Import of diff from JuliaGPU/julia.

base/reflection.jl

@@ -493,6 +493,15 @@ function uncompressed_ast(m::Method, s::CodeInfo)
     return s
 end
 
+# this type mirrors jl_cghooks_t (documented in julia.h)
+immutable CodegenHooks
+    module_setup::Any
+    module_activation::Any
+
+    CodegenHooks(;module_setup=nothing, module_activation=nothing) =
+        new(module_setup, module_activation)
+end
+
 # this type mirrors jl_cgparams_t (documented in julia.h)
 immutable CodegenParams
     cached::Cint
@@ -504,14 +513,18 @@ immutable CodegenParams
     static_alloc::Cint
     dynamic_alloc::Cint
 
+    hooks::CodegenHooks
+
     CodegenParams(;cached::Bool=true,
                    runtime::Int=1, exceptions::Int=1,
                    track_allocations::Int=1, code_coverage::Int=1,
-                   static_alloc::Int=1, dynamic_alloc::Int=1) =
+                   static_alloc::Int=1, dynamic_alloc::Int=1,
+                   hooks::CodegenHooks=CodegenHooks()) =
         new(Cint(cached),
             Cint(runtime), Cint(exceptions),
             Cint(track_allocations), Cint(code_coverage),
-            Cint(static_alloc), Cint(dynamic_alloc))
+            Cint(static_alloc), Cint(dynamic_alloc),
+            hooks)
 end
 
 # Printing code representations in IR and assembly
@@ -538,9 +551,9 @@ function _dump_function(linfo::Core.MethodInstance, native::Bool, wrapper::Bool,
         throw(ArgumentError("'syntax' must be either :intel or :att"))
     end
     if native
-        llvmf = ccall(:jl_get_llvmf_decl, Ptr{Void}, (Any, Bool, CodegenParams), linfo, wrapper, params)
+        llvmf = ccall(:jl_get_llvmf_decl, Ptr{Void}, (Any, Bool, CodegenParams, CodegenHooks), linfo, wrapper, params, params.hooks)
     else
-        llvmf = ccall(:jl_get_llvmf_defn, Ptr{Void}, (Any, Bool, Bool, CodegenParams), linfo, wrapper, optimize, params)
+        llvmf = ccall(:jl_get_llvmf_defn, Ptr{Void}, (Any, Bool, Bool, CodegenParams, CodegenHooks), linfo, wrapper, optimize, params, params.hooks)
     end
     if llvmf == C_NULL
         error("could not compile the specified method")

deps/Versions.make

@@ -1,4 +1,4 @@
-LLVM_VER = 3.7.1
+LLVM_VER = 3.9.0
 LLVM_LIB_SUFFIX =
 PCRE_VER = 10.22
 DSFMT_VER = 2.2.3

deps/llvm.mk

@@ -76,7 +76,11 @@ endif # LLVM_VER != svn
 LLVM_CXXFLAGS := $(CXXFLAGS)
 LLVM_CPPFLAGS := $(CPPFLAGS)
 LLVM_LDFLAGS := $(LDFLAGS)
-LLVM_TARGETS := host
+ifeq ($(LLVM_USE_CMAKE),1)
+	LLVM_TARGETS := host;NVPTX
+else
+	LLVM_TARGETS := host,nvptx
+endif
 LLVM_TARGET_FLAGS := --enable-targets=$(LLVM_TARGETS)
 LLVM_CMAKE += -DLLVM_TARGETS_TO_BUILD:STRING="$(LLVM_TARGETS)" -DCMAKE_BUILD_TYPE="$(LLVM_CMAKE_BUILDTYPE)"
 LLVM_CMAKE += -DLLVM_TOOLS_INSTALL_DIR=$(shell $(JULIAHOME)/contrib/relative_path.sh $(build_prefix) $(build_depsbindir))
@@ -466,6 +470,14 @@ $(eval $(call LLVM_PATCH,llvm-r282182)) # Remove for 4.0
 $(eval $(call LLVM_PATCH,llvm-arm-fix-prel31))
 endif # LLVM_VER
 
