static LLVMValueRef gen_digestof_value(compile_t* c, LLVMValueRef value)
{
LLVMTypeRef type = LLVMTypeOf(value);
switch(LLVMGetTypeKind(type))
{
case LLVMFloatTypeKind:
value = LLVMBuildBitCast(c->builder, value, c->i32, "");
return LLVMBuildZExt(c->builder, value, c->i64, "");
case LLVMDoubleTypeKind:
return LLVMBuildBitCast(c->builder, value, c->i64, "");
case LLVMIntegerTypeKind:
{
uint32_t width = LLVMGetIntTypeWidth(type);
if(width < 64)
{
value = LLVMBuildZExt(c->builder, value, c->i64, "");
} else if(width == 128) {
LLVMValueRef shift = LLVMConstInt(c->i128, 64, false);
LLVMValueRef high = LLVMBuildLShr(c->builder, value, shift, "");
high = LLVMBuildTrunc(c->builder, high, c->i64, "");
value = LLVMBuildTrunc(c->builder, value, c->i64, "");
value = LLVMBuildXor(c->builder, value, high, "");
}
return value;
}
case LLVMStructTypeKind:
{
uint32_t count = LLVMCountStructElementTypes(type);
LLVMValueRef result = LLVMConstInt(c->i64, 0, false);
for(uint32_t i = 0; i < count; i++)
{
LLVMValueRef elem = LLVMBuildExtractValue(c->builder, value, i, "");
elem = gen_digestof_value(c, elem);
result = LLVMBuildXor(c->builder, result, elem, "");
}
return result;
}
case LLVMPointerTypeKind:
return LLVMBuildPtrToInt(c->builder, value, c->i64, "");
default: {}
}
assert(0);
return NULL;
}
static void emit_pk2h(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;
struct lp_build_context *uint_bld = &bld_base->uint_bld;
LLVMTypeRef fp16, i16;
LLVMValueRef const16, comp[2];
unsigned i;
fp16 = LLVMHalfTypeInContext(context);
i16 = LLVMInt16TypeInContext(context);
const16 = lp_build_const_int32(uint_bld->gallivm, 16);
for (i = 0; i < 2; i++) {
comp[i] = LLVMBuildFPTrunc(builder, emit_data->args[i], fp16, "");
comp[i] = LLVMBuildBitCast(builder, comp[i], i16, "");
comp[i] = LLVMBuildZExt(builder, comp[i], uint_bld->elem_type, "");
}
comp[1] = LLVMBuildShl(builder, comp[1], const16, "");
comp[0] = LLVMBuildOr(builder, comp[0], comp[1], "");
emit_data->output[emit_data->chan] = comp[0];
}
/**
* Converts int16 half-float to float32
* Note this can be performed in 1 instruction if vcvtph2ps exists (f16c/cvt16)
* [llvm.x86.vcvtph2ps / _mm_cvtph_ps]
*
* @param src value to convert
*
*/
LLVMValueRef
lp_build_half_to_float(struct gallivm_state *gallivm,
LLVMValueRef src)
{
LLVMBuilderRef builder = gallivm->builder;
LLVMTypeRef src_type = LLVMTypeOf(src);
unsigned src_length = LLVMGetTypeKind(src_type) == LLVMVectorTypeKind ?
LLVMGetVectorSize(src_type) : 1;
struct lp_type f32_type = lp_type_float_vec(32, 32 * src_length);
struct lp_type i32_type = lp_type_int_vec(32, 32 * src_length);
LLVMTypeRef int_vec_type = lp_build_vec_type(gallivm, i32_type);
LLVMValueRef h;
if (util_cpu_caps.has_f16c && HAVE_LLVM >= 0x0301 &&
(src_length == 4 || src_length == 8)) {
const char *intrinsic = NULL;
if (src_length == 4) {
src = lp_build_pad_vector(gallivm, src, 8);
intrinsic = "llvm.x86.vcvtph2ps.128";
}
else {
intrinsic = "llvm.x86.vcvtph2ps.256";
}
return lp_build_intrinsic_unary(builder, intrinsic,
lp_build_vec_type(gallivm, f32_type), src);
}
/* Convert int16 vector to int32 vector by zero ext (might generate bad code) */
h = LLVMBuildZExt(builder, src, int_vec_type, "");
return lp_build_smallfloat_to_float(gallivm, f32_type, h, 10, 5, 0, true);
}
/**
* 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;
}
static LLVMValueRef
translateBinOp(SymbolTable *TyTable, SymbolTable *ValTable, ASTNode *Node) {
ASTNode *NodeE1 = (ASTNode*) ptrVectorGet(&(Node->Child), 0),
*NodeE2 = (ASTNode*) ptrVectorGet(&(Node->Child), 1);
LLVMValueRef ValueE1Ptr = translateExpr(TyTable, ValTable, NodeE1),
ValueE2Ptr = translateExpr(TyTable, ValTable, NodeE2);
LLVMValueRef ValueE1 = LLVMBuildLoad(Builder, ValueE1Ptr, "binop.ld.e1."),
ValueE2 = LLVMBuildLoad(Builder, ValueE2Ptr, "binop.ld.e2.");
LLVMValueRef ResultVal = NULL;
switch (LLVMGetTypeKind(LLVMTypeOf(ValueE1))) {
case LLVMIntegerTypeKind: ResultVal = translateIntBinOp (Node->Kind, ValueE1, ValueE2); break;
case LLVMFloatTypeKind: ResultVal = translateFloatBinOp (Node->Kind, ValueE1, ValueE2); break;
case LLVMStructTypeKind: ResultVal = translateStructBinOp(Node->Kind, ValueE1Ptr, ValueE2Ptr); break;
case LLVMPointerTypeKind: ResultVal = translateStringBinOp(Node->Kind, ValueE1, ValueE2); break;
default: return NULL;
}
switch (LLVMGetTypeKind(LLVMTypeOf(ResultVal))) {
case LLVMIntegerTypeKind: ResultVal = LLVMBuildZExt(Builder, ResultVal, LLVMInt32Type(), ""); break;
default: break;
}
return wrapValue(ResultVal);
}
LLVMValueRef gen_not(compile_t* c, ast_t* ast)
{
LLVMValueRef value = gen_expr(c, ast);
if(value == NULL)
return NULL;
ast_t* type = ast_type(ast);
if(is_bool(type))
{
if(LLVMIsAConstantInt(value))
{
if(is_always_true(value))
return LLVMConstInt(c->ibool, 0, false);
return LLVMConstInt(c->ibool, 1, false);
}
LLVMValueRef test = LLVMBuildICmp(c->builder, LLVMIntEQ, value,
LLVMConstInt(c->ibool, 0, false), "");
return LLVMBuildZExt(c->builder, test, c->ibool, "");
}
if(LLVMIsAConstantInt(value))
return LLVMConstNot(value);
return LLVMBuildNot(c->builder, value, "");
}
LLVMValueRef gendesc_fieldptr(compile_t* c, LLVMValueRef ptr,
LLVMValueRef field_info)
{
LLVMValueRef offset = LLVMBuildExtractValue(c->builder, field_info, 0, "");
offset = LLVMBuildZExt(c->builder, offset, c->intptr, "");
return LLVMBuildInBoundsGEP(c->builder, ptr, &offset, 1, "");
}
/**
* 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;
}
static void number_conversion(compile_t* c, num_conv_t* from, num_conv_t* to,
bool native128)
{
if(!native128 &&
((from->is_float && (to->size > 64)) ||
(to->is_float && (from->size > 64)))
)
{
return;
}
reach_type_t* t = reach_type_name(c->reach, from->type_name);
if(t == NULL)
return;
FIND_METHOD(to->fun_name);
start_function(c, m, to->type, &from->type, 1);
LLVMValueRef arg = LLVMGetParam(m->func, 0);
LLVMValueRef result;
if(from->is_float)
{
if(to->is_float)
{
if(from->size < to->size)
result = LLVMBuildFPExt(c->builder, arg, to->type, "");
else if(from->size > to->size)
result = LLVMBuildFPTrunc(c->builder, arg, to->type, "");
else
result = arg;
} else if(to->is_signed) {
result = LLVMBuildFPToSI(c->builder, arg, to->type, "");
} else {
result = LLVMBuildFPToUI(c->builder, arg, to->type, "");
}
} else if(to->is_float) {
if(from->is_signed)
result = LLVMBuildSIToFP(c->builder, arg, to->type, "");
else
result = LLVMBuildUIToFP(c->builder, arg, to->type, "");
} else if(from->size > to->size) {
result = LLVMBuildTrunc(c->builder, arg, to->type, "");
} else if(from->size < to->size) {
if(from->is_signed)
result = LLVMBuildSExt(c->builder, arg, to->type, "");
else
result = LLVMBuildZExt(c->builder, arg, to->type, "");
} else {
result = arg;
}
LLVMBuildRet(c->builder, result);
codegen_finishfun(c);
BOX_FUNCTION();
}
LLVMValueRef gendesc_ptr_to_fields(compile_t* c, LLVMValueRef object,
LLVMValueRef desc)
{
// Skip the descriptor.
LLVMValueRef offset = desc_field(c, desc, DESC_FIELD_OFFSET);
offset = LLVMBuildZExt(c->builder, offset, c->intptr, "");
LLVMValueRef base = LLVMBuildBitCast(c->builder, object, c->void_ptr, "");
return LLVMBuildInBoundsGEP(c->builder, base, &offset, 1, "");
}
LLVMValueRef gen_eq_rvalue(compile_t* c, ast_t* left, LLVMValueRef r_value)
{
ast_t* type = ast_type(left);
bool sign = is_signed(type);
LLVMValueRef l_value = gen_expr(c, left);
LLVMValueRef test = make_cmp_value(c, sign, l_value, r_value,
LLVMRealOEQ, LLVMIntEQ, LLVMIntEQ);
return LLVMBuildZExt(c->builder, test, c->ibool, "");
}
LLVMValueRef gen_not(struct node *ast)
{
LLVMValueRef truth;
truth = LLVMBuildICmp(builder,
LLVMIntEQ,
codegen(ast->one),
CONST(0),
"");
return LLVMBuildZExt(builder, truth, TYPE_INT, "");
}
LLVMValueRef gen_ge(struct node *ast)
{
LLVMValueRef truth;
truth = LLVMBuildICmp(builder,
LLVMIntSGE,
codegen(ast->one),
codegen(ast->two),
"");
return LLVMBuildZExt(builder, truth, TYPE_INT, "");
}
LLVMValueRef gendesc_ptr_to_fields(compile_t* c, LLVMValueRef object,
LLVMValueRef desc)
{
// Skip the descriptor.
