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 LLVMValueRef gen_digestof_int64(compile_t* c, LLVMValueRef value)
{
pony_assert(LLVMTypeOf(value) == c->i64);
if(target_is_ilp32(c->opt->triple))
{
LLVMValueRef shift = LLVMConstInt(c->i64, 32, false);
LLVMValueRef high = LLVMBuildLShr(c->builder, value, shift, "");
high = LLVMBuildTrunc(c->builder, high, c->i32, "");
value = LLVMBuildTrunc(c->builder, value, c->i32, "");
value = LLVMBuildXor(c->builder, value, high, "");
}
return value;
}
/**
* 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;
}
/**
* Return (scalar-cast)val ? true : false;
*/
LLVMValueRef
lp_build_any_true_range(struct lp_build_context *bld,
unsigned real_length,
LLVMValueRef val)
{
LLVMBuilderRef builder = bld->gallivm->builder;
LLVMTypeRef scalar_type;
LLVMTypeRef true_type;
assert(real_length <= bld->type.length);
true_type = LLVMIntTypeInContext(bld->gallivm->context,
bld->type.width * real_length);
scalar_type = LLVMIntTypeInContext(bld->gallivm->context,
bld->type.width * bld->type.length);
val = LLVMBuildBitCast(builder, val, scalar_type, "");
/*
* We're using always native types so we can use intrinsics.
* However, if we don't do per-element calculations, we must ensure
* the excess elements aren't used since they may contain garbage.
*/
if (real_length < bld->type.length) {
val = LLVMBuildTrunc(builder, val, true_type, "");
}
return LLVMBuildICmp(builder, LLVMIntNE,
val, LLVMConstNull(true_type), "");
}
/**
* 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();
}
static bool guard_match(compile_t* c, ast_t* guard,
LLVMBasicBlockRef next_block)
{
if(ast_id(guard) == TK_NONE)
return true;
LLVMValueRef value = gen_expr(c, guard);
if(value == NULL)
return false;
LLVMBasicBlockRef continue_block = codegen_block(c, "pattern_continue");
LLVMValueRef test = LLVMBuildTrunc(c->builder, value, c->i1, "");
LLVMBuildCondBr(c->builder, test, continue_block, next_block);
LLVMPositionBuilderAtEnd(c->builder, continue_block);
return true;
}
static bool check_value(compile_t* c, ast_t* pattern, ast_t* param_type,
LLVMValueRef value, LLVMBasicBlockRef next_block)
{
reach_type_t* t = reach_type(c->reach, param_type);
LLVMValueRef r_value = gen_assign_cast(c, t->use_type, value, param_type);
if(r_value == NULL)
return false;
LLVMValueRef result = gen_pattern_eq(c, pattern, r_value);
if(result == NULL)
return false;
LLVMBasicBlockRef continue_block = codegen_block(c, "pattern_continue");
LLVMValueRef test = LLVMBuildTrunc(c->builder, result, c->i1, "");
LLVMBuildCondBr(c->builder, test, continue_block, next_block);
LLVMPositionBuilderAtEnd(c->builder, continue_block);
return true;
}
/**
* 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;
}
static LLVMValueRef cast_ffi_arg(compile_t* c, LLVMValueRef arg,
LLVMTypeRef param)
{
if(arg == NULL)
return NULL;
LLVMTypeRef arg_type = LLVMTypeOf(arg);
if(param == c->i1)
{
// If the parameter is an i1, it must be from an LLVM intrinsic. In that
// case, the argument must be a Bool encoded as an ibool.
if(arg_type != c->ibool)
return NULL;
// Truncate the Bool's i8 representation to i1.