+$(eval $(call LLVM_PATCH,llvm-asan))
+ifeq ($(LLVM_VER_SHORT),3.9)
+$(eval $(call LLVM_PATCH,llvm-D9168_argument_alignment))
+endif
+ifeq ($(BUILD_LLVM_CLANG),1)
+$(eval $(call LLVM_PATCH,compiler-rt-asan))
+endif
+
 ifeq ($(LLVM_VER),3.7.1)
 ifeq ($(BUILD_LLDB),1)
 $(eval $(call LLVM_PATCH,lldb-3.7.1))

deps/patches/compiler-rt-asan.patch

@@ -0,0 +1,13 @@
+Index: lib/asan/asan_mapping.h
+===================================================================
+--- a/projects/compiler-rt/lib/asan/asan_mapping.h	(revision 256758)
++++ b/projects/compiler-rt/lib/asan/asan_mapping.h	(working copy)
+@@ -146,7 +146,7 @@
+ #  elif SANITIZER_IOS
+ #    define SHADOW_OFFSET kIosShadowOffset64
+ #  else
+-#   define SHADOW_OFFSET kDefaultShort64bitShadowOffset
++#   define SHADOW_OFFSET kDefaultShadowOffset64
+ #  endif
+ # endif
+ #endif

deps/patches/llvm-D9168_argument_alignment.patch

@@ -0,0 +1,98 @@
+Index: lib/Target/NVPTX/NVPTXISelLowering.h
+===================================================================
+--- a/lib/Target/NVPTX/NVPTXISelLowering.h
++++ b/lib/Target/NVPTX/NVPTXISelLowering.h
+@@ -539,7 +539,8 @@
+   SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const override;
+
+   unsigned getArgumentAlignment(SDValue Callee, const ImmutableCallSite *CS,
+-                                Type *Ty, unsigned Idx) const;
++                                Type *Ty, unsigned Idx,
++                                const DataLayout &DL) const;
+ };
+ } // namespace llvm
+
+Index: lib/Target/NVPTX/NVPTXISelLowering.cpp
+===================================================================
+--- a/lib/Target/NVPTX/NVPTXISelLowering.cpp
++++ b/lib/Target/NVPTX/NVPTXISelLowering.cpp
+@@ -1024,11 +1024,15 @@
+   return O.str();
+ }
+
+-unsigned
+-NVPTXTargetLowering::getArgumentAlignment(SDValue Callee,
+-                                          const ImmutableCallSite *CS,
+-                                          Type *Ty,
+-                                          unsigned Idx) const {
++unsigned NVPTXTargetLowering::getArgumentAlignment(SDValue Callee,
++                                                   const ImmutableCallSite *CS,
++                                                   Type *Ty, unsigned Idx,
++                                                   const DataLayout &DL) const {
++  if (!CS) {
++    // CallSite is zero, fallback to ABI type alignment
++    return DL.getABITypeAlignment(Ty);
++  }
++
+   unsigned Align = 0;
+   const Value *DirectCallee = CS->getCalledFunction();
+
+@@ -1046,7 +1050,7 @@
+
+       const Value *CalleeV = cast<CallInst>(CalleeI)->getCalledValue();
+       // Ignore any bitcast instructions
+-      while(isa<ConstantExpr>(CalleeV)) {
++      while (isa<ConstantExpr>(CalleeV)) {
+         const ConstantExpr *CE = cast<ConstantExpr>(CalleeV);
+         if (!CE->isCast())
+           break;
+@@ -1069,7 +1073,6 @@
+
+   // Call is indirect or alignment information is not available, fall back to
+   // the ABI type alignment
+-  auto &DL = CS->getCaller()->getParent()->getDataLayout();
+   return DL.getABITypeAlignment(Ty);
+ }
+
+@@ -1126,7 +1129,8 @@
+         ComputePTXValueVTs(*this, DAG.getDataLayout(), Ty, vtparts, &Offsets,
+                            0);
+
+-        unsigned align = getArgumentAlignment(Callee, CS, Ty, paramCount + 1);
++        unsigned align =
++            getArgumentAlignment(Callee, CS, Ty, paramCount + 1, DL);
+         // declare .param .align <align> .b8 .param<n>[<size>];
+         unsigned sz = DL.getTypeAllocSize(Ty);
+         SDVTList DeclareParamVTs = DAG.getVTList(MVT::Other, MVT::Glue);
+@@ -1166,7 +1170,8 @@
+       }
+       if (Ty->isVectorTy()) {
+         EVT ObjectVT = getValueType(DL, Ty);
+-        unsigned align = getArgumentAlignment(Callee, CS, Ty, paramCount + 1);
++        unsigned align =
++            getArgumentAlignment(Callee, CS, Ty, paramCount + 1, DL);
+         // declare .param .align <align> .b8 .param<n>[<size>];
+         unsigned sz = DL.getTypeAllocSize(Ty);
+         SDVTList DeclareParamVTs = DAG.getVTList(MVT::Other, MVT::Glue);
+@@ -1426,7 +1431,7 @@
+                           DeclareRetOps);
+       InFlag = Chain.getValue(1);
+     } else {
+-      retAlignment = getArgumentAlignment(Callee, CS, retTy, 0);
++      retAlignment = getArgumentAlignment(Callee, CS, retTy, 0, DL);
+       SDVTList DeclareRetVTs = DAG.getVTList(MVT::Other, MVT::Glue);
+       SDValue DeclareRetOps[] = { Chain,
+                                   DAG.getConstant(retAlignment, dl, MVT::i32),
+@@ -1633,9 +1638,10 @@
+     } else {
+       SmallVector<EVT, 16> VTs;
+       SmallVector<uint64_t, 16> Offsets;
+-      ComputePTXValueVTs(*this, DAG.getDataLayout(), retTy, VTs, &Offsets, 0);
++      auto &DL = DAG.getDataLayout();
++      ComputePTXValueVTs(*this, DL, retTy, VTs, &Offsets, 0);
+       assert(VTs.size() == Ins.size() && "Bad value decomposition");
+-      unsigned RetAlign = getArgumentAlignment(Callee, CS, retTy, 0);
++      unsigned RetAlign = getArgumentAlignment(Callee, CS, retTy, 0, DL);
+       for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
+         unsigned sz = VTs[i].getSizeInBits();
+         unsigned AlignI = GreatestCommonDivisor64(RetAlign, Offsets[i]);

