jitlayers: Use GlobalISel on AArch64 at -O0/-O1

deps/clang.version

@@ -3,4 +3,4 @@
 ## jll artifact
 # Clang (paired with LLVM, only here as a JLL download)
 CLANG_JLL_NAME := Clang
-CLANG_JLL_VER  := 20.1.8+0
+CLANG_JLL_VER  := 21.1.8+0

deps/lld.version

@@ -2,4 +2,4 @@
 
 ## jll artifact
 LLD_JLL_NAME := LLD
-LLD_JLL_VER := 20.1.8+0
+LLD_JLL_VER := 21.1.8+0

deps/llvm-tools.version

@@ -3,5 +3,5 @@
 ## jll artifact
 # LLVM_tools (downloads LLVM_jll to get things like `lit` and `opt`)
 LLVM_TOOLS_JLL_NAME := LLVM
-LLVM_TOOLS_JLL_VER := 20.1.8+0
-LLVM_TOOLS_ASSERT_JLL_VER := 20.1.8+0
+LLVM_TOOLS_JLL_VER := 21.1.8+0
+LLVM_TOOLS_ASSERT_JLL_VER := 21.1.8+0

deps/llvm.version

@@ -2,14 +2,14 @@
 
 ## jll artifact
 LLVM_JLL_NAME := libLLVM
-LLVM_ASSERT_JLL_VER := 20.1.8+0
+LLVM_ASSERT_JLL_VER := 21.1.8+0
 ## source build
 # Version number of LLVM
-LLVM_VER := 20.1.8
+LLVM_VER := 21.1.8
 # Git branch name in `LLVM_GIT_URL` repository
-LLVM_BRANCH=julia-20.1.8-0
+LLVM_BRANCH=julia-21.1.8-0
 # Git ref in `LLVM_GIT_URL` repository
-LLVM_SHA1=julia-20.1.8-0
+LLVM_SHA1=julia-21.1.8-0
 
 ## Following options are used to automatically fetch patchset from Julia's fork.  This is
 ## useful if you want to build an external LLVM while still applying Julia's patches.
@@ -18,6 +18,6 @@ LLVM_APPLY_JULIA_PATCHES := 0
 # GitHub repository to use for fetching the Julia patches to apply to LLVM source code.
 LLVM_JULIA_DIFF_GITHUB_REPO := https://github.com/llvm/llvm-project
 # Base GitHub ref for generating the diff.
-LLVM_BASE_REF := llvm:llvmorg-20.1.8
+LLVM_BASE_REF := llvm:llvmorg-21.1.8
 # Julia fork's GitHub ref for generating the diff.
-LLVM_JULIA_REF := JuliaLang:julia-20.1.8-0
+LLVM_JULIA_REF := JuliaLang:$(LLVM_SHA1)

src/Makefile

@@ -29,6 +29,8 @@ FLAGS_GCC += -Wno-unused-result
 # GCC is throwing warnings like `warning: 'int __builtin_memcmp_eq(const void*, const void*, long unsigned int)' specified bound 18446744073709551615 exceeds maximum object size 9223372036854775807` in llvm's StringRef.h == seemingly because it doesn't realise the size can't be 0.
 FLAGS_GCC += -Wno-stringop-overflow
 FLAGS_GCC += -Wno-stringop-overread
+# LLVM 21 causes maybe uninitialized warnings in include/llvm/ADT/DenseMap.h
+FLAGS_GCC += -Wno-maybe-uninitialized
 
 # Required to be allowed to use nullability extension and not be required to annotate all of everything
 JCFLAGS_CLANG += -Wno-atomic-alignment -Wno-nullability-extension -Wno-nullability-completeness
@@ -640,6 +642,9 @@ SA_EXCEPTIONS-jloptions.c                   := -Xanalyzer -analyzer-config -Xana
  # clang doesn't understand that e->vars has the same value in save_env (NULL) and restore_env (assumed non-NULL)
 SA_EXCEPTIONS-subtype.c                     := -Xanalyzer -analyzer-config -Xanalyzer silence-checkers="core.uninitialized.Assign;core.UndefinedBinaryOperatorResult"
 SA_EXCEPTIONS-codegen.c                     := -Xanalyzer -analyzer-config -Xanalyzer silence-checkers="core"
+ # core.FixedAddressDereference is new in clang 21 and false-positives on LLVM ilist headers and gc_read_stack
+SA_EXCEPTIONS-llvm-alloc-opt.cpp            := -Xanalyzer -analyzer-config -Xanalyzer silence-checkers="core.FixedAddressDereference"
+SA_EXCEPTIONS-gc-stock.c                    := -Xanalyzer -analyzer-config -Xanalyzer silence-checkers="core.FixedAddressDereference"
  # these need to be annotated (and possibly fixed)
 SKIP_GC_CHECK := codegen.cpp rtutils.c
 