return LLVMBuildTrunc(c->builder, arg, c->i1, "");
}
switch(LLVMGetTypeKind(param))
{
case LLVMPointerTypeKind:
{
if(LLVMGetTypeKind(arg_type) == LLVMIntegerTypeKind)
arg = LLVMBuildIntToPtr(c->builder, arg, param, "");
else
arg = LLVMBuildBitCast(c->builder, arg, param, "");
break;
}
default: {}
}
return arg;
}
static void emit_up2h(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, input, val;
unsigned i;
fp16 = LLVMHalfTypeInContext(context);
i16 = LLVMInt16TypeInContext(context);
const16 = lp_build_const_int32(uint_bld->gallivm, 16);
input = emit_data->args[0];
for (i = 0; i < 2; i++) {
val = i == 1 ? LLVMBuildLShr(builder, input, const16, "") : input;
val = LLVMBuildTrunc(builder, val, i16, "");
val = LLVMBuildBitCast(builder, val, fp16, "");
emit_data->output[i] =
LLVMBuildFPExt(builder, val, bld_base->base.elem_type, "");
}
}
LLVMValueRef make_short_circuit(compile_t* c, ast_t* left, ast_t* right,
bool is_and)
{
LLVMBasicBlockRef entry_block = LLVMGetInsertBlock(c->builder);
LLVMBasicBlockRef left_block = codegen_block(c, "sc_left");
LLVMValueRef branch = LLVMBuildBr(c->builder, left_block);
LLVMPositionBuilderAtEnd(c->builder, left_block);
LLVMValueRef l_value = gen_expr(c, left);
if(l_value == NULL)
return NULL;
if(LLVMIsAConstantInt(l_value))
{
LLVMInstructionEraseFromParent(branch);
LLVMDeleteBasicBlock(left_block);
LLVMPositionBuilderAtEnd(c->builder, entry_block);
if(is_and)
{
if(is_always_false(l_value))
return gen_expr(c, left);
} else {
if(is_always_true(l_value))
return gen_expr(c, left);
}
return gen_expr(c, right);
}
LLVMBasicBlockRef left_exit_block = LLVMGetInsertBlock(c->builder);
LLVMBasicBlockRef right_block = codegen_block(c, "sc_right");
LLVMBasicBlockRef post_block = codegen_block(c, "sc_post");
LLVMValueRef test = LLVMBuildTrunc(c->builder, l_value, c->i1, "");
if(is_and)
LLVMBuildCondBr(c->builder, test, right_block, post_block);
else
LLVMBuildCondBr(c->builder, test, post_block, right_block);
LLVMPositionBuilderAtEnd(c->builder, right_block);
LLVMValueRef r_value = gen_expr(c, right);
if(r_value == NULL)
return NULL;
LLVMBasicBlockRef right_exit_block = LLVMGetInsertBlock(c->builder);
LLVMBuildBr(c->builder, post_block);
LLVMPositionBuilderAtEnd(c->builder, post_block);
LLVMValueRef phi = LLVMBuildPhi(c->builder, c->ibool, "");
LLVMAddIncoming(phi, &l_value, &left_exit_block, 1);
LLVMAddIncoming(phi, &r_value, &right_exit_block, 1);
if(LLVMIsAConstantInt(r_value))
{
if(is_and)
{
if(is_always_false(r_value))
return r_value;
} else {
if(is_always_true(r_value))
return r_value;
}
return l_value;
}
return phi;
}
LLVMValueRef gen_repeat(compile_t* c, ast_t* ast)
{
bool needed = is_result_needed(ast);
AST_GET_CHILDREN(ast, body, cond, else_clause);
ast_t* type = ast_type(ast);
ast_t* body_type = ast_type(body);
ast_t* else_type = ast_type(else_clause);
reach_type_t* phi_type = NULL;
if(needed && !is_control_type(type))
phi_type = reach_type(c->reach, type);
LLVMBasicBlockRef body_block = codegen_block(c, "repeat_body");
LLVMBasicBlockRef cond_block = codegen_block(c, "repeat_cond");
LLVMBasicBlockRef else_block = codegen_block(c, "repeat_else");
LLVMBasicBlockRef post_block = NULL;
LLVMBuildBr(c->builder, body_block);
// start the post block so that a break can modify the phi node
LLVMValueRef phi = GEN_NOTNEEDED;
if(!is_control_type(type))
{
// Start the post block so that a break can modify the phi node.
post_block = codegen_block(c, "repeat_post");
LLVMPositionBuilderAtEnd(c->builder, post_block);
if(needed)
phi = LLVMBuildPhi(c->builder, phi_type->use_type, "");
}
// Push the loop status.
codegen_pushloop(c, cond_block, post_block, else_block);
// Body.
LLVMPositionBuilderAtEnd(c->builder, body_block);
LLVMValueRef value = gen_expr(c, body);
if(needed)
value = gen_assign_cast(c, phi_type->use_type, value, body_type);
if(value == NULL)
return NULL;
LLVMBasicBlockRef body_from = NULL;
// If the body can't result in a value, don't generate the conditional
// evaluation. This basic block for the body already has a terminator.
if(value != GEN_NOVALUE)
{
// The body evaluates the condition itself, jumping either back to the body
// or directly to the post block.
LLVMValueRef c_value = gen_expr(c, cond);
if(c_value == NULL)
return NULL;
body_from = LLVMGetInsertBlock(c->builder);
LLVMValueRef test = LLVMBuildTrunc(c->builder, c_value, c->i1, "");
LLVMBuildCondBr(c->builder, test, post_block, body_block);
}
// cond block
// This is only evaluated from a continue, jumping either back to the body
// or to the else block.
LLVMPositionBuilderAtEnd(c->builder, cond_block);
LLVMValueRef i_value = gen_expr(c, cond);
LLVMValueRef test = LLVMBuildTrunc(c->builder, i_value, c->i1, "");
LLVMBuildCondBr(c->builder, test, else_block, body_block);
// Don't need loop status for the else block.
codegen_poploop(c);
// else
// Only happens for a continue in the last iteration.