deps/patches/llvm-asan.patch

@@ -0,0 +1,13 @@
+diff --git a/lib/Transforms/Instrumentation/AddressSanitizer.cpp b/lib/Transforms/Instrumentation/AddressSanitizer.cpp
+index 43d1b37..3efe3d6 100644
+--- a/lib/Transforms/Instrumentation/AddressSanitizer.cpp
++++ b/lib/Transforms/Instrumentation/AddressSanitizer.cpp
+@@ -404,7 +404,7 @@ static ShadowMapping getShadowMapping(Triple &TargetTriple, int LongSize,
+       if (IsKasan)
+         Mapping.Offset = kLinuxKasan_ShadowOffset64;
+       else
+-        Mapping.Offset = kSmallX86_64ShadowOffset;
++        Mapping.Offset = kDefaultShadowOffset64;
+     } else if (IsWindows && IsX86_64) {
+       Mapping.Offset = kWindowsShadowOffset64;
+     } else if (IsMIPS64)

src/cgutils.cpp

@@ -45,6 +45,21 @@ static inline bool _feat_test(jl_codectx_t* ctx, const char *caller, int featval
 }
 
 
+// --- hook checks ---
+
+#define JL_HOOK_TEST(params,hook)           \
+    ((params)->hooks.hook != jl_nothing)
+
+// NOTE: this is just a throwing version of jl_call1...
+#define JL_HOOK_CALL1(params,hook,arg1)     \
+    jl_value_t **argv;                      \
+    JL_GC_PUSHARGS(argv, 2);                \
+    argv[0] = (params)->hooks.hook;         \
+    argv[1] = arg1;                         \
+    jl_apply(argv, 2);                      \
+    JL_GC_POP();
+
+
 // --- string constants ---
 static StringMap<GlobalVariable*> stringConstants;
 static Value *stringConstPtr(IRBuilder<> &builder, const std::string &txt)

src/codegen.cpp

@@ -923,25 +923,30 @@ jl_llvm_functions_t jl_compile_linfo(jl_method_instance_t *li, jl_code_info_t *s
     f = (Function*)decls.functionObject;
     specf = (Function*)decls.specFunctionObject;
 
