/**
* Gather one element from scatter positions in memory.
*
* @sa lp_build_gather()
*/
LLVMValueRef
lp_build_gather_elem(struct gallivm_state *gallivm,
unsigned length,
unsigned src_width,
unsigned dst_width,
LLVMValueRef base_ptr,
LLVMValueRef offsets,
unsigned i)
{
LLVMTypeRef src_type = LLVMIntTypeInContext(gallivm->context, src_width);
LLVMTypeRef src_ptr_type = LLVMPointerType(src_type, 0);
LLVMTypeRef dst_elem_type = LLVMIntTypeInContext(gallivm->context, dst_width);
LLVMValueRef ptr;
LLVMValueRef res;
assert(LLVMTypeOf(base_ptr) == LLVMPointerType(LLVMInt8TypeInContext(gallivm->context), 0));
ptr = lp_build_gather_elem_ptr(gallivm, length, base_ptr, offsets, i);
ptr = LLVMBuildBitCast(gallivm->builder, ptr, src_ptr_type, "");
res = LLVMBuildLoad(gallivm->builder, ptr, "");
assert(src_width <= dst_width);
if (src_width > dst_width)
res = LLVMBuildTrunc(gallivm->builder, res, dst_elem_type, "");
if (src_width < dst_width)
res = LLVMBuildZExt(gallivm->builder, res, dst_elem_type, "");
return res;
}
/**
* Generates LLVM IR to call debug_printf.
*/
static LLVMValueRef
lp_build_print_args(struct gallivm_state* gallivm,
int argcount,
LLVMValueRef* args)
{
LLVMBuilderRef builder = gallivm->builder;
LLVMContextRef context = gallivm->context;
LLVMValueRef func_printf;
LLVMTypeRef printf_type;
int i;
assert(args);
assert(argcount > 0);
assert(LLVMTypeOf(args[0]) == LLVMPointerType(LLVMInt8TypeInContext(context), 0));
/* Cast any float arguments to doubles as printf expects */
for (i = 1; i < argcount; i++) {
LLVMTypeRef type = LLVMTypeOf(args[i]);
if (LLVMGetTypeKind(type) == LLVMFloatTypeKind)
args[i] = LLVMBuildFPExt(builder, args[i], LLVMDoubleTypeInContext(context), "");
}
printf_type = LLVMFunctionType(LLVMInt32TypeInContext(context), NULL, 0, 1);
func_printf = lp_build_const_int_pointer(gallivm, func_to_pointer((func_pointer)debug_printf));
func_printf = LLVMBuildBitCast(builder, func_printf, LLVMPointerType(printf_type, 0), "debug_printf");
return LLVMBuildCall(builder, func_printf, args, argcount, "");
}
/**
* Gather one element from scatter positions in memory.
*
* @sa lp_build_gather()
*/
LLVMValueRef
lp_build_gather_elem(struct gallivm_state *gallivm,
unsigned length,
unsigned src_width,
unsigned dst_width,
boolean aligned,
LLVMValueRef base_ptr,
LLVMValueRef offsets,
unsigned i,
boolean vector_justify)
{
LLVMTypeRef src_type = LLVMIntTypeInContext(gallivm->context, src_width);
LLVMTypeRef src_ptr_type = LLVMPointerType(src_type, 0);
LLVMTypeRef dst_elem_type = LLVMIntTypeInContext(gallivm->context, dst_width);
LLVMValueRef ptr;
LLVMValueRef res;
assert(LLVMTypeOf(base_ptr) == LLVMPointerType(LLVMInt8TypeInContext(gallivm->context), 0));
ptr = lp_build_gather_elem_ptr(gallivm, length, base_ptr, offsets, i);
ptr = LLVMBuildBitCast(gallivm->builder, ptr, src_ptr_type, "");
res = LLVMBuildLoad(gallivm->builder, ptr, "");
/* XXX
* On some archs we probably really want to avoid having to deal
* with alignments lower than 4 bytes (if fetch size is a power of
* two >= 32). On x86 it doesn't matter, however.
* We should be able to guarantee full alignment for any kind of texture
* fetch (except ARB_texture_buffer_range, oops), but not vertex fetch
* (there's PIPE_CAP_VERTEX_BUFFER_OFFSET_4BYTE_ALIGNED_ONLY and friends
* but I don't think that's quite what we wanted).
* For ARB_texture_buffer_range, PIPE_CAP_TEXTURE_BUFFER_OFFSET_ALIGNMENT
* looks like a good fit, but it seems this cap bit (and OpenGL) aren't
* enforcing what we want (which is what d3d10 does, the offset needs to
* be aligned to element size, but GL has bytes regardless of element
* size which would only leave us with minimum alignment restriction of 16
* which doesn't make much sense if the type isn't 4x32bit). Due to
* translation of offsets to first_elem in sampler_views it actually seems
* gallium could not do anything else except 16 no matter what...
*/
if (!aligned) {
LLVMSetAlignment(res, 1);
}
assert(src_width <= dst_width);
if (src_width > dst_width) {
res = LLVMBuildTrunc(gallivm->builder, res, dst_elem_type, "");
} else if (src_width < dst_width) {
res = LLVMBuildZExt(gallivm->builder, res, dst_elem_type, "");
if (vector_justify) {
#ifdef PIPE_ARCH_BIG_ENDIAN
res = LLVMBuildShl(gallivm->builder, res,
LLVMConstInt(dst_elem_type, dst_width - src_width, 0), "");
#endif
}
}
return res;
}
LLVMTypeRef ty_class_indirect()
{
if(indirect_struct_type)
return indirect_struct_type;
indirect_struct_type = LLVMStructCreateNamed(utl_get_current_context(), "__egl_class_indirect");
LLVMTypeRef tys[] = {
LLVMInt32TypeInContext(utl_get_current_context()),
LLVMPointerType(LLVMPointerType(LLVMInt8TypeInContext(utl_get_current_context()), 0), 0),
LLVMPointerType(LLVMInt64TypeInContext(utl_get_current_context()), 0),
LLVMPointerType(LLVMPointerType(LLVMInt8TypeInContext(utl_get_current_context()), 0), 0)
};
LLVMStructSetBody(indirect_struct_type, tys, 4, 0);
return indirect_struct_type;
}
/* returns a LLVM representation corresponding to the C translation of the
* given IDL type.
*/
LLVMTypeRef llvm_value_type(struct llvm_ctx *ctx, IDL_tree type)
{
if(type == NULL) return LLVMVoidTypeInContext(ctx->ctx);
switch(IDL_NODE_TYPE(type)) {
case IDLN_TYPE_INTEGER: {
static short bitlens[] = {
[IDL_INTEGER_TYPE_SHORT] = 16,
[IDL_INTEGER_TYPE_LONG] = 32,
[IDL_INTEGER_TYPE_LONGLONG] = 64,
};
int t = IDL_TYPE_INTEGER(type).f_type;
assert(t < G_N_ELEMENTS(bitlens));
return LLVMIntTypeInContext(ctx->ctx, bitlens[t]);
}
case IDLN_NATIVE: {
/* each of these is the size of a single word, which is all LLVM
* wants to know.
*/
if(IS_WORD_TYPE(type) || IS_FPAGE_TYPE(type)
|| IS_TIME_TYPE(type))
{
return ctx->wordt;
} else {
fprintf(stderr, "%s: native type `%s' not supported\n",
__FUNCTION__, NATIVE_NAME(type));
abort();
}
break;
}
case IDLN_TYPE_FLOAT:
switch(IDL_TYPE_FLOAT(type).f_type) {
case IDL_FLOAT_TYPE_FLOAT:
return LLVMFloatTypeInContext(ctx->ctx);
case IDL_FLOAT_TYPE_DOUBLE:
return LLVMDoubleTypeInContext(ctx->ctx);
case IDL_FLOAT_TYPE_LONGDOUBLE:
return LLVMFP128TypeInContext(ctx->ctx);
}
g_assert_not_reached();
case IDLN_TYPE_BOOLEAN:
case IDLN_TYPE_OCTET:
case IDLN_TYPE_CHAR:
return LLVMInt8TypeInContext(ctx->ctx);
case IDLN_TYPE_WIDE_CHAR:
return ctx->i32t;
case IDLN_TYPE_ENUM: return LLVMInt16TypeInContext(ctx->ctx);
default:
NOTDEFINED(type);
}
}
LLVMValueRef
lp_build_const_string_variable(LLVMModuleRef module,
LLVMContextRef context,
const char *str, int len)
{
LLVMValueRef string = LLVMAddGlobal(module, LLVMArrayType(LLVMInt8TypeInContext(context), len + 1), "");
LLVMSetGlobalConstant(string, TRUE);
LLVMSetLinkage(string, LLVMInternalLinkage);
LLVMSetInitializer(string, LLVMConstStringInContext(context, str, len + 1, TRUE));
return string;
}
LLVMTypeRef ett_closure_type(EagleComplexType *type) { if(!ET_IS_CLOSURE(type)) return NULL;
EagleFunctionType *ft = (EagleFunctionType *)type;
LLVMTypeRef *tys = malloc(sizeof(LLVMTypeRef) * (ft->pct + 1));
int i;
for(i = 1; i < ft->pct + 1; i++)
tys[i] = ett_llvm_type(ft->params[i - 1]);
tys[0] = LLVMPointerType(LLVMInt8TypeInContext(utl_get_current_context()), 0);
LLVMTypeRef out = LLVMFunctionType(ett_llvm_type(ft->retType), tys, ft->pct + 1, 0);
free(tys);
return out;
}
/**
* @param n is the number of pixels processed
* @param packed is a <n x i32> vector with the packed YUYV blocks
* @param i is a <n x i32> vector with the x pixel coordinate (0 or 1)
* @return a <4*n x i8> vector with the pixel RGBA values in AoS
*/
LLVMValueRef
lp_build_fetch_subsampled_rgba_aos(struct gallivm_state *gallivm,
const struct util_format_description *format_desc,
unsigned n,
LLVMValueRef base_ptr,
LLVMValueRef offset,
LLVMValueRef i,
LLVMValueRef j)
{
LLVMValueRef packed;
LLVMValueRef rgba;
struct lp_type fetch_type;
assert(format_desc->layout == UTIL_FORMAT_LAYOUT_SUBSAMPLED);
assert(format_desc->block.bits == 32);
assert(format_desc->block.width == 2);
assert(format_desc->block.height == 1);
fetch_type = lp_type_uint(32);
packed = lp_build_gather(gallivm, n, 32, fetch_type, TRUE, base_ptr, offset, FALSE);
(void)j;
switch (format_desc->format) {
case PIPE_FORMAT_UYVY:
rgba = uyvy_to_rgba_aos(gallivm, n, packed, i);
break;
case PIPE_FORMAT_YUYV:
rgba = yuyv_to_rgba_aos(gallivm, n, packed, i);
break;
case PIPE_FORMAT_R8G8_B8G8_UNORM:
rgba = rgbg_to_rgba_aos(gallivm, n, packed, i);
break;
case PIPE_FORMAT_G8R8_G8B8_UNORM:
rgba = grgb_to_rgba_aos(gallivm, n, packed, i);
break;
case PIPE_FORMAT_G8R8_B8R8_UNORM:
rgba = grbr_to_rgba_aos(gallivm, n, packed, i);
break;
case PIPE_FORMAT_R8G8_R8B8_UNORM:
rgba = rgrb_to_rgba_aos(gallivm, n, packed, i);
break;
default:
assert(0);
rgba = LLVMGetUndef(LLVMVectorType(LLVMInt8TypeInContext(gallivm->context), 4*n));
break;
}
return rgba;
}
static LLVMValueRef
rgb_to_rgba_aos(struct gallivm_state *gallivm,
unsigned n,
LLVMValueRef r, LLVMValueRef g, LLVMValueRef b)
{
LLVMBuilderRef builder = gallivm->builder;
struct lp_type type;
LLVMValueRef a;
LLVMValueRef rgba;
memset(&type, 0, sizeof type);
type.sign = TRUE;
type.width = 32;
type.length = n;
assert(lp_check_value(type, r));
assert(lp_check_value(type, g));
assert(lp_check_value(type, b));
/*
* Make a 4 x unorm8 vector
*/
#ifdef PIPE_ARCH_LITTLE_ENDIAN
r = r;
g = LLVMBuildShl(builder, g, lp_build_const_int_vec(gallivm, type, 8), "");
b = LLVMBuildShl(builder, b, lp_build_const_int_vec(gallivm, type, 16), "");
a = lp_build_const_int_vec(gallivm, type, 0xff000000);
#else
r = LLVMBuildShl(builder, r, lp_build_const_int_vec(gallivm, type, 24), "");
g = LLVMBuildShl(builder, g, lp_build_const_int_vec(gallivm, type, 16), "");
b = LLVMBuildShl(builder, b, lp_build_const_int_vec(gallivm, type, 8), "");
a = lp_build_const_int_vec(gallivm, type, 0x000000ff);
#endif
rgba = r;
rgba = LLVMBuildOr(builder, rgba, g, "");
rgba = LLVMBuildOr(builder, rgba, b, "");
rgba = LLVMBuildOr(builder, rgba, a, "");
rgba = LLVMBuildBitCast(builder, rgba,
LLVMVectorType(LLVMInt8TypeInContext(gallivm->context), 4*n), "");
return rgba;
}
/* Initialize module-independent parts of the context.
*
* The caller is responsible for initializing ctx::module and ctx::builder.
*/
void
ac_llvm_context_init(struct ac_llvm_context *ctx, LLVMContextRef context)
{
LLVMValueRef args[1];
ctx->context = context;
ctx->module = NULL;
ctx->builder = NULL;
ctx->voidt = LLVMVoidTypeInContext(ctx->context);
ctx->i1 = LLVMInt1TypeInContext(ctx->context);
ctx->i8 = LLVMInt8TypeInContext(ctx->context);
ctx->i16 = LLVMIntTypeInContext(ctx->context, 16);
ctx->i32 = LLVMIntTypeInContext(ctx->context, 32);
ctx->i64 = LLVMIntTypeInContext(ctx->context, 64);
ctx->f16 = LLVMHalfTypeInContext(ctx->context);
ctx->f32 = LLVMFloatTypeInContext(ctx->context);
ctx->f64 = LLVMDoubleTypeInContext(ctx->context);
ctx->v4i32 = LLVMVectorType(ctx->i32, 4);
ctx->v4f32 = LLVMVectorType(ctx->f32, 4);
ctx->v8i32 = LLVMVectorType(ctx->i32, 8);
ctx->i32_0 = LLVMConstInt(ctx->i32, 0, false);
ctx->i32_1 = LLVMConstInt(ctx->i32, 1, false);
ctx->f32_0 = LLVMConstReal(ctx->f32, 0.0);
ctx->f32_1 = LLVMConstReal(ctx->f32, 1.0);
ctx->range_md_kind = LLVMGetMDKindIDInContext(ctx->context,
"range", 5);
ctx->invariant_load_md_kind = LLVMGetMDKindIDInContext(ctx->context,
"invariant.load", 14);
ctx->fpmath_md_kind = LLVMGetMDKindIDInContext(ctx->context, "fpmath", 6);
args[0] = LLVMConstReal(ctx->f32, 2.5);
ctx->fpmath_md_2p5_ulp = LLVMMDNodeInContext(ctx->context, args, 1);
ctx->uniform_md_kind = LLVMGetMDKindIDInContext(ctx->context,
"amdgpu.uniform", 14);
ctx->empty_md = LLVMMDNodeInContext(ctx->context, NULL, 0);
}
/**
* Get the pointer to one element from scatter positions in memory.
