Merge pull request #5046 from JuliaLang/kf/llvmcall

Add llvmcall

base/boot.jl

@@ -147,7 +147,7 @@ export
     JULIA_HOME, nothing, Main,
     # intrinsics module
     Intrinsics
-    #ccall, cglobal, abs_float, add_float, add_int, and_int, ashr_int,
+    #ccall, cglobal, llvmcall, abs_float, add_float, add_int, and_int, ashr_int,
     #box, bswap_int, checked_fptosi, checked_fptoui, checked_sadd,
     #checked_smul, checked_ssub, checked_uadd, checked_umul, checked_usub,
     #checked_trunc_sint, checked_trunc_uint,

base/inference.jl

@@ -117,6 +117,10 @@ t_func[nan_dom_err] = (2, 2, (a, b)->a)
 t_func[eval(Core.Intrinsics,:ccall)] =
     (3, Inf, (fptr, rt, at, a...)->(is(rt,Type{Void}) ? Nothing :
                                     isType(rt) ? rt.parameters[1] : Any))
+t_func[eval(Core.Intrinsics,:llvmcall)] =
+    (3, Inf, (fptr, rt, at, a...)->(is(rt,Type{Void}) ? Nothing :
+                                    isType(rt) ? rt.parameters[1] : 
+                                    isa(rt,Tuple) ? map(x->x.parameters[1],rt) : Any))
 t_func[eval(Core.Intrinsics,:cglobal)] =
     (1, 2, (fptr, t...)->(isempty(t) ? Ptr{Void} :
                           isType(t[1]) ? Ptr{t[1].parameters[1]} : Ptr))

src/ccall.cpp

@@ -568,6 +568,212 @@ static Value *emit_cglobal(jl_value_t **args, size_t nargs, jl_codectx_t *ctx)
     return mark_julia_type(res, rt);
 }
 