LLVMValueRef offset = desc_field(c, desc, DESC_FIELD_OFFSET);
offset = LLVMBuildZExt(c->builder, offset, c->intptr, "");
LLVMValueRef base = LLVMBuildPtrToInt(c->builder, object, c->intptr, "");
LLVMValueRef result = LLVMBuildAdd(c->builder, base, offset, "");
// Return as a c->intptr.
return result;
}
static LLVMValueRef make_cmp(compile_t* c, ast_t* left, ast_t* right,
LLVMRealPredicate cmp_f, LLVMIntPredicate cmp_si, LLVMIntPredicate cmp_ui)
{
ast_t* type = ast_type(left);
bool sign = is_signed(type);
LLVMValueRef l_value = gen_expr(c, left);
LLVMValueRef r_value = gen_expr(c, right);
LLVMValueRef test = make_cmp_value(c, sign, l_value, r_value,
cmp_f, cmp_si, cmp_ui);
return LLVMBuildZExt(c->builder, test, c->ibool, "");
}
static void maybe_is_none(compile_t* c, reach_type_t* t)
{
// Returns true if the receiver is null.
FIND_METHOD("is_none");
start_function(c, m, c->ibool, &t->use_type, 1);
LLVMValueRef receiver = LLVMGetParam(m->func, 0);
LLVMValueRef test = LLVMBuildIsNull(c->builder, receiver, "");
LLVMValueRef value = LLVMBuildZExt(c->builder, test, c->ibool, "");
LLVMBuildRet(c->builder, value);
codegen_finishfun(c);
BOX_FUNCTION();
}
/**
* Converts int16 half-float to float32
* Note this can be performed in 1 instruction if vcvtph2ps exists (sse5 i think?)
* [llvm.x86.vcvtph2ps / _mm_cvtph_ps]
*
* @param src_type <vector> type of int16
* @param src value to convert
*
* ref http://fgiesen.wordpress.com/2012/03/28/half-to-float-done-quic/
*/
LLVMValueRef
lp_build_half_to_float(struct gallivm_state *gallivm,
struct lp_type src_type,
LLVMValueRef src)
{
struct lp_type f32_type = lp_type_float_vec(32, 32 * src_type.length);
struct lp_type i32_type = lp_type_int_vec(32, 32 * src_type.length);
LLVMBuilderRef builder = gallivm->builder;
LLVMTypeRef int_vec_type = lp_build_vec_type(gallivm, i32_type);
LLVMTypeRef float_vec_type = lp_build_vec_type(gallivm, f32_type);
/* Constants */
LLVMValueRef i32_13 = lp_build_const_int_vec(gallivm, i32_type, 13);
LLVMValueRef i32_16 = lp_build_const_int_vec(gallivm, i32_type, 16);
LLVMValueRef i32_mask_nosign = lp_build_const_int_vec(gallivm, i32_type, 0x7fff);
LLVMValueRef i32_was_infnan = lp_build_const_int_vec(gallivm, i32_type, 0x7bff);
LLVMValueRef i32_exp_infnan = lp_build_const_int_vec(gallivm, i32_type, 0xff << 23);
LLVMValueRef f32_magic = LLVMBuildBitCast(builder,
lp_build_const_int_vec(gallivm, i32_type, (254 - 15) << 23),
float_vec_type, "");
/* Convert int16 vector to int32 vector by zero ext */
LLVMValueRef h = LLVMBuildZExt(builder, src, int_vec_type, "");
/* Exponent / mantissa bits */
LLVMValueRef expmant = LLVMBuildAnd(builder, i32_mask_nosign, h, "");
LLVMValueRef shifted = LLVMBuildBitCast(builder, LLVMBuildShl(builder, expmant, i32_13, ""), float_vec_type, "");
/* Exponent adjust */
LLVMValueRef scaled = LLVMBuildBitCast(builder, LLVMBuildFMul(builder, shifted, f32_magic, ""), int_vec_type, "");
/* Make sure Inf/NaN survive */
LLVMValueRef b_wasinfnan = lp_build_compare(gallivm, i32_type, PIPE_FUNC_GREATER, expmant, i32_was_infnan);
LLVMValueRef infnanexp = LLVMBuildAnd(builder, b_wasinfnan, i32_exp_infnan, "");
/* Sign bit */
LLVMValueRef justsign = LLVMBuildXor(builder, h, expmant, "");
LLVMValueRef sign = LLVMBuildShl(builder, justsign, i32_16, "");
/* Combine result */
LLVMValueRef sign_inf = LLVMBuildOr(builder, sign, infnanexp, "");
LLVMValueRef final = LLVMBuildOr(builder, scaled, sign_inf, "");
/* Cast from int32 vector to float32 vector */
return LLVMBuildBitCast(builder, final, float_vec_type, "");
}
/**
* lp_build_assert.
*
* Build an assertion in LLVM IR by building a function call to the
* lp_assert() function above.
*
* \param condition should be an 'i1' or 'i32' value
* \param msg a string to print if the assertion fails.
*/
LLVMValueRef
lp_build_assert(LLVMBuilderRef builder, LLVMValueRef condition,
const char *msg)
{
LLVMModuleRef module;
LLVMTypeRef arg_types[2];
LLVMValueRef msg_string, assert_func, params[2], r;
module = LLVMGetGlobalParent(LLVMGetBasicBlockParent(
LLVMGetInsertBlock(builder)));
msg_string = lp_build_const_string_variable(module, msg, strlen(msg) + 1);
arg_types[0] = LLVMInt32Type();
arg_types[1] = LLVMPointerType(LLVMInt8Type(), 0);
/* lookup the lp_assert function */
assert_func = LLVMGetNamedFunction(module, "lp_assert");
/* Create the assertion function if not found */
if (!assert_func) {
LLVMTypeRef func_type =
LLVMFunctionType(LLVMVoidType(), arg_types, 2, 0);
assert_func = LLVMAddFunction(module, "lp_assert", func_type);
LLVMSetFunctionCallConv(assert_func, LLVMCCallConv);
LLVMSetLinkage(assert_func, LLVMExternalLinkage);
LLVMAddGlobalMapping(lp_build_engine, assert_func,
func_to_pointer((func_pointer)lp_assert));
}
assert(assert_func);
/* build function call param list */
params[0] = LLVMBuildZExt(builder, condition, arg_types[0], "");
params[1] = LLVMBuildBitCast(builder, msg_string, arg_types[1], "");
/* check arg types */
assert(LLVMTypeOf(params[0]) == arg_types[0]);
assert(LLVMTypeOf(params[1]) == arg_types[1]);
r = LLVMBuildCall(builder, assert_func, params, 2, "");
return r;
}
LLVMValueRef gen_is(compile_t* c, ast_t* ast)
{
AST_GET_CHILDREN(ast, left, right);
ast_t* left_type = ast_type(left);
ast_t* right_type = ast_type(right);
LLVMValueRef l_value = gen_expr(c, left);
LLVMValueRef r_value = gen_expr(c, right);
if((l_value == NULL) || (r_value == NULL))
return NULL;
codegen_debugloc(c, ast);
LLVMValueRef result = gen_is_value(c, left_type, right_type,
l_value, r_value);
LLVMValueRef value = LLVMBuildZExt(c->builder, result, c->ibool, "");
codegen_debugloc(c, NULL);
return value;
}
/**
* Gather elements from scatter positions in memory into a single vector.
*
* @param src_width src element width
* @param dst_width result element width (source will be expanded to fit)
* @param length length of the offsets,
* @param base_ptr base pointer, should be a i8 pointer type.
* @param offsets vector with offsets
*/
LLVMValueRef
lp_build_gather(LLVMBuilderRef builder,
unsigned length,
unsigned src_width,
unsigned dst_width,
LLVMValueRef base_ptr,
LLVMValueRef offsets)
{
LLVMTypeRef src_type = LLVMIntType(src_width);
LLVMTypeRef src_ptr_type = LLVMPointerType(src_type, 0);
LLVMTypeRef dst_elem_type = LLVMIntType(dst_width);
LLVMTypeRef dst_vec_type = LLVMVectorType(dst_elem_type, length);
LLVMValueRef res;
unsigned i;
res = LLVMGetUndef(dst_vec_type);
for(i = 0; i < length; ++i) {
LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0);
LLVMValueRef elem_offset;
LLVMValueRef elem_ptr;
LLVMValueRef elem;
elem_offset = LLVMBuildExtractElement(builder, offsets, index, "");
elem_ptr = LLVMBuildGEP(builder, base_ptr, &elem_offset, 1, "");
elem_ptr = LLVMBuildBitCast(builder, elem_ptr, src_ptr_type, "");
elem = LLVMBuildLoad(builder, elem_ptr, "");
assert(src_width <= dst_width);
if(src_width > dst_width)
elem = LLVMBuildTrunc(builder, elem, dst_elem_type, "");
if(src_width < dst_width)
elem = LLVMBuildZExt(builder, elem, dst_elem_type, "");
res = LLVMBuildInsertElement(builder, res, elem, index, "");
}
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 gen_main(compile_t* c, gentype_t* main_g, gentype_t* env_g)
{
LLVMTypeRef params[3];
params[0] = c->i32;
params[1] = LLVMPointerType(LLVMPointerType(c->i8, 0), 0);
params[2] = LLVMPointerType(LLVMPointerType(c->i8, 0), 0);
LLVMTypeRef ftype = LLVMFunctionType(c->i32, params, 3, false);
LLVMValueRef func = LLVMAddFunction(c->module, "main", ftype);
codegen_startfun(c, func, false);
LLVMValueRef args[3];
args[0] = LLVMGetParam(func, 0);
LLVMSetValueName(args[0], "argc");
args[1] = LLVMGetParam(func, 1);
LLVMSetValueName(args[1], "argv");
args[2] = LLVMGetParam(func, 2);
LLVMSetValueName(args[1], "envp");
// Initialise the pony runtime with argc and argv, getting a new argc.
args[0] = gencall_runtime(c, "pony_init", args, 2, "argc");
// Create the main actor and become it.
LLVMValueRef main_actor = create_main(c, main_g);
// Create an Env on the main actor's heap.
const char* env_name = "Env";
const char* env_create = genname_fun(env_name, "_create", NULL);
LLVMValueRef env_args[4];
env_args[0] = gencall_alloc(c, env_g);
env_args[1] = LLVMBuildZExt(c->builder, args[0], c->i64, "");
env_args[2] = args[1];
env_args[3] = args[2];
LLVMValueRef env = gencall_runtime(c, env_create, env_args, 4, "env");
LLVMSetInstructionCallConv(env, GEN_CALLCONV);
// Run primitive initialisers using the main actor's heap.
primitive_call(c, stringtab("_init"), env);
// Create a type for the message.
