Fix #5459

Make.inc

@@ -42,6 +42,10 @@ USE_INTEL_JITEVENTS = 0
 # libc++ is standard on OS X 10.9, but not for earlier releases
 USE_LIBCPP = 0
 
+# Select the cpu architecture to target
+# Current available options are "native" and "core2"
+JULIA_CPU_TARGET = native
+
 # we include twice to pickup user definitions better
 ifeq (exists, $(shell [ -e $(JULIAHOME)/Make.user ] && echo exists ))
 include $(JULIAHOME)/Make.user
@@ -208,6 +212,7 @@ endif
 CC = $(CROSS_COMPILE)gcc
 CXX = $(CROSS_COMPILE)g++
 JCFLAGS = -std=gnu99 -pipe $(fPIC) -fno-strict-aliasing -D_FILE_OFFSET_BITS=64
+JCPPFLAGS =
 JCXXFLAGS = -pipe $(fPIC) -fno-rtti
 DEBUGFLAGS = -O0 -ggdb3 -DDEBUG -fstack-protector-all
 SHIPFLAGS = -O3 -falign-functions
@@ -217,6 +222,7 @@ ifeq ($(USECLANG),1)
 CC = $(CROSS_COMPILE)clang
 CXX = $(CROSS_COMPILE)clang++
 JCFLAGS = -pipe $(fPIC) -fno-strict-aliasing -D_FILE_OFFSET_BITS=64
+JCPPFLAGS =
 JCXXFLAGS = -pipe $(fPIC) -fno-rtti
 DEBUGFLAGS = -O0 -g -DDEBUG -fstack-protector-all
 SHIPFLAGS = -O3
@@ -229,7 +235,7 @@ else
 CC += $(STDLIBCPP_FLAG) -mmacosx-version-min=10.6
 CXX += $(STDLIBCPP_FLAG) -mmacosx-version-min=10.6
 endif
-JCFLAGS += -D_LARGEFILE_SOURCE -D_DARWIN_USE_64_BIT_INODE=1
+JCPPFLAGS += -D_LARGEFILE_SOURCE -D_DARWIN_USE_64_BIT_INODE=1
 endif
 endif
 
@@ -241,6 +247,14 @@ AS := $(CROSS_COMPILE)as
 LD := $(CROSS_COMPILE)ld
 RANLIB := $(CROSS_COMPILE)ranlib
 
+ifeq ($(JULIA_CPU_TARGET),native)
+  JCPPFLAGS += -DJULIA_TARGET_NATIVE
+else ifeq ($(JULIA_CPU_TARGET),core2)
+  JCPPFLAGS += -DJULIA_TARGET_CORE2
+else
+  $(error Unknown cpu target architecture)
+endif
+
 
 # Calculate relative paths to libdir and private_libdir
 build_libdir_rel = $(shell $(JULIAHOME)/contrib/relative_path.sh $(build_bindir) $(build_libdir))
@@ -250,7 +264,7 @@ build_private_libdir_rel = $(shell $(JULIAHOME)/contrib/relative_path.sh $(build
 private_libdir_rel = $(shell $(JULIAHOME)/contrib/relative_path.sh $(bindir) $(private_libdir))
 
 # if not absolute, then relative to the directory of the julia executable
-JCFLAGS += "-DJL_SYSTEM_IMAGE_PATH=\"$(build_private_libdir_rel)/sys.ji\""
+JCPPFLAGS += "-DJL_SYSTEM_IMAGE_PATH=\"$(build_private_libdir_rel)/sys.ji\""
 
