Merge branch 'master' into build-rm-Versions.make

Makefile

@@ -197,7 +197,7 @@ else
 JL_PRIVATE_LIBS-$(USE_SYSTEM_ZLIB) += libz
 endif
 ifeq ($(USE_LLVM_SHLIB),1)
-JL_PRIVATE_LIBS-$(USE_SYSTEM_LLVM) += libLLVM libLLVM-13jl
+JL_PRIVATE_LIBS-$(USE_SYSTEM_LLVM) += libLLVM libLLVM-14jl
 endif
 JL_PRIVATE_LIBS-$(USE_SYSTEM_LIBUNWIND) += libunwind
 

base/binaryplatforms.jl

@@ -903,7 +903,7 @@ function detect_cxxstring_abi()
  end
 
  function open_libllvm(f::Function)
- for lib_name in ("libLLVM-13jl", "libLLVM", "LLVM", "libLLVMSupport")
+ for lib_name in ("libLLVM-14jl", "libLLVM", "LLVM", "libLLVMSupport")
  hdl = Libdl.dlopen_e(lib_name)
  if hdl != C_NULL
  try

contrib/refresh_checksums.mk

@@ -26,7 +26,7 @@ NON_CLANG_TRIPLETS=$(filter-out %-darwin %-freebsd,$(TRIPLETS))
 # These are the projects currently using BinaryBuilder; both GCC-expanded and non-GCC-expanded:
 BB_PROJECTS=mbedtls libssh2 nghttp2 mpfr curl libgit2 pcre libuv unwind llvmunwind dsfmt objconv p7zip zlib libsuitesparse openlibm blastrampoline
 BB_GCC_EXPANDED_PROJECTS=openblas csl
-BB_CXX_EXPANDED_PROJECTS=gmp llvm clang llvm-tools
+BB_CXX_EXPANDED_PROJECTS=gmp llvm clang llvm-tools lld
 # These are non-BB source-only deps
 NON_BB_PROJECTS=patchelf mozillacert lapack libwhich utf8proc
 
@@ -80,8 +80,12 @@ $(foreach project,$(BB_CXX_EXPANDED_PROJECTS),$(foreach triplet,$(CLANG_TRIPLETS
 
 # Special libLLVM_asserts_jll/LLVM_assert_jll targets
 $(foreach triplet,$(NON_CLANG_TRIPLETS),$(foreach cxxstring_abi,cxx11 cxx03,$(eval $(call checksum_dep,llvm,$(triplet)-$(cxxstring_abi),assert))))
+$(foreach triplet,$(NON_CLANG_TRIPLETS),$(foreach cxxstring_abi,cxx11 cxx03,$(eval $(call checksum_dep,clang,$(triplet)-$(cxxstring_abi),assert))))
+$(foreach triplet,$(NON_CLANG_TRIPLETS),$(foreach cxxstring_abi,cxx11 cxx03,$(eval $(call checksum_dep,lld,$(triplet)-$(cxxstring_abi),assert))))
 $(foreach triplet,$(NON_CLANG_TRIPLETS),$(foreach cxxstring_abi,cxx11 cxx03,$(eval $(call checksum_dep,llvm-tools,$(triplet)-$(cxxstring_abi),assert))))
 $(foreach triplet,$(CLANG_TRIPLETS),$(eval $(call checksum_dep,llvm,$(triplet),assert)))
+$(foreach triplet,$(CLANG_TRIPLETS),$(eval $(call checksum_dep,clang,$(triplet),assert)))
+$(foreach triplet,$(CLANG_TRIPLETS),$(eval $(call checksum_dep,lld,$(triplet),assert)))
 $(foreach triplet,$(CLANG_TRIPLETS),$(eval $(call checksum_dep,llvm-tools,$(triplet),assert)))
 
