static LLVMValueRef
CreateFibFunction(LLVMModuleRef M, LLVMContextRef Context)
{
LLVMBuilderRef B = LLVMCreateBuilderInContext(Context);
// Create the fib function and insert it into module M. This function is said
// to return an int and take an int parameter.
LLVMTypeRef ParamTypes[] = {LLVMInt32TypeInContext(Context)};
LLVMTypeRef ReturnType = LLVMInt32TypeInContext(Context);
LLVMTypeRef FunctionTy = LLVMFunctionType(ReturnType, ParamTypes, 1, 0);
LLVMValueRef FibF = LLVMAddFunction(M, "fib", FunctionTy);
// Add a basic block to the function.
LLVMBasicBlockRef BB = LLVMAppendBasicBlockInContext(Context, FibF, "EntryBlock");
// Get pointers to the constants.
LLVMValueRef One = LLVMConstInt(LLVMInt32TypeInContext(Context), 1, 0);
LLVMValueRef Two = LLVMConstInt(LLVMInt32TypeInContext(Context), 2, 0);
// Get pointer to the integer argument of the add1 function...
LLVMValueRef ArgX = LLVMGetFirstParam(FibF); // Get the arg.
LLVMSetValueName(ArgX, "AnArg"); // Give it a nice symbolic name for fun.
// Create the true_block.
LLVMBasicBlockRef RetBB = LLVMAppendBasicBlockInContext(Context, FibF, "return");
// Create an exit block.
LLVMBasicBlockRef RecurseBB = LLVMAppendBasicBlockInContext(Context, FibF, "recurse");
// Create the "if (arg <= 2) goto exitbb"
LLVMPositionBuilderAtEnd(B, BB);
LLVMValueRef CondInst = LLVMBuildICmp(B, LLVMIntSLE, ArgX, Two, "cond");
LLVMBuildCondBr(B, CondInst, RetBB, RecurseBB);
// Create: ret int 1
LLVMPositionBuilderAtEnd(B, RetBB);
LLVMBuildRet(B, One);
// create fib(x-1)
LLVMPositionBuilderAtEnd(B, RecurseBB);
LLVMValueRef Sub = LLVMBuildSub(B, ArgX, One, "arg");
LLVMValueRef CallFibX1 = LLVMBuildCall(B, FibF, &Sub, 1, "fibx1");
LLVMSetTailCall(CallFibX1, 1);
// create fib(x-2)
LLVMPositionBuilderAtEnd(B, RecurseBB);
Sub = LLVMBuildSub(B, ArgX, Two, "arg");
LLVMValueRef CallFibX2 = LLVMBuildCall(B, FibF, &Sub, 1, "fibx2");
LLVMSetTailCall(CallFibX2, 1);
// fib(x-1)+fib(x-2)
LLVMPositionBuilderAtEnd(B, RecurseBB);
LLVMValueRef Sum = LLVMBuildAdd(B, CallFibX1, CallFibX2, "addresult");
// Create the return instruction and add it to the basic block
LLVMPositionBuilderAtEnd(B, RecurseBB);
LLVMBuildRet(B, Sum);
return FibF;
}
static LLVMValueRef
add_printf_test(struct gallivm_state *gallivm)
{
LLVMModuleRef module = gallivm->module;
LLVMTypeRef args[1] = { LLVMIntTypeInContext(gallivm->context, 32) };
LLVMValueRef func = LLVMAddFunction(module, "test_printf", LLVMFunctionType(LLVMVoidTypeInContext(gallivm->context), args, 1, 0));
LLVMBuilderRef builder = gallivm->builder;
LLVMBasicBlockRef block = LLVMAppendBasicBlockInContext(gallivm->context, func, "entry");
LLVMSetFunctionCallConv(func, LLVMCCallConv);
LLVMPositionBuilderAtEnd(builder, block);
lp_build_printf(gallivm, "hello, world\n");
lp_build_printf(gallivm, "print 5 6: %d %d\n", LLVMConstInt(LLVMInt32TypeInContext(gallivm->context), 5, 0),
LLVMConstInt(LLVMInt32TypeInContext(gallivm->context), 6, 0));
/* Also test lp_build_assert(). This should not fail. */
lp_build_assert(gallivm, LLVMConstInt(LLVMInt32TypeInContext(gallivm->context), 1, 0), "assert(1)");
LLVMBuildRetVoid(builder);
gallivm_verify_function(gallivm, func);
return func;
}
static LLVMValueRef
add_conv_test(struct gallivm_state *gallivm,
struct lp_type src_type, unsigned num_srcs,
struct lp_type dst_type, unsigned num_dsts)