+// llvmcall(ir, (rettypes...), (argtypes...), args...)
+static Value *emit_llvmcall(jl_value_t **args, size_t nargs, jl_codectx_t *ctx)
+{
+
+    JL_NARGSV(llvmcall, 3)
+    jl_value_t *rt = NULL, *at = NULL, *ir = NULL;
+    JL_GC_PUSH3(&ir, &rt, &at);
+    {
+    JL_TRY {
+        at  = jl_interpret_toplevel_expr_in(ctx->module, args[3],
+                                            &jl_tupleref(ctx->sp,0),
+                                            jl_tuple_len(ctx->sp)/2);
+    }
+    JL_CATCH {
+        jl_rethrow_with_add("error interpreting llvmcall return type");
+    }
+    }
+    {
+    JL_TRY {
+        rt  = jl_interpret_toplevel_expr_in(ctx->module, args[2],
+                                            &jl_tupleref(ctx->sp,0),
+                                            jl_tuple_len(ctx->sp)/2);
+    }
+    JL_CATCH {
+        jl_rethrow_with_add("error interpreting llvmcall argument tuple");
+    }
+    }
+    {
+    JL_TRY {
+        ir  = jl_interpret_toplevel_expr_in(ctx->module, args[1],
+                                            &jl_tupleref(ctx->sp,0),
+                                            jl_tuple_len(ctx->sp)/2);
+    }
+    JL_CATCH {
+        jl_rethrow_with_add("error interpreting IR argument");
+    }
+    }
+    int i = 1;
+    if (ir == NULL) {
+        jl_error("Cannot statically evaluate first argument to llvmcall");
+    }
+    bool isString = jl_is_byte_string(ir);
+    bool isPtr = jl_is_cpointer(ir);
+    if (!isString && !isPtr)
+    {
+        jl_error("First argument to llvmcall must be a string or pointer to an LLVM Function");
+    }
+
+    JL_TYPECHK(llvmcall, type, rt);
+    JL_TYPECHK(llvmcall, tuple, at);
+    JL_TYPECHK(llvmcall, type, at);
+
+    std::stringstream ir_stream;
+
+    jl_tuple_t *stt = jl_alloc_tuple(nargs - 3);
+
+    for (size_t i = 0; i < nargs-3; ++i)
+    {
+        jl_tupleset(stt,i,expr_type(args[4+i],ctx));
+    }
+
+    // Generate arguments
+    std::string arguments;
+    llvm::raw_string_ostream argstream(arguments);
+    jl_tuple_t *tt = (jl_tuple_t*)at;
+    jl_value_t *rtt = rt;
+
+    size_t nargt = jl_tuple_len(tt);
+    Value *argvals[nargt];
+    std::vector<llvm::Type*> argtypes;
+    /*
+     * Semantics for arguments are as follows:
+     * If the argument type is immutable (including bitstype), we pass the loaded llvm value
+     * type. Otherwise we pass a pointer to a jl_value_t.
+     */
+    for (size_t i = 0; i < nargt; ++i)
+    {
+        jl_value_t *tti = jl_tupleref(tt,i);
+        Type *t = julia_type_to_llvm(tti);
+        argtypes.push_back(t);
+        if (4+i > nargs)
+        {
+            jl_error("Missing arguments to llvmcall!");
+        }
+        jl_value_t *argi = args[4+i];
+        Value *arg;
+        bool needroot = false;
+        if (t == jl_pvalue_llvmt || !jl_isbits(tti)) {
+            arg = emit_expr(argi, ctx, true);
+            if (t == jl_pvalue_llvmt && arg->getType() != jl_pvalue_llvmt) {
+                arg = boxed(arg, ctx);
+                needroot = true;
+            }
+        }
+        else {
+            arg = emit_unboxed(argi, ctx);
+            if (jl_is_bitstype(expr_type(argi, ctx))) {
+                arg = emit_unbox(t, arg, tti);
+            }
+        }
+
+#ifdef JL_GC_MARKSWEEP
+        // make sure args are rooted
+        if (t == jl_pvalue_llvmt && (needroot || might_need_root(argi))) {
+            make_gcroot(arg, ctx);
+        }
+#endif
+        bool mightNeedTempSpace = false;
+        argvals[i] = julia_to_native(t,tti,arg,argi,false,i,ctx,&mightNeedTempSpace,&mightNeedTempSpace);
+    }
+
+    Function *f;
+    Type *rettype = julia_type_to_llvm(rtt);
+    if (isString) {
+        // Make sure to find a unique name
+        std::string ir_name;
+        while(true) {
+            std::stringstream name;
+            name << (ctx->f->getName().str()) << i++;
+            ir_name = name.str();
+            if(jl_Module->getFunction(ir_name) == NULL)
+                break;
+        }
+
+        bool first = true;
+        for (std::vector<Type *>::iterator it = argtypes.begin(); it != argtypes.end(); ++it) {
+            if(!first)
+                argstream << ",";
+            else 
+                first = false;
+            (*it)->print(argstream);
+            argstream << " ";
+        }
+
+        std::string rstring;
+        llvm::raw_string_ostream rtypename(rstring);
+        rettype->print(rtypename);
+
+        ir_stream << "; Number of arguments: " << nargt << "\n"
+        << "define "<<rtypename.str()<<" @\"" << ir_name << "\"("<<argstream.str()<<") {\n"
+        << jl_string_data(ir) << "\n}";
+        SMDiagnostic Err = SMDiagnostic();
+        std::string ir_string = ir_stream.str();
+        Module *m = ParseAssemblyString(ir_string.data(),jl_Module,Err,jl_LLVMContext);
+        if (m == NULL) {
+            std::string message = "Failed to parse LLVM Assembly: \n";
+            llvm::raw_string_ostream stream(message);
+            Err.print("julia",stream,true);
+            jl_error(stream.str().c_str());
+        }
+        f = m->getFunction(ir_name);
+    } else {
+        assert(isPtr);
+        // Create Function skeleton
+        f = (llvm::Function*)jl_unbox_voidpointer(ir);
+        assert(f->getReturnType() == rettype);
+        int i = 0;
+        for (std::vector<Type *>::iterator it = argtypes.begin();
+            it != argtypes.end(); ++it, ++i)
+            assert(*it == f->getFunctionType()->getParamType(i));
+
+#ifdef USE_MCJIT
+        if (f->getParent() != jl_Module)
+        {
+            FunctionMover mover(jl_Module,f->getParent());
+            f = (llvm::Function*)MapValue(f,mover.VMap,RF_None,NULL,&mover);
+        }
+#endif
+
+        //f->dump();
+        #ifndef LLVM35
+        if (verifyFunction(*f,PrintMessageAction)) {
+        #else
+        llvm::raw_fd_ostream out(1,false);
+        if (verifyFunction(*f,&out))
+        {
+        #endif
+            f->dump();
+            jl_error("Malformed LLVM Function");
+        }
+    }
+
+    /*
+     * It might be tempting to just try to set the Always inline attribute on the function
+     * and hope for the best. However, this doesn't work since that would require an inlining
+     * pass (which is a Call Graph pass and cannot be managed by a FunctionPassManager). Instead
+     * We are sneaky and call the inliner directly. This however doesn't work until we've actually
+     * generated the entire function, so we need to store it in the context until the end of the
+     * function. This also has the benefit of looking exactly like we cut/pasted it in in `code_llvm`.
+     */
+    f->setLinkage(GlobalValue::LinkOnceODRLinkage);
+
+    // the actual call
+    CallInst *inst = builder.CreateCall(prepare_call(f),ArrayRef<Value*>(&argvals[0],nargt));
+    ctx->to_inline.push_back(inst);
+
+    JL_GC_POP();
+
+    if(inst->getType() != rettype)
+    {
+        jl_error("Return type of llvmcall'ed function does not match declared return type");
+    }
+
+    return mark_julia_type(inst,rtt);
+}
+
 // --- code generator for ccall itself ---
 