LLVMTypeRef f_params[4];
f_params[0] = c->i32;
f_params[1] = c->i32;
f_params[2] = c->void_ptr;
f_params[3] = LLVMTypeOf(env);
LLVMTypeRef msg_type = LLVMStructTypeInContext(c->context, f_params, 4,
false);
LLVMTypeRef msg_type_ptr = LLVMPointerType(msg_type, 0);
// Allocate the message, setting its size and ID.
uint32_t index = genfun_vtable_index(c, main_g, stringtab("create"), NULL);
size_t msg_size = LLVMABISizeOfType(c->target_data, msg_type);
args[0] = LLVMConstInt(c->i32, pool_index(msg_size), false);
args[1] = LLVMConstInt(c->i32, index, false);
LLVMValueRef msg = gencall_runtime(c, "pony_alloc_msg", args, 2, "");
LLVMValueRef msg_ptr = LLVMBuildBitCast(c->builder, msg, msg_type_ptr, "");
// Set the message contents.
LLVMValueRef env_ptr = LLVMBuildStructGEP(c->builder, msg_ptr, 3, "");
LLVMBuildStore(c->builder, env, env_ptr);
// Trace the message.
gencall_runtime(c, "pony_gc_send", NULL, 0, "");
const char* env_trace = genname_trace(env_name);
args[0] = LLVMBuildBitCast(c->builder, env, c->object_ptr, "");
args[1] = LLVMGetNamedFunction(c->module, env_trace);
gencall_runtime(c, "pony_traceobject", args, 2, "");
gencall_runtime(c, "pony_send_done", NULL, 0, "");
// Send the message.
args[0] = main_actor;
args[1] = msg;
gencall_runtime(c, "pony_sendv", args, 2, "");
// Start the runtime.
LLVMValueRef zero = LLVMConstInt(c->i32, 0, false);
LLVMValueRef rc = gencall_runtime(c, "pony_start", &zero, 1, "");
// Run primitive finalisers. We create a new main actor as a context to run
// the finalisers in, but we do not initialise or schedule it.
LLVMValueRef final_actor = create_main(c, main_g);
primitive_call(c, stringtab("_final"), NULL);
args[0] = final_actor;
gencall_runtime(c, "pony_destroy", args, 1, "");
// Return the runtime exit code.
LLVMBuildRet(c->builder, rc);
codegen_finishfun(c);
// External linkage for main().
LLVMSetLinkage(func, LLVMExternalLinkage);
}
/**
* Load depth/stencil values.
* The stored values are linear, swizzle them.
*
* \param type the data type of the fragment depth/stencil values
* \param format_desc description of the depth/stencil surface
* \param loop_counter the current loop iteration
* \param depth_ptr pointer to the depth/stencil values of this 4x4 block
* \param depth_stride stride of the depth/stencil buffer
* \param z_fb contains z values loaded from fb (may include padding)
* \param s_fb contains s values loaded from fb (may include padding)
*/
void
lp_build_depth_stencil_load_swizzled(struct gallivm_state *gallivm,
struct lp_type z_src_type,
const struct util_format_description *format_desc,
LLVMValueRef depth_ptr,
LLVMValueRef depth_stride,
LLVMValueRef *z_fb,
LLVMValueRef *s_fb,
LLVMValueRef loop_counter)
{
LLVMBuilderRef builder = gallivm->builder;
LLVMValueRef shuffles[LP_MAX_VECTOR_LENGTH / 4];
LLVMValueRef zs_dst1, zs_dst2;
LLVMValueRef zs_dst_ptr;
LLVMValueRef depth_offset1, depth_offset2;
LLVMTypeRef load_ptr_type;
unsigned depth_bytes = format_desc->block.bits / 8;
struct lp_type zs_type = lp_depth_type(format_desc, z_src_type.length);
struct lp_type zs_load_type = zs_type;
zs_load_type.length = zs_load_type.length / 2;
load_ptr_type = LLVMPointerType(lp_build_vec_type(gallivm, zs_load_type), 0);
if (z_src_type.length == 4) {
unsigned i;
LLVMValueRef looplsb = LLVMBuildAnd(builder, loop_counter,
lp_build_const_int32(gallivm, 1), "");
LLVMValueRef loopmsb = LLVMBuildAnd(builder, loop_counter,
lp_build_const_int32(gallivm, 2), "");
LLVMValueRef offset2 = LLVMBuildMul(builder, loopmsb,
depth_stride, "");
depth_offset1 = LLVMBuildMul(builder, looplsb,
lp_build_const_int32(gallivm, depth_bytes * 2), "");
depth_offset1 = LLVMBuildAdd(builder, depth_offset1, offset2, "");
/* just concatenate the loaded 2x2 values into 4-wide vector */
for (i = 0; i < 4; i++) {
shuffles[i] = lp_build_const_int32(gallivm, i);
}
}
else {
unsigned i;
LLVMValueRef loopx2 = LLVMBuildShl(builder, loop_counter,
lp_build_const_int32(gallivm, 1), "");
assert(z_src_type.length == 8);
depth_offset1 = LLVMBuildMul(builder, loopx2, depth_stride, "");
/*
* We load 2x4 values, and need to swizzle them (order
* 0,1,4,5,2,3,6,7) - not so hot with avx unfortunately.
*/
for (i = 0; i < 8; i++) {
shuffles[i] = lp_build_const_int32(gallivm, (i&1) + (i&2) * 2 + (i&4) / 2);
}
}
depth_offset2 = LLVMBuildAdd(builder, depth_offset1, depth_stride, "");
/* Load current z/stencil values from z/stencil buffer */
zs_dst_ptr = LLVMBuildGEP(builder, depth_ptr, &depth_offset1, 1, "");
zs_dst_ptr = LLVMBuildBitCast(builder, zs_dst_ptr, load_ptr_type, "");
zs_dst1 = LLVMBuildLoad(builder, zs_dst_ptr, "");
zs_dst_ptr = LLVMBuildGEP(builder, depth_ptr, &depth_offset2, 1, "");
zs_dst_ptr = LLVMBuildBitCast(builder, zs_dst_ptr, load_ptr_type, "");
zs_dst2 = LLVMBuildLoad(builder, zs_dst_ptr, "");
*z_fb = LLVMBuildShuffleVector(builder, zs_dst1, zs_dst2,
LLVMConstVector(shuffles, zs_type.length), "");
*s_fb = *z_fb;
if (format_desc->block.bits < z_src_type.width) {
/* Extend destination ZS values (e.g., when reading from Z16_UNORM) */
*z_fb = LLVMBuildZExt(builder, *z_fb,
lp_build_int_vec_type(gallivm, z_src_type), "");
}
else if (format_desc->block.bits > 32) {
/* rely on llvm to handle too wide vector we have here nicely */
unsigned i;
struct lp_type typex2 = zs_type;
struct lp_type s_type = zs_type;
LLVMValueRef shuffles1[LP_MAX_VECTOR_LENGTH / 4];
LLVMValueRef shuffles2[LP_MAX_VECTOR_LENGTH / 4];
LLVMValueRef tmp;
typex2.width = typex2.width / 2;
typex2.length = typex2.length * 2;
s_type.width = s_type.width / 2;
s_type.floating = 0;
tmp = LLVMBuildBitCast(builder, *z_fb,
lp_build_vec_type(gallivm, typex2), "");
for (i = 0; i < zs_type.length; i++) {
shuffles1[i] = lp_build_const_int32(gallivm, i * 2);
shuffles2[i] = lp_build_const_int32(gallivm, i * 2 + 1);
}
*z_fb = LLVMBuildShuffleVector(builder, tmp, tmp,
LLVMConstVector(shuffles1, zs_type.length), "");
*s_fb = LLVMBuildShuffleVector(builder, tmp, tmp,
LLVMConstVector(shuffles2, zs_type.length), "");
*s_fb = LLVMBuildBitCast(builder, *s_fb,
lp_build_vec_type(gallivm, s_type), "");
lp_build_name(*s_fb, "s_dst");
}
lp_build_name(*z_fb, "z_dst");
lp_build_name(*s_fb, "s_dst");
lp_build_name(*z_fb, "z_dst");
}
/*
* Create a function that deforms a tuple of type desc up to natts columns.
*/
LLVMValueRef
slot_compile_deform(LLVMJitContext *context, TupleDesc desc, int natts)
{
char *funcname;
LLVMModuleRef mod;
LLVMBuilderRef b;
LLVMTypeRef deform_sig;
LLVMValueRef v_deform_fn;
LLVMBasicBlockRef b_entry;
LLVMBasicBlockRef b_adjust_unavail_cols;
LLVMBasicBlockRef b_find_start;
LLVMBasicBlockRef b_out;
LLVMBasicBlockRef b_dead;
LLVMBasicBlockRef *attcheckattnoblocks;
LLVMBasicBlockRef *attstartblocks;
LLVMBasicBlockRef *attisnullblocks;
LLVMBasicBlockRef *attcheckalignblocks;
LLVMBasicBlockRef *attalignblocks;
LLVMBasicBlockRef *attstoreblocks;
LLVMValueRef v_offp;
LLVMValueRef v_tupdata_base;
LLVMValueRef v_tts_values;
LLVMValueRef v_tts_nulls;
LLVMValueRef v_slotoffp;
LLVMValueRef v_slowp;
LLVMValueRef v_nvalidp;
LLVMValueRef v_nvalid;
LLVMValueRef v_maxatt;
LLVMValueRef v_slot;
LLVMValueRef v_tupleheaderp;
LLVMValueRef v_tuplep;
LLVMValueRef v_infomask1;
LLVMValueRef v_infomask2;
LLVMValueRef v_bits;
LLVMValueRef v_hoff;
LLVMValueRef v_hasnulls;
/* last column (0 indexed) guaranteed to exist */
int guaranteed_column_number = -1;
/* current known alignment */
int known_alignment = 0;
/* if true, known_alignment describes definite offset of column */
bool attguaranteedalign = true;
int attnum;
mod = llvm_mutable_module(context);
funcname = llvm_expand_funcname(context, "deform");
/*
* Check which columns do have to exist, so we don't have to check the
* rows natts unnecessarily.
*/
for (attnum = 0; attnum < desc->natts; attnum++)
{
Form_pg_attribute att = TupleDescAttr(desc, attnum);
/*
* If the column is possibly missing, we can't rely on its (or
* subsequent) NOT NULL constraints to indicate minimum attributes in
* the tuple, so stop here.
*/
if (att->atthasmissing)
break;
/*
* Column is NOT NULL and there've been no preceding missing columns,
* it's guaranteed that all columns up to here exist at least in the
* NULL bitmap.