{
LLVMModuleRef module = gallivm->module;
LLVMContextRef context = gallivm->context;
LLVMBuilderRef builder = gallivm->builder;
LLVMTypeRef args[2];
LLVMValueRef func;
LLVMValueRef src_ptr;
LLVMValueRef dst_ptr;
LLVMBasicBlockRef block;
LLVMValueRef src[LP_MAX_VECTOR_LENGTH];
LLVMValueRef dst[LP_MAX_VECTOR_LENGTH];
unsigned i;
args[0] = LLVMPointerType(lp_build_vec_type(gallivm, src_type), 0);
args[1] = LLVMPointerType(lp_build_vec_type(gallivm, dst_type), 0);
func = LLVMAddFunction(module, "test",
LLVMFunctionType(LLVMVoidTypeInContext(context),
args, 2, 0));
LLVMSetFunctionCallConv(func, LLVMCCallConv);
src_ptr = LLVMGetParam(func, 0);
dst_ptr = LLVMGetParam(func, 1);
block = LLVMAppendBasicBlockInContext(context, func, "entry");
LLVMPositionBuilderAtEnd(builder, block);
for(i = 0; i < num_srcs; ++i) {
LLVMValueRef index = LLVMConstInt(LLVMInt32TypeInContext(context), i, 0);
LLVMValueRef ptr = LLVMBuildGEP(builder, src_ptr, &index, 1, "");
src[i] = LLVMBuildLoad(builder, ptr, "");
}
lp_build_conv(gallivm, src_type, dst_type, src, num_srcs, dst, num_dsts);
for(i = 0; i < num_dsts; ++i) {
LLVMValueRef index = LLVMConstInt(LLVMInt32TypeInContext(context), i, 0);
LLVMValueRef ptr = LLVMBuildGEP(builder, dst_ptr, &index, 1, "");
LLVMBuildStore(builder, dst[i], ptr);
}
LLVMBuildRetVoid(builder);;
gallivm_verify_function(gallivm, func);
return func;
}
/**
* Unpack and broadcast packed aos values consisting of only the
* first value, i.e. x1 x2 _ _ will become x1 x1 x1 x1 x2 x2 x2 x2
*/
LLVMValueRef
lp_build_unpack_broadcast_aos_scalars(struct gallivm_state *gallivm,
struct lp_type src_type,
struct lp_type dst_type,
const LLVMValueRef src)
{
LLVMTypeRef i32t = LLVMInt32TypeInContext(gallivm->context);
LLVMValueRef shuffles[LP_MAX_VECTOR_LENGTH];
unsigned num_dst = dst_type.length;
unsigned num_src = dst_type.length / 4;
unsigned i;
assert(num_dst / 4 <= src_type.length);
for (i = 0; i < num_src; i++) {
shuffles[i*4] = LLVMConstInt(i32t, i, 0);
shuffles[i*4+1] = LLVMConstInt(i32t, i, 0);
shuffles[i*4+2] = LLVMConstInt(i32t, i, 0);
shuffles[i*4+3] = LLVMConstInt(i32t, i, 0);
}
if (num_src == 1) {
return lp_build_extract_broadcast(gallivm, src_type, dst_type,
src, shuffles[0]);
}
else {
return LLVMBuildShuffleVector(gallivm->builder, src, src,
LLVMConstVector(shuffles, num_dst), "");
}
}
/**
* Pack first element of aos values,
* pad out to destination size.
* i.e. x1 _ _ _ x2 _ _ _ will become x1 x2 _ _
*/
LLVMValueRef
lp_build_pack_aos_scalars(struct gallivm_state *gallivm,
struct lp_type src_type,
struct lp_type dst_type,
const LLVMValueRef src)
{
LLVMTypeRef i32t = LLVMInt32TypeInContext(gallivm->context);
LLVMValueRef undef = LLVMGetUndef(i32t);
LLVMValueRef shuffles[LP_MAX_VECTOR_LENGTH];
unsigned num_src = src_type.length / 4;
unsigned num_dst = dst_type.length;
unsigned i;
assert(num_src <= num_dst);
for (i = 0; i < num_src; i++) {
shuffles[i] = LLVMConstInt(i32t, i * 4, 0);
}
for (i = num_src; i < num_dst; i++) {
shuffles[i] = undef;
}
if (num_dst == 1) {
return LLVMBuildExtractElement(gallivm->builder, src, shuffles[0], "");
}
else {
return LLVMBuildShuffleVector(gallivm->builder, src, src,
LLVMConstVector(shuffles, num_dst), "");
}
}
/**
* 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, "");
}
/**
* Extract and broadcast texture size.