*
* @sa lp_build_gather()
*/
LLVMValueRef
lp_build_gather_elem_ptr(struct gallivm_state *gallivm,
unsigned length,
LLVMValueRef base_ptr,
LLVMValueRef offsets,
unsigned i)
{
LLVMValueRef offset;
LLVMValueRef ptr;
assert(LLVMTypeOf(base_ptr) == LLVMPointerType(LLVMInt8TypeInContext(gallivm->context), 0));
if (length == 1) {
assert(i == 0);
offset = offsets;
} else {
LLVMValueRef index = lp_build_const_int32(gallivm, i);
offset = LLVMBuildExtractElement(gallivm->builder, offsets, index, "");
}
ptr = LLVMBuildGEP(gallivm->builder, base_ptr, &offset, 1, "");
return ptr;
}
static void init_runtime(compile_t* c)
{
c->str_builtin = stringtab("$0");
c->str_Bool = stringtab("Bool");
c->str_I8 = stringtab("I8");
c->str_I16 = stringtab("I16");
c->str_I32 = stringtab("I32");
c->str_I64 = stringtab("I64");
c->str_I128 = stringtab("I128");
c->str_ILong = stringtab("ILong");
c->str_ISize = stringtab("ISize");
c->str_U8 = stringtab("U8");
c->str_U16 = stringtab("U16");
c->str_U32 = stringtab("U32");
c->str_U64 = stringtab("U64");
c->str_U128 = stringtab("U128");
c->str_ULong = stringtab("ULong");
c->str_USize = stringtab("USize");
c->str_F32 = stringtab("F32");
c->str_F64 = stringtab("F64");
c->str_Pointer = stringtab("Pointer");
c->str_Maybe = stringtab("MaybePointer");
c->str_DoNotOptimise = stringtab("DoNotOptimise");
c->str_Array = stringtab("Array");
c->str_String = stringtab("String");
c->str_Platform = stringtab("Platform");
c->str_Main = stringtab("Main");
c->str_Env = stringtab("Env");
c->str_add = stringtab("add");
c->str_sub = stringtab("sub");
c->str_mul = stringtab("mul");
c->str_div = stringtab("div");
c->str_mod = stringtab("mod");
c->str_neg = stringtab("neg");
c->str_and = stringtab("op_and");
c->str_or = stringtab("op_or");
c->str_xor = stringtab("op_xor");
c->str_not = stringtab("op_not");
c->str_shl = stringtab("shl");
c->str_shr = stringtab("shr");
c->str_eq = stringtab("eq");
c->str_ne = stringtab("ne");
c->str_lt = stringtab("lt");
c->str_le = stringtab("le");
c->str_ge = stringtab("ge");
c->str_gt = stringtab("gt");
c->str_this = stringtab("this");
c->str_create = stringtab("create");
c->str__create = stringtab("_create");
c->str__init = stringtab("_init");
c->str__final = stringtab("_final");
c->str__event_notify = stringtab("_event_notify");
LLVMTypeRef type;
LLVMTypeRef params[5];
LLVMValueRef value;
c->void_type = LLVMVoidTypeInContext(c->context);
c->ibool = LLVMInt8TypeInContext(c->context);
c->i1 = LLVMInt1TypeInContext(c->context);
c->i8 = LLVMInt8TypeInContext(c->context);
c->i16 = LLVMInt16TypeInContext(c->context);
c->i32 = LLVMInt32TypeInContext(c->context);
c->i64 = LLVMInt64TypeInContext(c->context);
c->i128 = LLVMIntTypeInContext(c->context, 128);
c->f32 = LLVMFloatTypeInContext(c->context);
c->f64 = LLVMDoubleTypeInContext(c->context);
c->intptr = LLVMIntPtrTypeInContext(c->context, c->target_data);
// i8*
c->void_ptr = LLVMPointerType(c->i8, 0);
// forward declare object
c->object_type = LLVMStructCreateNamed(c->context, "__object");
c->object_ptr = LLVMPointerType(c->object_type, 0);
// padding required in an actor between the descriptor and fields
c->actor_pad = LLVMArrayType(c->i8, PONY_ACTOR_PAD_SIZE);
// message
params[0] = c->i32; // size
params[1] = c->i32; // id
c->msg_type = LLVMStructCreateNamed(c->context, "__message");
c->msg_ptr = LLVMPointerType(c->msg_type, 0);
LLVMStructSetBody(c->msg_type, params, 2, false);
// trace
// void (*)(i8*, __object*)
params[0] = c->void_ptr;
params[1] = c->object_ptr;
c->trace_type = LLVMFunctionType(c->void_type, params, 2, false);
c->trace_fn = LLVMPointerType(c->trace_type, 0);
// serialise
// void (*)(i8*, __object*, i8*, intptr, i32)
params[0] = c->void_ptr;
params[1] = c->object_ptr;
params[2] = c->void_ptr;
params[3] = c->intptr;
params[4] = c->i32;
c->serialise_type = LLVMFunctionType(c->void_type, params, 5, false);
c->serialise_fn = LLVMPointerType(c->serialise_type, 0);
// dispatch
// void (*)(i8*, __object*, $message*)
params[0] = c->void_ptr;
params[1] = c->object_ptr;
params[2] = c->msg_ptr;
c->dispatch_type = LLVMFunctionType(c->void_type, params, 3, false);
c->dispatch_fn = LLVMPointerType(c->dispatch_type, 0);
// void (*)(__object*)
params[0] = c->object_ptr;
c->final_fn = LLVMPointerType(
LLVMFunctionType(c->void_type, params, 1, false), 0);
// descriptor, opaque version
// We need this in order to build our own structure.
const char* desc_name = genname_descriptor(NULL);
c->descriptor_type = LLVMStructCreateNamed(c->context, desc_name);
c->descriptor_ptr = LLVMPointerType(c->descriptor_type, 0);
// field descriptor
// Also needed to build a descriptor structure.
params[0] = c->i32;
params[1] = c->descriptor_ptr;
c->field_descriptor = LLVMStructTypeInContext(c->context, params, 2, false);
// descriptor, filled in
gendesc_basetype(c, c->descriptor_type);
// define object
params[0] = c->descriptor_ptr;
LLVMStructSetBody(c->object_type, params, 1, false);
// $i8* pony_ctx()
type = LLVMFunctionType(c->void_ptr, NULL, 0, false);
value = LLVMAddFunction(c->module, "pony_ctx", type);
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
LLVMAddFunctionAttr(value, LLVMReadNoneAttribute);
// __object* pony_create(i8*, __Desc*)
params[0] = c->void_ptr;
params[1] = c->descriptor_ptr;
type = LLVMFunctionType(c->object_ptr, params, 2, false);
value = LLVMAddFunction(c->module, "pony_create", type);
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
LLVMSetReturnNoAlias(value);
LLVMSetDereferenceable(value, 0, PONY_ACTOR_PAD_SIZE);
// void ponyint_destroy(__object*)
params[0] = c->object_ptr;
type = LLVMFunctionType(c->void_type, params, 1, false);
value = LLVMAddFunction(c->module, "ponyint_destroy", type);
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
// void pony_sendv(i8*, __object*, $message*);
params[0] = c->void_ptr;
params[1] = c->object_ptr;
params[2] = c->msg_ptr;
type = LLVMFunctionType(c->void_type, params, 3, false);
value = LLVMAddFunction(c->module, "pony_sendv", type);
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
// i8* pony_alloc(i8*, intptr)
params[0] = c->void_ptr;
params[1] = c->intptr;
type = LLVMFunctionType(c->void_ptr, params, 2, false);
value = LLVMAddFunction(c->module, "pony_alloc", type);
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
LLVMSetReturnNoAlias(value);
#if PONY_LLVM >= 307
LLVMSetDereferenceableOrNull(value, 0, HEAP_MIN);
#endif
// i8* pony_alloc_small(i8*, i32)
params[0] = c->void_ptr;
params[1] = c->i32;
type = LLVMFunctionType(c->void_ptr, params, 2, false);
value = LLVMAddFunction(c->module, "pony_alloc_small", type);
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
LLVMSetReturnNoAlias(value);
LLVMSetDereferenceable(value, 0, HEAP_MIN);
// i8* pony_alloc_large(i8*, intptr)
params[0] = c->void_ptr;
params[1] = c->intptr;
type = LLVMFunctionType(c->void_ptr, params, 2, false);
value = LLVMAddFunction(c->module, "pony_alloc_large", type);
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
LLVMSetReturnNoAlias(value);
LLVMSetDereferenceable(value, 0, HEAP_MAX + 1);
// i8* pony_realloc(i8*, i8*, intptr)
params[0] = c->void_ptr;
params[1] = c->void_ptr;
params[2] = c->intptr;
type = LLVMFunctionType(c->void_ptr, params, 3, false);
value = LLVMAddFunction(c->module, "pony_realloc", type);
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
LLVMSetReturnNoAlias(value);
#if PONY_LLVM >= 307
LLVMSetDereferenceableOrNull(value, 0, HEAP_MIN);
#endif
// i8* pony_alloc_final(i8*, intptr, c->final_fn)
params[0] = c->void_ptr;
params[1] = c->intptr;
params[2] = c->final_fn;
type = LLVMFunctionType(c->void_ptr, params, 3, false);
value = LLVMAddFunction(c->module, "pony_alloc_final", type);
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
LLVMSetReturnNoAlias(value);
#if PONY_LLVM >= 307
LLVMSetDereferenceableOrNull(value, 0, HEAP_MIN);
#endif
// $message* pony_alloc_msg(i32, i32)
params[0] = c->i32;
params[1] = c->i32;
type = LLVMFunctionType(c->msg_ptr, params, 2, false);
value = LLVMAddFunction(c->module, "pony_alloc_msg", type);
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
LLVMSetReturnNoAlias(value);
// void pony_trace(i8*, i8*)
params[0] = c->void_ptr;
params[1] = c->void_ptr;
type = LLVMFunctionType(c->void_type, params, 2, false);
value = LLVMAddFunction(c->module, "pony_trace", type);
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
// i8* pony_traceobject(i8*, __object*, __Desc*, i32)
params[0] = c->void_ptr;
params[1] = c->object_ptr;
params[2] = c->descriptor_ptr;
params[3] = c->i32;
type = LLVMFunctionType(c->void_ptr, params, 4, false);
value = LLVMAddFunction(c->module, "pony_traceknown", type);
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
// i8* pony_traceunknown(i8*, __object*, i32)
params[0] = c->void_ptr;
params[1] = c->object_ptr;
params[2] = c->i32;
type = LLVMFunctionType(c->void_ptr, params, 3, false);
value = LLVMAddFunction(c->module, "pony_traceunknown", type);
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
// void pony_gc_send(i8*)
params[0] = c->void_ptr;
type = LLVMFunctionType(c->void_type, params, 1, false);
value = LLVMAddFunction(c->module, "pony_gc_send", type);
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
// void pony_gc_recv(i8*)
params[0] = c->void_ptr;
type = LLVMFunctionType(c->void_type, params, 1, false);
value = LLVMAddFunction(c->module, "pony_gc_recv", type);
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
// void pony_send_done(i8*)
params[0] = c->void_ptr;
type = LLVMFunctionType(c->void_type, params, 1, false);
value = LLVMAddFunction(c->module, "pony_send_done", type);
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
// void pony_recv_done(i8*)
params[0] = c->void_ptr;
type = LLVMFunctionType(c->void_type, params, 1, false);
value = LLVMAddFunction(c->module, "pony_recv_done", type);
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
// void pony_serialise_reserve(i8*, i8*, intptr)
params[0] = c->void_ptr;
params[1] = c->void_ptr;
params[2] = c->intptr;
type = LLVMFunctionType(c->void_type, params, 3, false);
value = LLVMAddFunction(c->module, "pony_serialise_reserve", type);
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
// intptr pony_serialise_offset(i8*, i8*)
params[0] = c->void_ptr;
params[1] = c->void_ptr;
type = LLVMFunctionType(c->intptr, params, 2, false);
value = LLVMAddFunction(c->module, "pony_serialise_offset", type);
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
// i8* pony_deserialise_offset(i8*, __desc*, intptr)
params[0] = c->void_ptr;
params[1] = c->descriptor_ptr;
params[2] = c->intptr;
type = LLVMFunctionType(c->void_ptr, params, 3, false);
value = LLVMAddFunction(c->module, "pony_deserialise_offset", type);
// i8* pony_deserialise_block(i8*, intptr, intptr)
params[0] = c->void_ptr;
params[1] = c->intptr;
params[2] = c->intptr;
type = LLVMFunctionType(c->void_ptr, params, 3, false);
value = LLVMAddFunction(c->module, "pony_deserialise_block", type);
// i32 pony_init(i32, i8**)
params[0] = c->i32;
params[1] = LLVMPointerType(c->void_ptr, 0);
type = LLVMFunctionType(c->i32, params, 2, false);
value = LLVMAddFunction(c->module, "pony_init", type);
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
// void pony_become(i8*, __object*)
params[0] = c->void_ptr;
params[1] = c->object_ptr;
type = LLVMFunctionType(c->void_type, params, 2, false);
value = LLVMAddFunction(c->module, "pony_become", type);
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
// i32 pony_start(i32)
params[0] = c->i32;
type = LLVMFunctionType(c->i32, params, 1, false);
value = LLVMAddFunction(c->module, "pony_start", type);
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
// void pony_throw()
type = LLVMFunctionType(c->void_type, NULL, 0, false);
LLVMAddFunction(c->module, "pony_throw", type);
// i32 pony_personality_v0(...)