*/
if (att->attnotnull)
guaranteed_column_number = attnum;
}
/* Create the signature and function */
{
LLVMTypeRef param_types[1];
param_types[0] = l_ptr(StructTupleTableSlot);
deform_sig = LLVMFunctionType(LLVMVoidType(), param_types,
lengthof(param_types), 0);
}
v_deform_fn = LLVMAddFunction(mod, funcname, deform_sig);
LLVMSetLinkage(v_deform_fn, LLVMInternalLinkage);
LLVMSetParamAlignment(LLVMGetParam(v_deform_fn, 0), MAXIMUM_ALIGNOF);
llvm_copy_attributes(AttributeTemplate, v_deform_fn);
b_entry =
LLVMAppendBasicBlock(v_deform_fn, "entry");
b_adjust_unavail_cols =
LLVMAppendBasicBlock(v_deform_fn, "adjust_unavail_cols");
b_find_start =
LLVMAppendBasicBlock(v_deform_fn, "find_startblock");
b_out =
LLVMAppendBasicBlock(v_deform_fn, "outblock");
b_dead =
LLVMAppendBasicBlock(v_deform_fn, "deadblock");
b = LLVMCreateBuilder();
attcheckattnoblocks = palloc(sizeof(LLVMBasicBlockRef) * natts);
attstartblocks = palloc(sizeof(LLVMBasicBlockRef) * natts);
attisnullblocks = palloc(sizeof(LLVMBasicBlockRef) * natts);
attcheckalignblocks = palloc(sizeof(LLVMBasicBlockRef) * natts);
attalignblocks = palloc(sizeof(LLVMBasicBlockRef) * natts);
attstoreblocks = palloc(sizeof(LLVMBasicBlockRef) * natts);
known_alignment = 0;
LLVMPositionBuilderAtEnd(b, b_entry);
/* perform allocas first, llvm only converts those to registers */
v_offp = LLVMBuildAlloca(b, TypeSizeT, "v_offp");
v_slot = LLVMGetParam(v_deform_fn, 0);
v_tts_values =
l_load_struct_gep(b, v_slot, FIELDNO_TUPLETABLESLOT_VALUES,
"tts_values");
v_tts_nulls =
l_load_struct_gep(b, v_slot, FIELDNO_TUPLETABLESLOT_ISNULL,
"tts_ISNULL");
v_slotoffp = LLVMBuildStructGEP(b, v_slot, FIELDNO_TUPLETABLESLOT_OFF, "");
v_slowp = LLVMBuildStructGEP(b, v_slot, FIELDNO_TUPLETABLESLOT_SLOW, "");
v_nvalidp = LLVMBuildStructGEP(b, v_slot, FIELDNO_TUPLETABLESLOT_NVALID, "");
v_tupleheaderp =
l_load_struct_gep(b, v_slot, FIELDNO_TUPLETABLESLOT_TUPLE,
"tupleheader");
v_tuplep =
l_load_struct_gep(b, v_tupleheaderp, FIELDNO_HEAPTUPLEDATA_DATA,
"tuple");
v_bits =
LLVMBuildBitCast(b,
LLVMBuildStructGEP(b, v_tuplep,
FIELDNO_HEAPTUPLEHEADERDATA_BITS,
""),
l_ptr(LLVMInt8Type()),
"t_bits");
v_infomask1 =
l_load_struct_gep(b, v_tuplep,
FIELDNO_HEAPTUPLEHEADERDATA_INFOMASK,
"infomask1");
v_infomask2 =
l_load_struct_gep(b,
v_tuplep, FIELDNO_HEAPTUPLEHEADERDATA_INFOMASK2,
"infomask2");
/* t_infomask & HEAP_HASNULL */
v_hasnulls =
LLVMBuildICmp(b, LLVMIntNE,
LLVMBuildAnd(b,
l_int16_const(HEAP_HASNULL),
v_infomask1, ""),
l_int16_const(0),
"hasnulls");
/* t_infomask2 & HEAP_NATTS_MASK */
v_maxatt = LLVMBuildAnd(b,
l_int16_const(HEAP_NATTS_MASK),
v_infomask2,
"maxatt");
v_hoff =
l_load_struct_gep(b, v_tuplep,
FIELDNO_HEAPTUPLEHEADERDATA_HOFF,
"t_hoff");
v_tupdata_base =
LLVMBuildGEP(b,
LLVMBuildBitCast(b,
v_tuplep,
l_ptr(LLVMInt8Type()),
""),
&v_hoff, 1,
"v_tupdata_base");
/*
* Load tuple start offset from slot. Will be reset below in case there's
* no existing deformed columns in slot.
*/
{
LLVMValueRef v_off_start;
v_off_start = LLVMBuildLoad(b, v_slotoffp, "v_slot_off");
v_off_start = LLVMBuildZExt(b, v_off_start, TypeSizeT, "");
LLVMBuildStore(b, v_off_start, v_offp);
}
/* build the basic block for each attribute, need them as jump target */
for (attnum = 0; attnum < natts; attnum++)
{
attcheckattnoblocks[attnum] =
l_bb_append_v(v_deform_fn, "block.attr.%d.attcheckattno", attnum);
attstartblocks[attnum] =
l_bb_append_v(v_deform_fn, "block.attr.%d.start", attnum);
attisnullblocks[attnum] =
l_bb_append_v(v_deform_fn, "block.attr.%d.attisnull", attnum);
attcheckalignblocks[attnum] =
l_bb_append_v(v_deform_fn, "block.attr.%d.attcheckalign", attnum);
attalignblocks[attnum] =
l_bb_append_v(v_deform_fn, "block.attr.%d.align", attnum);
attstoreblocks[attnum] =
l_bb_append_v(v_deform_fn, "block.attr.%d.store", attnum);
}
/*
* Check if's guaranteed the all the desired attributes are available in
* tuple. If so, we can start deforming. If not, need to make sure to
* fetch the missing columns.
*/
if ((natts - 1) <= guaranteed_column_number)
{
/* just skip through unnecessary blocks */
LLVMBuildBr(b, b_adjust_unavail_cols);
LLVMPositionBuilderAtEnd(b, b_adjust_unavail_cols);
LLVMBuildBr(b, b_find_start);
}
else
{
LLVMValueRef v_params[3];
/* branch if not all columns available */
LLVMBuildCondBr(b,
LLVMBuildICmp(b, LLVMIntULT,
v_maxatt,
l_int16_const(natts),
""),
b_adjust_unavail_cols,
b_find_start);
/* if not, memset tts_isnull of relevant cols to true */
LLVMPositionBuilderAtEnd(b, b_adjust_unavail_cols);
v_params[0] = v_slot;
v_params[1] = LLVMBuildZExt(b, v_maxatt, LLVMInt32Type(), "");
v_params[2] = l_int32_const(natts);
LLVMBuildCall(b, llvm_get_decl(mod, FuncSlotGetmissingattrs),
v_params, lengthof(v_params), "");
LLVMBuildBr(b, b_find_start);
}
LLVMPositionBuilderAtEnd(b, b_find_start);
v_nvalid = LLVMBuildLoad(b, v_nvalidp, "");
/*
* Build switch to go from nvalid to the right startblock. Callers
* currently don't have the knowledge, but it'd be good for performance to
* avoid this check when it's known that the slot is empty (e.g. in scan
* nodes).
*/
if (true)
{
LLVMValueRef v_switch = LLVMBuildSwitch(b, v_nvalid,
b_dead, natts);
for (attnum = 0; attnum < natts; attnum++)
{
LLVMValueRef v_attno = l_int32_const(attnum);
LLVMAddCase(v_switch, v_attno, attcheckattnoblocks[attnum]);
}
}
else
{
/* jump from entry block to first block */
LLVMBuildBr(b, attcheckattnoblocks[0]);
}
LLVMPositionBuilderAtEnd(b, b_dead);
LLVMBuildUnreachable(b);
/*
* Iterate over each attribute that needs to be deformed, build code to
* deform it.
*/
for (attnum = 0; attnum < natts; attnum++)
{
Form_pg_attribute att = TupleDescAttr(desc, attnum);
LLVMValueRef v_incby;
int alignto;
LLVMValueRef l_attno = l_int16_const(attnum);
LLVMValueRef v_attdatap;
LLVMValueRef v_resultp;
/* build block checking whether we did all the necessary attributes */
LLVMPositionBuilderAtEnd(b, attcheckattnoblocks[attnum]);
/*
* If this is the first attribute, slot->tts_nvalid was 0. Therefore
* reset offset to 0 to, it be from a previous execution.
*/
if (attnum == 0)
{
LLVMBuildStore(b, l_sizet_const(0), v_offp);
}
/*
* Build check whether column is available (i.e. whether the tuple has
* that many columns stored). We can avoid the branch if we know
* there's a subsequent NOT NULL column.
*/
if (attnum <= guaranteed_column_number)
{
LLVMBuildBr(b, attstartblocks[attnum]);
}
else
{
LLVMValueRef v_islast;
v_islast = LLVMBuildICmp(b, LLVMIntUGE,
l_attno,
v_maxatt,
"heap_natts");
LLVMBuildCondBr(b, v_islast, b_out, attstartblocks[attnum]);
}
LLVMPositionBuilderAtEnd(b, attstartblocks[attnum]);
/*
* Check for nulls if necessary. No need to take missing attributes
* into account, because in case they're present the heaptuple's natts
* would have indicated that a slot_getmissingattrs() is needed.
*/
if (!att->attnotnull)
{
LLVMBasicBlockRef b_ifnotnull;
LLVMBasicBlockRef b_ifnull;
LLVMBasicBlockRef b_next;
LLVMValueRef v_attisnull;
LLVMValueRef v_nullbyteno;
LLVMValueRef v_nullbytemask;
LLVMValueRef v_nullbyte;
LLVMValueRef v_nullbit;
b_ifnotnull = attcheckalignblocks[attnum];
b_ifnull = attisnullblocks[attnum];
if (attnum + 1 == natts)
b_next = b_out;
else
b_next = attcheckattnoblocks[attnum + 1];
v_nullbyteno = l_int32_const(attnum >> 3);
v_nullbytemask = l_int8_const(1 << ((attnum) & 0x07));
v_nullbyte = l_load_gep1(b, v_bits, v_nullbyteno, "attnullbyte");
v_nullbit = LLVMBuildICmp(b,
LLVMIntEQ,
LLVMBuildAnd(b, v_nullbyte, v_nullbytemask, ""),
l_int8_const(0),
"attisnull");
v_attisnull = LLVMBuildAnd(b, v_hasnulls, v_nullbit, "");
LLVMBuildCondBr(b, v_attisnull, b_ifnull, b_ifnotnull);
LLVMPositionBuilderAtEnd(b, b_ifnull);
/* store null-byte */
LLVMBuildStore(b,
l_int8_const(1),
LLVMBuildGEP(b, v_tts_nulls, &l_attno, 1, ""));
/* store zero datum */
LLVMBuildStore(b,
l_sizet_const(0),
LLVMBuildGEP(b, v_tts_values, &l_attno, 1, ""));
LLVMBuildBr(b, b_next);
attguaranteedalign = false;
}
else
{
/**
* Gather elements from scatter positions in memory into a single vector.
* Use for fetching texels from a texture.
* For SSE, typical values are length=4, src_width=32, dst_width=32.