*
* @param size_type type of the texture size vector (either
* bld->int_size_type or bld->float_size_type)
* @param coord_type type of the texture size vector (either
* bld->int_coord_type or bld->coord_type)
* @param int_size vector with the integer texture size (width, height,
* depth)
*/
void
lp_build_extract_image_sizes(struct lp_build_sample_context *bld,
struct lp_type size_type,
struct lp_type coord_type,
LLVMValueRef size,
LLVMValueRef *out_width,
LLVMValueRef *out_height,
LLVMValueRef *out_depth)
{
const unsigned dims = bld->dims;
LLVMTypeRef i32t = LLVMInt32TypeInContext(bld->gallivm->context);
*out_width = lp_build_extract_broadcast(bld->gallivm,
size_type,
coord_type,
size,
LLVMConstInt(i32t, 0, 0));
if (dims >= 2) {
*out_height = lp_build_extract_broadcast(bld->gallivm,
size_type,
coord_type,
size,
LLVMConstInt(i32t, 1, 0));
if (dims == 3) {
*out_depth = lp_build_extract_broadcast(bld->gallivm,
size_type,
coord_type,
size,
LLVMConstInt(i32t, 2, 0));
}
}
}
/**
* Swizzle a vector consisting of an array of XYZW structs.
*
* This fills a vector of dst_len length with the swizzled channels from src.
*
* e.g. with swizzles = { 2, 1, 0 } and swizzle_count = 6 results in
* RGBA RGBA = BGR BGR BG
*
* @param swizzles the swizzle array
* @param num_swizzles the number of elements in swizzles
* @param dst_len the length of the result
*/
LLVMValueRef
lp_build_swizzle_aos_n(struct gallivm_state* gallivm,
LLVMValueRef src,
const unsigned char* swizzles,
unsigned num_swizzles,
unsigned dst_len)
{
LLVMBuilderRef builder = gallivm->builder;
LLVMValueRef shuffles[LP_MAX_VECTOR_WIDTH];
unsigned i;
assert(dst_len < LP_MAX_VECTOR_WIDTH);
for (i = 0; i < dst_len; ++i) {
int swizzle = swizzles[i % num_swizzles];
if (swizzle == LP_BLD_SWIZZLE_DONTCARE) {
shuffles[i] = LLVMGetUndef(LLVMInt32TypeInContext(gallivm->context));
} else {
shuffles[i] = lp_build_const_int32(gallivm, swizzle);
}
}
return LLVMBuildShuffleVector(builder, src, LLVMGetUndef(LLVMTypeOf(src)), LLVMConstVector(shuffles, dst_len), "");
}
LLVMValueRef
lp_build_broadcast(struct gallivm_state *gallivm,
LLVMTypeRef vec_type,
LLVMValueRef scalar)
{
LLVMValueRef res;
if (LLVMGetTypeKind(vec_type) != LLVMVectorTypeKind) {
/* scalar */
assert(vec_type == LLVMTypeOf(scalar));
res = scalar;
} else {
LLVMBuilderRef builder = gallivm->builder;
const unsigned length = LLVMGetVectorSize(vec_type);
LLVMValueRef undef = LLVMGetUndef(vec_type);
/* The shuffle vector is always made of int32 elements */
LLVMTypeRef i32_type = LLVMInt32TypeInContext(gallivm->context);
LLVMTypeRef i32_vec_type = LLVMVectorType(i32_type, length);
assert(LLVMGetElementType(vec_type) == LLVMTypeOf(scalar));
res = LLVMBuildInsertElement(builder, undef, scalar, LLVMConstNull(i32_type), "");
res = LLVMBuildShuffleVector(builder, res, undef, LLVMConstNull(i32_vec_type), "");
}
return res;
}
static void emit_icmp(const struct lp_build_tgsi_action *action,
struct lp_build_tgsi_context *bld_base,
struct lp_build_emit_data *emit_data)
{
unsigned pred;
LLVMBuilderRef builder = bld_base->base.gallivm->builder;
LLVMContextRef context = bld_base->base.gallivm->context;
switch (emit_data->inst->Instruction.Opcode) {
case TGSI_OPCODE_USEQ:
case TGSI_OPCODE_U64SEQ: pred = LLVMIntEQ; break;
case TGSI_OPCODE_USNE:
case TGSI_OPCODE_U64SNE: pred = LLVMIntNE; break;
case TGSI_OPCODE_USGE:
case TGSI_OPCODE_U64SGE: pred = LLVMIntUGE; break;
case TGSI_OPCODE_USLT:
case TGSI_OPCODE_U64SLT: pred = LLVMIntULT; break;
case TGSI_OPCODE_ISGE:
case TGSI_OPCODE_I64SGE: pred = LLVMIntSGE; break;
case TGSI_OPCODE_ISLT:
