast ("as_decay") so the store is
well-typed. Currently uses Builder.CreateAddrSpaceCast; TODO:
migrate to cheerp_typed_ptrcast (the preferred form used by the
later CGCall boundary code) so the AS transition is explicit in
IR and the upstream analyses see a single canonical intrinsic
rather than a mix of addrspacecast and typed_ptrcast.
- VisitCastExpr: don't trip the "different address spaces" assert
in Cheerp mode; emit PointerBitCastOrAddrSpaceCast in the
fallback path. Same TODO applies — prefer cheerp_typed_ptrcast
over a plain addrspacecast once the call sites are consistent.
- GetUserCastIntrinsic: take the concrete llvm::Type of the source
pointer so the intrinsic's src type preserves the caller's AS
instead of being re-derived from the QualType.
- EmitPointerWithAlignment / EmitCastLValue: thread the source
pointer's LLVM type into GetUserCastIntrinsic.
Thread Cheerp address spaces through blocks codegen:
- _NSConcrete{Global,Stack}Block in correct AS
- block descriptor GV + BlockDescriptorType pointer AS
- generated block invoke / copy-helper / destroy-helper functions
get the right AS and an asmjs section when targeting wasm
- pointer cast to generic block literal uses the right AS
- arrangeBuiltinFunction{Call,Declaration} threads isCheerpWasm
- DeclClonePragmaWeak propagates the Cheerp mode attr
TODO: split out the sanitizer ctor/dtor AS changes
(AddressSanitizer.cpp, ModuleUtils.cpp) into their own commit;
they were drive-by and are not about blocks.
We just do the bare minimum to make the tests pass
Replace plain addrspacecast with the cheerp_typed_ptrcast intrinsic on the boundaries where a call-site pointer AS can differ from the callee's expected AS: - CGCall: call arguments and direct-return values now go through cheerp_typed_ptrcast when the AS mismatches; drop the sret-arg addrspacecast that is already handled upstream. - CGExprScalar: user cast lowering uses cheerp_typed_ptrcast when the src/dst AS differ, falling back to the user-cast intrinsic otherwise. - CodeGenFunction: give the "result.ptr" alloca the same AS as the return value so the stored pointer matches.
TLS data lives in wasm linear memory and must be in CheerpAS::Wasm. A few helpers were still being emitted in AS 0, crashing later casts. - ItaniumCXXABI: _ZTW thread wrappers now use the function AS that matches the variable's data AS; __tls_guard is always in Wasm AS with section "asmjs" (wasm-AS is the only option that works in TUs mixing js and wasm TLs, since js can read wasm memory but not the reverse; pure-js TLs are stripped by CheerpLowerAtomicPass). - ThreadLocalLowering: __getThreadLocalAddress wrapper now returns i8* in Wasm AS and lives in the Wasm function AS. Assert instead of bitcasting when the intrinsic's AS doesn't match.
Asan was inserting allocas and memory-intrinsic callbacks in AS 0, which then crashed AllocaLowering / IdenticalCodeFolding when the surrounding wasm code was in CheerpAS::Wasm. - createAllocaForLayout and createDynamicAllocasInitStorage: emit MyAlloca / DynamicAllocaLayout in Wasm AS for cheerp-wasm. - handleDynamicAllocaCall: replacement alloca inherits the original alloca's AS so the downstream inttoptr/uses stay consistent. - instrumentMemIntrinsic + initializeCallbacks: __asan_memcpy / memmove / memset are declared with i8* in Wasm AS, and the argument pointer casts target the same AS, removing the bogus addrspacecast to AS 0 on every memcpy/memmove/memset.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Just a draft to track the status of the AS codegen work