 # External stdlibs

deps/clang.version

@@ -1,4 +1,4 @@
 ## jll artifact
 # Clang (paired with LLVM, only here as a JLL download)
 CLANG_JLL_NAME := Clang
-CLANG_JLL_VER := 13.0.1+0
+CLANG_JLL_VER := 14.0.2+1

deps/lld.version

@@ -0,0 +1,3 @@
+## jll artifact
+LLD_JLL_NAME := LLD
+LLD_JLL_VER := 14.0.2+1

deps/llvm-tools.version

@@ -1,5 +1,5 @@
 ## jll artifact
 # LLVM_tools (downloads LLVM_jll to get things like `lit` and `opt`)
 LLVM_TOOLS_JLL_NAME := LLVM
-LLVM_TOOLS_JLL_VER := 13.0.1+0
-LLVM_TOOLS_ASSERT_JLL_VER := 13.0.1+0
+LLVM_TOOLS_JLL_VER := 14.0.2+1
+LLVM_TOOLS_ASSERT_JLL_VER := 14.0.2+1

deps/llvm.mk

@@ -79,6 +79,9 @@ LLVM_CMAKE += -DLLVM_EXTERNAL_RV_SOURCE_DIR=$(LLVM_MONOSRC_DIR)/rv
 LLVM_CMAKE += -DLLVM_CXX_STD=c++14
 endif
 
+# Otherwise LLVM will translate \\ to / on mingw
+LLVM_CMAKE += -DLLVM_WINDOWS_PREFER_FORWARD_SLASH=False
+
 # Allow adding LLVM specific flags
 LLVM_CFLAGS += $(CFLAGS)
 LLVM_CXXFLAGS += $(CXXFLAGS)
@@ -286,16 +289,26 @@ else # USE_BINARYBUILDER_LLVM
 
 # We provide a way to subversively swap out which LLVM JLL we pull artifacts from
 ifeq ($(LLVM_ASSERTIONS), 1)
-LLVM_JLL_DOWNLOAD_NAME := libLLVM_assert
-LLVM_JLL_VER := $(LLVM_ASSERT_JLL_VER)
-LLVM_TOOLS_JLL_DOWNLOAD_NAME := LLVM_assert
-LLVM_TOOLS_JLL_VER := $(LLVM_TOOLS_ASSERT_JLL_VER)
+# LLVM_JLL_DOWNLOAD_NAME := libLLVM_assert
+# LLVM_JLL_VER := $(LLVM_ASSERT_JLL_VER)
+# LLVM_TOOLS_JLL_DOWNLOAD_NAME := LLVM_assert
+# LLVM_TOOLS_JLL_VER := $(LLVM_TOOLS_ASSERT_JLL_VER)
+LLVM_JLL_TAGS := -llvm_version+$(LLVM_VER_MAJ).asserts
+CLANG_JLL_TAGS := -llvm_version+$(LLVM_VER_MAJ).asserts
+LLD_JLL_TAGS := -llvm_version+$(LLVM_VER_MAJ).asserts
+LLVM_TOOLS_JLL_TAGS := -llvm_version+$(LLVM_VER_MAJ).asserts
+else
+LLVM_JLL_TAGS := -llvm_version+$(LLVM_VER_MAJ)
+CLANG_JLL_TAGS := -llvm_version+$(LLVM_VER_MAJ)
+LLD_JLL_TAGS := -llvm_version+$(LLVM_VER_MAJ)
+LLVM_TOOLS_JLL_TAGS := -llvm_version+$(LLVM_VER_MAJ)
 endif
 
 $(eval $(call bb-install,llvm,LLVM,false,true))
 $(eval $(call bb-install,clang,CLANG,false,true))
+$(eval $(call bb-install,lld,LLD,false,true))
 $(eval $(call bb-install,llvm-tools,LLVM_TOOLS,false,true))
 
-install-clang install-llvm-tools: install-llvm
+install-lld install-clang install-llvm-tools: install-llvm
 
 endif # USE_BINARYBUILDER_LLVM

deps/llvm.version

@@ -1,8 +1,8 @@
 ## jll artifact
 LLVM_JLL_NAME := libLLVM
-LLVM_ASSERT_JLL_VER := 13.0.1+0
+LLVM_ASSERT_JLL_VER := 14.0.2+1
 