type = LLVMFunctionType(c->i32, NULL, 0, true);
c->personality = LLVMAddFunction(c->module, "pony_personality_v0", type);
// void llvm.memcpy.*(i8*, i8*, i32/64, i32, i1)
params[0] = c->void_ptr;
params[1] = c->void_ptr;
params[3] = c->i32;
params[4] = c->i1;
if(target_is_ilp32(c->opt->triple))
{
params[2] = c->i32;
type = LLVMFunctionType(c->void_type, params, 5, false);
value = LLVMAddFunction(c->module, "llvm.memcpy.p0i8.p0i8.i32", type);
} else {
params[2] = c->i64;
type = LLVMFunctionType(c->void_type, params, 5, false);
value = LLVMAddFunction(c->module, "llvm.memcpy.p0i8.p0i8.i64", type);
}
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
// void llvm.memmove.*(i8*, i8*, i32/64, i32, i1)
params[0] = c->void_ptr;
params[1] = c->void_ptr;
params[3] = c->i32;
params[4] = c->i1;
if(target_is_ilp32(c->opt->triple))
{
params[2] = c->i32;
type = LLVMFunctionType(c->void_type, params, 5, false);
value = LLVMAddFunction(c->module, "llvm.memmove.p0i8.p0i8.i32", type);
} else {
params[2] = c->i64;
type = LLVMFunctionType(c->void_type, params, 5, false);
value = LLVMAddFunction(c->module, "llvm.memmove.p0i8.p0i8.i64", type);
}
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
}
LLVMTypeRef ett_llvm_type(EagleComplexType *type)
{
switch(type->type)
{
case ETVoid:
return LLVMVoidTypeInContext(utl_get_current_context());
case ETFloat:
return LLVMFloatTypeInContext(utl_get_current_context());
case ETDouble:
return LLVMDoubleTypeInContext(utl_get_current_context());
case ETInt1:
return LLVMInt1TypeInContext(utl_get_current_context());
case ETGeneric: // In practice this doesn't matter
case ETAny:
case ETInt8:
case ETUInt8:
return LLVMInt8TypeInContext(utl_get_current_context());
case ETInt16:
case ETUInt16:
return LLVMInt16TypeInContext(utl_get_current_context());
case ETInt32:
case ETUInt32:
return LLVMInt32TypeInContext(utl_get_current_context());
case ETInt64:
case ETUInt64:
return LLVMInt64TypeInContext(utl_get_current_context());
case ETCString:
return LLVMPointerType(LLVMInt8TypeInContext(utl_get_current_context()), 0);
case ETEnum:
return LLVMInt64TypeInContext(utl_get_current_context());
case ETGenerator:
{
if(generator_type)
return generator_type;
LLVMTypeRef ptmp[2];
ptmp[0] = LLVMPointerType(LLVMInt8TypeInContext(utl_get_current_context()), 0);
ptmp[1] = LLVMPointerType(LLVMInt8TypeInContext(utl_get_current_context()), 0);
generator_type = LLVMStructCreateNamed(utl_get_current_context(), "__egl_gen_strct");
LLVMStructSetBody(generator_type, ptmp, 2, 0);
return generator_type;
}
case ETClass:
case ETStruct:
{
EagleStructType *st = (EagleStructType *)type;
LLVMTypeRef loaded = LLVMGetTypeByName(the_module, st->name);
if(loaded)
return loaded;
return NULL;
// LLVMTypeRef ty = LLVMStructTypeInContext(utl_get_current_context(),
}
case ETInterface:
{
return LLVMInt8TypeInContext(utl_get_current_context());
}
case ETPointer:
{
EaglePointerType *pt = (EaglePointerType *)type;
if(pt->counted || pt->weak)
{
LLVMTypeRef ptmp[2];
ptmp[0] = LLVMPointerType(LLVMInt8TypeInContext(utl_get_current_context()), 0);
ptmp[1] = LLVMInt1TypeInContext(utl_get_current_context());
LLVMTypeRef tys[6];
tys[0] = LLVMInt64TypeInContext(utl_get_current_context());
tys[1] = LLVMInt16TypeInContext(utl_get_current_context());
tys[2] = LLVMInt16TypeInContext(utl_get_current_context());
tys[3] = LLVMPointerType(LLVMInt8TypeInContext(utl_get_current_context()), 0);
tys[4] = LLVMPointerType(LLVMFunctionType(LLVMVoidTypeInContext(utl_get_current_context()), ptmp, 2, 0), 0);
tys[5] = ett_llvm_type(pt->to);
return LLVMPointerType(ty_get_counted(LLVMStructTypeInContext(utl_get_current_context(), tys, 6, 0)), 0);
}
return LLVMPointerType(ett_llvm_type(((EaglePointerType *)type)->to), 0);
}
case ETArray:
{
EagleArrayType *at = (EagleArrayType *)type;
if(at->ct < 0)
return LLVMPointerType(ett_llvm_type(at->of), 0);
else
return LLVMArrayType(ett_llvm_type(at->of), at->ct);
}
case ETFunction:
{
EagleFunctionType *ft = (EagleFunctionType *)type;
if(ET_IS_CLOSURE(type))
{
LLVMTypeRef tys[2];
tys[0] = LLVMPointerType(LLVMInt8TypeInContext(utl_get_current_context()), 0);
tys[1] = LLVMPointerType(LLVMInt8TypeInContext(utl_get_current_context()), 0);
return LLVMStructTypeInContext(utl_get_current_context(), tys, 2, 0);
}
LLVMTypeRef *tys = malloc(sizeof(LLVMTypeRef) * ft->pct);
int i;
for(i = 0; i < ft->pct; i++)
tys[i] = ett_llvm_type(ft->params[i]);
LLVMTypeRef out = LLVMFunctionType(ett_llvm_type(ft->retType), tys, ft->pct, 0);
free(tys);
return out;
}
default:
return NULL;
}
}
static void
lp_jit_create_types(struct lp_fragment_shader_variant *lp)
{
struct gallivm_state *gallivm = lp->gallivm;
LLVMContextRef lc = gallivm->context;
LLVMTypeRef viewport_type, texture_type, sampler_type;
/* struct lp_jit_viewport */
{
LLVMTypeRef elem_types[LP_JIT_VIEWPORT_NUM_FIELDS];
elem_types[LP_JIT_VIEWPORT_MIN_DEPTH] =
elem_types[LP_JIT_VIEWPORT_MAX_DEPTH] = LLVMFloatTypeInContext(lc);
viewport_type = LLVMStructTypeInContext(lc, elem_types,
Elements(elem_types), 0);
LP_CHECK_MEMBER_OFFSET(struct lp_jit_viewport, min_depth,
gallivm->target, viewport_type,
LP_JIT_VIEWPORT_MIN_DEPTH);
LP_CHECK_MEMBER_OFFSET(struct lp_jit_viewport, max_depth,
gallivm->target, viewport_type,
LP_JIT_VIEWPORT_MAX_DEPTH);
LP_CHECK_STRUCT_SIZE(struct lp_jit_viewport,
gallivm->target, viewport_type);
}
/* struct lp_jit_texture */
{
LLVMTypeRef elem_types[LP_JIT_TEXTURE_NUM_FIELDS];
elem_types[LP_JIT_TEXTURE_WIDTH] =
elem_types[LP_JIT_TEXTURE_HEIGHT] =
elem_types[LP_JIT_TEXTURE_DEPTH] =
elem_types[LP_JIT_TEXTURE_FIRST_LEVEL] =
elem_types[LP_JIT_TEXTURE_LAST_LEVEL] = LLVMInt32TypeInContext(lc);
elem_types[LP_JIT_TEXTURE_BASE] = LLVMPointerType(LLVMInt8TypeInContext(lc), 0);
elem_types[LP_JIT_TEXTURE_ROW_STRIDE] =
elem_types[LP_JIT_TEXTURE_IMG_STRIDE] =
elem_types[LP_JIT_TEXTURE_MIP_OFFSETS] =
LLVMArrayType(LLVMInt32TypeInContext(lc), LP_MAX_TEXTURE_LEVELS);
texture_type = LLVMStructTypeInContext(lc, elem_types,
Elements(elem_types), 0);
LP_CHECK_MEMBER_OFFSET(struct lp_jit_texture, width,
gallivm->target, texture_type,
LP_JIT_TEXTURE_WIDTH);
LP_CHECK_MEMBER_OFFSET(struct lp_jit_texture, height,
gallivm->target, texture_type,
LP_JIT_TEXTURE_HEIGHT);
LP_CHECK_MEMBER_OFFSET(struct lp_jit_texture, depth,
gallivm->target, texture_type,
LP_JIT_TEXTURE_DEPTH);
LP_CHECK_MEMBER_OFFSET(struct lp_jit_texture, first_level,
gallivm->target, texture_type,
LP_JIT_TEXTURE_FIRST_LEVEL);
LP_CHECK_MEMBER_OFFSET(struct lp_jit_texture, last_level,
gallivm->target, texture_type,
LP_JIT_TEXTURE_LAST_LEVEL);
LP_CHECK_MEMBER_OFFSET(struct lp_jit_texture, base,
gallivm->target, texture_type,
LP_JIT_TEXTURE_BASE);
LP_CHECK_MEMBER_OFFSET(struct lp_jit_texture, row_stride,
gallivm->target, texture_type,
LP_JIT_TEXTURE_ROW_STRIDE);
LP_CHECK_MEMBER_OFFSET(struct lp_jit_texture, img_stride,
gallivm->target, texture_type,
LP_JIT_TEXTURE_IMG_STRIDE);
LP_CHECK_MEMBER_OFFSET(struct lp_jit_texture, mip_offsets,
gallivm->target, texture_type,
LP_JIT_TEXTURE_MIP_OFFSETS);
LP_CHECK_STRUCT_SIZE(struct lp_jit_texture,
gallivm->target, texture_type);
}
/* struct lp_jit_sampler */
{
LLVMTypeRef elem_types[LP_JIT_SAMPLER_NUM_FIELDS];
elem_types[LP_JIT_SAMPLER_MIN_LOD] =
elem_types[LP_JIT_SAMPLER_MAX_LOD] =
elem_types[LP_JIT_SAMPLER_LOD_BIAS] = LLVMFloatTypeInContext(lc);
elem_types[LP_JIT_SAMPLER_BORDER_COLOR] =
LLVMArrayType(LLVMFloatTypeInContext(lc), 4);
sampler_type = LLVMStructTypeInContext(lc, elem_types,
Elements(elem_types), 0);
LP_CHECK_MEMBER_OFFSET(struct lp_jit_sampler, min_lod,
gallivm->target, sampler_type,
LP_JIT_SAMPLER_MIN_LOD);
LP_CHECK_MEMBER_OFFSET(struct lp_jit_sampler, max_lod,
gallivm->target, sampler_type,
LP_JIT_SAMPLER_MAX_LOD);
LP_CHECK_MEMBER_OFFSET(struct lp_jit_sampler, lod_bias,
gallivm->target, sampler_type,
LP_JIT_SAMPLER_LOD_BIAS);
LP_CHECK_MEMBER_OFFSET(struct lp_jit_sampler, border_color,
gallivm->target, sampler_type,
LP_JIT_SAMPLER_BORDER_COLOR);
LP_CHECK_STRUCT_SIZE(struct lp_jit_sampler,
gallivm->target, sampler_type);
}
/* struct lp_jit_context */
{
LLVMTypeRef elem_types[LP_JIT_CTX_COUNT];
LLVMTypeRef context_type;
elem_types[LP_JIT_CTX_CONSTANTS] =
LLVMArrayType(LLVMPointerType(LLVMFloatTypeInContext(lc), 0), LP_MAX_TGSI_CONST_BUFFERS);
elem_types[LP_JIT_CTX_NUM_CONSTANTS] =
LLVMArrayType(LLVMInt32TypeInContext(lc), LP_MAX_TGSI_CONST_BUFFERS);
elem_types[LP_JIT_CTX_ALPHA_REF] = LLVMFloatTypeInContext(lc);
elem_types[LP_JIT_CTX_STENCIL_REF_FRONT] =
elem_types[LP_JIT_CTX_STENCIL_REF_BACK] = LLVMInt32TypeInContext(lc);
elem_types[LP_JIT_CTX_U8_BLEND_COLOR] = LLVMPointerType(LLVMInt8TypeInContext(lc), 0);
elem_types[LP_JIT_CTX_F_BLEND_COLOR] = LLVMPointerType(LLVMFloatTypeInContext(lc), 0);
elem_types[LP_JIT_CTX_VIEWPORTS] = LLVMPointerType(viewport_type, 0);
elem_types[LP_JIT_CTX_TEXTURES] = LLVMArrayType(texture_type,
PIPE_MAX_SHADER_SAMPLER_VIEWS);
elem_types[LP_JIT_CTX_SAMPLERS] = LLVMArrayType(sampler_type,
PIPE_MAX_SAMPLERS);
context_type = LLVMStructTypeInContext(lc, elem_types,
Elements(elem_types), 0);
LP_CHECK_MEMBER_OFFSET(struct lp_jit_context, constants,
gallivm->target, context_type,
LP_JIT_CTX_CONSTANTS);
LP_CHECK_MEMBER_OFFSET(struct lp_jit_context, num_constants,
gallivm->target, context_type,
LP_JIT_CTX_NUM_CONSTANTS);
LP_CHECK_MEMBER_OFFSET(struct lp_jit_context, alpha_ref_value,
gallivm->target, context_type,
LP_JIT_CTX_ALPHA_REF);
LP_CHECK_MEMBER_OFFSET(struct lp_jit_context, stencil_ref_front,
gallivm->target, context_type,
LP_JIT_CTX_STENCIL_REF_FRONT);
LP_CHECK_MEMBER_OFFSET(struct lp_jit_context, stencil_ref_back,
gallivm->target, context_type,
LP_JIT_CTX_STENCIL_REF_BACK);
LP_CHECK_MEMBER_OFFSET(struct lp_jit_context, u8_blend_color,
gallivm->target, context_type,
LP_JIT_CTX_U8_BLEND_COLOR);
LP_CHECK_MEMBER_OFFSET(struct lp_jit_context, f_blend_color,
gallivm->target, context_type,
LP_JIT_CTX_F_BLEND_COLOR);
LP_CHECK_MEMBER_OFFSET(struct lp_jit_context, viewports,
gallivm->target, context_type,
LP_JIT_CTX_VIEWPORTS);
LP_CHECK_MEMBER_OFFSET(struct lp_jit_context, textures,
gallivm->target, context_type,
LP_JIT_CTX_TEXTURES);
LP_CHECK_MEMBER_OFFSET(struct lp_jit_context, samplers,
gallivm->target, context_type,
LP_JIT_CTX_SAMPLERS);
LP_CHECK_STRUCT_SIZE(struct lp_jit_context,
gallivm->target, context_type);
lp->jit_context_ptr_type = LLVMPointerType(context_type, 0);
}
/* struct lp_jit_thread_data */
{
LLVMTypeRef elem_types[LP_JIT_THREAD_DATA_COUNT];
LLVMTypeRef thread_data_type;
elem_types[LP_JIT_THREAD_DATA_COUNTER] = LLVMInt64TypeInContext(lc);
elem_types[LP_JIT_THREAD_DATA_RASTER_STATE_VIEWPORT_INDEX] =
LLVMInt32TypeInContext(lc);
thread_data_type = LLVMStructTypeInContext(lc, elem_types,
Elements(elem_types), 0);
lp->jit_thread_data_ptr_type = LLVMPointerType(thread_data_type, 0);
}
if (gallivm_debug & GALLIVM_DEBUG_IR) {
LLVMDumpModule(gallivm->module);
}
}
static void init_runtime(compile_t* c)
{
c->str_builtin = stringtab("$0");
c->str_Bool = stringtab("Bool");
c->str_I8 = stringtab("I8");
c->str_I16 = stringtab("I16");
c->str_I32 = stringtab("I32");
c->str_I64 = stringtab("I64");
c->str_I128 = stringtab("I128");
c->str_ILong = stringtab("ILong");
c->str_ISize = stringtab("ISize");
c->str_U8 = stringtab("U8");
c->str_U16 = stringtab("U16");
c->str_U32 = stringtab("U32");
c->str_U64 = stringtab("U64");
c->str_U128 = stringtab("U128");
c->str_ULong = stringtab("ULong");
c->str_USize = stringtab("USize");
c->str_F32 = stringtab("F32");
c->str_F64 = stringtab("F64");
c->str_Pointer = stringtab("Pointer");
c->str_Maybe = stringtab("MaybePointer");
c->str_DoNotOptimise = stringtab("DoNotOptimise");
c->str_Array = stringtab("Array");
c->str_String = stringtab("String");
c->str_Platform = stringtab("Platform");
c->str_Main = stringtab("Main");
c->str_Env = stringtab("Env");
c->str_add = stringtab("add");
c->str_sub = stringtab("sub");
c->str_mul = stringtab("mul");
c->str_div = stringtab("div");
c->str_mod = stringtab("mod");
c->str_neg = stringtab("neg");
c->str_add_unsafe = stringtab("add_unsafe");
c->str_sub_unsafe = stringtab("sub_unsafe");
c->str_mul_unsafe = stringtab("mul_unsafe");
c->str_div_unsafe = stringtab("div_unsafe");
c->str_mod_unsafe = stringtab("mod_unsafe");
c->str_neg_unsafe = stringtab("neg_unsafe");
c->str_and = stringtab("op_and");
c->str_or = stringtab("op_or");
c->str_xor = stringtab("op_xor");
c->str_not = stringtab("op_not");
c->str_shl = stringtab("shl");
c->str_shr = stringtab("shr");
c->str_shl_unsafe = stringtab("shl_unsafe");
c->str_shr_unsafe = stringtab("shr_unsafe");
c->str_eq = stringtab("eq");
c->str_ne = stringtab("ne");
c->str_lt = stringtab("lt");
c->str_le = stringtab("le");
c->str_ge = stringtab("ge");
c->str_gt = stringtab("gt");
c->str_eq_unsafe = stringtab("eq_unsafe");
c->str_ne_unsafe = stringtab("ne_unsafe");
c->str_lt_unsafe = stringtab("lt_unsafe");
c->str_le_unsafe = stringtab("le_unsafe");
c->str_ge_unsafe = stringtab("ge_unsafe");
c->str_gt_unsafe = stringtab("gt_unsafe");
c->str_this = stringtab("this");
c->str_create = stringtab("create");
c->str__create = stringtab("_create");
c->str__init = stringtab("_init");
c->str__final = stringtab("_final");
c->str__event_notify = stringtab("_event_notify");
c->str__serialise_space = stringtab("_serialise_space");
c->str__serialise = stringtab("_serialise");
c->str__deserialise = stringtab("_deserialise");
LLVMTypeRef type;
LLVMTypeRef params[5];
LLVMValueRef value;
c->void_type = LLVMVoidTypeInContext(c->context);
c->i1 = LLVMInt1TypeInContext(c->context);
c->i8 = LLVMInt8TypeInContext(c->context);
c->i16 = LLVMInt16TypeInContext(c->context);
c->i32 = LLVMInt32TypeInContext(c->context);
c->i64 = LLVMInt64TypeInContext(c->context);
c->i128 = LLVMIntTypeInContext(c->context, 128);
c->f32 = LLVMFloatTypeInContext(c->context);
c->f64 = LLVMDoubleTypeInContext(c->context);
c->intptr = LLVMIntPtrTypeInContext(c->context, c->target_data);
// i8*
c->void_ptr = LLVMPointerType(c->i8, 0);
// forward declare object
c->object_type = LLVMStructCreateNamed(c->context, "__object");
c->object_ptr = LLVMPointerType(c->object_type, 0);
// padding required in an actor between the descriptor and fields
c->actor_pad = LLVMArrayType(c->i8, PONY_ACTOR_PAD_SIZE);
// message
params[0] = c->i32; // size
params[1] = c->i32; // id
c->msg_type = LLVMStructCreateNamed(c->context, "__message");
c->msg_ptr = LLVMPointerType(c->msg_type, 0);
LLVMStructSetBody(c->msg_type, params, 2, false);
// trace
// void (*)(i8*, __object*)
params[0] = c->void_ptr;
params[1] = c->object_ptr;
c->trace_type = LLVMFunctionType(c->void_type, params, 2, false);
c->trace_fn = LLVMPointerType(c->trace_type, 0);
// serialise
// void (*)(i8*, __object*, i8*, intptr, i32)
params[0] = c->void_ptr;
params[1] = c->object_ptr;
params[2] = c->void_ptr;
params[3] = c->intptr;
params[4] = c->i32;
c->serialise_type = LLVMFunctionType(c->void_type, params, 5, false);
c->serialise_fn = LLVMPointerType(c->serialise_type, 0);
// serialise_space
// i64 (__object*)
params[0] = c->object_ptr;
c->custom_serialise_space_fn = LLVMPointerType(
LLVMFunctionType(c->i64, params, 1, false), 0);
// custom_deserialise
// void (*)(__object*, void*)
params[0] = c->object_ptr;
params[1] = c->void_ptr;
c->custom_deserialise_fn = LLVMPointerType(
LLVMFunctionType(c->void_type, params, 2, false), 0);
// dispatch
// void (*)(i8*, __object*, $message*)
params[0] = c->void_ptr;
params[1] = c->object_ptr;
params[2] = c->msg_ptr;
c->dispatch_type = LLVMFunctionType(c->void_type, params, 3, false);
c->dispatch_fn = LLVMPointerType(c->dispatch_type, 0);
// void (*)(__object*)
params[0] = c->object_ptr;
c->final_fn = LLVMPointerType(
LLVMFunctionType(c->void_type, params, 1, false), 0);
// descriptor, opaque version
// We need this in order to build our own structure.