src/abi_win32.cpp

@@ -53,9 +53,20 @@ bool needPassByRef(jl_datatype_t *dt, AttrBuilder &ab, LLVMContext &ctx, Type *T
 {
     // Use pass by reference for all structs
     if (dt->layout->nfields > 0 || dt->layout->npointers) {
+        // Avoid using i128 directly as a byval type since LLVM gives it 16-byte
+        // stack alignment on x86-32, while the C ABI uses 4-byte alignment.
+        if (Ty->isIntegerTy() && Ty->getIntegerBitWidth() > 64)
+            Ty = ArrayType::get(Type::getInt32Ty(ctx), jl_datatype_size(dt) / 4);
         ab.addByValAttr(Ty);
         return true;
     }
+    // Large primitive types (like Int128) must also be passed by reference
+    // to avoid i128's 16-byte stack alignment in LLVM.
+    size_t size = jl_datatype_size(dt);
+    if (jl_is_primitivetype(dt) && size > 8) {
+        ab.addByValAttr(ArrayType::get(Type::getInt32Ty(ctx), size / 4));
+        return true;
+    }
     return false;
 }
 

src/abi_x86.cpp

@@ -72,6 +72,11 @@ bool needPassByRef(jl_datatype_t *dt, AttrBuilder &ab, LLVMContext &ctx, Type *T
     size_t size = jl_datatype_size(dt);
     if (is_complex64(dt) || is_complex128(dt) || (jl_is_primitivetype(dt) && size <= 8))
         return false;
+    // Avoid using i128 directly as a byval type since LLVM gives it 16-byte
+    // stack alignment on x86-32, while the C ABI uses 4-byte alignment.
+    // Use an array of i32 instead, which naturally has 4-byte alignment.
+    if (Ty->isIntegerTy() && Ty->getIntegerBitWidth() > 64)
+        Ty = ArrayType::get(Type::getInt32Ty(ctx), size / 4);
     ab.addByValAttr(Ty);
     return true;
 }

src/aotcompile.cpp

@@ -338,10 +338,15 @@ class egal_set {
     egal_set() = default;
     void insert(jl_value_t *val)
     {
+        // list/keyset are GC-rooted by the caller via JL_GC_PUSH
+        JL_GC_PROMISE_ROOTED(val);
+        JL_GC_PROMISE_ROOTED(list);
+        JL_GC_PROMISE_ROOTED(keyset);
         jl_value_t *rval = jl_idset_get(list, keyset, val);
         if (rval == NULL) {
             ssize_t idx;
             list = jl_idset_put_key(list, val, &idx);
+            JL_GC_PROMISE_ROOTED(list);
             keyset = jl_idset_put_idx(list, keyset, idx);
         }
     }

src/clangsa/GCChecker.cpp

@@ -1050,7 +1050,7 @@ bool GCChecker::processAllocationOfResult(const CallEvent &Call,
   SymbolRef Sym = Call.getReturnValue().getAsSymbol();
   if (!Sym) {
     SVal S = C.getSValBuilder().conjureSymbolVal(
-        Call.getOriginExpr(), C.getLocationContext(), QT, C.blockCount());
+        C.getCFGElementRef(), C.getLocationContext(), QT, C.blockCount());
     State = State->BindExpr(Call.getOriginExpr(), C.getLocationContext(), S);
     Sym = S.getAsSymbol();
   }
@@ -1146,8 +1146,8 @@ SymbolRef GCChecker::getSymbolForResult(const Expr *Result,
   if (Loaded.isUnknown() || !Loaded.getAsSymbol()) {
     if (OldValS || GCChecker::isGCTracked(Result)) {
       Loaded = C.getSValBuilder().conjureSymbolVal(
-          nullptr, Result, C.getLocationContext(), Result->getType(),
-          C.blockCount());
+          nullptr, C.getCFGElementRef(), C.getLocationContext(),
+          Result->getType(), C.blockCount());
       State = State->bindLoc(*ValLoc, Loaded, C.getLocationContext());
       // State = State->BindExpr(Result, C.getLocationContext(),
       // State->getSVal(*ValLoc));
@@ -1201,7 +1201,7 @@ void GCChecker::checkDerivingExpr(const Expr *Result, const Expr *Parent,
       return;
     }
     ResultVal = C.getSValBuilder().conjureSymbolVal(
-        Result, C.getLocationContext(), Result->getType(),
+        C.getCFGElementRef(), C.getLocationContext(), Result->getType(),
         C.blockCount());
     State = State->BindExpr(Result, C.getLocationContext(), ResultVal);
   }
@@ -1544,7 +1544,8 @@ bool GCChecker::evalCall(const CallEvent &Call, CheckerContext &C) const {
       SVal Items = State->getSVal(ItemsLoc);
       if (Items.isUnknown()) {
         Items = C.getSValBuilder().conjureSymbolVal(
-            CE, C.getLocationContext(), FD->getType(), C.blockCount());
+            C.getCFGElementRef(), C.getLocationContext(), FD->getType(),
+            C.blockCount());
         State = State->bindLoc(ItemsLoc, Items, C.getLocationContext());
       }
       assert(Items.getAsRegion());