case TGSI_OPCODE_I64SLT: pred = LLVMIntSLT; break;
default:
assert(!"unknown instruction");
pred = 0;
break;
}
LLVMValueRef v = LLVMBuildICmp(builder, pred,
emit_data->args[0], emit_data->args[1],"");
v = LLVMBuildSExtOrBitCast(builder, v,
LLVMInt32TypeInContext(context), "");
emit_data->output[emit_data->chan] = v;
}
static void
store_cached_block(struct gallivm_state *gallivm,
LLVMValueRef *col,
LLVMValueRef tag_value,
LLVMValueRef hash_index,
LLVMValueRef cache)
{
LLVMBuilderRef builder = gallivm->builder;
LLVMValueRef ptr, indices[3];
LLVMTypeRef type_ptr4x32;
unsigned count;
type_ptr4x32 = LLVMPointerType(LLVMVectorType(LLVMInt32TypeInContext(gallivm->context), 4), 0);
indices[0] = lp_build_const_int32(gallivm, 0);
indices[1] = lp_build_const_int32(gallivm, LP_BUILD_FORMAT_CACHE_MEMBER_TAGS);
indices[2] = hash_index;
ptr = LLVMBuildGEP(builder, cache, indices, ARRAY_SIZE(indices), "");
LLVMBuildStore(builder, tag_value, ptr);
indices[1] = lp_build_const_int32(gallivm, LP_BUILD_FORMAT_CACHE_MEMBER_DATA);
hash_index = LLVMBuildMul(builder, hash_index,
lp_build_const_int32(gallivm, 16), "");
for (count = 0; count < 4; count++) {
indices[2] = hash_index;
ptr = LLVMBuildGEP(builder, cache, indices, ARRAY_SIZE(indices), "");
ptr = LLVMBuildBitCast(builder, ptr, type_ptr4x32, "");
LLVMBuildStore(builder, col[count], ptr);
hash_index = LLVMBuildAdd(builder, hash_index,
lp_build_const_int32(gallivm, 4), "");
}
}
static LLVMValueRef
emit_fetch_immediate(
struct lp_build_tgsi_context *bld_base,
const struct tgsi_full_src_register *reg,
enum tgsi_opcode_type type,
unsigned swizzle)
{
LLVMTypeRef ctype;
LLVMContextRef ctx = bld_base->base.gallivm->context;
switch (type) {
case TGSI_TYPE_UNSIGNED:
case TGSI_TYPE_SIGNED:
ctype = LLVMInt32TypeInContext(ctx);
break;
case TGSI_TYPE_UNTYPED:
case TGSI_TYPE_FLOAT:
ctype = LLVMFloatTypeInContext(ctx);
break;
default:
ctype = 0;
break;
}
struct lp_build_tgsi_soa_context *bld = lp_soa_context(bld_base);
return LLVMConstBitCast(bld->immediates[reg->Register.Index][swizzle], ctype);
}
static void emit_dcmp(const struct lp_build_tgsi_action *action,
struct lp_build_tgsi_context *bld_base,
struct lp_build_emit_data *emit_data)
{
LLVMBuilderRef builder = bld_base->base.gallivm->builder;
LLVMContextRef context = bld_base->base.gallivm->context;
LLVMRealPredicate pred;
/* Use ordered for everything but NE (which is usual for
* float comparisons)
*/
switch (emit_data->inst->Instruction.Opcode) {
case TGSI_OPCODE_DSEQ: pred = LLVMRealOEQ; break;
case TGSI_OPCODE_DSGE: pred = LLVMRealOGE; break;
case TGSI_OPCODE_DSLT: pred = LLVMRealOLT; break;
case TGSI_OPCODE_DSNE: pred = LLVMRealUNE; break;
default: assert(!"unknown instruction"); pred = 0; break;
}
LLVMValueRef v = LLVMBuildFCmp(builder, pred,
emit_data->args[0], emit_data->args[1],"");
v = LLVMBuildSExtOrBitCast(builder, v,
LLVMInt32TypeInContext(context), "");
emit_data->output[emit_data->chan] = v;
}
/**
* Combined extract and broadcast (mere shuffle in most cases)
*/
LLVMValueRef
lp_build_extract_broadcast(struct gallivm_state *gallivm,
struct lp_type src_type,
struct lp_type dst_type,
LLVMValueRef vector,
LLVMValueRef index)
{
LLVMTypeRef i32t = LLVMInt32TypeInContext(gallivm->context);
LLVMValueRef res;
assert(src_type.floating == dst_type.floating);
assert(src_type.width == dst_type.width);
assert(lp_check_value(src_type, vector));
assert(LLVMTypeOf(index) == i32t);
if (src_type.length == 1) {
if (dst_type.length == 1) {
/*
* Trivial scalar -> scalar.
*/
res = vector;
}
else {
/*
* Broadcast scalar -> vector.