const char* desc_name = genname_descriptor(NULL);
c->descriptor_type = LLVMStructCreateNamed(c->context, desc_name);
c->descriptor_ptr = LLVMPointerType(c->descriptor_type, 0);
// field descriptor
// Also needed to build a descriptor structure.
params[0] = c->i32;
params[1] = c->descriptor_ptr;
c->field_descriptor = LLVMStructTypeInContext(c->context, params, 2, false);
// descriptor, filled in
gendesc_basetype(c, c->descriptor_type);
// define object
params[0] = c->descriptor_ptr;
LLVMStructSetBody(c->object_type, params, 1, false);
#if PONY_LLVM >= 309
LLVM_DECLARE_ATTRIBUTEREF(nounwind_attr, nounwind, 0);
LLVM_DECLARE_ATTRIBUTEREF(readnone_attr, readnone, 0);
LLVM_DECLARE_ATTRIBUTEREF(readonly_attr, readonly, 0);
LLVM_DECLARE_ATTRIBUTEREF(inacc_or_arg_mem_attr,
inaccessiblemem_or_argmemonly, 0);
LLVM_DECLARE_ATTRIBUTEREF(noalias_attr, noalias, 0);
LLVM_DECLARE_ATTRIBUTEREF(noreturn_attr, noreturn, 0);
LLVM_DECLARE_ATTRIBUTEREF(deref_actor_attr, dereferenceable,
PONY_ACTOR_PAD_SIZE + (target_is_ilp32(c->opt->triple) ? 4 : 8));
LLVM_DECLARE_ATTRIBUTEREF(align_pool_attr, align, ponyint_pool_size(0));
LLVM_DECLARE_ATTRIBUTEREF(align_heap_attr, align, HEAP_MIN);
LLVM_DECLARE_ATTRIBUTEREF(deref_or_null_alloc_attr, dereferenceable_or_null,
HEAP_MIN);
LLVM_DECLARE_ATTRIBUTEREF(deref_alloc_small_attr, dereferenceable, HEAP_MIN);
LLVM_DECLARE_ATTRIBUTEREF(deref_alloc_large_attr, dereferenceable,
HEAP_MAX << 1);
#endif
// i8* pony_ctx()
type = LLVMFunctionType(c->void_ptr, NULL, 0, false);
value = LLVMAddFunction(c->module, "pony_ctx", type);
#if PONY_LLVM >= 309
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex, nounwind_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex, readnone_attr);
#else
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
LLVMAddFunctionAttr(value, LLVMReadNoneAttribute);
#endif
// __object* pony_create(i8*, __Desc*)
params[0] = c->void_ptr;
params[1] = c->descriptor_ptr;
type = LLVMFunctionType(c->object_ptr, params, 2, false);
value = LLVMAddFunction(c->module, "pony_create", type);
#if PONY_LLVM >= 309
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex, nounwind_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex,
inacc_or_arg_mem_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeReturnIndex, noalias_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeReturnIndex, deref_actor_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeReturnIndex, align_pool_attr);
#else
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
# if PONY_LLVM >= 308
LLVMSetInaccessibleMemOrArgMemOnly(value);
# endif
LLVMSetReturnNoAlias(value);
LLVMSetDereferenceable(value, 0, PONY_ACTOR_PAD_SIZE +
(target_is_ilp32(c->opt->triple) ? 4 : 8));
#endif
// void ponyint_destroy(__object*)
params[0] = c->object_ptr;
type = LLVMFunctionType(c->void_type, params, 1, false);
value = LLVMAddFunction(c->module, "ponyint_destroy", type);
#if PONY_LLVM >= 309
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex, nounwind_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex,
inacc_or_arg_mem_attr);
#else
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
# if PONY_LLVM >= 308
LLVMSetInaccessibleMemOrArgMemOnly(value);
# endif
#endif
// void pony_sendv(i8*, __object*, $message*, $message*)
params[0] = c->void_ptr;
params[1] = c->object_ptr;
params[2] = c->msg_ptr;
params[3] = c->msg_ptr;
type = LLVMFunctionType(c->void_type, params, 4, false);
value = LLVMAddFunction(c->module, "pony_sendv", type);
#if PONY_LLVM >= 309
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex, nounwind_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex,
inacc_or_arg_mem_attr);
#else
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
# if PONY_LLVM >= 308
LLVMSetInaccessibleMemOrArgMemOnly(value);
# endif
#endif
// void pony_sendv_single(i8*, __object*, $message*, $message*)
params[0] = c->void_ptr;
params[1] = c->object_ptr;
params[2] = c->msg_ptr;
params[3] = c->msg_ptr;
type = LLVMFunctionType(c->void_type, params, 4, false);
value = LLVMAddFunction(c->module, "pony_sendv_single", type);
#if PONY_LLVM >= 309
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex, nounwind_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex,
inacc_or_arg_mem_attr);
#else
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
# if PONY_LLVM >= 308
LLVMSetInaccessibleMemOrArgMemOnly(value);
# endif
#endif
// i8* pony_alloc(i8*, intptr)
params[0] = c->void_ptr;
params[1] = c->intptr;
type = LLVMFunctionType(c->void_ptr, params, 2, false);
value = LLVMAddFunction(c->module, "pony_alloc", type);
#if PONY_LLVM >= 309
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex, nounwind_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex,
inacc_or_arg_mem_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeReturnIndex, noalias_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeReturnIndex,
deref_or_null_alloc_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeReturnIndex, align_heap_attr);
#else
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
# if PONY_LLVM >= 308
LLVMSetInaccessibleMemOrArgMemOnly(value);
# endif
LLVMSetReturnNoAlias(value);
LLVMSetDereferenceableOrNull(value, 0, HEAP_MIN);
#endif
// i8* pony_alloc_small(i8*, i32)
params[0] = c->void_ptr;
params[1] = c->i32;
type = LLVMFunctionType(c->void_ptr, params, 2, false);
value = LLVMAddFunction(c->module, "pony_alloc_small", type);
#if PONY_LLVM >= 309
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex, nounwind_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex,
inacc_or_arg_mem_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeReturnIndex, noalias_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeReturnIndex,
deref_alloc_small_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeReturnIndex, align_heap_attr);
#else
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
# if PONY_LLVM >= 308
LLVMSetInaccessibleMemOrArgMemOnly(value);
# endif
LLVMSetReturnNoAlias(value);
LLVMSetDereferenceable(value, 0, HEAP_MIN);
#endif
// i8* pony_alloc_large(i8*, intptr)
params[0] = c->void_ptr;
params[1] = c->intptr;
type = LLVMFunctionType(c->void_ptr, params, 2, false);
value = LLVMAddFunction(c->module, "pony_alloc_large", type);
#if PONY_LLVM >= 309
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex, nounwind_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex,
inacc_or_arg_mem_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeReturnIndex, noalias_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeReturnIndex,
deref_alloc_large_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeReturnIndex, align_heap_attr);
#else
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
# if PONY_LLVM >= 308
LLVMSetInaccessibleMemOrArgMemOnly(value);
# endif
LLVMSetReturnNoAlias(value);
LLVMSetDereferenceable(value, 0, HEAP_MAX << 1);
#endif
// i8* pony_realloc(i8*, i8*, intptr)
params[0] = c->void_ptr;
params[1] = c->void_ptr;
params[2] = c->intptr;
type = LLVMFunctionType(c->void_ptr, params, 3, false);
value = LLVMAddFunction(c->module, "pony_realloc", type);
#if PONY_LLVM >= 309
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex, nounwind_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex,
inacc_or_arg_mem_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeReturnIndex, noalias_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeReturnIndex,
deref_or_null_alloc_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeReturnIndex, align_heap_attr);
#else
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
# if PONY_LLVM >= 308
LLVMSetInaccessibleMemOrArgMemOnly(value);
# endif
LLVMSetReturnNoAlias(value);
LLVMSetDereferenceableOrNull(value, 0, HEAP_MIN);
#endif
// i8* pony_alloc_final(i8*, intptr)
params[0] = c->void_ptr;
params[1] = c->intptr;
type = LLVMFunctionType(c->void_ptr, params, 2, false);
value = LLVMAddFunction(c->module, "pony_alloc_final", type);
#if PONY_LLVM >= 309
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex, nounwind_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex,
inacc_or_arg_mem_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeReturnIndex, noalias_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeReturnIndex,
deref_or_null_alloc_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeReturnIndex, align_heap_attr);
#else
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
# if PONY_LLVM >= 308
LLVMSetInaccessibleMemOrArgMemOnly(value);
# endif
LLVMSetReturnNoAlias(value);
LLVMSetDereferenceableOrNull(value, 0, HEAP_MIN);
#endif
// i8* pony_alloc_small_final(i8*, i32)
params[0] = c->void_ptr;
params[1] = c->i32;
type = LLVMFunctionType(c->void_ptr, params, 2, false);
value = LLVMAddFunction(c->module, "pony_alloc_small_final", type);
#if PONY_LLVM >= 309
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex, nounwind_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex,
inacc_or_arg_mem_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeReturnIndex, noalias_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeReturnIndex,
deref_alloc_small_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeReturnIndex, align_heap_attr);
#else
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
# if PONY_LLVM >= 308
LLVMSetInaccessibleMemOrArgMemOnly(value);
# endif
LLVMSetReturnNoAlias(value);
LLVMSetDereferenceable(value, 0, HEAP_MIN);
#endif
// i8* pony_alloc_large_final(i8*, intptr)
params[0] = c->void_ptr;
params[1] = c->intptr;
type = LLVMFunctionType(c->void_ptr, params, 2, false);
value = LLVMAddFunction(c->module, "pony_alloc_large_final", type);
#if PONY_LLVM >= 309
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex, nounwind_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex,
inacc_or_arg_mem_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeReturnIndex, noalias_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeReturnIndex,
deref_alloc_large_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeReturnIndex, align_heap_attr);
#else
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
# if PONY_LLVM >= 308
LLVMSetInaccessibleMemOrArgMemOnly(value);
# endif
LLVMSetReturnNoAlias(value);
LLVMSetDereferenceable(value, 0, HEAP_MAX << 1);
#endif
// $message* pony_alloc_msg(i32, i32)
params[0] = c->i32;
params[1] = c->i32;
type = LLVMFunctionType(c->msg_ptr, params, 2, false);
value = LLVMAddFunction(c->module, "pony_alloc_msg", type);
#if PONY_LLVM >= 309
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex, nounwind_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex,
inacc_or_arg_mem_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeReturnIndex, noalias_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeReturnIndex, align_pool_attr);
#else
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
# if PONY_LLVM >= 308
LLVMSetInaccessibleMemOrArgMemOnly(value);
# endif
LLVMSetReturnNoAlias(value);
#endif
// void pony_trace(i8*, i8*)
params[0] = c->void_ptr;
params[1] = c->void_ptr;
type = LLVMFunctionType(c->void_type, params, 2, false);
value = LLVMAddFunction(c->module, "pony_trace", type);
#if PONY_LLVM >= 309
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex, nounwind_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex,
inacc_or_arg_mem_attr);
LLVMAddAttributeAtIndex(value, 2, readnone_attr);
#else
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
# if PONY_LLVM >= 308
LLVMSetInaccessibleMemOrArgMemOnly(value);
# endif
value = LLVMGetParam(value, 1);
LLVMAddAttribute(value, LLVMReadNoneAttribute);
#endif
// void pony_traceknown(i8*, __object*, __Desc*, i32)
params[0] = c->void_ptr;
params[1] = c->object_ptr;
params[2] = c->descriptor_ptr;
params[3] = c->i32;
type = LLVMFunctionType(c->void_type, params, 4, false);
value = LLVMAddFunction(c->module, "pony_traceknown", type);
#if PONY_LLVM >= 309
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex, nounwind_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex,
inacc_or_arg_mem_attr);
LLVMAddAttributeAtIndex(value, 2, readonly_attr);
#else
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
# if PONY_LLVM >= 308
LLVMSetInaccessibleMemOrArgMemOnly(value);
# endif
value = LLVMGetParam(value, 1);
LLVMAddAttribute(value, LLVMReadOnlyAttribute);
#endif
// void pony_traceunknown(i8*, __object*, i32)
params[0] = c->void_ptr;
params[1] = c->object_ptr;
params[2] = c->i32;
type = LLVMFunctionType(c->void_type, params, 3, false);
value = LLVMAddFunction(c->module, "pony_traceunknown", type);
#if PONY_LLVM >= 309
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex, nounwind_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex,
inacc_or_arg_mem_attr);
LLVMAddAttributeAtIndex(value, 2, readonly_attr);
#else
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
# if PONY_LLVM >= 308
LLVMSetInaccessibleMemOrArgMemOnly(value);
# endif
value = LLVMGetParam(value, 1);
LLVMAddAttribute(value, LLVMReadOnlyAttribute);
#endif
// void pony_gc_send(i8*)
params[0] = c->void_ptr;
type = LLVMFunctionType(c->void_type, params, 1, false);
value = LLVMAddFunction(c->module, "pony_gc_send", type);
#if PONY_LLVM >= 309
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex, nounwind_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex,
inacc_or_arg_mem_attr);
#else
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
# if PONY_LLVM >= 308
LLVMSetInaccessibleMemOrArgMemOnly(value);
# endif
#endif
// void pony_gc_recv(i8*)
params[0] = c->void_ptr;
type = LLVMFunctionType(c->void_type, params, 1, false);
value = LLVMAddFunction(c->module, "pony_gc_recv", type);
#if PONY_LLVM >= 309
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex, nounwind_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex,
inacc_or_arg_mem_attr);
#else
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
# if PONY_LLVM >= 308
LLVMSetInaccessibleMemOrArgMemOnly(value);
# endif
#endif
// void pony_send_done(i8*)
params[0] = c->void_ptr;
type = LLVMFunctionType(c->void_type, params, 1, false);
value = LLVMAddFunction(c->module, "pony_send_done", type);
#if PONY_LLVM >= 309
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex, nounwind_attr);
#else
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
#endif
// void pony_recv_done(i8*)
params[0] = c->void_ptr;
type = LLVMFunctionType(c->void_type, params, 1, false);
value = LLVMAddFunction(c->module, "pony_recv_done", type);
#if PONY_LLVM >= 309
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex, nounwind_attr);
#else
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
#endif
// void pony_serialise_reserve(i8*, i8*, intptr)
params[0] = c->void_ptr;
params[1] = c->void_ptr;
params[2] = c->intptr;
type = LLVMFunctionType(c->void_type, params, 3, false);
value = LLVMAddFunction(c->module, "pony_serialise_reserve", type);
#if PONY_LLVM >= 309
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex, nounwind_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex,
inacc_or_arg_mem_attr);
LLVMAddAttributeAtIndex(value, 2, readnone_attr);
#else
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
# if PONY_LLVM >= 308
LLVMSetInaccessibleMemOrArgMemOnly(value);
# endif
value = LLVMGetParam(value, 1);
LLVMAddAttribute(value, LLVMReadNoneAttribute);
#endif
// intptr pony_serialise_offset(i8*, i8*)
params[0] = c->void_ptr;
params[1] = c->void_ptr;
type = LLVMFunctionType(c->intptr, params, 2, false);
value = LLVMAddFunction(c->module, "pony_serialise_offset", type);
#if PONY_LLVM >= 309
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex, nounwind_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex,
inacc_or_arg_mem_attr);
LLVMAddAttributeAtIndex(value, 2, readonly_attr);
#else
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
# if PONY_LLVM >= 308
LLVMSetInaccessibleMemOrArgMemOnly(value);
# endif
value = LLVMGetParam(value, 1);
LLVMAddAttribute(value, LLVMReadOnlyAttribute);
#endif
// i8* pony_deserialise_offset(i8*, __desc*, intptr)
params[0] = c->void_ptr;
params[1] = c->descriptor_ptr;
params[2] = c->intptr;
type = LLVMFunctionType(c->void_ptr, params, 3, false);
value = LLVMAddFunction(c->module, "pony_deserialise_offset", type);
#if PONY_LLVM >= 309
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex,
inacc_or_arg_mem_attr);
#elif PONY_LLVM == 308
LLVMSetInaccessibleMemOrArgMemOnly(value);
#endif
// i8* pony_deserialise_block(i8*, intptr, intptr)
params[0] = c->void_ptr;
params[1] = c->intptr;
params[2] = c->intptr;
type = LLVMFunctionType(c->void_ptr, params, 3, false);
value = LLVMAddFunction(c->module, "pony_deserialise_block", type);
#if PONY_LLVM >= 309
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex,
inacc_or_arg_mem_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeReturnIndex, noalias_attr);
#else
# if PONY_LLVM >= 308
LLVMSetInaccessibleMemOrArgMemOnly(value);
# endif
LLVMSetReturnNoAlias(value);
#endif
// i32 pony_init(i32, i8**)
params[0] = c->i32;
params[1] = LLVMPointerType(c->void_ptr, 0);
type = LLVMFunctionType(c->i32, params, 2, false);
value = LLVMAddFunction(c->module, "pony_init", type);
#if PONY_LLVM >= 309
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex, nounwind_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex,
inacc_or_arg_mem_attr);
#else
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
# if PONY_LLVM >= 308
LLVMSetInaccessibleMemOrArgMemOnly(value);
# endif
#endif
// void pony_become(i8*, __object*)
params[0] = c->void_ptr;
params[1] = c->object_ptr;
type = LLVMFunctionType(c->void_type, params, 2, false);
value = LLVMAddFunction(c->module, "pony_become", type);
#if PONY_LLVM >= 309
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex, nounwind_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex,
inacc_or_arg_mem_attr);
#else
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
# if PONY_LLVM >= 308
LLVMSetInaccessibleMemOrArgMemOnly(value);
# endif
#endif
// i32 pony_start(i32, i32)
params[0] = c->i32;
params[1] = c->i32;
type = LLVMFunctionType(c->i32, params, 2, false);
value = LLVMAddFunction(c->module, "pony_start", type);
#if PONY_LLVM >= 309
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex, nounwind_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex,
inacc_or_arg_mem_attr);
#else
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
# if PONY_LLVM >= 308
LLVMSetInaccessibleMemOrArgMemOnly(value);
# endif
#endif
// i32 pony_get_exitcode()
type = LLVMFunctionType(c->i32, NULL, 0, false);
value = LLVMAddFunction(c->module, "pony_get_exitcode", type);
#if PONY_LLVM >= 309
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex, nounwind_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex, readonly_attr);
#else
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
LLVMAddFunctionAttr(value, LLVMReadOnlyAttribute);
#endif
// void pony_throw()
type = LLVMFunctionType(c->void_type, NULL, 0, false);
value = LLVMAddFunction(c->module, "pony_throw", type);
#if PONY_LLVM >= 309
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex, noreturn_attr);
#else
LLVMAddFunctionAttr(value, LLVMNoReturnAttribute);
#endif
// i32 pony_personality_v0(...)