src/jitlayers.cpp

@@ -1688,7 +1688,17 @@ void optimizeDLSyms(Module &M) JL_NOTSAFEPOINT_LEAVE JL_NOTSAFEPOINT_ENTER {
 void fixupTM(TargetMachine &TM) {
     auto TheTriple = TM.getTargetTriple();
     if (jl_options.opt_level < 2) {
-        if (!TheTriple.isARM() && !TheTriple.isPPC64() && !TheTriple.isAArch64())
+        // Try GlobalISel on AArch64 - it's the default in LLVM at -O0 and
+        // is apparently generally faster than SelectionDAG while producing good code.
+        // Use fallback mode so unsupported patterns fall back to SelectionDAG.
+        // Note: Requires RemoveJuliaAddrspacesPass to run before codegen
+        // because GlobalISel doesn't handle Julia's custom address spaces.
+        if (TheTriple.isAArch64()) {
+            TM.setGlobalISel(true);
+            TM.setGlobalISelAbort(GlobalISelAbortMode::Disable);
+            TM.setFastISel(false);
+        }
+        else if (!TheTriple.isARM() && !TheTriple.isPPC64())
             TM.setFastISel(true);
         else    // FastISel seems to be buggy Ref #13321
             TM.setFastISel(false);
@@ -1736,8 +1746,9 @@ JuliaOJIT::JuliaOJIT()
     ObjectLayer.addPlugin(std::make_unique<EHFrameRegistrationPlugin>(
         ExecutorAddr::fromPtr(JLEHFrames::registerEHFrameSectionAllocAction),
         ExecutorAddr::fromPtr(JLEHFrames::deregisterEHFrameSectionAllocAction)));
-#endif
+# else
     ObjectLayer.addPlugin(cantFail(EHFrameRegistrationPlugin::Create(ES)));
+# endif
 #endif
 
     ObjectLayer.addPlugin(DebuginfoPlugin);

stdlib/LLD_jll/Project.toml

@@ -1,6 +1,6 @@
 name = "LLD_jll"
 uuid = "d55e3150-da41-5e91-b323-ecfd1eec6109"
-version = "20.1.8+0"
+version = "21.1.8+0"
 
 [deps]
 Zlib_jll = "83775a58-1f1d-513f-b197-d71354ab007a"
@@ -9,8 +9,8 @@ libLLVM_jll = "8f36deef-c2a5-5394-99ed-8e07531fb29a"
 Artifacts = "56f22d72-fd6d-98f1-02f0-08ddc0907c33"
 
 [compat]
-julia = "1.13"
-libLLVM_jll = "20.1.8"
+julia = "1.14"
+libLLVM_jll = "21.1.8"
 
 [extras]
 Test = "8dfed614-e22c-5e08-85e1-65c5234f0b40"

stdlib/libLLVM_jll/Project.toml

@@ -1,6 +1,6 @@
 name = "libLLVM_jll"
 uuid = "8f36deef-c2a5-5394-99ed-8e07531fb29a"
-version = "20.1.8+0"
+version = "21.1.8+0"
 
 [deps]
 Artifacts = "56f22d72-fd6d-98f1-02f0-08ddc0907c33"
@@ -13,7 +13,7 @@ Zstd_jll = "3161d3a3-bdf6-5164-811a-617609db77b4"
 CompilerSupportLibraries_jll = "1.3.0"
 Zlib_jll = "1"
 Zstd_jll = "1.5.7"
-julia = "1.13"
+julia = "1.14"
 
 [extras]
 Test = "8dfed614-e22c-5e08-85e1-65c5234f0b40"

test/atomics.jl

@@ -1115,3 +1115,15 @@ function add_one57190!()
 end
 
 @test add_one57190!() == 1
+
+# Test atomic Float16 operations at all optimization levels (GlobalISel miscompile on AArch64)
+# See https://github.com/JuliaLang/julia/pull/54140#issuecomment-2855794363
+for opt in 0:3
+    @test success(run(```$(Base.julia_cmd()) --startup-file=no -O$opt -e '
+        a = Threads.Atomic{Float16}(Float16(0))
+        a[] = Float16(1.5)
+        @assert a[] === Float16(1.5) "atomic Float16 store failed: got \$(a[]) expected 1.5 (opt level = -O$(Base.JLOptions().opt_level))"
+        a[] = Float16(3.25)
+        @assert a[] === Float16(3.25) "atomic Float16 store failed: got \$(a[]) expected 3.25 (opt level = -O$(Base.JLOptions().opt_level))"
+    '```))
+end