*/
res = lp_build_broadcast(gallivm,
lp_build_vec_type(gallivm, dst_type),
vector);
}
}
else {
if (dst_type.length > 1) {
/*
* shuffle - result can be of different length.
*/
LLVMValueRef shuffle;
shuffle = lp_build_broadcast(gallivm,
LLVMVectorType(i32t, dst_type.length),
index);
res = LLVMBuildShuffleVector(gallivm->builder, vector,
LLVMGetUndef(lp_build_vec_type(gallivm, src_type)),
shuffle, "");
}
else {
/*
* Trivial extract scalar from vector.
*/
res = LLVMBuildExtractElement(gallivm->builder, vector, index, "");
}
}
return res;
}
/**
* Return mask ? a : b;
*
* mask is a TGSI_WRITEMASK_xxx.
*/
LLVMValueRef
lp_build_select_aos(struct lp_build_context *bld,
unsigned mask,
LLVMValueRef a,
LLVMValueRef b,
unsigned num_channels)
{
LLVMBuilderRef builder = bld->gallivm->builder;
const struct lp_type type = bld->type;
const unsigned n = type.length;
unsigned i, j;
assert((mask & ~0xf) == 0);
assert(lp_check_value(type, a));
assert(lp_check_value(type, b));
if(a == b)
return a;
if((mask & 0xf) == 0xf)
return a;
if((mask & 0xf) == 0x0)
return b;
if(a == bld->undef || b == bld->undef)
return bld->undef;
/*
* There are two major ways of accomplishing this:
* - with a shuffle
* - with a select
*
* The flip between these is empirical and might need to be adjusted.
*/
if (n <= 4) {
/*
* Shuffle.
*/
LLVMTypeRef elem_type = LLVMInt32TypeInContext(bld->gallivm->context);
LLVMValueRef shuffles[LP_MAX_VECTOR_LENGTH];
for(j = 0; j < n; j += num_channels)
for(i = 0; i < num_channels; ++i)
shuffles[j + i] = LLVMConstInt(elem_type,
(mask & (1 << i) ? 0 : n) + j + i,
0);
return LLVMBuildShuffleVector(builder, a, b, LLVMConstVector(shuffles, n), "");
}
else {
LLVMValueRef mask_vec = lp_build_const_mask_aos(bld->gallivm, type, mask, num_channels);
return lp_build_select(bld, mask_vec, a, b);
}
}
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;
}
int
main(int argc, char **argv)
{
int n = argc > 1 ? atol(argv[1]) : 24;
LLVMInitializeNativeTarget();
LLVMLinkInInterpreter();
LLVMContextRef Context = LLVMContextCreate();
// Create some module to put our function into it.
LLVMModuleRef M = LLVMModuleCreateWithNameInContext("test", Context);
// We are about to create the "fib" function:
LLVMValueRef FibF = CreateFibFunction(M, Context);
// Now we going to create JIT
LLVMExecutionEngineRef EE;
char * outError;
if (LLVMCreateInterpreterForModule(&EE, M, &outError) != 0) {
printf("%s\n", outError);
return 1;
}
printf("verifying...\n");
if (LLVMVerifyModule(M, LLVMReturnStatusAction, &outError) != 0) {
printf("%s\n", outError);
return 1;
}
printf("OK\n");
printf("We just constructed this LLVM module:\n\n---------\n");
printf("%s\n", LLVMPrintModuleToString(M));
LLVMGenericValueRef Args = LLVMCreateGenericValueOfInt(LLVMInt32TypeInContext(Context), n, 0);
LLVMGenericValueRef Result = LLVMRunFunction(EE, FibF, 1, &Args);
printf("Result: %llu\n", LLVMGenericValueToInt(Result, 0));
return 0;
}
static LLVMValueRef emit_swizzle(
struct lp_build_tgsi_context * bld_base,
LLVMValueRef value,
unsigned swizzle_x,
unsigned swizzle_y,
unsigned swizzle_z,
unsigned swizzle_w)
{
LLVMValueRef swizzles[4];
LLVMTypeRef i32t =
LLVMInt32TypeInContext(bld_base->base.gallivm->context);
swizzles[0] = LLVMConstInt(i32t, swizzle_x, 0);
swizzles[1] = LLVMConstInt(i32t, swizzle_y, 0);
swizzles[2] = LLVMConstInt(i32t, swizzle_z, 0);
swizzles[3] = LLVMConstInt(i32t, swizzle_w, 0);
return LLVMBuildShuffleVector(bld_base->base.gallivm->builder,
value,
LLVMGetUndef(LLVMTypeOf(value)),
LLVMConstVector(swizzles, 4), "");
}
static LLVMValueRef
emit_fetch_immediate(
struct lp_build_tgsi_context *bld_base,
const struct tgsi_full_src_register *reg,