type = LLVMFunctionType(c->i32, NULL, 0, true);
c->personality = LLVMAddFunction(c->module, "pony_personality_v0", type);
// i32 memcmp(i8*, i8*, intptr)
params[0] = c->void_ptr;
params[1] = c->void_ptr;
params[2] = c->intptr;
type = LLVMFunctionType(c->i32, params, 3, false);
value = LLVMAddFunction(c->module, "memcmp", type);
#if PONY_LLVM >= 309
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex, nounwind_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex, readonly_attr);
LLVMAddAttributeAtIndex(value, 1, readonly_attr);
LLVMAddAttributeAtIndex(value, 2, readonly_attr);
#else
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
LLVMAddFunctionAttr(value, LLVMReadOnlyAttribute);
LLVMValueRef param = LLVMGetParam(value, 0);
LLVMAddAttribute(param, LLVMReadOnlyAttribute);
param = LLVMGetParam(value, 1);
LLVMAddAttribute(param, LLVMReadOnlyAttribute);
#endif
}
static void init_runtime(compile_t* c)
{
c->str_1 = stringtab("$1");
c->str_Bool = stringtab("Bool");
c->str_I8 = stringtab("I8");
c->str_I16 = stringtab("I16");
c->str_I32 = stringtab("I32");
c->str_I64 = stringtab("I64");
c->str_I128 = stringtab("I128");
c->str_U8 = stringtab("U8");
c->str_U16 = stringtab("U16");
c->str_U32 = stringtab("U32");
c->str_U64 = stringtab("U64");
c->str_U128 = stringtab("U128");
c->str_F32 = stringtab("F32");
c->str_F64 = stringtab("F64");
c->str_Pointer = stringtab("Pointer");
c->str_Array = stringtab("Array");
c->str_Platform = stringtab("Platform");
c->str_add = stringtab("add");
c->str_sub = stringtab("sub");
c->str_mul = stringtab("mul");
c->str_div = stringtab("div");
c->str_mod = stringtab("mod");
c->str_neg = stringtab("neg");
c->str_and = stringtab("op_and");
c->str_or = stringtab("op_or");
c->str_xor = stringtab("op_xor");
c->str_not = stringtab("op_not");
c->str_shl = stringtab("shl");
c->str_shr = stringtab("shr");
c->str_eq = stringtab("eq");
c->str_ne = stringtab("ne");
c->str_lt = stringtab("lt");
c->str_le = stringtab("le");
c->str_ge = stringtab("ge");
c->str_gt = stringtab("gt");
LLVMTypeRef type;
LLVMTypeRef params[4];
LLVMValueRef value;
c->void_type = LLVMVoidTypeInContext(c->context);
c->i1 = LLVMInt1TypeInContext(c->context);
c->i8 = LLVMInt8TypeInContext(c->context);
c->i16 = LLVMInt16TypeInContext(c->context);
c->i32 = LLVMInt32TypeInContext(c->context);
c->i64 = LLVMInt64TypeInContext(c->context);
c->i128 = LLVMIntTypeInContext(c->context, 128);
c->f32 = LLVMFloatTypeInContext(c->context);
c->f64 = LLVMDoubleTypeInContext(c->context);
c->intptr = LLVMIntPtrTypeInContext(c->context, c->target_data);
// i8*
c->void_ptr = LLVMPointerType(c->i8, 0);
// forward declare object
c->object_type = LLVMStructCreateNamed(c->context, "$object");
c->object_ptr = LLVMPointerType(c->object_type, 0);
// padding required in an actor between the descriptor and fields
c->actor_pad = LLVMArrayType(c->i8, PONY_ACTOR_PAD_SIZE);
// message
params[0] = c->i32; // size
params[1] = c->i32; // id
c->msg_type = LLVMStructCreateNamed(c->context, "$message");
c->msg_ptr = LLVMPointerType(c->msg_type, 0);
LLVMStructSetBody(c->msg_type, params, 2, false);
// trace
// void (*)($object*)
params[0] = c->object_ptr;
c->trace_type = LLVMFunctionType(c->void_type, params, 1, false);
c->trace_fn = LLVMPointerType(c->trace_type, 0);
// dispatch
// void (*)($object*, $message*)
params[0] = c->object_ptr;
params[1] = c->msg_ptr;
c->dispatch_type = LLVMFunctionType(c->void_type, params, 2, false);
c->dispatch_fn = LLVMPointerType(c->dispatch_type, 0);
// void (*)($object*)
params[0] = c->object_ptr;
c->final_fn = LLVMPointerType(
LLVMFunctionType(c->void_type, params, 1, false), 0);
// descriptor, opaque version
// We need this in order to build our own structure.
const char* desc_name = genname_descriptor(NULL);
c->descriptor_type = LLVMStructCreateNamed(c->context, desc_name);
c->descriptor_ptr = LLVMPointerType(c->descriptor_type, 0);
// field descriptor
// Also needed to build a descriptor structure.
params[0] = c->i32;
params[1] = c->descriptor_ptr;
c->field_descriptor = LLVMStructTypeInContext(c->context, params, 2, false);
// descriptor, filled in
c->descriptor_type = gendesc_type(c, NULL);
// define object
params[0] = c->descriptor_ptr;
LLVMStructSetBody(c->object_type, params, 1, false);
// $object* pony_create($desc*)
params[0] = c->descriptor_ptr;
type = LLVMFunctionType(c->object_ptr, params, 1, false);
value = LLVMAddFunction(c->module, "pony_create", type);
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
LLVMSetReturnNoAlias(value);
// void pony_destroy($object*)
params[0] = c->object_ptr;
type = LLVMFunctionType(c->void_type, params, 1, false);
value = LLVMAddFunction(c->module, "pony_destroy", type);
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
//LLVMSetReturnNoAlias(value);
// void pony_sendv($object*, $message*);
params[0] = c->object_ptr;
params[1] = c->msg_ptr;
type = LLVMFunctionType(c->void_type, params, 2, false);
value = LLVMAddFunction(c->module, "pony_sendv", type);
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
// i8* pony_alloc(i64)
params[0] = c->i64;
type = LLVMFunctionType(c->void_ptr, params, 1, false);
value = LLVMAddFunction(c->module, "pony_alloc", type);
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
LLVMSetReturnNoAlias(value);
// i8* pony_realloc(i8*, i64)
params[0] = c->void_ptr;
params[1] = c->i64;
type = LLVMFunctionType(c->void_ptr, params, 2, false);
value = LLVMAddFunction(c->module, "pony_realloc", type);
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
LLVMSetReturnNoAlias(value);
// i8* pony_alloc_final(i64, c->final_fn)
params[0] = c->i64;
params[1] = c->final_fn;
type = LLVMFunctionType(c->void_ptr, params, 2, false);
value = LLVMAddFunction(c->module, "pony_alloc_final", type);
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
LLVMSetReturnNoAlias(value);
// $message* pony_alloc_msg(i32, i32)
params[0] = c->i32;
params[1] = c->i32;
type = LLVMFunctionType(c->msg_ptr, params, 2, false);
value = LLVMAddFunction(c->module, "pony_alloc_msg", type);
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
LLVMSetReturnNoAlias(value);
// void pony_trace(i8*)
params[0] = c->void_ptr;
type = LLVMFunctionType(c->void_type, params, 1, false);
value = LLVMAddFunction(c->module, "pony_trace", type);
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
// void pony_traceactor($object*)
params[0] = c->object_ptr;
type = LLVMFunctionType(c->void_type, params, 1, false);
value = LLVMAddFunction(c->module, "pony_traceactor", type);
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
// void pony_traceobject($object*, trace_fn)
params[0] = c->object_ptr;
params[1] = c->trace_fn;
type = LLVMFunctionType(c->void_type, params, 2, false);
value = LLVMAddFunction(c->module, "pony_traceobject", type);
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
// void pony_traceunknown($object*)
params[0] = c->object_ptr;
type = LLVMFunctionType(c->void_type, params, 1, false);
value = LLVMAddFunction(c->module, "pony_traceunknown", type);
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
// void pony_trace_tag_or_actor($object*)
params[0] = c->object_ptr;
type = LLVMFunctionType(c->void_type, params, 1, false);
value = LLVMAddFunction(c->module, "pony_trace_tag_or_actor", type);
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
// void pony_gc_send()
type = LLVMFunctionType(c->void_type, NULL, 0, false);
value = LLVMAddFunction(c->module, "pony_gc_send", type);
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
// void pony_gc_recv()
type = LLVMFunctionType(c->void_type, NULL, 0, false);
value = LLVMAddFunction(c->module, "pony_gc_recv", type);
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
// void pony_send_done()
type = LLVMFunctionType(c->void_type, NULL, 0, false);
value = LLVMAddFunction(c->module, "pony_send_done", type);
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
// void pony_recv_done()
type = LLVMFunctionType(c->void_type, NULL, 0, false);
value = LLVMAddFunction(c->module, "pony_recv_done", type);
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
// i32 pony_init(i32, i8**)
params[0] = c->i32;
params[1] = LLVMPointerType(c->void_ptr, 0);
type = LLVMFunctionType(c->i32, params, 2, false);
value = LLVMAddFunction(c->module, "pony_init", type);
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
// void pony_become($object*)
params[0] = c->object_ptr;
type = LLVMFunctionType(c->void_type, params, 1, false);
value = LLVMAddFunction(c->module, "pony_become", type);
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
// i32 pony_start(i32)
params[0] = c->i32;
type = LLVMFunctionType(c->i32, params, 1, false);
value = LLVMAddFunction(c->module, "pony_start", type);
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
// void pony_throw()
type = LLVMFunctionType(c->void_type, NULL, 0, false);
LLVMAddFunction(c->module, "pony_throw", type);
// i32 pony_personality_v0(...)
type = LLVMFunctionType(c->i32, NULL, 0, true);
c->personality = LLVMAddFunction(c->module, "pony_personality_v0", type);
// i8* memcpy(...)
type = LLVMFunctionType(c->void_ptr, NULL, 0, true);
value = LLVMAddFunction(c->module, "memcpy", type);
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
// i8* memmove(...)
type = LLVMFunctionType(c->void_ptr, NULL, 0, true);
value = LLVMAddFunction(c->module, "memmove", type);
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
}
static LLVMValueRef
lp_build_gather_avx2(struct gallivm_state *gallivm,
unsigned length,
unsigned src_width,
struct lp_type dst_type,
LLVMValueRef base_ptr,
LLVMValueRef offsets)
{
LLVMBuilderRef builder = gallivm->builder;
LLVMTypeRef src_type, src_vec_type;
LLVMValueRef res;
struct lp_type res_type = dst_type;
res_type.length *= length;
if (dst_type.floating) {
src_type = src_width == 64 ? LLVMDoubleTypeInContext(gallivm->context) :
LLVMFloatTypeInContext(gallivm->context);
} else {
src_type = LLVMIntTypeInContext(gallivm->context, src_width);
}
src_vec_type = LLVMVectorType(src_type, length);
/* XXX should allow hw scaling (can handle i8, i16, i32, i64 for x86) */
assert(LLVMTypeOf(base_ptr) == LLVMPointerType(LLVMInt8TypeInContext(gallivm->context), 0));
if (0) {
/*
* XXX: This will cause LLVM pre 3.7 to hang; it works on LLVM 3.8 but
* will not use the AVX2 gather instrinsics (even with llvm 4.0), at
* least with Haswell. See
* http://lists.llvm.org/pipermail/llvm-dev/2016-January/094448.html
* And the generated code doing the emulation is quite a bit worse
* than what we get by doing it ourselves too.
