[pull] master from GaijinEntertainment:master

include/daScript/simulate/aot_builtin.h

@@ -179,4 +179,7 @@ namespace das {
     void builtin_main_loop ( const TBlock<bool> & block, Context * context, LineInfoArg * at );
 
     vec4f _builtin_hash ( Context & context, SimNode_CallBase * call, vec4f * args );
+
+    const char * das_get_platform_name();
+    const char * das_get_architecture_name();
 }

modules/dasLLVM/daslib/llvm_boost.das

@@ -176,12 +176,9 @@ def LLVMIsVector3(typ : LLVMOpaqueType?) : bool {
 def LLVMBuildLoadVector3(builder : LLVMOpaqueBuilder?; var types : PrimitiveTypes?; var typ : LLVMOpaqueType?; ptr : LLVMOpaqueValue?; name : string) {
     verify(LLVMGetVectorSize(typ) == 3u)
     let elemT = LLVMGetElementType(typ)
-    var ld = LLVMBuildLoad2(builder, LLVMVectorType(elemT, 4u), ptr, name)
-    LLVMSetVolatile(ld, 1)
-    LLVMSetAlignment(ld, 4u)
-    var x = LLVMBuildExtractElement(builder, ld, types.ConstI32(0ul), "")
-    var y = LLVMBuildExtractElement(builder, ld, types.ConstI32(1ul), "")
-    var z = LLVMBuildExtractElement(builder, ld, types.ConstI32(2ul), "")
+    var x = LLVMBuildLoad2(builder, elemT, ptr, "")
+    var y = LLVMBuildLoad2(builder, elemT, LLVMBuildGEP2(builder, elemT, ptr, types.ConstI32(1ul), "", true), "")
+    var z = LLVMBuildLoad2(builder, elemT, LLVMBuildGEP2(builder, elemT, ptr, types.ConstI32(2ul), "", true), "")
     var xyz = LLVMGetUndef(LLVMVectorType(elemT, 3u))
     xyz = LLVMBuildInsertElement(builder, xyz, x, types.ConstI32(0ul), "")
     xyz = LLVMBuildInsertElement(builder, xyz, y, types.ConstI32(1ul), "")

modules/dasLLVM/daslib/llvm_jit.das

@@ -240,7 +240,7 @@ class LlvmJitVisitor : AstVisitor {
     option_no_alias : bool = false
     option_no_capture : bool = false
     own_di : bool = false
-    ldu_hint : array<bool>
+    ldu_hint : bool
     attributes : Attributes?
 
     def LlvmJitVisitor(ctx : Context?; var types_ : PrimitiveTypes?; uids : UidNodes?; need_di : bool; dib : LLVMOpaqueDIBuilder?; attrs : Attributes?) {
@@ -698,7 +698,7 @@ class LlvmJitVisitor : AstVisitor {
         assert(g_builder != null, "missing builder")
         thisFunc = get_ptr(fun)
         uid.reset(get_ptr(fun))
-        ldu_hint |> push(LLVM_JIT_ALLOW_UNALIGNED_VECTOR_READ_OUT_OF_BOUNDS)
+        ldu_hint = LLVM_JIT_ALLOW_UNALIGNED_VECTOR_READ_OUT_OF_BOUNDS
         let fnmna = uid.get_dll_fn_name(fun)
         ffunc = LLVMGetNamedFunction(g_mod, fnmna.impl())
         wfunc = LLVMGetNamedFunction(g_mod, fnmna.publ())
@@ -731,7 +731,7 @@ class LlvmJitVisitor : AstVisitor {
         LLVMPositionBuilderAtEnd(g_builder, function_entry)
         LLVMBuildBr(g_builder, function_body)
         debug_after_function()
-        ldu_hint |> pop()
+        LLVM_JIT_ALLOW_UNALIGNED_VECTOR_READ_OUT_OF_BOUNDS = ldu_hint
         ffunc = null
         thisFunc = null
         uid.reset(null)
@@ -6190,6 +6190,9 @@ class JIT_LLVM : AstSimulateMacro {
         assume path_to_shared_lib = string(prog.policies.jit_path_to_shared_lib)
         assume path_to_linker = string(prog.policies.jit_path_to_linker)
 
+        // Set global options
+        LLVM_JIT_ALLOW_UNALIGNED_VECTOR_READ_OUT_OF_BOUNDS = prog._options |> find_arg("jit_vec3_ldu") ?as tBool ?? LLVM_JIT_ALLOW_UNALIGNED_VECTOR_READ_OUT_OF_BOUNDS_DEFAULT
+
         prog |> for_each_module <| $(mod) {
             mod |> for_each_function("") <| $(fun) {
                 var rqj = fun.moreFlags.requestJit
@@ -6325,3 +6328,9 @@ class JIT_LLVM : AstSimulateMacro {
 }
 