enum tgsi_opcode_type type,
unsigned swizzle)
{
LLVMTypeRef ctype;
LLVMContextRef ctx = bld_base->base.gallivm->context;
switch (type) {
case TGSI_TYPE_UNSIGNED:
case TGSI_TYPE_SIGNED:
ctype = LLVMInt32TypeInContext(ctx);
break;
case TGSI_TYPE_UNTYPED:
case TGSI_TYPE_FLOAT:
ctype = LLVMFloatTypeInContext(ctx);
break;
default:
ctype = 0;
break;
}
struct lp_build_tgsi_soa_context *bld = lp_soa_context(bld_base);
if (swizzle == ~0) {
LLVMValueRef values[TGSI_NUM_CHANNELS] = {};
unsigned chan;
for (chan = 0; chan < TGSI_NUM_CHANNELS; chan++) {
values[chan] = LLVMConstBitCast(bld->immediates[reg->Register.Index][chan], ctype);
}
return lp_build_gather_values(bld_base->base.gallivm, values,
TGSI_NUM_CHANNELS);
} else {
return LLVMConstBitCast(bld->immediates[reg->Register.Index][swizzle], ctype);
}
}
LLVMValueRef
lp_build_broadcast(struct gallivm_state *gallivm,
LLVMTypeRef vec_type,
LLVMValueRef scalar)
{
LLVMValueRef res;
if (LLVMGetTypeKind(vec_type) != LLVMVectorTypeKind) {
/* scalar */
assert(vec_type == LLVMTypeOf(scalar));
res = scalar;
} else {
LLVMBuilderRef builder = gallivm->builder;
const unsigned length = LLVMGetVectorSize(vec_type);
LLVMValueRef undef = LLVMGetUndef(vec_type);
LLVMTypeRef i32_type = LLVMInt32TypeInContext(gallivm->context);
assert(LLVMGetElementType(vec_type) == LLVMTypeOf(scalar));
if (HAVE_LLVM >= 0x207) {
/* The shuffle vector is always made of int32 elements */
LLVMTypeRef i32_vec_type = LLVMVectorType(i32_type, length);
res = LLVMBuildInsertElement(builder, undef, scalar, LLVMConstNull(i32_type), "");
res = LLVMBuildShuffleVector(builder, res, undef, LLVMConstNull(i32_vec_type), "");
} else {
/* XXX: The above path provokes a bug in LLVM 2.6 */
unsigned i;
res = undef;
for(i = 0; i < length; ++i) {
LLVMValueRef index = lp_build_const_int32(gallivm, i);
res = LLVMBuildInsertElement(builder, res, scalar, index, "");
}
}
}
return res;
}
/**
* Generate code to compute coordinate gradient (rho).
* \param ddx partial derivatives of (s, t, r, q) with respect to X
* \param ddy partial derivatives of (s, t, r, q) with respect to Y
*
* XXX: The resulting rho is scalar, so we ignore all but the first element of
* derivatives that are passed by the shader.
*/
static LLVMValueRef
lp_build_rho(struct lp_build_sample_context *bld,
unsigned unit,
const LLVMValueRef ddx[4],
const LLVMValueRef ddy[4])
{
struct lp_build_context *int_size_bld = &bld->int_size_bld;
struct lp_build_context *float_size_bld = &bld->float_size_bld;
struct lp_build_context *float_bld = &bld->float_bld;
const unsigned dims = bld->dims;
LLVMBuilderRef builder = bld->gallivm->builder;
LLVMTypeRef i32t = LLVMInt32TypeInContext(bld->gallivm->context);
LLVMValueRef index0 = LLVMConstInt(i32t, 0, 0);
LLVMValueRef index1 = LLVMConstInt(i32t, 1, 0);
LLVMValueRef index2 = LLVMConstInt(i32t, 2, 0);
LLVMValueRef dsdx, dsdy, dtdx, dtdy, drdx, drdy;
LLVMValueRef rho_x, rho_y;
LLVMValueRef rho_vec;
LLVMValueRef int_size, float_size;
LLVMValueRef rho;
LLVMValueRef first_level, first_level_vec;
dsdx = ddx[0];
dsdy = ddy[0];
if (dims <= 1) {
rho_x = dsdx;
rho_y = dsdy;
}
else {
rho_x = float_size_bld->undef;
rho_y = float_size_bld->undef;
rho_x = LLVMBuildInsertElement(builder, rho_x, dsdx, index0, "");
rho_y = LLVMBuildInsertElement(builder, rho_y, dsdy, index0, "");
dtdx = ddx[1];
dtdy = ddy[1];
rho_x = LLVMBuildInsertElement(builder, rho_x, dtdx, index1, "");
rho_y = LLVMBuildInsertElement(builder, rho_y, dtdy, index1, "");
if (dims >= 3) {
drdx = ddx[2];
drdy = ddy[2];
rho_x = LLVMBuildInsertElement(builder, rho_x, drdx, index2, "");
rho_y = LLVMBuildInsertElement(builder, rho_y, drdy, index2, "");
}
}