*/
LLVMTypeRef i32_type = LLVMIntTypeInContext(gallivm->context, 32);
LLVMTypeRef i32_vec_type = LLVMVectorType(i32_type, length);
LLVMTypeRef i1_type = LLVMIntTypeInContext(gallivm->context, 1);
LLVMTypeRef i1_vec_type = LLVMVectorType(i1_type, length);
LLVMTypeRef src_ptr_type = LLVMPointerType(src_type, 0);
LLVMValueRef src_ptr;
base_ptr = LLVMBuildBitCast(builder, base_ptr, src_ptr_type, "");
/* Rescale offsets from bytes to elements */
LLVMValueRef scale = LLVMConstInt(i32_type, src_width/8, 0);
scale = lp_build_broadcast(gallivm, i32_vec_type, scale);
assert(LLVMTypeOf(offsets) == i32_vec_type);
offsets = LLVMBuildSDiv(builder, offsets, scale, "");
src_ptr = LLVMBuildGEP(builder, base_ptr, &offsets, 1, "vector-gep");
char intrinsic[64];
util_snprintf(intrinsic, sizeof intrinsic, "llvm.masked.gather.v%u%s%u",
length, dst_type.floating ? "f" : "i", src_width);
LLVMValueRef alignment = LLVMConstInt(i32_type, src_width/8, 0);
LLVMValueRef mask = LLVMConstAllOnes(i1_vec_type);
LLVMValueRef passthru = LLVMGetUndef(src_vec_type);
LLVMValueRef args[] = { src_ptr, alignment, mask, passthru };
res = lp_build_intrinsic(builder, intrinsic, src_vec_type, args, 4, 0);
} else {
LLVMTypeRef i8_type = LLVMIntTypeInContext(gallivm->context, 8);
const char *intrinsic = NULL;
unsigned l_idx = 0;
assert(src_width == 32 || src_width == 64);
if (src_width == 32) {
assert(length == 4 || length == 8);
} else {
assert(length == 2 || length == 4);
}
static const char *intrinsics[2][2][2] = {
{{"llvm.x86.avx2.gather.d.d",
"llvm.x86.avx2.gather.d.d.256"},
{"llvm.x86.avx2.gather.d.q",
"llvm.x86.avx2.gather.d.q.256"}},
{{"llvm.x86.avx2.gather.d.ps",
"llvm.x86.avx2.gather.d.ps.256"},
{"llvm.x86.avx2.gather.d.pd",
"llvm.x86.avx2.gather.d.pd.256"}},
};
if ((src_width == 32 && length == 8) ||
(src_width == 64 && length == 4)) {
l_idx = 1;
}
intrinsic = intrinsics[dst_type.floating][src_width == 64][l_idx];
LLVMValueRef passthru = LLVMGetUndef(src_vec_type);
LLVMValueRef mask = LLVMConstAllOnes(src_vec_type);
mask = LLVMConstBitCast(mask, src_vec_type);
LLVMValueRef scale = LLVMConstInt(i8_type, 1, 0);
LLVMValueRef args[] = { passthru, base_ptr, offsets, mask, scale };
res = lp_build_intrinsic(builder, intrinsic, src_vec_type, args, 5, 0);
}
res = LLVMBuildBitCast(builder, res, lp_build_vec_type(gallivm, res_type), "");
return res;
}
/**
* Return mask ? a : b;
*
* mask is a bitwise mask, composed of 0 or ~0 for each element. Any other value
* will yield unpredictable results.
*/
LLVMValueRef
lp_build_select(struct lp_build_context *bld,
LLVMValueRef mask,
LLVMValueRef a,
LLVMValueRef b)
{
LLVMBuilderRef builder = bld->gallivm->builder;
LLVMContextRef lc = bld->gallivm->context;
struct lp_type type = bld->type;
LLVMValueRef res;
assert(lp_check_value(type, a));
assert(lp_check_value(type, b));
if(a == b)
return a;
if (type.length == 1) {
mask = LLVMBuildTrunc(builder, mask, LLVMInt1TypeInContext(lc), "");
res = LLVMBuildSelect(builder, mask, a, b, "");
}
else if (0) {
/* Generate a vector select.
*
* XXX: Using vector selects would avoid emitting intrinsics, but they aren't
* properly supported yet.
*
* LLVM 3.0 includes experimental support provided the -promote-elements
* options is passed to LLVM's command line (e.g., via
* llvm::cl::ParseCommandLineOptions), but resulting code quality is much
* worse, probably because some optimization passes don't know how to
* handle vector selects.
*
* See also:
* - http://lists.cs.uiuc.edu/pipermail/llvmdev/2011-October/043659.html
*/
/* Convert the mask to a vector of booleans.
* XXX: There are two ways to do this. Decide what's best.
*/
if (1) {
LLVMTypeRef bool_vec_type = LLVMVectorType(LLVMInt1TypeInContext(lc), type.length);
mask = LLVMBuildTrunc(builder, mask, bool_vec_type, "");
} else {
mask = LLVMBuildICmp(builder, LLVMIntNE, mask, LLVMConstNull(bld->int_vec_type), "");
}
res = LLVMBuildSelect(builder, mask, a, b, "");
}
else if (((util_cpu_caps.has_sse4_1 &&
type.width * type.length == 128) ||
(util_cpu_caps.has_avx &&
type.width * type.length == 256 && type.width >= 32)) &&
!LLVMIsConstant(a) &&
!LLVMIsConstant(b) &&
!LLVMIsConstant(mask)) {
const char *intrinsic;
LLVMTypeRef arg_type;
LLVMValueRef args[3];
/*
* There's only float blend in AVX but can just cast i32/i64
* to float.
*/
if (type.width * type.length == 256) {
if (type.width == 64) {
intrinsic = "llvm.x86.avx.blendv.pd.256";
arg_type = LLVMVectorType(LLVMDoubleTypeInContext(lc), 4);
}
else {
intrinsic = "llvm.x86.avx.blendv.ps.256";
arg_type = LLVMVectorType(LLVMFloatTypeInContext(lc), 8);
}
}
else if (type.floating &&
type.width == 64) {
intrinsic = "llvm.x86.sse41.blendvpd";
arg_type = LLVMVectorType(LLVMDoubleTypeInContext(lc), 2);
} else if (type.floating &&
type.width == 32) {
intrinsic = "llvm.x86.sse41.blendvps";
arg_type = LLVMVectorType(LLVMFloatTypeInContext(lc), 4);
} else {
intrinsic = "llvm.x86.sse41.pblendvb";
arg_type = LLVMVectorType(LLVMInt8TypeInContext(lc), 16);
}
if (arg_type != bld->int_vec_type) {
mask = LLVMBuildBitCast(builder, mask, arg_type, "");
}
if (arg_type != bld->vec_type) {
a = LLVMBuildBitCast(builder, a, arg_type, "");
b = LLVMBuildBitCast(builder, b, arg_type, "");
}
args[0] = b;
args[1] = a;
args[2] = mask;
res = lp_build_intrinsic(builder, intrinsic,
arg_type, args, Elements(args));
if (arg_type != bld->vec_type) {
res = LLVMBuildBitCast(builder, res, bld->vec_type, "");
}
}
else {
res = lp_build_select_bitwise(bld, mask, a, b);
}
return res;
}
/*
* Do a cached lookup.
*
* Returns (vectors of) 4x8 rgba aos value
*/
LLVMValueRef
lp_build_fetch_cached_texels(struct gallivm_state *gallivm,
const struct util_format_description *format_desc,
unsigned n,
LLVMValueRef base_ptr,
LLVMValueRef offset,
LLVMValueRef i,
LLVMValueRef j,
LLVMValueRef cache)
{
LLVMBuilderRef builder = gallivm->builder;
unsigned count, low_bit, log2size;
LLVMValueRef color, offset_stored, addr, ptr_addrtrunc, tmp;
LLVMValueRef ij_index, hash_index, hash_mask, block_index;
LLVMTypeRef i8t = LLVMInt8TypeInContext(gallivm->context);
LLVMTypeRef i32t = LLVMInt32TypeInContext(gallivm->context);
LLVMTypeRef i64t = LLVMInt64TypeInContext(gallivm->context);
struct lp_type type;
struct lp_build_context bld32;
memset(&type, 0, sizeof type);
type.width = 32;
type.length = n;
assert(format_desc->block.width == 4);
assert(format_desc->block.height == 4);
lp_build_context_init(&bld32, gallivm, type);
/*
* compute hash - we use direct mapped cache, the hash function could
* be better but it needs to be simple
* per-element:
* compare offset with offset stored at tag (hash)
* if not equal decode/store block, update tag
* extract color from cache
* assemble result vector
*/
/* TODO: not ideal with 32bit pointers... */
low_bit = util_logbase2(format_desc->block.bits / 8);
log2size = util_logbase2(LP_BUILD_FORMAT_CACHE_SIZE);
addr = LLVMBuildPtrToInt(builder, base_ptr, i64t, "");
ptr_addrtrunc = LLVMBuildPtrToInt(builder, base_ptr, i32t, "");
ptr_addrtrunc = lp_build_broadcast_scalar(&bld32, ptr_addrtrunc);
/* For the hash function, first mask off the unused lowest bits. Then just
do some xor with address bits - only use lower 32bits */
ptr_addrtrunc = LLVMBuildAdd(builder, offset, ptr_addrtrunc, "");
ptr_addrtrunc = LLVMBuildLShr(builder, ptr_addrtrunc,
lp_build_const_int_vec(gallivm, type, low_bit), "");
/* This only really makes sense for size 64,128,256 */
hash_index = ptr_addrtrunc;
ptr_addrtrunc = LLVMBuildLShr(builder, ptr_addrtrunc,
lp_build_const_int_vec(gallivm, type, 2*log2size), "");
hash_index = LLVMBuildXor(builder, ptr_addrtrunc, hash_index, "");
tmp = LLVMBuildLShr(builder, hash_index,
lp_build_const_int_vec(gallivm, type, log2size), "");
hash_index = LLVMBuildXor(builder, hash_index, tmp, "");
hash_mask = lp_build_const_int_vec(gallivm, type, LP_BUILD_FORMAT_CACHE_SIZE - 1);
hash_index = LLVMBuildAnd(builder, hash_index, hash_mask, "");
ij_index = LLVMBuildShl(builder, i, lp_build_const_int_vec(gallivm, type, 2), "");
ij_index = LLVMBuildAdd(builder, ij_index, j, "");
block_index = LLVMBuildShl(builder, hash_index,
lp_build_const_int_vec(gallivm, type, 4), "");
block_index = LLVMBuildAdd(builder, ij_index, block_index, "");
if (n > 1) {
color = LLVMGetUndef(LLVMVectorType(i32t, n));
for (count = 0; count < n; count++) {
LLVMValueRef index, cond, colorx;
LLVMValueRef block_indexx, hash_indexx, addrx, offsetx, ptr_addrx;
struct lp_build_if_state if_ctx;
index = lp_build_const_int32(gallivm, count);
offsetx = LLVMBuildExtractElement(builder, offset, index, "");
addrx = LLVMBuildZExt(builder, offsetx, i64t, "");
addrx = LLVMBuildAdd(builder, addrx, addr, "");
block_indexx = LLVMBuildExtractElement(builder, block_index, index, "");
hash_indexx = LLVMBuildLShr(builder, block_indexx,
lp_build_const_int32(gallivm, 4), "");
offset_stored = lookup_tag_data(gallivm, cache, hash_indexx);
cond = LLVMBuildICmp(builder, LLVMIntNE, offset_stored, addrx, "");
lp_build_if(&if_ctx, gallivm, cond);
{
ptr_addrx = LLVMBuildIntToPtr(builder, addrx,
LLVMPointerType(i8t, 0), "");
update_cached_block(gallivm, format_desc, ptr_addrx, hash_indexx, cache);
#if LP_BUILD_FORMAT_CACHE_DEBUG
update_cache_access(gallivm, cache, 1,
LP_BUILD_FORMAT_CACHE_MEMBER_ACCESS_MISS);
#endif
}
lp_build_endif(&if_ctx);
colorx = lookup_cached_pixel(gallivm, cache, block_indexx);
color = LLVMBuildInsertElement(builder, color, colorx,
lp_build_const_int32(gallivm, count), "");
}
}
else {
LLVMValueRef cond;
struct lp_build_if_state if_ctx;
tmp = LLVMBuildZExt(builder, offset, i64t, "");
addr = LLVMBuildAdd(builder, tmp, addr, "");
offset_stored = lookup_tag_data(gallivm, cache, hash_index);
cond = LLVMBuildICmp(builder, LLVMIntNE, offset_stored, addr, "");
lp_build_if(&if_ctx, gallivm, cond);
{
tmp = LLVMBuildIntToPtr(builder, addr, LLVMPointerType(i8t, 0), "");
update_cached_block(gallivm, format_desc, tmp, hash_index, cache);
#if LP_BUILD_FORMAT_CACHE_DEBUG
update_cache_access(gallivm, cache, 1,
LP_BUILD_FORMAT_CACHE_MEMBER_ACCESS_MISS);
#endif
}
lp_build_endif(&if_ctx);
color = lookup_cached_pixel(gallivm, cache, block_index);
}
#if LP_BUILD_FORMAT_CACHE_DEBUG
update_cache_access(gallivm, cache, n,
LP_BUILD_FORMAT_CACHE_MEMBER_ACCESS_TOTAL);
#endif
return LLVMBuildBitCast(builder, color, LLVMVectorType(i8t, n * 4), "");
}
static void
update_cached_block(struct gallivm_state *gallivm,
const struct util_format_description *format_desc,
LLVMValueRef ptr_addr,
LLVMValueRef hash_index,
LLVMValueRef cache)
{
LLVMBuilderRef builder = gallivm->builder;
LLVMTypeRef i8t = LLVMInt8TypeInContext(gallivm->context);
LLVMTypeRef pi8t = LLVMPointerType(i8t, 0);
LLVMTypeRef i32t = LLVMInt32TypeInContext(gallivm->context);
LLVMTypeRef i32x4 = LLVMVectorType(LLVMInt32TypeInContext(gallivm->context), 4);
LLVMValueRef function;
LLVMValueRef tag_value, tmp_ptr;
LLVMValueRef col[4];
unsigned i, j;
/*
* Use format_desc->fetch_rgba_8unorm() for each pixel in the block.
* This doesn't actually make any sense whatsoever, someone would need
* to write a function doing this for all pixels in a block (either as
* an external c function or with generated code). Don't ask.
*/
{
/*
* Function to call looks like:
* fetch(uint8_t *dst, const uint8_t *src, unsigned i, unsigned j)
*/
LLVMTypeRef ret_type;
LLVMTypeRef arg_types[4];
LLVMTypeRef function_type;
assert(format_desc->fetch_rgba_8unorm);
ret_type = LLVMVoidTypeInContext(gallivm->context);
arg_types[0] = pi8t;
arg_types[1] = pi8t;
arg_types[2] = i32t;
arg_types[3] = i32t;
function_type = LLVMFunctionType(ret_type, arg_types,
ARRAY_SIZE(arg_types), 0);
/* make const pointer for the C fetch_rgba_8unorm function */
function = lp_build_const_int_pointer(gallivm,
func_to_pointer((func_pointer) format_desc->fetch_rgba_8unorm));
/* cast the callee pointer to the function's type */
function = LLVMBuildBitCast(builder, function,
LLVMPointerType(function_type, 0),
"cast callee");
}
tmp_ptr = lp_build_array_alloca(gallivm, i32x4,
lp_build_const_int32(gallivm, 16),
"tmp_decode_store");
tmp_ptr = LLVMBuildBitCast(builder, tmp_ptr, pi8t, "");
/*
* Invoke format_desc->fetch_rgba_8unorm() for each pixel.
* This is going to be really really slow.
* Note: the block store format is actually
* x0y0x0y1x0y2x0y3 x1y0x1y1x1y2x1y3 ...
*/
for (i = 0; i < 4; ++i) {
for (j = 0; j < 4; ++j) {
LLVMValueRef args[4];
LLVMValueRef dst_offset = lp_build_const_int32(gallivm, (i * 4 + j) * 4);
/*
* Note we actually supply a pointer to the start of the block,
* not the start of the texture.
*/
args[0] = LLVMBuildGEP(gallivm->builder, tmp_ptr, &dst_offset, 1, "");
args[1] = ptr_addr;
args[2] = LLVMConstInt(i32t, i, 0);
args[3] = LLVMConstInt(i32t, j, 0);
LLVMBuildCall(builder, function, args, ARRAY_SIZE(args), "");
}
}
/* Finally store the block - pointless mem copy + update tag. */
tmp_ptr = LLVMBuildBitCast(builder, tmp_ptr, LLVMPointerType(i32x4, 0), "");
for (i = 0; i < 4; ++i) {
LLVMValueRef tmp_offset = lp_build_const_int32(gallivm, i);
LLVMValueRef ptr = LLVMBuildGEP(gallivm->builder, tmp_ptr, &tmp_offset, 1, "");
col[i] = LLVMBuildLoad(builder, ptr, "");
}
tag_value = LLVMBuildPtrToInt(gallivm->builder, ptr_addr,
LLVMInt64TypeInContext(gallivm->context), "");
store_cached_block(gallivm, col, tag_value, hash_index, cache);
}
// Generates the function prototype but overrides the name. This is used by
// the metadata to create external APIs to C.
//
// node - The node.
// module - The compilation unit this node is a part of.
// function_name - The name of the function to generate.
// is_external - A flag stating if this is an external prototype.
// value - A pointer to where the LLVM value should be returned to.