 
+[_macro, macro_function]
+def init_llvm_jit_module_options() {
+    if (is_compiling_macros_in_module("llvm_jit")) {
+        this_module() |> add_module_option("jit_vec3_ldu", Type.tBool)
+    }
+}

modules/dasLLVM/daslib/llvm_jit_common.das

@@ -30,7 +30,8 @@ let public LLVM_DEBUG_LINE_TRACES = false
 
 let public LLVM_JIT_LOG = false || LLVM_DEBUG_EVERYTHING
 
-var public LLVM_JIT_ALLOW_UNALIGNED_VECTOR_READ_OUT_OF_BOUNDS = false
+let public LLVM_JIT_ALLOW_UNALIGNED_VECTOR_READ_OUT_OF_BOUNDS_DEFAULT = false   // NOTE - this should be true, to match interpreter and AOT behavior
+var public LLVM_JIT_ALLOW_UNALIGNED_VECTOR_READ_OUT_OF_BOUNDS = LLVM_JIT_ALLOW_UNALIGNED_VECTOR_READ_OUT_OF_BOUNDS_DEFAULT
 
 let public LLVM_ENABLE_OPT_PASS = LLVM_DEBUG_EVERYTHING ? false : true
 let public LLVM_LOG_RESULT = LLVM_DEBUG_EVERYTHING || LLVM_DEBUG_RESULT
@@ -869,9 +870,8 @@ def public build_broadcast_vector(builder : LLVMOpaqueBuilder?; opType : TypeDec
     }
 }
 
-    def public LLVMBuildLoadData2Aligned(builder : LLVMOpaqueBuilder?; var typ : LLVMOpaqueType?; ptr : LLVMOpaqueValue?; alignment : int; name : string) {
+def public LLVMBuildLoadData2Aligned(builder : LLVMOpaqueBuilder?; var typ : LLVMOpaqueType?; ptr : LLVMOpaqueValue?; alignment : int; name : string) {
     if (!LLVM_JIT_ALLOW_UNALIGNED_VECTOR_READ_OUT_OF_BOUNDS && alignment != 16 && LLVMIsVector3(typ)) {
-        // slow path
         return LLVMBuildLoadVector3(builder, g_prim_t, typ, ptr, name)
     } else {
         return LLVMBuildLoad2Aligned(builder, typ, ptr, alignment |> uint, name)

modules/dasLLVM/profile/nbodies.das

@@ -46,7 +46,6 @@ let {
     nbodies = 5
 }
 
-[jit]
 def offset_momentum(var bodies : body[5]) {
     var px : float3
     for (b in bodies) {
@@ -55,7 +54,7 @@ def offset_momentum(var bodies : body[5]) {
     bodies[0].v = px / SOLAR_MASS
 }
 
-[jit, hint(alwaysinline, hot, noalias=bodies, unsafe_range_check)]
+[hint(vec3_ldu)]
 def advance(var bodies : body[5]) {
     unroll <| $() {
         for (i in range(nbodies)) {
@@ -74,7 +73,6 @@ def advance(var bodies : body[5]) {
     }
 }
 
-[jit]
 def energy(var bodies : body[5]) {
     var e = 0.0
     var i = 0
@@ -89,15 +87,13 @@ def energy(var bodies : body[5]) {
     return e
 }
 
-[jit]
 def scale_bodies(scale; var bodies : body[5]) {
     for (b in bodies) {
         b.mass *= scale * scale
         b.v *= scale
     }
 }
 
-[jit]
 def nbodies(n : int) {
     scale_bodies(0.01, g_bodies)
     for (i in range(n)) {

src/builtin/module_builtin_runtime.cpp

@@ -1603,6 +1603,38 @@ namespace das
         idpi->noPointerCast = true;
     }
 
+    // windows, darwin, linux, etc
+    const char * das_get_platform_name() {
+        #if defined(_WIN32) || defined(_WIN64)
+            return "windows";
+        #elif defined(__APPLE__) || defined(__MACH__)
+            return "darwin";
+        #elif defined(__linux__)
+            return "linux";
+        #elif defined(__EMSCRIPTEN__)
+            return "emscripten";
+        #else
+            return "unknown";
+        #endif
+    }
+
+    // x86, arm, etc
+    const char * das_get_architecture_name() {
+        #if defined(__x86_64__) || defined(_M_X64)
+            return "x86_64";
+        #elif defined(__i386) || defined(_M_IX86)
+            return "x86";
+        #elif defined(__aarch64__)
+            return "arm64";
+        #elif defined(__arm__) || defined(_M_ARM)
+            return "arm";
+        #elif defined(__EMSCRIPTEN__)
+            return "wasm32";
+        #else
+            return "unknown";
+        #endif
+    }
+
     void Module_BuiltIn::addRuntime(ModuleLibrary & lib) {
         // printer flags
         addAlias(makePrintFlags());
@@ -2133,5 +2165,10 @@ namespace das
         addExtern<DAS_BIND_FUN(__bit_set64)>(*this, lib, "__bit_set",
             SideEffects::modifyArgument, "__bit_set64")
                 ->args({"value","mask","on"});
+        // platform and architecture
+        addExtern<DAS_BIND_FUN(das_get_platform_name)>(*this, lib, "get_platform_name",
+            SideEffects::none, "das_get_platform_name");
+        addExtern<DAS_BIND_FUN(das_get_architecture_name)>(*this, lib, "get_architecture_name",
+            SideEffects::none, "das_get_architecture_name");
     }
 }