rho_x = lp_build_abs(float_size_bld, rho_x);
rho_y = lp_build_abs(float_size_bld, rho_y);
rho_vec = lp_build_max(float_size_bld, rho_x, rho_y);
first_level = bld->dynamic_state->first_level(bld->dynamic_state,
bld->gallivm, unit);
first_level_vec = lp_build_broadcast_scalar(&bld->int_size_bld, first_level);
int_size = lp_build_minify(int_size_bld, bld->int_size, first_level_vec);
float_size = lp_build_int_to_float(float_size_bld, int_size);
rho_vec = lp_build_mul(float_size_bld, rho_vec, float_size);
if (dims <= 1) {
rho = rho_vec;
}
else {
if (dims >= 2) {
LLVMValueRef rho_s, rho_t, rho_r;
rho_s = LLVMBuildExtractElement(builder, rho_vec, index0, "");
rho_t = LLVMBuildExtractElement(builder, rho_vec, index1, "");
rho = lp_build_max(float_bld, rho_s, rho_t);
if (dims >= 3) {
rho_r = LLVMBuildExtractElement(builder, rho_vec, index2, "");
rho = lp_build_max(float_bld, rho, rho_r);
}
}
}
return rho;
}
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);
}
/**
* Pack a single pixel.
*
* @param rgba 4 float vector with the unpacked components.
*
* XXX: This is mostly for reference and testing -- operating a single pixel at
* a time is rarely if ever needed.
*/
LLVMValueRef
lp_build_pack_rgba_aos(struct gallivm_state *gallivm,
const struct util_format_description *desc,
LLVMValueRef rgba)
{
LLVMBuilderRef builder = gallivm->builder;
LLVMTypeRef type;
LLVMValueRef packed = NULL;
LLVMValueRef swizzles[4];
LLVMValueRef shifted, casted, scaled, unswizzled;
LLVMValueRef shifts[4];
LLVMValueRef scales[4];
boolean normalized;
unsigned shift;
unsigned i, j;
assert(desc->layout == UTIL_FORMAT_LAYOUT_PLAIN);
assert(desc->block.width == 1);
assert(desc->block.height == 1);
type = LLVMIntTypeInContext(gallivm->context, desc->block.bits);
/* Unswizzle the color components into the source vector. */
for (i = 0; i < 4; ++i) {
for (j = 0; j < 4; ++j) {
if (desc->swizzle[j] == i)
break;
}
if (j < 4)
swizzles[i] = lp_build_const_int32(gallivm, j);
else
swizzles[i] = LLVMGetUndef(LLVMInt32TypeInContext(gallivm->context));
}
unswizzled = LLVMBuildShuffleVector(builder, rgba,
LLVMGetUndef(LLVMVectorType(LLVMFloatTypeInContext(gallivm->context), 4)),
LLVMConstVector(swizzles, 4), "");
normalized = FALSE;
shift = 0;
for (i = 0; i < 4; ++i) {
unsigned bits = desc->channel[i].size;
if (desc->channel[i].type == UTIL_FORMAT_TYPE_VOID) {
shifts[i] = LLVMGetUndef(LLVMInt32TypeInContext(gallivm->context));
scales[i] = LLVMGetUndef(LLVMFloatTypeInContext(gallivm->context));
}
else {
unsigned mask = (1 << bits) - 1;
assert(desc->channel[i].type == UTIL_FORMAT_TYPE_UNSIGNED);
assert(bits < 32);
shifts[i] = lp_build_const_int32(gallivm, shift);
if (desc->channel[i].normalized) {
scales[i] = lp_build_const_float(gallivm, mask);
normalized = TRUE;
}
else
scales[i] = lp_build_const_float(gallivm, 1.0);
}
shift += bits;
}
if (normalized)
scaled = LLVMBuildFMul(builder, unswizzled, LLVMConstVector(scales, 4), "");
else
scaled = unswizzled;
casted = LLVMBuildFPToSI(builder, scaled, LLVMVectorType(LLVMInt32TypeInContext(gallivm->context), 4), "");
shifted = LLVMBuildShl(builder, casted, LLVMConstVector(shifts, 4), "");
/* Bitwise or all components */
for (i = 0; i < 4; ++i) {
if (desc->channel[i].type == UTIL_FORMAT_TYPE_UNSIGNED) {
LLVMValueRef component = LLVMBuildExtractElement(builder, shifted,
lp_build_const_int32(gallivm, i), "");
if (packed)
packed = LLVMBuildOr(builder, packed, component, "");
else
packed = component;
}
}
if (!packed)
packed = LLVMGetUndef(LLVMInt32TypeInContext(gallivm->context));
if (desc->block.bits < 32)
packed = LLVMBuildTrunc(builder, packed, type, "");
return packed;
}
/**
* Unpack a single pixel into its RGBA components.