*
* When src_width < dst_width, the return value can be justified in
* one of two ways:
* "integer justification" is used when the caller treats the destination
* as a packed integer bitmask, as described by the channels' "shift" and
* "width" fields;
* "vector justification" is used when the caller casts the destination
* to a vector and needs channel X to be in vector element 0.
*
* @param length length of the offsets
* @param src_width src element width in bits
* @param dst_type result element type (src will be expanded to fit,
* but truncation is not allowed)
* (this may be a vector, must be pot sized)
* @param aligned whether the data is guaranteed to be aligned (to src_width)
* @param base_ptr base pointer, needs to be a i8 pointer type.
* @param offsets vector with offsets
* @param vector_justify select vector rather than integer justification
*/
LLVMValueRef
lp_build_gather(struct gallivm_state *gallivm,
unsigned length,
unsigned src_width,
struct lp_type dst_type,
boolean aligned,
LLVMValueRef base_ptr,
LLVMValueRef offsets,
boolean vector_justify)
{
LLVMValueRef res;
boolean need_expansion = src_width < dst_type.width * dst_type.length;
boolean vec_fetch;
struct lp_type fetch_type, fetch_dst_type;
LLVMTypeRef src_type;
assert(src_width <= dst_type.width * dst_type.length);
/*
* This is quite a mess...
* Figure out if the fetch should be done as:
* a) scalar or vector
* b) float or int
*
* As an example, for a 96bit fetch expanded into 4x32bit, it is better
* to use (3x32bit) vector type (then pad the vector). Otherwise, the
* zext will cause extra instructions.
* However, the same isn't true for 3x16bit (the codegen for that is
* completely worthless on x86 simd, and for 3x8bit is is way worse
* still, don't try that... (To get really good code out of llvm for
* these cases, the only way is to decompose the fetches manually
* into 1x32bit/1x16bit, or 1x16/1x8bit respectively, although the latter
* case requires sse41, otherwise simple scalar zext is way better.
* But probably not important enough, so don't bother.)
* Also, we try to honor the floating bit of destination (but isn't
* possible if caller asks for instance for 2x32bit dst_type with
* 48bit fetch - the idea would be to use 3x16bit fetch, pad and
* cast to 2x32f type, so the fetch is always int and on top of that
* we avoid the vec pad and use scalar zext due the above mentioned
* issue).
* Note this is optimized for x86 sse2 and up backend. Could be tweaked
* for other archs if necessary...
*/
if (((src_width % 32) == 0) && ((src_width % dst_type.width) == 0) &&
(dst_type.length > 1)) {
/* use vector fetch (if dst_type is vector) */
vec_fetch = TRUE;
if (dst_type.floating) {
fetch_type = lp_type_float_vec(dst_type.width, src_width);
} else {
fetch_type = lp_type_int_vec(dst_type.width, src_width);
}
/* intentionally not using lp_build_vec_type here */
src_type = LLVMVectorType(lp_build_elem_type(gallivm, fetch_type),
fetch_type.length);
fetch_dst_type = fetch_type;
fetch_dst_type.length = dst_type.length;
} else {
/* use scalar fetch */
vec_fetch = FALSE;
if (dst_type.floating && ((src_width == 32) || (src_width == 64))) {
fetch_type = lp_type_float(src_width);
} else {
fetch_type = lp_type_int(src_width);
}
src_type = lp_build_vec_type(gallivm, fetch_type);
fetch_dst_type = fetch_type;
fetch_dst_type.width = dst_type.width * dst_type.length;
}
if (length == 1) {
/* Scalar */
res = lp_build_gather_elem_vec(gallivm, length,
src_width, src_type, fetch_dst_type,
aligned, base_ptr, offsets, 0,
vector_justify);
return LLVMBuildBitCast(gallivm->builder, res,
lp_build_vec_type(gallivm, dst_type), "");
/*
* Excluding expansion from these paths because if you need it for
* 32bit/64bit fetches you're doing it wrong (this is gather, not
* conversion) and it would be awkward for floats.
*/
} else if (util_cpu_caps.has_avx2 && !need_expansion &&
src_width == 32 && (length == 4 || length == 8)) {
return lp_build_gather_avx2(gallivm, length, src_width, dst_type,
base_ptr, offsets);
/*
* This looks bad on paper wrt throughtput/latency on Haswell.
* Even on Broadwell it doesn't look stellar.
* Albeit no measurements were done (but tested to work).
* Should definitely enable on Skylake.
* (In general, should be more of a win if the fetch is 256bit wide -
* this is true for the 32bit case above too.)
*/
} else if (0 && util_cpu_caps.has_avx2 && !need_expansion &&
src_width == 64 && (length == 2 || length == 4)) {
return lp_build_gather_avx2(gallivm, length, src_width, dst_type,
base_ptr, offsets);
} else {
/* Vector */
LLVMValueRef elems[LP_MAX_VECTOR_WIDTH / 8];
unsigned i;
boolean vec_zext = FALSE;
struct lp_type res_type, gather_res_type;
LLVMTypeRef res_t, gather_res_t;
res_type = fetch_dst_type;
res_type.length *= length;
gather_res_type = res_type;
if (src_width == 16 && dst_type.width == 32 && dst_type.length == 1) {
/*
* Note that llvm is never able to optimize zext/insert combos
* directly (i.e. zero the simd reg, then place the elements into
* the appropriate place directly). (I think this has to do with
* scalar/vector transition.) And scalar 16->32bit zext simd loads
* aren't possible (instead loading to scalar reg first).
* No idea about other archs...
* We could do this manually, but instead we just use a vector
* zext, which is simple enough (and, in fact, llvm might optimize
* this away).
* (We're not trying that with other bit widths as that might not be
* easier, in particular with 8 bit values at least with only sse2.)
*/
assert(vec_fetch == FALSE);
gather_res_type.width /= 2;
fetch_dst_type = fetch_type;
src_type = lp_build_vec_type(gallivm, fetch_type);
vec_zext = TRUE;
}
res_t = lp_build_vec_type(gallivm, res_type);
gather_res_t = lp_build_vec_type(gallivm, gather_res_type);
res = LLVMGetUndef(gather_res_t);
for (i = 0; i < length; ++i) {
LLVMValueRef index = lp_build_const_int32(gallivm, i);
elems[i] = lp_build_gather_elem_vec(gallivm, length,
src_width, src_type, fetch_dst_type,
aligned, base_ptr, offsets, i,
vector_justify);
if (!vec_fetch) {
res = LLVMBuildInsertElement(gallivm->builder, res, elems[i], index, "");
}
}
if (vec_zext) {
res = LLVMBuildZExt(gallivm->builder, res, res_t, "");
if (vector_justify) {
#ifdef PIPE_ARCH_BIG_ENDIAN
unsigned sv = dst_type.width - src_width;
res = LLVMBuildShl(gallivm->builder, res,
lp_build_const_int_vec(gallivm, res_type, sv), "");
#endif
}
}
if (vec_fetch) {
/*
* Do bitcast now otherwise llvm might get some funny ideas wrt
* float/int types...
*/
for (i = 0; i < length; i++) {
elems[i] = LLVMBuildBitCast(gallivm->builder, elems[i],
lp_build_vec_type(gallivm, dst_type), "");
}
res = lp_build_concat(gallivm, elems, dst_type, length);
} else {
struct lp_type really_final_type = dst_type;
assert(res_type.length * res_type.width ==
dst_type.length * dst_type.width * length);
really_final_type.length *= length;
res = LLVMBuildBitCast(gallivm->builder, res,
lp_build_vec_type(gallivm, really_final_type), "");
}
}
return res;
}
static LLVMValueRef box_is_box(compile_t* c, ast_t* left_type,
LLVMValueRef l_value, LLVMValueRef r_value, int possible_boxes)
{
pony_assert(LLVMGetTypeKind(LLVMTypeOf(l_value)) == LLVMPointerTypeKind);
pony_assert(LLVMGetTypeKind(LLVMTypeOf(r_value)) == LLVMPointerTypeKind);
LLVMBasicBlockRef this_block = LLVMGetInsertBlock(c->builder);
LLVMBasicBlockRef checkbox_block = codegen_block(c, "is_checkbox");
LLVMBasicBlockRef box_block = codegen_block(c, "is_box");
LLVMBasicBlockRef num_block = NULL;
if((possible_boxes & BOXED_SUBTYPES_NUMERIC) != 0)
num_block = codegen_block(c, "is_num");
LLVMBasicBlockRef tuple_block = NULL;
if((possible_boxes & BOXED_SUBTYPES_TUPLE) != 0)
tuple_block = codegen_block(c, "is_tuple");
LLVMBasicBlockRef post_block = codegen_block(c, "is_post");
LLVMValueRef eq_addr = LLVMBuildICmp(c->builder, LLVMIntEQ, l_value, r_value,
"");
LLVMBuildCondBr(c->builder, eq_addr, post_block, checkbox_block);
// Check whether we have two boxed objects of the same type.
LLVMPositionBuilderAtEnd(c->builder, checkbox_block);
LLVMValueRef l_desc = gendesc_fetch(c, l_value);
LLVMValueRef r_desc = gendesc_fetch(c, r_value);
LLVMValueRef same_type = LLVMBuildICmp(c->builder, LLVMIntEQ, l_desc, r_desc,
"");
LLVMValueRef l_typeid = NULL;
if((possible_boxes & BOXED_SUBTYPES_UNBOXED) != 0)
{
l_typeid = gendesc_typeid(c, l_value);
LLVMValueRef boxed_mask = LLVMConstInt(c->i32, 1, false);
LLVMValueRef left_boxed = LLVMBuildAnd(c->builder, l_typeid, boxed_mask,
"");
LLVMValueRef zero = LLVMConstInt(c->i32, 0, false);
left_boxed = LLVMBuildICmp(c->builder, LLVMIntEQ, left_boxed, zero, "");
LLVMValueRef both_boxed = LLVMBuildAnd(c->builder, same_type, left_boxed,
"");
LLVMBuildCondBr(c->builder, both_boxed, box_block, post_block);
} else {
LLVMBuildCondBr(c->builder, same_type, box_block, post_block);
}
// Check whether it's a numeric primitive or a tuple.
LLVMPositionBuilderAtEnd(c->builder, box_block);
if((possible_boxes & BOXED_SUBTYPES_BOXED) == BOXED_SUBTYPES_BOXED)
{
if(l_typeid == NULL)
l_typeid = gendesc_typeid(c, l_value);
LLVMValueRef num_mask = LLVMConstInt(c->i32, 2, false);
LLVMValueRef boxed_num = LLVMBuildAnd(c->builder, l_typeid, num_mask, "");
LLVMValueRef zero = LLVMConstInt(c->i32, 0, false);
boxed_num = LLVMBuildICmp(c->builder, LLVMIntEQ, boxed_num, zero, "");
LLVMBuildCondBr(c->builder, boxed_num, num_block, tuple_block);
} else if((possible_boxes & BOXED_SUBTYPES_NUMERIC) != 0) {
LLVMBuildBr(c->builder, num_block);
} else {
pony_assert((possible_boxes & BOXED_SUBTYPES_TUPLE) != 0);
LLVMBuildBr(c->builder, tuple_block);
}
LLVMValueRef args[3];
LLVMValueRef is_num = NULL;
if(num_block != NULL)
{
// Get the machine word size and memcmp without unboxing.