 // ccall(pointer, rettype, (argtypes...), args...)

src/codegen.cpp

@@ -48,6 +48,7 @@
 #include "llvm/Target/TargetMachine.h"
 #else
 #include "llvm/Analysis/Verifier.h"
+#include "llvm/Assembly/Parser.h"
 #endif
 #include "llvm/DebugInfo/DIContext.h"
 #if defined(LLVM_VERSION_MAJOR) && LLVM_VERSION_MAJOR == 3 && LLVM_VERSION_MINOR >= 4

src/intrinsics.cpp

@@ -39,7 +39,7 @@ namespace JL_I {
         // pointer access
         pointerref, pointerset, pointertoref,
         // c interface
-        ccall, cglobal, jl_alloca
+        ccall, cglobal, jl_alloca, llvmcall
     };
 };
 
@@ -831,6 +831,7 @@ static Value *emit_intrinsic(intrinsic f, jl_value_t **args, size_t nargs,
     switch (f) {
     case ccall: return emit_ccall(args, nargs, ctx);
     case cglobal: return emit_cglobal(args, nargs, ctx);
+    case llvmcall: return emit_llvmcall(args, nargs, ctx);
 
     HANDLE(box,2)         return generic_box(args[1], args[2], ctx);
     HANDLE(unbox,2)       return generic_unbox(args[1], args[2], ctx);
@@ -1459,4 +1460,5 @@ extern "C" void jl_init_intrinsic_functions(void)
     ADD_I(nan_dom_err);
     ADD_I(ccall); ADD_I(cglobal);
     ADD_I(jl_alloca);
+    ADD_I(llvmcall);
 }

test/Makefile

@@ -8,7 +8,7 @@ TESTS = all core keywordargs numbers strings unicode collections hashing	\
 	git pkg resolve suitesparse complex version pollfd mpfr	broadcast       \
 	socket floatapprox priorityqueue readdlm regex float16 combinatorics    \
 	sysinfo rounding ranges mod2pi euler show lineedit      \
-	replcompletions backtrace repl test goto
+	replcompletions backtrace repl test goto llvmcall
 
 default: all
 

test/llvmcall.jl

@@ -0,0 +1,33 @@
+using Base.llvmcall
+
+function add1234(x::(Int32,Int32,Int32,Int32))
+    llvmcall("""%3 = add <4 x i32> %1, %0
+      ret <4 x i32> %3""",(Int32,Int32,Int32,Int32),
+      ((Int32,Int32,Int32,Int32),(Int32,Int32,Int32,Int32)),
+        (int32(1),int32(2),int32(3),int32(4)),
+      x)
+end
+
+function add1234(x::NTuple{4,Int64})
+    llvmcall("""%3 = add <4 x i64> %1, %0
+      ret <4 x i64> %3""",NTuple{4,Int64},
+      (NTuple{4,Int64},NTuple{4,Int64}),
+        (int64(1),int64(2),int64(3),int64(4)),
+      x)
+end
+
+@test add1234(map(int32,(2,3,4,5))) === map(int32,(3,5,7,9))
+@test add1234(map(int64,(2,3,4,5))) === map(int64,(3,5,7,9))
+
+# Test whether llvmcall escapes the function name correctly
+baremodule PlusTest
+    using Base.llvmcall
+    using Base.Test
+    using Base
+
+    function +(x::Int32, y::Int32)
+        llvmcall("""%3 = add i32 %1, %0
+                    ret i32 %3""", Int32, (Int32, Int32), x, y)
+    end
+    @test int32(1)+int32(2)==int32(3)
+end

test/runtests.jl

@@ -8,7 +8,7 @@ testnames = [
     "resolve", "pollfd", "mpfr", "broadcast", "complex", "socket",
     "floatapprox", "readdlm", "regex", "float16", "combinatorics",
     "sysinfo", "rounding", "ranges", "mod2pi", "euler", "show",
-    "lineedit", "replcompletions", "repl", "test", "examples", "goto"
+    "lineedit", "replcompletions", "repl", "test", "examples", "goto", "llvmcall"
 ]
 @unix_only push!(testnames, "unicode")