 # On Windows, we want shared library files to end up in $(build_bindir), instead of $(build_libdir)
 ifeq ($(OS),WINNT)
@@ -321,7 +335,7 @@ endif
 else
 ifeq ($(OS),Darwin)
 LIBUNWIND=$(build_libdir)/libosxunwind.a
-JCFLAGS+=-DLIBOSXUNWIND
+JCPPFLAGS+=-DLIBOSXUNWIND
 else
 LIBUNWIND=$(build_libdir)/libunwind-generic.a $(build_libdir)/libunwind.a
 endif
@@ -485,8 +499,7 @@ endif
 ifeq ($(OS), WINNT)
 OSLIBS += -Wl,--export-all-symbols -Wl,--version-script=$(JULIAHOME)/src/julia.expmap \
 	$(NO_WHOLE_ARCHIVE) -lpsapi -lkernel32 -lws2_32 -liphlpapi -lwinmm -ldbghelp -lssp
-JCFLAGS += -D_WIN32_WINNT=0x0600
-JCXXFLAGS += -D_WIN32_WINNT=0x0600
+JCPPFLAGS += -D_WIN32_WINNT=0x0600
 JLDFLAGS = -Wl,--stack,8388608
 ifeq ($(ARCH),i686)
 JLDFLAGS += -Wl,--large-address-aware

Makefile

@@ -65,10 +65,7 @@ $(build_sysconfdir)/julia/juliarc.jl: contrib/windows/juliarc.jl
 endif
 
 # use sys.ji if it exists, otherwise run two stages
-$(build_private_libdir)/sys%ji: $(build_private_libdir)/sys%bc
-
-$(build_private_libdir)/sys%o: $(build_private_libdir)/sys%bc
-	$(call spawn,$(LLVM_LLC)) -filetype=obj -relocation-model=pic -mattr=-bmi2,-avx2 -o $(call cygpath_w,$@) $(call cygpath_w,$<)
+$(build_private_libdir)/sys%ji: $(build_private_libdir)/sys%o
 
 .PRECIOUS: $(build_private_libdir)/sys%o
 
@@ -77,11 +74,11 @@ $(build_private_libdir)/sys%$(SHLIB_EXT): $(build_private_libdir)/sys%o
 		$$([ $(OS) = Darwin ] && echo -Wl,-undefined,dynamic_lookup || echo -Wl,--unresolved-symbols,ignore-all ) \
 		$$([ $(OS) = WINNT ] && echo -ljulia -lssp)
 
-$(build_private_libdir)/sys0.bc:
+$(build_private_libdir)/sys0.o:
 	@$(QUIET_JULIA) cd base && \
 	$(call spawn,$(JULIA_EXECUTABLE)) --build $(call cygpath_w,$(build_private_libdir)/sys0) sysimg.jl
 
-$(build_private_libdir)/sys.bc: VERSION base/*.jl base/pkg/*.jl base/linalg/*.jl base/sparse/*.jl $(build_datarootdir)/julia/helpdb.jl $(build_datarootdir)/man/man1/julia.1 $(build_private_libdir)/sys0.$(SHLIB_EXT)
+$(build_private_libdir)/sys.o: VERSION base/*.jl base/pkg/*.jl base/linalg/*.jl base/sparse/*.jl $(build_datarootdir)/julia/helpdb.jl $(build_datarootdir)/man/man1/julia.1 $(build_private_libdir)/sys0.$(SHLIB_EXT)
 	@$(QUIET_JULIA) cd base && \
 	$(call spawn,$(JULIA_EXECUTABLE)) --build $(call cygpath_w,$(build_private_libdir)/sys) \
 		-J$(call cygpath_w,$(build_private_libdir))/$$([ -e $(build_private_libdir)/sys.ji ] && echo sys.ji || echo sys0.ji) -f sysimg.jl \

src/Makefile

@@ -4,6 +4,7 @@ include $(JULIAHOME)/deps/Versions.make
 
 override CFLAGS += $(JCFLAGS)
 override CXXFLAGS += $(JCXXFLAGS)
+override CPPFLAGS += $(JCPPFLAGS)
 