LLVMPositionBuilderAtEnd(c->builder, num_block);
if(l_typeid == NULL)
l_typeid = gendesc_typeid(c, l_value);
LLVMValueRef num_sizes = LLVMBuildBitCast(c->builder, c->numeric_sizes,
c->void_ptr, "");
args[0] = LLVMBuildZExt(c->builder, l_typeid, c->intptr, "");
LLVMValueRef size = LLVMBuildInBoundsGEP(c->builder, num_sizes, args, 1,
"");
size = LLVMBuildBitCast(c->builder, size, LLVMPointerType(c->i32, 0), "");
size = LLVMBuildLoad(c->builder, size, "");
LLVMSetAlignment(size, 4);
LLVMValueRef one = LLVMConstInt(c->i32, 1, false);
args[0] = LLVMBuildInBoundsGEP(c->builder, l_value, &one, 1, "");
args[0] = LLVMBuildBitCast(c->builder, args[0], c->void_ptr, "");
args[1] = LLVMBuildInBoundsGEP(c->builder, r_value, &one, 1, "");
args[1] = LLVMBuildBitCast(c->builder, args[1], c->void_ptr, "");
args[2] = LLVMBuildZExt(c->builder, size, c->intptr, "");
is_num = gencall_runtime(c, "memcmp", args, 3, "");
is_num = LLVMBuildICmp(c->builder, LLVMIntEQ, is_num,
LLVMConstInt(c->i32, 0, false), "");
LLVMBuildBr(c->builder, post_block);
}
LLVMValueRef is_tuple = NULL;
if(tuple_block != NULL)
{
// Call the type-specific __is function, which will unbox the tuples.
LLVMPositionBuilderAtEnd(c->builder, tuple_block);
reach_type_t* r_left = reach_type(c->reach, left_type);
reach_method_t* is_fn = reach_method(r_left, TK_BOX, stringtab("__is"),
NULL);
pony_assert(is_fn != NULL);
LLVMValueRef func = gendesc_vtable(c, l_value, is_fn->vtable_index);
LLVMTypeRef params[2];
params[0] = c->object_ptr;
params[1] = c->object_ptr;
LLVMTypeRef type = LLVMFunctionType(c->i1, params, 2, false);
func = LLVMBuildBitCast(c->builder, func, LLVMPointerType(type, 0), "");
args[0] = l_value;
args[1] = r_value;
is_tuple = codegen_call(c, func, args, 2);
LLVMBuildBr(c->builder, post_block);
}
LLVMPositionBuilderAtEnd(c->builder, post_block);
LLVMValueRef phi = LLVMBuildPhi(c->builder, c->i1, "");
LLVMValueRef one = LLVMConstInt(c->i1, 1, false);
LLVMValueRef zero = LLVMConstInt(c->i1, 0, false);
LLVMAddIncoming(phi, &one, &this_block, 1);
if(is_num != NULL)
LLVMAddIncoming(phi, &is_num, &num_block, 1);
if(is_tuple != NULL)
LLVMAddIncoming(phi, &is_tuple, &tuple_block, 1);
LLVMAddIncoming(phi, &zero, &checkbox_block, 1);
return phi;
}
static LLVMValueRef gen_digestof_value(compile_t* c, ast_t* type,
LLVMValueRef value)
{
LLVMTypeRef impl_type = LLVMTypeOf(value);
switch(LLVMGetTypeKind(impl_type))
{
case LLVMFloatTypeKind:
value = LLVMBuildBitCast(c->builder, value, c->i32, "");
return LLVMBuildZExt(c->builder, value, c->intptr, "");
case LLVMDoubleTypeKind:
value = LLVMBuildBitCast(c->builder, value, c->i64, "");
return gen_digestof_int64(c, value);
case LLVMIntegerTypeKind:
{
uint32_t width = LLVMGetIntTypeWidth(impl_type);
if(width < 64)
{
return LLVMBuildZExt(c->builder, value, c->intptr, "");
} else if(width == 64) {
return gen_digestof_int64(c, value);
} else if(width == 128) {
LLVMValueRef shift = LLVMConstInt(c->i128, 64, false);
LLVMValueRef high = LLVMBuildLShr(c->builder, value, shift, "");
high = LLVMBuildTrunc(c->builder, high, c->i64, "");
value = LLVMBuildTrunc(c->builder, value, c->i64, "");
high = gen_digestof_int64(c, high);
value = gen_digestof_int64(c, value);
return LLVMBuildXor(c->builder, value, high, "");
}
break;
}
case LLVMStructTypeKind:
{
uint32_t count = LLVMCountStructElementTypes(impl_type);
LLVMValueRef result = LLVMConstInt(c->intptr, 0, false);
ast_t* child = ast_child(type);
for(uint32_t i = 0; i < count; i++)
{
LLVMValueRef elem = LLVMBuildExtractValue(c->builder, value, i, "");
elem = gen_digestof_value(c, child, elem);
result = LLVMBuildXor(c->builder, result, elem, "");
child = ast_sibling(child);
}
pony_assert(child == NULL);
return result;
}
case LLVMPointerTypeKind:
if(!is_known(type))
{
reach_type_t* t = reach_type(c->reach, type);
int sub_kind = subtype_kind(t);
if((sub_kind & SUBTYPE_KIND_BOXED) != 0)
return gen_digestof_box(c, t, value, sub_kind);
}
return LLVMBuildPtrToInt(c->builder, value, c->intptr, "");
default: {}
}
pony_assert(0);
return NULL;
}
/**
* 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;
}
/**
* Truncate or expand the bitwidth.
*
* NOTE: Getting the right sign flags is crucial here, as we employ some
* intrinsics that do saturation.
*/
void
lp_build_resize(struct gallivm_state *gallivm,
struct lp_type src_type,
struct lp_type dst_type,
const LLVMValueRef *src, unsigned num_srcs,
LLVMValueRef *dst, unsigned num_dsts)
{
LLVMBuilderRef builder = gallivm->builder;
LLVMValueRef tmp[LP_MAX_VECTOR_LENGTH];
unsigned i;
/*
* We don't support float <-> int conversion here. That must be done
* before/after calling this function.
*/
assert(src_type.floating == dst_type.floating);
/*
* We don't support double <-> float conversion yet, although it could be
* added with little effort.
*/
assert((!src_type.floating && !dst_type.floating) ||
src_type.width == dst_type.width);
/* We must not loose or gain channels. Only precision */
assert(src_type.length * num_srcs == dst_type.length * num_dsts);
/* We don't support M:N conversion, only 1:N, M:1, or 1:1 */
assert(num_srcs == 1 || num_dsts == 1);
assert(src_type.length <= LP_MAX_VECTOR_LENGTH);
assert(dst_type.length <= LP_MAX_VECTOR_LENGTH);
assert(num_srcs <= LP_MAX_VECTOR_LENGTH);
assert(num_dsts <= LP_MAX_VECTOR_LENGTH);
if (src_type.width > dst_type.width) {
/*
* Truncate bit width.
*/
assert(num_dsts == 1);
if (src_type.width * src_type.length == dst_type.width * dst_type.length) {
/*
* Register width remains constant -- use vector packing intrinsics
*/
tmp[0] = lp_build_pack(gallivm, src_type, dst_type, TRUE, src, num_srcs);
}
else {
if (src_type.width / dst_type.width > num_srcs) {
/*
* First change src vectors size (with shuffle) so they have the
* same size as the destination vector, then pack normally.
* Note: cannot use cast/extract because llvm generates atrocious code.
*/
unsigned size_ratio = (src_type.width * src_type.length) /
(dst_type.length * dst_type.width);
unsigned new_length = src_type.length / size_ratio;
for (i = 0; i < size_ratio * num_srcs; i++) {
unsigned start_index = (i % size_ratio) * new_length;
tmp[i] = lp_build_extract_range(gallivm, src[i / size_ratio],
start_index, new_length);
}
num_srcs *= size_ratio;
src_type.length = new_length;
tmp[0] = lp_build_pack(gallivm, src_type, dst_type, TRUE, tmp, num_srcs);
}
else {
/*
* Truncate bit width but expand vector size - first pack
* then expand simply because this should be more AVX-friendly
* for the cases we probably hit.
*/
unsigned size_ratio = (dst_type.width * dst_type.length) /
(src_type.length * src_type.width);
unsigned num_pack_srcs = num_srcs / size_ratio;
dst_type.length = dst_type.length / size_ratio;
for (i = 0; i < size_ratio; i++) {
tmp[i] = lp_build_pack(gallivm, src_type, dst_type, TRUE,
&src[i*num_pack_srcs], num_pack_srcs);
}
tmp[0] = lp_build_concat(gallivm, tmp, dst_type, size_ratio);
}
}
}
else if (src_type.width < dst_type.width) {
/*
* Expand bit width.
*/
assert(num_srcs == 1);
if (src_type.width * src_type.length == dst_type.width * dst_type.length) {
/*
* Register width remains constant -- use vector unpack intrinsics
*/
lp_build_unpack(gallivm, src_type, dst_type, src[0], tmp, num_dsts);
}
else {
/*
* Do it element-wise.