 ## source build
-LLVM_VER := 13.0.1
-LLVM_BRANCH=julia-13.0.1-0
-LLVM_SHA1=julia-13.0.1-0
+LLVM_VER := 14.0.2
+LLVM_BRANCH=julia-14.0.2-1
+LLVM_SHA1=julia-14.0.2-1

deps/tools/bb-install.mk

@@ -27,10 +27,13 @@ $(2)_JLL_VER ?= $$(shell [ -f $$($(2)_STDLIB_PATH)/Project.toml ] && grep "^vers
 # Allow things to override which JLL we pull from, e.g. libLLVM_jll vs. libLLVM_assert_jll
 $(2)_JLL_DOWNLOAD_NAME ?= $$($(2)_JLL_NAME)
 
+# Allow things to provide platform tags
+$(2)_JLL_TAGS ?=
+
 $(2)_BB_TRIPLET := $$($$(TRIPLET_VAR))
 $(2)_JLL_VER_NOPLUS := $$(firstword $$(subst +,$(SPACE),$$($(2)_JLL_VER)))
-$(2)_JLL_BASENAME := $$($(2)_JLL_DOWNLOAD_NAME).v$$($(2)_JLL_VER).$$($(2)_BB_TRIPLET).tar.gz
-$(2)_BB_URL := https://github.com/JuliaBinaryWrappers/$$($(2)_JLL_DOWNLOAD_NAME)_jll.jl/releases/download/$$($(2)_JLL_DOWNLOAD_NAME)-v$$($(2)_JLL_VER)/$$($(2)_JLL_DOWNLOAD_NAME).v$$($(2)_JLL_VER_NOPLUS).$$($(2)_BB_TRIPLET).tar.gz
+$(2)_JLL_BASENAME := $$($(2)_JLL_DOWNLOAD_NAME).v$$($(2)_JLL_VER).$$($(2)_BB_TRIPLET)$$($(2)_JLL_TAGS).tar.gz
+$(2)_BB_URL := https://github.com/JuliaBinaryWrappers/$$($(2)_JLL_DOWNLOAD_NAME)_jll.jl/releases/download/$$($(2)_JLL_DOWNLOAD_NAME)-v$$($(2)_JLL_VER)/$$($(2)_JLL_DOWNLOAD_NAME).v$$($(2)_JLL_VER_NOPLUS).$$($(2)_BB_TRIPLET)$$($(2)_JLL_TAGS).tar.gz
 
 $$(SRCCACHE)/$$($(2)_JLL_BASENAME): | $$(SRCCACHE)
  $$(JLDOWNLOAD) $$@ $$($(2)_BB_URL)

src/Makefile

@@ -126,7 +126,7 @@ else
 ifeq ($(OS), Darwin)
 CG_LLVMLINK += $(LLVM_LDFLAGS) -lLLVM
 else
-CG_LLVMLINK += $(LLVM_LDFLAGS) -lLLVM-13jl
+CG_LLVMLINK += $(LLVM_LDFLAGS) -lLLVM-14jl
 endif
 endif
 endif

src/aotcompile.cpp

@@ -631,7 +631,19 @@ void addOptimizationPasses(legacy::PassManagerBase *PM, int opt_level,
  // to merge allocations and sometimes eliminate them,
  // since AllocOpt does not handle PhiNodes.
  // Enable this instruction hoisting because of this and Union benchmarks.
- auto simplifyCFGOptions = SimplifyCFGOptions().hoistCommonInsts(true);
+ auto basicSimplifyCFGOptions = SimplifyCFGOptions()
+ .convertSwitchRangeToICmp(true)
+ .convertSwitchToLookupTable(true)
+ .forwardSwitchCondToPhi(true);
+ auto aggressiveSimplifyCFGOptions = SimplifyCFGOptions()
+ .convertSwitchRangeToICmp(true)
+ .convertSwitchToLookupTable(true)
+ .forwardSwitchCondToPhi(true)
+ //These mess with loop rotation, so only do them after that
+ .hoistCommonInsts(true)
+ // Causes an SRET assertion error in late-gc-lowering
+ // .sinkCommonInsts(true)
+ ;
 #ifdef JL_DEBUG_BUILD
  PM->add(createGCInvariantVerifierPass(true));
  PM->add(createVerifierPass());
@@ -646,7 +658,7 @@ void addOptimizationPasses(legacy::PassManagerBase *PM, int opt_level,
  if (opt_level == 1)
  PM->add(createInstSimplifyLegacyPass());
  }
- PM->add(createCFGSimplificationPass(simplifyCFGOptions));
+ PM->add(createCFGSimplificationPass(basicSimplifyCFGOptions));
  if (opt_level == 1) {
  PM->add(createSROAPass());
  PM->add(createInstructionCombiningPass());
@@ -676,7 +688,7 @@ void addOptimizationPasses(legacy::PassManagerBase *PM, int opt_level,
  // minimal clean-up to get rid of CPU feature checks
  if (opt_level == 1) {
  PM->add(createInstSimplifyLegacyPass());
- PM->add(createCFGSimplificationPass(simplifyCFGOptions));
+ PM->add(createCFGSimplificationPass(basicSimplifyCFGOptions));
  }
  }
 #if defined(_COMPILER_ASAN_ENABLED_)
@@ -697,7 +709,7 @@ void addOptimizationPasses(legacy::PassManagerBase *PM, int opt_level,
  PM->add(createBasicAAWrapperPass());
  }
 