-    // Step 4. Prepare debug info to receive this function
-    // record that this function name came from this linfo,
-    // so we can build a reverse mapping for debug-info.
-    bool toplevel = li->def == NULL;
-    if (!toplevel) {
-        const DataLayout &DL =
-#if JL_LLVM_VERSION >= 30500
-            m->getDataLayout();
-#else
-            *jl_data_layout;
-#endif
-        // but don't remember toplevel thunks because
-        // they may not be rooted in the gc for the life of the program,
-        // and the runtime doesn't notify us when the code becomes unreachable :(
-        jl_add_linfo_in_flight((specf ? specf : f)->getName(), li, DL);
-    }
 
-    // Step 5. Add the result to the execution engine now
-    jl_finalize_module(m.release(), !toplevel);
+    if (JL_HOOK_TEST(params, module_activation)) {
+        JL_HOOK_CALL1(params, module_activation, jl_box_voidpointer(wrap(m.release())));
+    } else {
+        // Step 4. Prepare debug info to receive this function
+        // record that this function name came from this linfo,
+        // so we can build a reverse mapping for debug-info.
+        bool toplevel = li->def == NULL;
+        if (!toplevel) {
+            const DataLayout &DL =
+    #if JL_LLVM_VERSION >= 30500
+                m->getDataLayout();
+    #else
+                *jl_data_layout;
+    #endif
+            // but don't remember toplevel thunks because
+            // they may not be rooted in the gc for the life of the program,
+            // and the runtime doesn't notify us when the code becomes unreachable :(
+            jl_add_linfo_in_flight((specf ? specf : f)->getName(), li, DL);
+        }
+
+        // Step 5. Add the result to the execution engine now
+        jl_finalize_module(m.release(), !toplevel);
+    }
 
     if (li->jlcall_api != 2) {
         // if not inlineable, code won't be needed again
@@ -988,8 +993,13 @@ static Value *getModuleFlag(Module *m, StringRef Key)
 #define getModuleFlag(m,str) m->getModuleFlag(str)
 #endif
 
-static void jl_setup_module(Module *m)
+static void jl_setup_module(Module *m, jl_cgparams_t *params = &jl_default_cgparams)
 {
+    if (JL_HOOK_TEST(params, module_setup)) {
+        JL_HOOK_CALL1(params, module_setup, jl_box_voidpointer(wrap(m)));
+        return;
+    }
+
     // Some linkers (*cough* OS X) don't understand DWARF v4, so we use v2 in
     // imaging mode. The structure of v4 is slightly nicer for debugging JIT
     // code.
@@ -1256,8 +1266,9 @@ void jl_extern_c(jl_function_t *f, jl_value_t *rt, jl_value_t *argt, char *name)
 // this is paired with jl_dump_function_ir and jl_dump_function_asm in particular ways:
 // misuse will leak memory or cause read-after-free
 extern "C" JL_DLLEXPORT
-void *jl_get_llvmf_defn(jl_method_instance_t *linfo, bool getwrapper, bool optimize, jl_cgparams_t params)
+void *jl_get_llvmf_defn(jl_method_instance_t *linfo, bool getwrapper, bool optimize, jl_cgparams_t params, jl_cghooks_t _hooks)
 {
+    params.hooks = _hooks;   // ccall doesn't know how to pass the nested jl_cghooks_t
     // `source` is `NULL` for generated functions.
     // The `isstaged` check can be removed if that is not the case anymore.
     if (linfo->def && linfo->def->source == NULL && !linfo->def->isstaged) {
@@ -1338,8 +1349,9 @@ void *jl_get_llvmf_defn(jl_method_instance_t *linfo, bool getwrapper, bool optim
 