 SRCS = \
 	jltypes gf ast builtins module codegen interpreter \

src/cgutils.cpp

@@ -215,6 +215,11 @@ extern "C" int32_t jl_get_llvm_gv(jl_value_t *p)
         return 0;
     return it->second.index;
 }
+
+extern "C" {
+    extern void jl_cpuid(int32_t CPUInfo[4], int32_t InfoType);
+}
+
 static void jl_gen_llvm_gv_array()
 {
     // emit the variable table into the code image (can only call this once)
@@ -234,6 +239,29 @@ static void jl_gen_llvm_gv_array()
             GlobalVariable::ExternalLinkage,
             ConstantInt::get(T_size,globalUnique+1),
             "jl_globalUnique");
+
+    Constant *feature_string = ConstantDataArray::getString(jl_LLVMContext,jl_cpu_string);
+    new GlobalVariable(
+            *jl_Module,
+            feature_string->getType(),
+            true,
+            GlobalVariable::ExternalLinkage,
+            feature_string,
+            "jl_sysimg_cpu_target");
+
+    // For native also store the cpuid
+    if(strcmp(jl_cpu_string,"native") == 0) {
+        uint32_t info[4];
+
+        jl_cpuid((int32_t*)info, 1);
+        new GlobalVariable(
+            *jl_Module,
+            T_int64,
+            true,
+            GlobalVariable::ExternalLinkage,
+            ConstantInt::get(T_int64,((uint64_t)info[2])|(((uint64_t)info[3])<<32)),
+            "jl_sysimg_cpu_cpuid");
+    }
 }
 
 static int32_t jl_assign_functionID(Function *functionObject)

src/codegen.cpp

@@ -30,11 +30,14 @@
 #include "llvm/ExecutionEngine/JITEventListener.h"
 #include "llvm/ExecutionEngine/JITMemoryManager.h"
 #include "llvm/PassManager.h"
+#include "llvm/Target/TargetLibraryInfo.h"
+#include "llvm/Support/TargetRegistry.h"
 #include "llvm/Analysis/Passes.h"
 #include "llvm/Bitcode/ReaderWriter.h"
 #if defined(LLVM_VERSION_MAJOR) && LLVM_VERSION_MAJOR == 3 && LLVM_VERSION_MINOR >= 5
 #define LLVM35 1
 #include "llvm/IR/Verifier.h"
+#include "llvm/Object/ObjectFile.h"
 #include "llvm/IR/DIBuilder.h"
 #include "llvm/AsmParser/Parser.h"
 #else
@@ -80,6 +83,7 @@
 #include "llvm/Target/TargetOptions.h"
 #include "llvm/Transforms/Scalar.h"
 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
+#include "llvm/Transforms/Instrumentation.h"
 #if defined(LLVM_VERSION_MAJOR) && LLVM_VERSION_MAJOR == 3 && LLVM_VERSION_MINOR >= 1
 #include "llvm/Transforms/Vectorize.h"
 #endif
@@ -127,6 +131,7 @@ static LLVMContext &jl_LLVMContext = getGlobalContext();
 static IRBuilder<> builder(getGlobalContext());
 static bool nested_compile=false;
 static ExecutionEngine *jl_ExecutionEngine;
+static TargetMachine *jl_TargetMachine;
 #ifdef USE_MCJIT
 static Module *shadow_module;
 static RTDyldMemoryManager *jl_mcjmm;
@@ -302,6 +307,47 @@ void jl_dump_bitcode(char* fname)
 #endif
 }
 
+extern "C"
+void jl_dump_objfile(char* fname, int jit_model)
+{
+    std::string err;
+    raw_fd_ostream OS(fname, err);
+    formatted_raw_ostream FOS(OS);
+    jl_gen_llvm_gv_array();
+
+    // We don't want to use MCJIT's target machine because
+    // it uses the large code model and we may potentially
+    // want less optimizations there.
+    OwningPtr<TargetMachine>
+    TM(jl_TargetMachine->getTarget().createTargetMachine(
+        jl_TargetMachine->getTargetTriple(),
+        jl_TargetMachine->getTargetCPU(),
+        jl_TargetMachine->getTargetFeatureString(),
+        jl_TargetMachine->Options,
+#ifdef _OS_LINUX_
+        Reloc::PIC_,
+#else
+        jit_model ? Reloc::PIC_ : Reloc::Default,
+#endif
+        jit_model ? CodeModel::JITDefault : CodeModel::Default,
+        CodeGenOpt::Aggressive // -O3
+        ));
+
+    PassManager PM;
+    PM.add(new TargetLibraryInfo(Triple(jl_TargetMachine->getTargetTriple())));
+    PM.add(new DataLayout(*jl_ExecutionEngine->getDataLayout()));
+    if (TM->addPassesToEmitFile(PM, FOS,
+            TargetMachine::CGFT_ObjectFile, false)) {
+        jl_error("Could not generate obj file for this target");
+    }
+
+#ifdef USE_MCJIT
+    PM.run(*shadow_module);
+#else
+    PM.run(*jl_Module);
+#endif
+}
+
 // aggregate of array metadata
 typedef struct {
     Value *dataptr;
@@ -366,13 +412,20 @@ static Type *NoopType;
 