*/
assert(src_type.length * num_srcs == dst_type.length * num_dsts);
for (i = 0; i < num_dsts; i++) {
tmp[i] = lp_build_undef(gallivm, dst_type);
}
for (i = 0; i < src_type.length; ++i) {
unsigned j = i / dst_type.length;
LLVMValueRef srcindex = lp_build_const_int32(gallivm, i);
LLVMValueRef dstindex = lp_build_const_int32(gallivm, i % dst_type.length);
LLVMValueRef val = LLVMBuildExtractElement(builder, src[0], srcindex, "");
if (src_type.sign && dst_type.sign) {
val = LLVMBuildSExt(builder, val, lp_build_elem_type(gallivm, dst_type), "");
} else {
val = LLVMBuildZExt(builder, val, lp_build_elem_type(gallivm, dst_type), "");
}
tmp[j] = LLVMBuildInsertElement(builder, tmp[j], val, dstindex, "");
}
}
}
else {
/*
* No-op
*/
assert(num_srcs == 1);
assert(num_dsts == 1);
tmp[0] = src[0];
}
for(i = 0; i < num_dsts; ++i)
dst[i] = tmp[i];
}
struct cl2llvm_val_t *llvm_type_cast(struct cl2llvm_val_t * original_val,
struct cl2llvmTypeWrap *totype_w_sign)
{
struct cl2llvm_val_t *llvm_val = cl2llvm_val_create();
int i;
struct cl2llvmTypeWrap *elem_type;
struct cl2llvm_val_t *cast_original_val;
LLVMValueRef index;
LLVMValueRef vector_addr;
LLVMValueRef vector;
LLVMValueRef const_elems[16];
LLVMTypeRef fromtype = cl2llvmTypeWrapGetLlvmType(original_val->type);
LLVMTypeRef totype = cl2llvmTypeWrapGetLlvmType(totype_w_sign);
int fromsign = cl2llvmTypeWrapGetSign(original_val->type);
int tosign = cl2llvmTypeWrapGetSign(totype_w_sign);
/*By default the return value is the same as the original_val*/
llvm_val->val = original_val->val;
cl2llvmTypeWrapSetLlvmType(llvm_val->type, cl2llvmTypeWrapGetLlvmType(original_val->type));
cl2llvmTypeWrapSetSign(llvm_val->type, cl2llvmTypeWrapGetSign(original_val->type));
snprintf(temp_var_name, sizeof temp_var_name,
"tmp_%d", temp_var_count++);
/* Check that fromtype is not a vector, unless both types are identical. */
if (LLVMGetTypeKind(fromtype) == LLVMVectorTypeKind)
{
if ((LLVMGetVectorSize(fromtype) != LLVMGetVectorSize(totype)
|| LLVMGetElementType(fromtype)
!= LLVMGetElementType(totype))
|| fromsign != tosign)
{
if (LLVMGetTypeKind(totype) == LLVMVectorTypeKind)
cl2llvm_yyerror("Casts between vector types are forbidden");
cl2llvm_yyerror("A vector may not be cast to any other type.");
}
}
/* If totype is a vector, create a vector whose components are equal to
original_val */
if (LLVMGetTypeKind(totype) == LLVMVectorTypeKind
&& LLVMGetTypeKind(fromtype) != LLVMVectorTypeKind)
{
/*Go to entry block and declare vector*/
LLVMPositionBuilder(cl2llvm_builder, cl2llvm_current_function->entry_block,
cl2llvm_current_function->branch_instr);
snprintf(temp_var_name, sizeof temp_var_name,
"tmp_%d", temp_var_count++);
vector_addr = LLVMBuildAlloca(cl2llvm_builder,
totype, temp_var_name);
LLVMPositionBuilderAtEnd(cl2llvm_builder, current_basic_block);
/* Load vector */
snprintf(temp_var_name, sizeof temp_var_name,
"tmp_%d", temp_var_count++);
vector = LLVMBuildLoad(cl2llvm_builder, vector_addr, temp_var_name);
/* Create object to represent element type of totype */
elem_type = cl2llvmTypeWrapCreate(LLVMGetElementType(totype), tosign);
/* If original_val is constant create a constant vector */
if (LLVMIsConstant(original_val->val))
{
cast_original_val = llvm_type_cast(original_val, elem_type);
for (i = 0; i < LLVMGetVectorSize(totype); i++)
const_elems[i] = cast_original_val->val;
vector = LLVMConstVector(const_elems,
LLVMGetVectorSize(totype));
llvm_val->val = vector;
cl2llvm_val_free(cast_original_val);
}
/* If original value is not constant insert elements */
else
{
for (i = 0; i < LLVMGetVectorSize(totype); i++)
{
index = LLVMConstInt(LLVMInt32Type(), i, 0);
cast_original_val = llvm_type_cast(original_val, elem_type);
snprintf(temp_var_name, sizeof temp_var_name,
"tmp_%d", temp_var_count++);
vector = LLVMBuildInsertElement(cl2llvm_builder,
vector, cast_original_val->val, index, temp_var_name);
cl2llvm_val_free(cast_original_val);
}
}
cl2llvmTypeWrapFree(elem_type);
llvm_val->val = vector;
}
if (fromtype == LLVMInt64Type())
{
if (totype == LLVMDoubleType())
{
if (fromsign)
{
llvm_val->val =
LLVMBuildSIToFP(cl2llvm_builder,
original_val->val, LLVMDoubleType(),
temp_var_name);
}
else
{
llvm_val->val =
LLVMBuildUIToFP(cl2llvm_builder,
original_val->val, LLVMDoubleType(),
temp_var_name);
}
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
}
else if (totype == LLVMFloatType())
{
if (fromsign)
{
llvm_val->val =
LLVMBuildSIToFP(cl2llvm_builder,
original_val->val, LLVMFloatType(),
temp_var_name);
}
else
{
llvm_val->val =
LLVMBuildUIToFP(cl2llvm_builder,
original_val->val, LLVMFloatType(),
temp_var_name);
}
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
}
else if (totype == LLVMHalfType())
{
if (fromsign)
{
llvm_val->val =
LLVMBuildSIToFP(cl2llvm_builder,
original_val->val, LLVMHalfType(),
temp_var_name);
}
else
{
llvm_val->val =
LLVMBuildUIToFP(cl2llvm_builder,
original_val->val, LLVMHalfType(),
temp_var_name);
}
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
}
else if (totype == LLVMInt64Type())
{
if (tosign)
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
else
cl2llvmTypeWrapSetSign(llvm_val->type, 0);
temp_var_count--;
}
else if (totype == LLVMInt32Type())
{
llvm_val->val = LLVMBuildTrunc(cl2llvm_builder,
original_val->val, LLVMInt32Type(), temp_var_name);
if(tosign)
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
else
cl2llvmTypeWrapSetSign(llvm_val->type, 0);
}
else if (totype == LLVMInt16Type())
{
llvm_val->val = LLVMBuildTrunc(cl2llvm_builder,
original_val->val, LLVMInt16Type(), temp_var_name);
if(tosign)
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
else
cl2llvmTypeWrapSetSign(llvm_val->type, 0);
}
else if (totype == LLVMInt8Type())
{
llvm_val->val = LLVMBuildTrunc(cl2llvm_builder,
original_val->val, LLVMInt8Type(), temp_var_name);
if(tosign)
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
else
cl2llvmTypeWrapSetSign(llvm_val->type, 0);
}
else if (totype == LLVMInt1Type())
{
llvm_val->val = LLVMBuildTrunc(cl2llvm_builder,
original_val->val, LLVMInt1Type(), temp_var_name);
if(tosign)
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
else
cl2llvmTypeWrapSetSign(llvm_val->type, 0);
}
}
else if (fromtype == LLVMInt32Type())
{
if (totype == LLVMDoubleType())
{
if (fromsign)
{
llvm_val->val =
LLVMBuildSIToFP(cl2llvm_builder,
original_val->val, LLVMDoubleType(),
temp_var_name);
}
else
{
llvm_val->val =
LLVMBuildUIToFP(cl2llvm_builder,
original_val->val, LLVMDoubleType(),
temp_var_name);
}
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
}
else if (totype == LLVMFloatType())
{
if (fromsign)
{
llvm_val->val =
LLVMBuildSIToFP(cl2llvm_builder,
original_val->val, LLVMFloatType(),
temp_var_name);
}
else
{
llvm_val->val =
LLVMBuildUIToFP(cl2llvm_builder,
original_val->val, LLVMFloatType(),
temp_var_name);
}
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
}
else if (totype == LLVMHalfType())
{
if (fromsign)
{
llvm_val->val =
LLVMBuildSIToFP(cl2llvm_builder,
original_val->val, LLVMHalfType(),
temp_var_name);
}
else
{
llvm_val->val =
LLVMBuildUIToFP(cl2llvm_builder,
original_val->val, LLVMHalfType(),
temp_var_name);
}
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
}
else if (totype == LLVMInt64Type())
{
if (fromsign)
{
llvm_val->val = LLVMBuildSExt(cl2llvm_builder,
original_val->val, LLVMInt64Type(),
temp_var_name);
}
else
{
llvm_val->val = LLVMBuildZExt(cl2llvm_builder,
original_val->val, LLVMInt64Type(),
temp_var_name);
}
if (tosign)
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
else
cl2llvmTypeWrapSetSign(llvm_val->type, 0);
}
else if (totype == LLVMInt32Type())
{
if(tosign)
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
else
cl2llvmTypeWrapSetSign(llvm_val->type, 0);
temp_var_count--;
}
else if (totype == LLVMInt16Type())
{
llvm_val->val = LLVMBuildTrunc(cl2llvm_builder,
original_val->val, LLVMInt16Type(), temp_var_name);
if(tosign)
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
else
cl2llvmTypeWrapSetSign(llvm_val->type, 0);
}
else if (totype == LLVMInt8Type())
{
llvm_val->val = LLVMBuildTrunc(cl2llvm_builder,
original_val->val, LLVMInt8Type(), temp_var_name);
if(tosign)
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
else
cl2llvmTypeWrapSetSign(llvm_val->type, 0);
}
else if (totype == LLVMInt1Type())
{
llvm_val->val = LLVMBuildTrunc(cl2llvm_builder,
original_val->val, LLVMInt1Type(), temp_var_name);
if(tosign)
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
else
cl2llvmTypeWrapSetSign(llvm_val->type, 0);
}
}
else if (fromtype == LLVMInt16Type())
{
if (totype == LLVMDoubleType())
{
if (fromsign)
{
llvm_val->val =
LLVMBuildSIToFP(cl2llvm_builder,
original_val->val, LLVMDoubleType(),
temp_var_name);
}
else
{
llvm_val->val =
LLVMBuildUIToFP(cl2llvm_builder,
original_val->val, LLVMDoubleType(),
temp_var_name);
}
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
}
else if (totype == LLVMFloatType())
{
if (fromsign)
{
llvm_val->val =
LLVMBuildSIToFP(cl2llvm_builder,
original_val->val, LLVMFloatType(),
temp_var_name);
}
else
{
llvm_val->val =
LLVMBuildUIToFP(cl2llvm_builder,
original_val->val, LLVMFloatType(),
temp_var_name);
}
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
}
else if (totype == LLVMHalfType())
{
if (fromsign)
{
llvm_val->val =
LLVMBuildSIToFP(cl2llvm_builder,
original_val->val, LLVMHalfType(),
temp_var_name);
}
else
{
llvm_val->val =
LLVMBuildUIToFP(cl2llvm_builder,