- PM->add(createCFGSimplificationPass(simplifyCFGOptions));
+ PM->add(createCFGSimplificationPass(basicSimplifyCFGOptions));
  PM->add(createDeadCodeEliminationPass());
  PM->add(createSROAPass());
 
@@ -711,7 +723,7 @@ void addOptimizationPasses(legacy::PassManagerBase *PM, int opt_level,
  PM->add(createAllocOptPass());
  // consider AggressiveInstCombinePass at optlevel > 2
  PM->add(createInstructionCombiningPass());
- PM->add(createCFGSimplificationPass(simplifyCFGOptions));
+ PM->add(createCFGSimplificationPass(basicSimplifyCFGOptions));
  if (dump_native)
  PM->add(createMultiVersioningPass(external_use));
  PM->add(createCPUFeaturesPass());
@@ -781,27 +793,23 @@ void addOptimizationPasses(legacy::PassManagerBase *PM, int opt_level,
  PM->add(createGVNPass()); // Must come after JumpThreading and before LoopVectorize
  }
  PM->add(createDeadStoreEliminationPass());
+ // see if all of the constant folding has exposed more loops
+ // to simplification and deletion
+ // this helps significantly with cleaning up iteration
+ PM->add(createCFGSimplificationPass(aggressiveSimplifyCFGOptions));
 
  // More dead allocation (store) deletion before loop optimization
  // consider removing this:
+ // Moving this after aggressive CFG simplification helps deallocate when allocations are hoisted
  PM->add(createAllocOptPass());
- // see if all of the constant folding has exposed more loops
- // to simplification and deletion
- // this helps significantly with cleaning up iteration
- PM->add(createCFGSimplificationPass()); // See note above, don't hoist instructions before LV
  PM->add(createLoopDeletionPass());
  PM->add(createInstructionCombiningPass());
  PM->add(createLoopVectorizePass());
  PM->add(createLoopLoadEliminationPass());
  // Cleanup after LV pass
  PM->add(createInstructionCombiningPass());
  PM->add(createCFGSimplificationPass( // Aggressive CFG simplification
- SimplifyCFGOptions()
- .forwardSwitchCondToPhi(true)
- .convertSwitchToLookupTable(true)
- .needCanonicalLoops(false)
- .hoistCommonInsts(true)
- // .sinkCommonInsts(true) // FIXME: Causes assertion in llvm-late-lowering
+ aggressiveSimplifyCFGOptions
  ));
  PM->add(createSLPVectorizerPass());
  // might need this after LLVM 11:
@@ -812,7 +820,7 @@ void addOptimizationPasses(legacy::PassManagerBase *PM, int opt_level,
  if (lower_intrinsics) {
  // LowerPTLS removes an indirect call. As a result, it is likely to trigger
  // LLVM's devirtualization heuristics, which would result in the entire
- // pass pipeline being re-exectuted. Prevent this by inserting a barrier.
+ // pass pipeline being re-executed. Prevent this by inserting a barrier.
  PM->add(createBarrierNoopPass());
  PM->add(createLowerExcHandlersPass());
  PM->add(createGCInvariantVerifierPass(false));