 // --- utilities ---
 
+#if defined(JULIA_TARGET_CORE2)
+const char *jl_cpu_string = "core2";
+#elif defined(JULIA_TARGET_NATIVE)
+const char *jl_cpu_string = "native";
+#else
+#error "Must select julia cpu target"
+#endif
+
 extern "C" {
     int globalUnique = 0;
 }
 
 #include "cgutils.cpp"
 
-
 static void jl_rethrow_with_add(const char *fmt, ...)
 {
     if (jl_typeis(jl_exception_in_transit, jl_errorexception_type)) {
@@ -4082,23 +4135,30 @@ extern "C" void jl_init_codegen(void)
 #endif
 #ifdef USE_MCJIT
     jl_mcjmm = new SectionMemoryManager();
-#endif
+#else
     // Temporarily disable Haswell BMI2 features due to LLVM bug.
     const char *mattr[] = {"-bmi2", "-avx2"};
+#endif
     std::vector<std::string> attrvec (mattr, mattr+2);
-    jl_ExecutionEngine = EngineBuilder(engine_module)
+    EngineBuilder eb = EngineBuilder(engine_module)
         .setEngineKind(EngineKind::JIT)
 #if defined(_OS_WINDOWS_) && defined(_CPU_X86_64_)
         .setJITMemoryManager(new JITMemoryManagerWin())
 #endif
         .setTargetOptions(options)
+#if defined(JULIA_TARGET_NATIVE)
+        .setMCPU("")
+#else
+        .setMCPU(jl_cpu_string)
+#endif
 #ifdef USE_MCJIT
         .setUseMCJIT(true)
-        .setMAttrs(attrvec)
+        .setMAttrs(attrvec);
 #else
-        .setMAttrs(attrvec)
+        .setMAttrs(attrvec);
 #endif
-        .create();
+    jl_TargetMachine = eb.selectTarget();
+    jl_ExecutionEngine = eb.create(jl_TargetMachine);
 #endif // LLVM VERSION
     jl_ExecutionEngine->DisableLazyCompilation();
 

src/dump.c

@@ -95,9 +95,11 @@ static void write_as_tag(ios_t *s, uint8_t tag)
 static void jl_serialize_value_(ios_t *s, jl_value_t *v);
 static jl_value_t *jl_deserialize_value(ios_t *s);
 static jl_value_t *jl_deserialize_value_internal(ios_t *s);
-static jl_value_t ***sysimg_gvars = NULL;
+jl_value_t ***sysimg_gvars = NULL;
 
 extern int globalUnique;
+extern void jl_cpuid(int32_t CPUInfo[4], int32_t InfoType);
+extern const char *jl_cpu_string;
 