 
 extern "C" JL_DLLEXPORT
-void *jl_get_llvmf_decl(jl_method_instance_t *linfo, bool getwrapper, jl_cgparams_t params)
+void *jl_get_llvmf_decl(jl_method_instance_t *linfo, bool getwrapper, jl_cgparams_t params, jl_cghooks_t _hooks)
 {
+    params.hooks = _hooks;   // ccall doesn't know how to pass the nested jl_cghooks_t
     // `source` is `NULL` for generated functions.
     // The `isstaged` check can be removed if that is not the case anymore.
     if (linfo->def && linfo->def->source == NULL && !linfo->def->isstaged) {
@@ -1397,9 +1409,9 @@ void *jl_get_llvmf(jl_tupletype_t *tt, bool getwrapper, bool getdeclarations)
     }
     void *f;
     if (getdeclarations)
-        f = jl_get_llvmf_decl(linfo, getwrapper, jl_default_cgparams);
+        f = jl_get_llvmf_decl(linfo, getwrapper, jl_default_cgparams, jl_no_cghooks);
     else
-        f = jl_get_llvmf_defn(linfo, getwrapper, true, jl_default_cgparams);
+        f = jl_get_llvmf_defn(linfo, getwrapper, true, jl_default_cgparams, jl_no_cghooks);
     JL_GC_POP();
     return f;
 }
@@ -4207,7 +4219,7 @@ static std::unique_ptr<Module> emit_function(jl_method_instance_t *lam, jl_code_
 
     ctx.sret = false;
     Module *M = new Module(ctx.name, jl_LLVMContext);
-    jl_setup_module(M);
+    jl_setup_module(M, params);
     if (specsig) { // assumes !va and !needsparams
         std::vector<Type*> fsig(0);
         Type *rt;

src/jltypes.c

@@ -68,6 +68,7 @@ JL_DLLEXPORT jl_value_t *jl_emptytuple=NULL;
 jl_svec_t *jl_emptysvec;
 jl_value_t *jl_nothing;
 
+jl_cghooks_t jl_no_cghooks;
 jl_cgparams_t jl_default_cgparams;
 
 // --- type properties and predicates ---
@@ -3531,7 +3532,8 @@ void jl_init_types(void)
     jl_type_type_mt = jl_new_method_table(jl_type_type->name->name, ptls->current_module);
     jl_type_type->name->mt = jl_type_type_mt;
 
-    jl_default_cgparams = (jl_cgparams_t){1, 1, 1, 1, 1, 1, 1};
+    jl_no_cghooks = (jl_cghooks_t){jl_nothing, jl_nothing};
+    jl_default_cgparams = (jl_cgparams_t){1, 1, 1, 1, 1, 1, 1, jl_no_cghooks};
 
     // initialize them. lots of cycles.
     jl_datatype_type->name = jl_new_typename(jl_symbol("DataType"));

src/julia.h

@@ -1757,6 +1757,19 @@ typedef struct {
 
 // codegen interface ----------------------------------------------------------
 
+typedef struct {
+    // module setup: prepare a module for code emission (data layout, DWARF version, ...)
+    // parameters: LLVMModuleRef as Ptr{Void}
+    // return value: none
+    jl_value_t *module_setup;
+
+    // module activation: registers debug info, adds module to JIT
+    // parameters: LLVMModuleRef as Ptr{Void}
+    // return value: none
+    jl_value_t *module_activation;
+} jl_cghooks_t;
+extern JL_DLLEXPORT jl_cghooks_t jl_no_cghooks;
+
 typedef struct {
     int cached;         // can the compiler consult/mutate the compilation cache?
 
@@ -1767,6 +1780,8 @@ typedef struct {
     int code_coverage;      // can we measure coverage (fallback: don't)?
     int static_alloc;       // is the compiler allowed to allocate statically?
     int dynamic_alloc;      // is the compiler allowed to allocate dynamically (requires runtime)?
+
+    jl_cghooks_t hooks;
 } jl_cgparams_t;
 extern JL_DLLEXPORT jl_cgparams_t jl_default_cgparams;