*
* @param desc the pixel format for the packed pixel value
* @param packed integer pixel in a format such as PIPE_FORMAT_B8G8R8A8_UNORM
*
* @return RGBA in a float[4] or ubyte[4] or ushort[4] vector.
*/
static INLINE LLVMValueRef
lp_build_unpack_arith_rgba_aos(struct gallivm_state *gallivm,
const struct util_format_description *desc,
LLVMValueRef packed)
{
LLVMBuilderRef builder = gallivm->builder;
LLVMValueRef shifted, casted, scaled, masked;
LLVMValueRef shifts[4];
LLVMValueRef masks[4];
LLVMValueRef scales[4];
boolean normalized;
boolean needs_uitofp;
unsigned shift;
unsigned i;
/* TODO: Support more formats */
assert(desc->layout == UTIL_FORMAT_LAYOUT_PLAIN);
assert(desc->block.width == 1);
assert(desc->block.height == 1);
assert(desc->block.bits <= 32);
/* Do the intermediate integer computations with 32bit integers since it
* matches floating point size */
assert (LLVMTypeOf(packed) == LLVMInt32TypeInContext(gallivm->context));
/* Broadcast the packed value to all four channels
* before: packed = BGRA
* after: packed = {BGRA, BGRA, BGRA, BGRA}
*/
packed = LLVMBuildInsertElement(builder,
LLVMGetUndef(LLVMVectorType(LLVMInt32TypeInContext(gallivm->context), 4)),
packed,
LLVMConstNull(LLVMInt32TypeInContext(gallivm->context)),
"");
packed = LLVMBuildShuffleVector(builder,
packed,
LLVMGetUndef(LLVMVectorType(LLVMInt32TypeInContext(gallivm->context), 4)),
LLVMConstNull(LLVMVectorType(LLVMInt32TypeInContext(gallivm->context), 4)),
"");
/* Initialize vector constants */
normalized = FALSE;
needs_uitofp = FALSE;
shift = 0;
/* Loop over 4 color components */
for (i = 0; i < 4; ++i) {
unsigned bits = desc->channel[i].size;
if (desc->channel[i].type == UTIL_FORMAT_TYPE_VOID) {
shifts[i] = LLVMGetUndef(LLVMInt32TypeInContext(gallivm->context));
masks[i] = LLVMConstNull(LLVMInt32TypeInContext(gallivm->context));
scales[i] = LLVMConstNull(LLVMFloatTypeInContext(gallivm->context));
}
else {
unsigned long long mask = (1ULL << bits) - 1;
assert(desc->channel[i].type == UTIL_FORMAT_TYPE_UNSIGNED);
if (bits == 32) {
needs_uitofp = TRUE;
}
shifts[i] = lp_build_const_int32(gallivm, shift);
masks[i] = lp_build_const_int32(gallivm, mask);
if (desc->channel[i].normalized) {
scales[i] = lp_build_const_float(gallivm, 1.0 / mask);
normalized = TRUE;
}
else
scales[i] = lp_build_const_float(gallivm, 1.0);
}
shift += bits;
}
/* Ex: convert packed = {BGRA, BGRA, BGRA, BGRA}
* into masked = {B, G, R, A}
*/
shifted = LLVMBuildLShr(builder, packed, LLVMConstVector(shifts, 4), "");
masked = LLVMBuildAnd(builder, shifted, LLVMConstVector(masks, 4), "");
if (!needs_uitofp) {
/* UIToFP can't be expressed in SSE2 */
casted = LLVMBuildSIToFP(builder, masked, LLVMVectorType(LLVMFloatTypeInContext(gallivm->context), 4), "");
} else {
casted = LLVMBuildUIToFP(builder, masked, LLVMVectorType(LLVMFloatTypeInContext(gallivm->context), 4), "");
}
/* At this point 'casted' may be a vector of floats such as
* {255.0, 255.0, 255.0, 255.0}. Next, if the pixel values are normalized
* we'll scale this to {1.0, 1.0, 1.0, 1.0}.
*/
if (normalized)
scaled = LLVMBuildFMul(builder, casted, LLVMConstVector(scales, 4), "");
else
scaled = casted;
return scaled;
}
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);
}
/**
* 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);
}
/*
* 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), "");
}