 static void jl_load_sysimg_so(char *fname)
 {
@@ -108,6 +110,26 @@ static void jl_load_sysimg_so(char *fname)
     if (sysimg_handle != 0) {
         sysimg_gvars = (jl_value_t***)jl_dlsym(sysimg_handle, "jl_sysimg_gvars");
         globalUnique = *(size_t*)jl_dlsym(sysimg_handle, "jl_globalUnique");
+        const char *cpu_target = (const char*)jl_dlsym(sysimg_handle, "jl_sysimg_cpu_target");
+        if (strcmp(cpu_target,jl_cpu_string) != 0)
+            jl_error("Julia and the system image were compiled for different architectures.\n"
+                     "Please delete or regenerate sys.{so,dll,dylib}.");
+        uint32_t info[4];
+        jl_cpuid((int32_t*)info, 1);
+        if (strcmp(cpu_target,"native") == 0)
+        {
+            uint64_t saved_cpuid = *(uint64_t*)jl_dlsym(sysimg_handle, "jl_sysimg_cpu_cpuid");
+            if (saved_cpuid != (((uint64_t)info[2])|(((uint64_t)info[3])<<32)))
+                jl_error("Target architecture mismatch. Please delete or regenerate sys.{so,dll,dylib}.");
+        } else if(strcmp(cpu_target,"core2") == 0) {
+            int HasSSSE3 = (info[3] & 1<<9);
+            if (!HasSSSE3)
+                jl_error("The current host does not support SSSE3, but the system image was compiled for Core2.\n"
+                         "Please delete or regenerate sys.{so,dll,dylib}.");
+        } else {
+            jl_error("System image has unknown target cpu architecture.\n"
+                     "Please delete or regenerate sys.{so,dll,dylib}.");
+        }
     }
     else {
         sysimg_gvars = 0;

src/init.c

@@ -867,13 +867,13 @@ DLLEXPORT int julia_trampoline(int argc, char **argv, int (*pmain)(int ac,char *
         if (asprintf(&build_ji, "%s.ji",build_path) > 0) {
             jl_save_system_image(build_ji);
             free(build_ji);
-            char *build_bc;
-            if (asprintf(&build_bc, "%s.bc",build_path) > 0) {
-                jl_dump_bitcode(build_bc);
-                free(build_bc);
+            char *build_o;
+            if (asprintf(&build_o, "%s.o",build_path) > 0) {
+                jl_dump_objfile(build_o,0);
+                free(build_o);
             }
             else {
-                ios_printf(ios_stderr,"FATAL: failed to create string for .bc build path");
+                ios_printf(ios_stderr,"FATAL: failed to create string for .o build path");
             }
         }
         else {

src/julia_internal.h

@@ -87,6 +87,7 @@ void jl_init_tasks(void *stack, size_t ssize);
 void jl_init_serializer(void);
 
 void jl_dump_bitcode(char *fname);
+void jl_dump_objfile(char *fname, int jit_model);
 int32_t jl_get_llvm_gv(jl_value_t *p);
 
 #ifdef _OS_LINUX_

src/sys.c

@@ -409,15 +409,9 @@ JL_STREAM *jl_stdin_stream(void)  { return (JL_STREAM*)JL_STDIN; }
 JL_STREAM *jl_stdout_stream(void) { return (JL_STREAM*)JL_STDOUT; }
 JL_STREAM *jl_stderr_stream(void) { return (JL_STREAM*)JL_STDERR; }
 
-// -- set/clear the FZ/DAZ flags on x86 & x86-64 --
-
-#ifdef __SSE__
+// CPUID
 
-#ifdef _OS_WINDOWS_
-#define cpuid    __cpuid
-#else
-
-void cpuid(int32_t CPUInfo[4], int32_t InfoType)
+DLLEXPORT void jl_cpuid(int32_t CPUInfo[4], int32_t InfoType)
 {
     __asm__ __volatile__ (
         #if defined(__i386__) && defined(__PIC__)
@@ -436,16 +430,17 @@ void cpuid(int32_t CPUInfo[4], int32_t InfoType)
     );
 }
 
-#endif
+// -- set/clear the FZ/DAZ flags on x86 & x86-64 --
+#ifdef __SSE__
 
 DLLEXPORT uint8_t jl_zero_subnormals(uint8_t isZero)
 {
     uint32_t flags = 0x00000000;
     int32_t info[4];
 
-    cpuid(info, 0);
+    jl_cpuid(info, 0);
     if (info[0] >= 1) {
-        cpuid(info, 0x00000001);
+        jl_cpuid(info, 0x00000001);
         if ((info[3] & ((int)1 << 26)) != 0) {
             // SSE2 supports both FZ and DAZ
             flags = 0x00008040;