//
// Returns 0 if successful, otherwise returns -1.
int qip_ast_function_codegen_prototype_with_name(qip_ast_node *node,
qip_module *module,
bstring function_name,
bool is_external,
LLVMValueRef *value)
{
int rc;
unsigned int i;
LLVMValueRef func = NULL;
bstring arg_type_name = NULL;
bstring return_type_name = NULL;
LLVMContextRef context = LLVMGetModuleContext(module->llvm_module);
// Determine the number of arguments. External calls always have an
// additional first argument that is the module reference.
unsigned int offset = (is_external ? 1 : 0);
unsigned int total_arg_count = node->function.arg_count + offset;
// Check for an existing prototype.
func = LLVMGetNamedFunction(module->llvm_module, bdata(function_name));
// If a prototype exists then simply verify it matches and return it.
if(func != NULL) {
check(LLVMCountBasicBlocks(func) == 0, "Illegal function redefinition");
check(LLVMCountParams(func) == total_arg_count, "Function prototype already exists with different arguments");
}
// If there is no prototype then create one.
else {
// Dynamically generate the return type of the function if it is missing.
if(node->function.return_type == NULL) {
rc = qip_ast_function_generate_return_type(node, module);
check(rc == 0, "Unable to generate return type for function");
}
// Create a list of function argument types.
qip_ast_node *arg;
LLVMTypeRef *params = malloc(sizeof(LLVMTypeRef) * total_arg_count);
// Create module argument for external calls.
if(is_external) {
params[0] = LLVMPointerType(LLVMInt8TypeInContext(context), 0);
}
// Create arguments.
for(i=0; i<node->function.arg_count; i++) {
qip_ast_node *arg = node->function.args[i];
LLVMTypeRef param = NULL;
rc = qip_module_get_type_ref(module, arg->farg.var_decl->var_decl.type, NULL, ¶m);
check(rc == 0, "Unable to determine function argument type");
// Pass argument as reference if this is a complex type.
if(qip_module_is_complex_type(module, param)) {
params[i+offset] = LLVMPointerType(param, 0);
}
// Otherwise pass it by value.
else {
params[i+offset] = param;
}
}
// Determine return type.
LLVMTypeRef return_type;
rc = qip_module_get_type_ref(module, node->function.return_type, NULL, &return_type);
check(rc == 0, "Unable to determine function return type");
if(qip_module_is_complex_type(module, return_type)) {
return_type = LLVMPointerType(return_type, 0);
}
// Create function type.
LLVMTypeRef funcType = LLVMFunctionType(return_type, params, total_arg_count, false);
check(funcType != NULL, "Unable to create function type");
// Create function.
func = LLVMAddFunction(module->llvm_module, bdata(function_name), funcType);
check(func != NULL, "Unable to create function");
// Assign module argument name.
if(is_external) {
LLVMSetValueName(LLVMGetParam(func, 0), "module");
}
// Assign names to function arguments.
for(i=0; i<node->function.arg_count; i++) {
arg = node->function.args[i];
LLVMValueRef param = LLVMGetParam(func, i+offset);
LLVMSetValueName(param, bdata(arg->farg.var_decl->var_decl.name));
}
}
// Return function prototype;
*value = func;
bdestroy(arg_type_name);
bdestroy(return_type_name);
return 0;
error:
bdestroy(arg_type_name);
bdestroy(return_type_name);
if(func) LLVMDeleteFunction(func);
*value = NULL;
return -1;
}
/**
* Perform the occlusion test and increase the counter.
* Test the depth mask. Add the number of channel which has none zero mask
* into the occlusion counter. e.g. maskvalue is {-1, -1, -1, -1}.
* The counter will add 4.
*
* \param type holds element type of the mask vector.
* \param maskvalue is the depth test mask.
* \param counter is a pointer of the uint32 counter.
*/
void
lp_build_occlusion_count(struct gallivm_state *gallivm,
struct lp_type type,
LLVMValueRef maskvalue,
LLVMValueRef counter)
{
LLVMBuilderRef builder = gallivm->builder;
LLVMContextRef context = gallivm->context;
LLVMValueRef countmask = lp_build_const_int_vec(gallivm, type, 1);
LLVMValueRef count, newcount;
assert(type.length <= 16);
assert(type.floating);
if(util_cpu_caps.has_sse && type.length == 4) {
const char *movmskintr = "llvm.x86.sse.movmsk.ps";
const char *popcntintr = "llvm.ctpop.i32";
LLVMValueRef bits = LLVMBuildBitCast(builder, maskvalue,
lp_build_vec_type(gallivm, type), "");
bits = lp_build_intrinsic_unary(builder, movmskintr,
LLVMInt32TypeInContext(context), bits);
count = lp_build_intrinsic_unary(builder, popcntintr,
LLVMInt32TypeInContext(context), bits);
}
else if(util_cpu_caps.has_avx && type.length == 8) {
const char *movmskintr = "llvm.x86.avx.movmsk.ps.256";
const char *popcntintr = "llvm.ctpop.i32";
LLVMValueRef bits = LLVMBuildBitCast(builder, maskvalue,
lp_build_vec_type(gallivm, type), "");
bits = lp_build_intrinsic_unary(builder, movmskintr,
LLVMInt32TypeInContext(context), bits);
count = lp_build_intrinsic_unary(builder, popcntintr,
LLVMInt32TypeInContext(context), bits);
}
else {
unsigned i;
LLVMValueRef countv = LLVMBuildAnd(builder, maskvalue, countmask, "countv");
LLVMTypeRef counttype = LLVMIntTypeInContext(context, type.length * 8);
LLVMTypeRef i8vntype = LLVMVectorType(LLVMInt8TypeInContext(context), type.length * 4);
LLVMValueRef shufflev, countd;
LLVMValueRef shuffles[16];
const char *popcntintr = NULL;
countv = LLVMBuildBitCast(builder, countv, i8vntype, "");
for (i = 0; i < type.length; i++) {
shuffles[i] = lp_build_const_int32(gallivm, 4*i);
}
shufflev = LLVMConstVector(shuffles, type.length);
countd = LLVMBuildShuffleVector(builder, countv, LLVMGetUndef(i8vntype), shufflev, "");
countd = LLVMBuildBitCast(builder, countd, counttype, "countd");
/*
* XXX FIXME
* this is bad on cpus without popcount (on x86 supported by intel
* nehalem, amd barcelona, and up - not tied to sse42).
* Would be much faster to just sum the 4 elements of the vector with
* some horizontal add (shuffle/add/shuffle/add after the initial and).
*/
switch (type.length) {
case 4:
popcntintr = "llvm.ctpop.i32";
break;
case 8:
popcntintr = "llvm.ctpop.i64";
break;
case 16:
popcntintr = "llvm.ctpop.i128";
break;
default:
assert(0);
}
count = lp_build_intrinsic_unary(builder, popcntintr, counttype, countd);
if (type.length > 4) {
count = LLVMBuildTrunc(builder, count, LLVMIntTypeInContext(context, 32), "");
}
}
newcount = LLVMBuildLoad(builder, counter, "origcount");
newcount = LLVMBuildAdd(builder, newcount, count, "newcount");
LLVMBuildStore(builder, newcount, counter);
}
/**
* Fetch a pixel into a 4 float AoS.
*
* \param format_desc describes format of the image we're fetching from
* \param ptr address of the pixel block (or the texel if uncompressed)
* \param i, j the sub-block pixel coordinates. For non-compressed formats
* these will always be (0, 0).
* \return a 4 element vector with the pixel's RGBA values.
*/
LLVMValueRef
lp_build_fetch_rgba_aos(struct gallivm_state *gallivm,
const struct util_format_description *format_desc,
struct lp_type type,
LLVMValueRef base_ptr,
LLVMValueRef offset,
LLVMValueRef i,
LLVMValueRef j)
{
LLVMBuilderRef builder = gallivm->builder;
unsigned num_pixels = type.length / 4;
struct lp_build_context bld;
assert(type.length <= LP_MAX_VECTOR_LENGTH);
assert(type.length % 4 == 0);
lp_build_context_init(&bld, gallivm, type);
/*
* Trivial case
*
* The format matches the type (apart of a swizzle) so no need for
* scaling or converting.
*/
if (format_matches_type(format_desc, type) &&
format_desc->block.bits <= type.width * 4 &&
util_is_power_of_two(format_desc->block.bits)) {
LLVMValueRef packed;
/*
* The format matches the type (apart of a swizzle) so no need for
* scaling or converting.
*/
packed = lp_build_gather(gallivm, type.length/4,
format_desc->block.bits, type.width*4,
base_ptr, offset);
assert(format_desc->block.bits <= type.width * type.length);
packed = LLVMBuildBitCast(gallivm->builder, packed,
lp_build_vec_type(gallivm, type), "");
return lp_build_format_swizzle_aos(format_desc, &bld, packed);
}
/*
* Bit arithmetic
*/
if (format_desc->layout == UTIL_FORMAT_LAYOUT_PLAIN &&
(format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB ||
format_desc->colorspace == UTIL_FORMAT_COLORSPACE_ZS) &&
format_desc->block.width == 1 &&
format_desc->block.height == 1 &&
util_is_power_of_two(format_desc->block.bits) &&
format_desc->block.bits <= 32 &&
format_desc->is_bitmask &&
!format_desc->is_mixed &&
(format_desc->channel[0].type == UTIL_FORMAT_TYPE_UNSIGNED ||
format_desc->channel[1].type == UTIL_FORMAT_TYPE_UNSIGNED)) {
LLVMValueRef tmps[LP_MAX_VECTOR_LENGTH/4];
LLVMValueRef res;
unsigned k;
/*
* Unpack a pixel at a time into a <4 x float> RGBA vector
*/
for (k = 0; k < num_pixels; ++k) {
LLVMValueRef packed;
packed = lp_build_gather_elem(gallivm, num_pixels,
format_desc->block.bits, 32,
base_ptr, offset, k);
tmps[k] = lp_build_unpack_arith_rgba_aos(gallivm,
format_desc,
packed);
}
/*
* Type conversion.
*
* TODO: We could avoid floating conversion for integer to
* integer conversions.
*/
if (gallivm_debug & GALLIVM_DEBUG_PERF && !type.floating) {
debug_printf("%s: unpacking %s with floating point\n",
__FUNCTION__, format_desc->short_name);
}
lp_build_conv(gallivm,
lp_float32_vec4_type(),
type,
tmps, num_pixels, &res, 1);
return lp_build_format_swizzle_aos(format_desc, &bld, res);
}
/*
* YUV / subsampled formats
*/
if (format_desc->layout == UTIL_FORMAT_LAYOUT_SUBSAMPLED) {
struct lp_type tmp_type;
LLVMValueRef tmp;
memset(&tmp_type, 0, sizeof tmp_type);
tmp_type.width = 8;
tmp_type.length = num_pixels * 4;
tmp_type.norm = TRUE;
tmp = lp_build_fetch_subsampled_rgba_aos(gallivm,
format_desc,
num_pixels,
base_ptr,
offset,
i, j);
lp_build_conv(gallivm,
tmp_type, type,
&tmp, 1, &tmp, 1);
return tmp;
}
/*
* Fallback to util_format_description::fetch_rgba_8unorm().
*/
if (format_desc->fetch_rgba_8unorm &&
!type.floating && type.width == 8 && !type.sign && type.norm) {
/*
* Fallback to calling util_format_description::fetch_rgba_8unorm.
*
* This is definitely not the most efficient way of fetching pixels, as
* we miss the opportunity to do vectorization, but this it is a
* convenient for formats or scenarios for which there was no opportunity
* or incentive to optimize.
*/
LLVMModuleRef module = LLVMGetGlobalParent(LLVMGetBasicBlockParent(LLVMGetInsertBlock(gallivm->builder)));
char name[256];
LLVMTypeRef i8t = LLVMInt8TypeInContext(gallivm->context);
LLVMTypeRef pi8t = LLVMPointerType(i8t, 0);
LLVMTypeRef i32t = LLVMInt32TypeInContext(gallivm->context);
LLVMValueRef function;
LLVMValueRef tmp_ptr;
LLVMValueRef tmp;
LLVMValueRef res;
LLVMValueRef callee;
unsigned k;
util_snprintf(name, sizeof name, "util_format_%s_fetch_rgba_8unorm",
format_desc->short_name);
if (gallivm_debug & GALLIVM_DEBUG_PERF) {
debug_printf("%s: falling back to %s\n", __FUNCTION__, name);
}
/*
* Declare and bind format_desc->fetch_rgba_8unorm().
*/
function = LLVMGetNamedFunction(module, name);
if (!function) {
/*
* Function to call looks like:
* fetch(uint8_t *dst, const uint8_t *src, unsigned i, unsigned j)
*/
LLVMTypeRef ret_type;
LLVMTypeRef arg_types[4];
LLVMTypeRef function_type;
ret_type = LLVMVoidTypeInContext(gallivm->context);
arg_types[0] = pi8t;
arg_types[1] = pi8t;
arg_types[2] = i32t;
arg_types[3] = i32t;
function_type = LLVMFunctionType(ret_type, arg_types,
Elements(arg_types), 0);
function = LLVMAddFunction(module, name, function_type);
LLVMSetFunctionCallConv(function, LLVMCCallConv);
LLVMSetLinkage(function, LLVMExternalLinkage);
assert(LLVMIsDeclaration(function));
}
/* make const pointer for the C fetch_rgba_float function */
callee = lp_build_const_int_pointer(gallivm,
func_to_pointer((func_pointer) format_desc->fetch_rgba_8unorm));
/* cast the callee pointer to the function's type */
function = LLVMBuildBitCast(builder, callee,
LLVMTypeOf(function), "cast callee");
tmp_ptr = lp_build_alloca(gallivm, i32t, "");
res = LLVMGetUndef(LLVMVectorType(i32t, num_pixels));
/*
* Invoke format_desc->fetch_rgba_8unorm() for each pixel and insert the result
* in the SoA vectors.
*/
for (k = 0; k < num_pixels; ++k) {
LLVMValueRef index = lp_build_const_int32(gallivm, k);
LLVMValueRef args[4];
args[0] = LLVMBuildBitCast(builder, tmp_ptr, pi8t, "");
args[1] = lp_build_gather_elem_ptr(gallivm, num_pixels,
base_ptr, offset, k);
if (num_pixels == 1) {
args[2] = i;
args[3] = j;
}
else {
args[2] = LLVMBuildExtractElement(builder, i, index, "");
args[3] = LLVMBuildExtractElement(builder, j, index, "");
}
LLVMBuildCall(builder, function, args, Elements(args), "");
tmp = LLVMBuildLoad(builder, tmp_ptr, "");
if (num_pixels == 1) {
res = tmp;
}
else {
res = LLVMBuildInsertElement(builder, res, tmp, index, "");
}
}
/* Bitcast from <n x i32> to <4n x i8> */
res = LLVMBuildBitCast(builder, res, bld.vec_type, "");
return res;
}
/*
* Fallback to util_format_description::fetch_rgba_float().
*/
if (format_desc->fetch_rgba_float) {
/*
* Fallback to calling util_format_description::fetch_rgba_float.
*
* This is definitely not the most efficient way of fetching pixels, as
* we miss the opportunity to do vectorization, but this it is a
* convenient for formats or scenarios for which there was no opportunity
* or incentive to optimize.
*/
LLVMModuleRef module = LLVMGetGlobalParent(LLVMGetBasicBlockParent(LLVMGetInsertBlock(builder)));
char name[256];
LLVMTypeRef f32t = LLVMFloatTypeInContext(gallivm->context);
LLVMTypeRef f32x4t = LLVMVectorType(f32t, 4);
LLVMTypeRef pf32t = LLVMPointerType(f32t, 0);
LLVMTypeRef pi8t = LLVMPointerType(LLVMInt8TypeInContext(gallivm->context), 0);
LLVMTypeRef i32t = LLVMInt32TypeInContext(gallivm->context);
LLVMValueRef function;
LLVMValueRef tmp_ptr;
LLVMValueRef tmps[LP_MAX_VECTOR_LENGTH/4];
LLVMValueRef res;
LLVMValueRef callee;
unsigned k;
util_snprintf(name, sizeof name, "util_format_%s_fetch_rgba_float",
format_desc->short_name);
if (gallivm_debug & GALLIVM_DEBUG_PERF) {
debug_printf("%s: falling back to %s\n", __FUNCTION__, name);
}
/*
* Declare and bind format_desc->fetch_rgba_float().