src/cgutils.cpp

@@ -605,7 +605,7 @@ static Type *bitstype_to_llvm(jl_value_t *bt, LLVMContext &ctxt, bool llvmcall =
 {
  assert(jl_is_primitivetype(bt));
  if (bt == (jl_value_t*)jl_bool_type)
- return getInt8Ty(ctxt);
+ return llvmcall ? getInt1Ty(ctxt) : getInt8Ty(ctxt);
  if (bt == (jl_value_t*)jl_int32_type)
  return getInt32Ty(ctxt);
  if (bt == (jl_value_t*)jl_int64_type)

src/codegen.cpp

@@ -2090,12 +2090,14 @@ static void jl_init_function(Function *F)
  // i686 Windows (which uses a 4-byte-aligned stack)
 #if JL_LLVM_VERSION >= 140000
  AttrBuilder attr(F->getContext());
+ attr.addStackAlignmentAttr(16);
+ F->addFnAttrs(attr);
 #else
  AttrBuilder attr;
-#endif
  attr.addStackAlignmentAttr(16);
  F->addAttributes(AttributeList::FunctionIndex, attr);
 #endif
+#endif
 #if defined(_OS_WINDOWS_) && defined(_CPU_X86_64_)
  F->setHasUWTable(); // force NeedsWinEH
 #endif

src/gc.c

@@ -190,22 +190,7 @@ NOINLINE uintptr_t gc_get_stack_ptr(void)
 
 #define should_timeout() 0
 
-static void jl_gc_wait_for_the_world(void)
-{
- if (jl_n_threads > 1)
- jl_wake_libuv();
- for (int i = 0; i < jl_n_threads; i++) {
- jl_ptls_t ptls2 = jl_all_tls_states[i];
- // This acquire load pairs with the release stores
- // in the signal handler of safepoint so we are sure that
- // all the stores on those threads are visible.
- // We're currently also using atomic store release in mutator threads
- // (in jl_gc_state_set), but we may want to use signals to flush the
- // memory operations on those threads lazily instead.
- while (!jl_atomic_load_relaxed(&ptls2->gc_state) || !jl_atomic_load_acquire(&ptls2->gc_state))
- jl_cpu_pause(); // yield?
- }
-}
+void jl_gc_wait_for_the_world(void);
 
 // malloc wrappers, aligned allocation
 
@@ -3258,7 +3243,7 @@ static int _jl_gc_collect(jl_ptls_t ptls, jl_gc_collection_t collection)
 
  _report_gc_finished(pause, gc_num.freed, sweep_full, recollect);
 
- gc_final_pause_end(t0, gc_end_time);
+ gc_final_pause_end(gc_start_time, gc_end_time);
  gc_time_sweep_pause(gc_end_time, actual_allocd, live_bytes,
  estimate_freed, sweep_full);
  gc_num.full_sweep += sweep_full;
@@ -3762,6 +3747,7 @@ static void *gc_perm_alloc_large(size_t sz, int zero, unsigned align, unsigned o
 #endif
  errno = last_errno;
  jl_may_leak(base);
+ assert(align > 0);
  unsigned diff = (offset - base) % align;
  return (void*)(base + diff);
 }

src/gf.c

@@ -109,7 +109,7 @@ static jl_method_instance_t *jl_specializations_get_linfo_(jl_method_t *m JL_PRO
  for (int locked = 0; ; locked++) {
  jl_array_t *speckeyset = jl_atomic_load_acquire(&m->speckeyset);
  jl_svec_t *specializations = jl_atomic_load_relaxed(&m->specializations);
- size_t i, cl = jl_svec_len(specializations);
+ size_t i = -1, cl = jl_svec_len(specializations);
  if (hv) {
  ssize_t idx = jl_smallintset_lookup(speckeyset, speccache_eq, type, specializations, hv);
  if (idx != -1) {

src/init.c

@@ -101,8 +101,14 @@ void jl_init_stack_limits(int ismaster, void **stack_lo, void **stack_hi)
  struct rlimit rl;
  getrlimit(RLIMIT_STACK, &rl);
  size_t stacksize = rl.rlim_cur;
+// We intentionally leak a stack address here core.StackAddressEscape
+# ifndef __clang_analyzer__
  *stack_hi = (void*)&stacksize;
  *stack_lo = (void*)((char*)*stack_hi - stacksize);
+# else
+ *stack_hi = 0;
+ *stack_lo = 0;
+# endif
 #endif
 }