original_val->val, LLVMHalfType(),
temp_var_name);
}
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
}
else if (totype == LLVMInt64Type())
{
if (fromsign)
{
llvm_val->val = LLVMBuildSExt(cl2llvm_builder,
original_val->val, LLVMInt64Type(),
temp_var_name);
}
else
{
llvm_val->val = LLVMBuildZExt(cl2llvm_builder,
original_val->val, LLVMInt64Type(),
temp_var_name);
}
if (tosign)
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
else
cl2llvmTypeWrapSetSign(llvm_val->type, 0);
}
else if (totype == LLVMInt32Type())
{
if (fromsign)
{
llvm_val->val = LLVMBuildSExt(cl2llvm_builder,
original_val->val, LLVMInt32Type(),
temp_var_name);
}
else
{
llvm_val->val = LLVMBuildZExt(cl2llvm_builder,
original_val->val, LLVMInt32Type(),
temp_var_name);
}
if(tosign)
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
else
cl2llvmTypeWrapSetSign(llvm_val->type, 0);
}
else if (totype == LLVMInt16Type())
{
if(tosign)
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
else
cl2llvmTypeWrapSetSign(llvm_val->type, 0);
temp_var_count--;
}
else if (totype == LLVMInt8Type())
{
llvm_val->val = LLVMBuildTrunc(cl2llvm_builder,
original_val->val, LLVMInt8Type(), temp_var_name);
if(tosign)
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
else
cl2llvmTypeWrapSetSign(llvm_val->type, 0);
}
else if (totype == LLVMInt1Type())
{
llvm_val->val = LLVMBuildTrunc(cl2llvm_builder,
original_val->val, LLVMInt1Type(), temp_var_name);
if(tosign)
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
else
cl2llvmTypeWrapSetSign(llvm_val->type, 0);
}
}
else if (fromtype == LLVMInt8Type())
{
if (totype == LLVMDoubleType())
{
if (fromsign)
{
llvm_val->val =
LLVMBuildSIToFP(cl2llvm_builder,
original_val->val, LLVMDoubleType(),
temp_var_name);
}
else
{
llvm_val->val =
LLVMBuildUIToFP(cl2llvm_builder,
original_val->val, LLVMDoubleType(),
temp_var_name);
}
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
}
else if (totype == LLVMFloatType())
{
if (fromsign)
{
llvm_val->val =
LLVMBuildSIToFP(cl2llvm_builder,
original_val->val, LLVMFloatType(),
temp_var_name);
}
else
{
llvm_val->val =
LLVMBuildUIToFP(cl2llvm_builder,
original_val->val, LLVMFloatType(),
temp_var_name);
}
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
}
else if (totype == LLVMHalfType())
{
if (fromsign)
{
llvm_val->val =
LLVMBuildSIToFP(cl2llvm_builder,
original_val->val, LLVMHalfType(),
temp_var_name);
}
else
{
llvm_val->val =
LLVMBuildUIToFP(cl2llvm_builder,
original_val->val, LLVMHalfType(),
temp_var_name);
}
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
}
else if (totype == LLVMInt64Type())
{
if (fromsign)
{
llvm_val->val = LLVMBuildSExt(cl2llvm_builder,
original_val->val, LLVMInt64Type(),
temp_var_name);
}
else
{
llvm_val->val = LLVMBuildZExt(cl2llvm_builder,
original_val->val, LLVMInt64Type(),
temp_var_name);
}
if (tosign)
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
else
cl2llvmTypeWrapSetSign(llvm_val->type, 0);
}
else if (totype == LLVMInt32Type())
{
if (fromsign)
{
llvm_val->val = LLVMBuildSExt(cl2llvm_builder,
original_val->val, LLVMInt32Type(),
temp_var_name);
}
else
{
llvm_val->val = LLVMBuildZExt(cl2llvm_builder,
original_val->val, LLVMInt32Type(),
temp_var_name);
}
if(tosign)
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
else
cl2llvmTypeWrapSetSign(llvm_val->type, 0);
}
else if (totype == LLVMInt16Type())
{
if (fromsign)
{
llvm_val->val = LLVMBuildSExt(cl2llvm_builder,
original_val->val, LLVMInt16Type(),
temp_var_name);
}
else
{
llvm_val->val = LLVMBuildZExt(cl2llvm_builder,
original_val->val, LLVMInt16Type(),
temp_var_name);
}
if(tosign)
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
else
cl2llvmTypeWrapSetSign(llvm_val->type, 0);
}
else if (totype == LLVMInt8Type())
{
if(tosign)
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
else
cl2llvmTypeWrapSetSign(llvm_val->type, 0);
temp_var_count--;
}
else if (totype == LLVMInt1Type())
{
llvm_val->val = LLVMBuildTrunc(cl2llvm_builder,
original_val->val, LLVMInt1Type(), temp_var_name);
if(tosign)
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
else
cl2llvmTypeWrapSetSign(llvm_val->type, 0);
}
}
else if (fromtype == LLVMInt1Type())
{
if (totype == LLVMDoubleType())
{
if (fromsign)
{
llvm_val->val =
LLVMBuildSIToFP(cl2llvm_builder,
original_val->val, LLVMDoubleType(),
temp_var_name);
}
else
{
llvm_val->val =
LLVMBuildUIToFP(cl2llvm_builder,
original_val->val, LLVMDoubleType(),
temp_var_name);
}
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
}
else if (totype == LLVMFloatType())
{
if (fromsign)
{
llvm_val->val =
LLVMBuildSIToFP(cl2llvm_builder,
original_val->val, LLVMFloatType(),
temp_var_name);
}
else
{
llvm_val->val =
LLVMBuildUIToFP(cl2llvm_builder,
original_val->val, LLVMFloatType(),
temp_var_name);
}
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
}
else if (totype == LLVMHalfType())
{
if (fromsign)
{
llvm_val->val =
LLVMBuildSIToFP(cl2llvm_builder,
original_val->val, LLVMHalfType(),
temp_var_name);
}
else
{
llvm_val->val =
LLVMBuildUIToFP(cl2llvm_builder,
original_val->val, LLVMHalfType(),
temp_var_name);
}
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
}
else if (totype == LLVMInt64Type())
{
if (fromsign)
{
llvm_val->val = LLVMBuildSExt(cl2llvm_builder,
original_val->val, LLVMInt64Type(),
temp_var_name);
}
else
{
llvm_val->val = LLVMBuildZExt(cl2llvm_builder,
original_val->val, LLVMInt64Type(),
temp_var_name);
}
if (tosign)
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
else
cl2llvmTypeWrapSetSign(llvm_val->type, 0);
}
else if (totype == LLVMInt32Type())
{
if (fromsign)
{
llvm_val->val = LLVMBuildSExt(cl2llvm_builder,
original_val->val, LLVMInt32Type(),
temp_var_name);
}
else
{
llvm_val->val = LLVMBuildZExt(cl2llvm_builder,
original_val->val, LLVMInt32Type(),
temp_var_name);
}
if(tosign)
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
else
cl2llvmTypeWrapSetSign(llvm_val->type, 0);
}
else if (totype == LLVMInt16Type())
{
if (fromsign)
{
llvm_val->val = LLVMBuildSExt(cl2llvm_builder,
original_val->val, LLVMInt16Type(),
temp_var_name);
}
else
{
llvm_val->val = LLVMBuildZExt(cl2llvm_builder,
original_val->val, LLVMInt16Type(),
temp_var_name);
}
if(tosign)
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
else
cl2llvmTypeWrapSetSign(llvm_val->type, 0);
}
else if (totype == LLVMInt8Type())
{
if (fromsign)
{
llvm_val->val = LLVMBuildSExt(cl2llvm_builder,
original_val->val, LLVMInt8Type(),
temp_var_name);
}
else
{
llvm_val->val = LLVMBuildZExt(cl2llvm_builder,
original_val->val, LLVMInt8Type(),
temp_var_name);
}
if(tosign)
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
else
cl2llvmTypeWrapSetSign(llvm_val->type, 0);
}
else if (totype == LLVMInt1Type())
{
if(tosign)
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
else
cl2llvmTypeWrapSetSign(llvm_val->type, 0);
temp_var_count--;
}
}
/*We now know that from type must be a floating point.*/
/*Floating point to signed integer conversions*/
else if (tosign && LLVMGetTypeKind(totype) == 8)
{
if (totype == LLVMInt64Type())
{
llvm_val->val = LLVMBuildFPToSI(cl2llvm_builder,
original_val->val, LLVMInt64Type(), temp_var_name);
}
else if (totype == LLVMInt32Type())
{
llvm_val->val = LLVMBuildFPToSI(cl2llvm_builder,
original_val->val, LLVMInt32Type(), temp_var_name);
}
else if (totype == LLVMInt16Type())
{
llvm_val->val = LLVMBuildFPToSI(cl2llvm_builder,
original_val->val, LLVMInt16Type(), temp_var_name);
}
else if (totype == LLVMInt8Type())
{
llvm_val->val = LLVMBuildFPToSI(cl2llvm_builder,
original_val->val, LLVMInt8Type(), temp_var_name);
}
else if (totype == LLVMInt1Type())
{
llvm_val->val = LLVMBuildFPToSI(cl2llvm_builder,
original_val->val, LLVMInt1Type(), temp_var_name);
}
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
}
/*Floating point to unsigned integer conversions*/
else if (!tosign)
{
if (totype == LLVMInt64Type())
{
llvm_val->val = LLVMBuildFPToUI(cl2llvm_builder,
original_val->val, LLVMInt64Type(), temp_var_name);
}
else if (totype == LLVMInt32Type())
{
llvm_val->val = LLVMBuildFPToUI(cl2llvm_builder,
original_val->val, LLVMInt32Type(), temp_var_name);
}
else if (totype == LLVMInt16Type())
{
llvm_val->val = LLVMBuildFPToUI(cl2llvm_builder,
original_val->val, LLVMInt16Type(), temp_var_name);
}
else if (totype == LLVMInt8Type())
{
llvm_val->val = LLVMBuildFPToUI(cl2llvm_builder,
original_val->val, LLVMInt8Type(), temp_var_name);
}
else if (totype == LLVMInt1Type())
{
llvm_val->val = LLVMBuildFPToUI(cl2llvm_builder,
original_val->val, LLVMInt1Type(), temp_var_name);
}
cl2llvmTypeWrapSetSign(llvm_val->type, 0);
}
else if (totype == LLVMDoubleType())
{
llvm_val->val = LLVMBuildFPExt(cl2llvm_builder,
original_val->val, LLVMDoubleType(), temp_var_name);
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
}
else if (totype == LLVMFloatType())
{
if (fromtype == LLVMDoubleType())
{
llvm_val->val = LLVMBuildFPTrunc(cl2llvm_builder,
original_val->val, LLVMFloatType(), temp_var_name);
}
else if (fromtype == LLVMHalfType())
{
llvm_val->val = LLVMBuildFPExt(cl2llvm_builder,
original_val->val, LLVMFloatType(), temp_var_name);
}
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
}
else if (totype == LLVMHalfType())
{
llvm_val->val = LLVMBuildFPTrunc(cl2llvm_builder,
original_val->val, LLVMHalfType(), temp_var_name);
cl2llvmTypeWrapSetSign(llvm_val->type, 1);
}
cl2llvmTypeWrapSetLlvmType(llvm_val->type, totype);
cl2llvmTypeWrapSetSign(llvm_val->type, tosign);
return llvm_val;
}