*/
function = LLVMGetNamedFunction(module, name);
if (!function) {
/*
* Function to call looks like:
* fetch(float *dst, const uint8_t *src, unsigned i, unsigned j)
*/
LLVMTypeRef ret_type;
LLVMTypeRef arg_types[4];
LLVMTypeRef function_type;
ret_type = LLVMVoidTypeInContext(gallivm->context);
arg_types[0] = pf32t;
arg_types[1] = pi8t;
arg_types[2] = i32t;
arg_types[3] = i32t;
function_type = LLVMFunctionType(ret_type, arg_types,
Elements(arg_types), 0);
function = LLVMAddFunction(module, name, function_type);
LLVMSetFunctionCallConv(function, LLVMCCallConv);
LLVMSetLinkage(function, LLVMExternalLinkage);
assert(LLVMIsDeclaration(function));
}
/* Note: we're using this casting here instead of LLVMAddGlobalMapping()
* to work around a bug in LLVM 2.6.
*/
/* make const pointer for the C fetch_rgba_float function */
callee = lp_build_const_int_pointer(gallivm,
func_to_pointer((func_pointer) format_desc->fetch_rgba_float));
/* cast the callee pointer to the function's type */
function = LLVMBuildBitCast(builder, callee,
LLVMTypeOf(function), "cast callee");
tmp_ptr = lp_build_alloca(gallivm, f32x4t, "");
/*
* Invoke format_desc->fetch_rgba_float() for each pixel and insert the result
* in the SoA vectors.
*/
for (k = 0; k < num_pixels; ++k) {
LLVMValueRef args[4];
args[0] = LLVMBuildBitCast(builder, tmp_ptr, pf32t, "");
args[1] = lp_build_gather_elem_ptr(gallivm, num_pixels,
base_ptr, offset, k);
if (num_pixels == 1) {
args[2] = i;
args[3] = j;
}
else {
LLVMValueRef index = lp_build_const_int32(gallivm, k);
args[2] = LLVMBuildExtractElement(builder, i, index, "");
args[3] = LLVMBuildExtractElement(builder, j, index, "");
}
LLVMBuildCall(builder, function, args, Elements(args), "");
tmps[k] = LLVMBuildLoad(builder, tmp_ptr, "");
}
lp_build_conv(gallivm,
lp_float32_vec4_type(),
type,
tmps, num_pixels, &res, 1);
return res;
}
assert(0);
return lp_build_undef(gallivm, type);
}
/**
* Build code to compare two values 'a' and 'b' of 'type' using the given func.
* \param func one of PIPE_FUNC_x
* The result values will be 0 for false or ~0 for true.
*/
LLVMValueRef
lp_build_compare(struct gallivm_state *gallivm,
const struct lp_type type,
unsigned func,
LLVMValueRef a,
LLVMValueRef b)
{
LLVMBuilderRef builder = gallivm->builder;
LLVMTypeRef int_vec_type = lp_build_int_vec_type(gallivm, type);
LLVMValueRef zeros = LLVMConstNull(int_vec_type);
LLVMValueRef ones = LLVMConstAllOnes(int_vec_type);
LLVMValueRef cond;
LLVMValueRef res;
assert(func >= PIPE_FUNC_NEVER);
assert(func <= PIPE_FUNC_ALWAYS);
assert(lp_check_value(type, a));
assert(lp_check_value(type, b));
if(func == PIPE_FUNC_NEVER)
return zeros;
if(func == PIPE_FUNC_ALWAYS)
return ones;
#if defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64)
/*
* There are no unsigned integer comparison instructions in SSE.
*/
if (!type.floating && !type.sign &&
type.width * type.length == 128 &&
util_cpu_caps.has_sse2 &&
(func == PIPE_FUNC_LESS ||
func == PIPE_FUNC_LEQUAL ||
func == PIPE_FUNC_GREATER ||
func == PIPE_FUNC_GEQUAL) &&
(gallivm_debug & GALLIVM_DEBUG_PERF)) {
debug_printf("%s: inefficient <%u x i%u> unsigned comparison\n",
__FUNCTION__, type.length, type.width);
}
#endif
#if HAVE_LLVM < 0x0207
#if defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64)
if(type.width * type.length == 128) {
if(type.floating && util_cpu_caps.has_sse) {
/* float[4] comparison */
LLVMTypeRef vec_type = lp_build_vec_type(gallivm, type);
LLVMValueRef args[3];
unsigned cc;
boolean swap;
swap = FALSE;
switch(func) {
case PIPE_FUNC_EQUAL:
cc = 0;
break;
case PIPE_FUNC_NOTEQUAL:
cc = 4;
break;
case PIPE_FUNC_LESS:
cc = 1;
break;
case PIPE_FUNC_LEQUAL:
cc = 2;
break;
case PIPE_FUNC_GREATER:
cc = 1;
swap = TRUE;
break;
case PIPE_FUNC_GEQUAL:
cc = 2;
swap = TRUE;
break;
default:
assert(0);
return lp_build_undef(gallivm, type);
}
if(swap) {
args[0] = b;
args[1] = a;
}
else {
args[0] = a;
args[1] = b;
}
args[2] = LLVMConstInt(LLVMInt8TypeInContext(gallivm->context), cc, 0);
res = lp_build_intrinsic(builder,
"llvm.x86.sse.cmp.ps",
vec_type,
args, 3);
res = LLVMBuildBitCast(builder, res, int_vec_type, "");
return res;
}
else if(util_cpu_caps.has_sse2) {
/* int[4] comparison */
static const struct {
unsigned swap:1;
unsigned eq:1;
unsigned gt:1;
unsigned not:1;
} table[] = {
{0, 0, 0, 1}, /* PIPE_FUNC_NEVER */
{1, 0, 1, 0}, /* PIPE_FUNC_LESS */
{0, 1, 0, 0}, /* PIPE_FUNC_EQUAL */
{0, 0, 1, 1}, /* PIPE_FUNC_LEQUAL */
{0, 0, 1, 0}, /* PIPE_FUNC_GREATER */
{0, 1, 0, 1}, /* PIPE_FUNC_NOTEQUAL */
{1, 0, 1, 1}, /* PIPE_FUNC_GEQUAL */
{0, 0, 0, 0} /* PIPE_FUNC_ALWAYS */
};
const char *pcmpeq;
const char *pcmpgt;
LLVMValueRef args[2];
LLVMValueRef res;
LLVMTypeRef vec_type = lp_build_vec_type(gallivm, type);
switch (type.width) {
case 8:
pcmpeq = "llvm.x86.sse2.pcmpeq.b";
pcmpgt = "llvm.x86.sse2.pcmpgt.b";
break;
case 16:
pcmpeq = "llvm.x86.sse2.pcmpeq.w";
pcmpgt = "llvm.x86.sse2.pcmpgt.w";
break;
case 32:
pcmpeq = "llvm.x86.sse2.pcmpeq.d";
pcmpgt = "llvm.x86.sse2.pcmpgt.d";
break;
default:
assert(0);
return lp_build_undef(gallivm, type);
}
/* There are no unsigned comparison instructions. So flip the sign bit
* so that the results match.
*/
if (table[func].gt && !type.sign) {
LLVMValueRef msb = lp_build_const_int_vec(gallivm, type, (unsigned long long)1 << (type.width - 1));
a = LLVMBuildXor(builder, a, msb, "");
b = LLVMBuildXor(builder, b, msb, "");
}
if(table[func].swap) {
args[0] = b;
args[1] = a;
}
else {
args[0] = a;
args[1] = b;
}
if(table[func].eq)
res = lp_build_intrinsic(builder, pcmpeq, vec_type, args, 2);
else if (table[func].gt)
res = lp_build_intrinsic(builder, pcmpgt, vec_type, args, 2);
else
res = LLVMConstNull(vec_type);
if(table[func].not)
res = LLVMBuildNot(builder, res, "");
return res;
}
} /* if (type.width * type.length == 128) */
#endif
#endif /* HAVE_LLVM < 0x0207 */
/* XXX: It is not clear if we should use the ordered or unordered operators */
if(type.floating) {
LLVMRealPredicate op;
switch(func) {
case PIPE_FUNC_NEVER:
op = LLVMRealPredicateFalse;
break;
case PIPE_FUNC_ALWAYS:
op = LLVMRealPredicateTrue;
break;
case PIPE_FUNC_EQUAL:
op = LLVMRealUEQ;
break;
case PIPE_FUNC_NOTEQUAL:
op = LLVMRealUNE;
break;
case PIPE_FUNC_LESS:
op = LLVMRealULT;
break;
case PIPE_FUNC_LEQUAL:
op = LLVMRealULE;
break;
case PIPE_FUNC_GREATER:
op = LLVMRealUGT;
break;
case PIPE_FUNC_GEQUAL:
op = LLVMRealUGE;
break;
default:
assert(0);
return lp_build_undef(gallivm, type);
}
#if HAVE_LLVM >= 0x0207
cond = LLVMBuildFCmp(builder, op, a, b, "");
res = LLVMBuildSExt(builder, cond, int_vec_type, "");
#else
if (type.length == 1) {
cond = LLVMBuildFCmp(builder, op, a, b, "");
res = LLVMBuildSExt(builder, cond, int_vec_type, "");
}
else {
unsigned i;
res = LLVMGetUndef(int_vec_type);
debug_printf("%s: warning: using slow element-wise float"
" vector comparison\n", __FUNCTION__);
for (i = 0; i < type.length; ++i) {
LLVMValueRef index = lp_build_const_int32(gallivm, i);
cond = LLVMBuildFCmp(builder, op,
LLVMBuildExtractElement(builder, a, index, ""),
LLVMBuildExtractElement(builder, b, index, ""),
"");
cond = LLVMBuildSelect(builder, cond,
LLVMConstExtractElement(ones, index),
LLVMConstExtractElement(zeros, index),
"");
res = LLVMBuildInsertElement(builder, res, cond, index, "");
}
}
#endif
}
else {
LLVMIntPredicate op;
switch(func) {
case PIPE_FUNC_EQUAL:
op = LLVMIntEQ;
break;
case PIPE_FUNC_NOTEQUAL:
op = LLVMIntNE;
break;
case PIPE_FUNC_LESS:
op = type.sign ? LLVMIntSLT : LLVMIntULT;
break;
case PIPE_FUNC_LEQUAL:
op = type.sign ? LLVMIntSLE : LLVMIntULE;
break;
case PIPE_FUNC_GREATER:
op = type.sign ? LLVMIntSGT : LLVMIntUGT;
break;
case PIPE_FUNC_GEQUAL:
op = type.sign ? LLVMIntSGE : LLVMIntUGE;
break;
default:
assert(0);
return lp_build_undef(gallivm, type);
}
#if HAVE_LLVM >= 0x0207
cond = LLVMBuildICmp(builder, op, a, b, "");
res = LLVMBuildSExt(builder, cond, int_vec_type, "");
#else
if (type.length == 1) {
cond = LLVMBuildICmp(builder, op, a, b, "");
res = LLVMBuildSExt(builder, cond, int_vec_type, "");
}
else {
unsigned i;
res = LLVMGetUndef(int_vec_type);
if (gallivm_debug & GALLIVM_DEBUG_PERF) {
debug_printf("%s: using slow element-wise int"
" vector comparison\n", __FUNCTION__);
}
for(i = 0; i < type.length; ++i) {
LLVMValueRef index = lp_build_const_int32(gallivm, i);
cond = LLVMBuildICmp(builder, op,
LLVMBuildExtractElement(builder, a, index, ""),
LLVMBuildExtractElement(builder, b, index, ""),
"");
cond = LLVMBuildSelect(builder, cond,
LLVMConstExtractElement(ones, index),
LLVMConstExtractElement(zeros, index),
"");
res = LLVMBuildInsertElement(builder, res, cond, index, "");
}
}
#endif
}
return res;
}
/**
* Gather one element from scatter positions in memory.
* Nearly the same as above, however the individual elements
* may be vectors themselves, and fetches may be float type.
* Can also do pad vector instead of ZExt.
*
* @sa lp_build_gather()
*/
static LLVMValueRef
lp_build_gather_elem_vec(struct gallivm_state *gallivm,
unsigned length,
unsigned src_width,
LLVMTypeRef src_type,
struct lp_type dst_type,
boolean aligned,
LLVMValueRef base_ptr,
LLVMValueRef offsets,
unsigned i,
boolean vector_justify)
{
LLVMValueRef ptr, res;
LLVMTypeRef src_ptr_type = LLVMPointerType(src_type, 0);
assert(LLVMTypeOf(base_ptr) == LLVMPointerType(LLVMInt8TypeInContext(gallivm->context), 0));
ptr = lp_build_gather_elem_ptr(gallivm, length, base_ptr, offsets, i);
ptr = LLVMBuildBitCast(gallivm->builder, ptr, src_ptr_type, "");
res = LLVMBuildLoad(gallivm->builder, ptr, "");
/* XXX
* On some archs we probably really want to avoid having to deal
* with alignments lower than 4 bytes (if fetch size is a power of
* two >= 32). On x86 it doesn't matter, however.
* We should be able to guarantee full alignment for any kind of texture
* fetch (except ARB_texture_buffer_range, oops), but not vertex fetch
* (there's PIPE_CAP_VERTEX_BUFFER_OFFSET_4BYTE_ALIGNED_ONLY and friends
* but I don't think that's quite what we wanted).
* For ARB_texture_buffer_range, PIPE_CAP_TEXTURE_BUFFER_OFFSET_ALIGNMENT
* looks like a good fit, but it seems this cap bit (and OpenGL) aren't
* enforcing what we want (which is what d3d10 does, the offset needs to
* be aligned to element size, but GL has bytes regardless of element
* size which would only leave us with minimum alignment restriction of 16
* which doesn't make much sense if the type isn't 4x32bit). Due to
* translation of offsets to first_elem in sampler_views it actually seems
* gallium could not do anything else except 16 no matter what...
*/
if (!aligned) {
LLVMSetAlignment(res, 1);
} else if (!util_is_power_of_two(src_width)) {
/*
* Full alignment is impossible, assume the caller really meant
* the individual elements were aligned (e.g. 3x32bit format).
* And yes the generated code may otherwise crash, llvm will
* really assume 128bit alignment with a 96bit fetch (I suppose
* that makes sense as it can just assume the upper 32bit to be
* whatever).
* Maybe the caller should be able to explicitly set this, but
* this should cover all the 3-channel formats.
*/
if (((src_width / 24) * 24 == src_width) &&
util_is_power_of_two(src_width / 24)) {
LLVMSetAlignment(res, src_width / 24);
} else {
LLVMSetAlignment(res, 1);
}
}
assert(src_width <= dst_type.width * dst_type.length);
if (src_width < dst_type.width * dst_type.length) {
if (dst_type.length > 1) {
res = lp_build_pad_vector(gallivm, res, dst_type.length);
/*
* vector_justify hopefully a non-issue since we only deal
* with src_width >= 32 here?
*/
} else {
LLVMTypeRef dst_elem_type = lp_build_vec_type(gallivm, dst_type);
/*
* Only valid if src_ptr_type is int type...
*/
res = LLVMBuildZExt(gallivm->builder, res, dst_elem_type, "");
#ifdef PIPE_ARCH_BIG_ENDIAN
if (vector_justify) {
res = LLVMBuildShl(gallivm->builder, res,
LLVMConstInt(dst_elem_type,
dst_type.width - src_width, 0), "");
}
if (src_width == 48) {
/* Load 3x16 bit vector.
* The sequence of loads on big-endian hardware proceeds as follows.
* 16-bit fields are denoted by X, Y, Z, and 0. In memory, the sequence
* of three fields appears in the order X, Y, Z.
*
* Load 32-bit word: 0.0.X.Y
* Load 16-bit halfword: 0.0.0.Z
* Rotate left: 0.X.Y.0
* Bitwise OR: 0.X.Y.Z
*
* The order in which we need the fields in the result is 0.Z.Y.X,
* the same as on little-endian; permute 16-bit fields accordingly
* within 64-bit register:
*/
LLVMValueRef shuffles[4] = {
lp_build_const_int32(gallivm, 2),
lp_build_const_int32(gallivm, 1),
lp_build_const_int32(gallivm, 0),
lp_build_const_int32(gallivm, 3),
};
res = LLVMBuildBitCast(gallivm->builder, res,
lp_build_vec_type(gallivm, lp_type_uint_vec(16, 4*16)), "");
res = LLVMBuildShuffleVector(gallivm->builder, res, res, LLVMConstVector(shuffles, 4), "");
res = LLVMBuildBitCast(gallivm->builder, res, dst_elem_type, "");
}
#endif
}
}
return res;
}