LLVMPositionBuilderAtEnd(c->builder, else_block);
LLVMValueRef else_value = gen_expr(c, else_clause);
LLVMBasicBlockRef else_from = NULL;
if(else_value == NULL)
return NULL;
if(needed)
else_value = gen_assign_cast(c, phi_type->use_type, else_value, else_type);
if(else_value != GEN_NOVALUE)
{
else_from = LLVMGetInsertBlock(c->builder);
LLVMBuildBr(c->builder, post_block);
}
if(is_control_type(type))
return GEN_NOVALUE;
// post
LLVMPositionBuilderAtEnd(c->builder, post_block);
if(needed)
{
if(value != GEN_NOVALUE)
LLVMAddIncoming(phi, &value, &body_from, 1);
if(else_value != GEN_NOVALUE)
LLVMAddIncoming(phi, &else_value, &else_from, 1);
return phi;
}
return GEN_NOTNEEDED;
}
LLVMValueRef gen_if(compile_t* c, ast_t* ast)
{
bool needed = is_result_needed(ast);
ast_t* type = ast_type(ast);
AST_GET_CHILDREN(ast, cond, left, right);
ast_t* left_type = ast_type(left);
ast_t* right_type = ast_type(right);
// We will have no type if both branches have return statements.
reach_type_t* phi_type = NULL;
if(!is_control_type(type))
phi_type = reach_type(c->reach, type);
LLVMValueRef c_value = gen_expr(c, cond);
if(c_value == NULL)
return NULL;
// If the conditional is constant, generate only one branch.
bool gen_left = true;
bool gen_right = true;
if(LLVMIsAConstantInt(c_value))
{
int value = (int)LLVMConstIntGetZExtValue(c_value);
if(value == 0)
gen_left = false;
else
gen_right = false;
}
LLVMBasicBlockRef then_block = codegen_block(c, "if_then");
LLVMBasicBlockRef else_block = codegen_block(c, "if_else");
LLVMBasicBlockRef post_block = NULL;
// If both branches return, we have no post block.
if(!is_control_type(type))
post_block = codegen_block(c, "if_post");
LLVMValueRef test = LLVMBuildTrunc(c->builder, c_value, c->i1, "");
LLVMBuildCondBr(c->builder, test, then_block, else_block);
// Left branch.
LLVMPositionBuilderAtEnd(c->builder, then_block);
LLVMValueRef l_value;
if(gen_left)
{
l_value = gen_expr(c, left);
} else if(phi_type != NULL) {
l_value = LLVMConstNull(phi_type->use_type);
} else {
LLVMBuildUnreachable(c->builder);
l_value = GEN_NOVALUE;
}
if(l_value != GEN_NOVALUE)
{
if(needed)
l_value = gen_assign_cast(c, phi_type->use_type, l_value, left_type);
if(l_value == NULL)
return NULL;
then_block = LLVMGetInsertBlock(c->builder);
LLVMBuildBr(c->builder, post_block);
}
// Right branch.
LLVMPositionBuilderAtEnd(c->builder, else_block);
LLVMValueRef r_value;
if(gen_right)
{
r_value = gen_expr(c, right);
} else if(phi_type != NULL) {
r_value = LLVMConstNull(phi_type->use_type);
} else {
LLVMBuildUnreachable(c->builder);
r_value = GEN_NOVALUE;
}
// If the right side returns, we don't branch to the post block.
if(r_value != GEN_NOVALUE)
{
if(needed)
r_value = gen_assign_cast(c, phi_type->use_type, r_value, right_type);
if(r_value == NULL)
return NULL;
else_block = LLVMGetInsertBlock(c->builder);
LLVMBuildBr(c->builder, post_block);
}
// If both sides return, we return a sentinal value.
if(is_control_type(type))
return GEN_NOVALUE;
// Continue in the post block.
LLVMPositionBuilderAtEnd(c->builder, post_block);
if(needed)
{
LLVMValueRef phi = LLVMBuildPhi(c->builder, phi_type->use_type, "");
if(l_value != GEN_NOVALUE)
LLVMAddIncoming(phi, &l_value, &then_block, 1);
if(r_value != GEN_NOVALUE)
LLVMAddIncoming(phi, &r_value, &else_block, 1);
return phi;
}
return GEN_NOTNEEDED;
}
/**
* Return mask ? a : b;
*
* mask is a bitwise mask, composed of 0 or ~0 for each element. Any other value
* will yield unpredictable results.
*/
LLVMValueRef
lp_build_select(struct lp_build_context *bld,
LLVMValueRef mask,
LLVMValueRef a,
LLVMValueRef b)
{
LLVMBuilderRef builder = bld->gallivm->builder;
LLVMContextRef lc = bld->gallivm->context;
struct lp_type type = bld->type;
LLVMValueRef res;
assert(lp_check_value(type, a));
assert(lp_check_value(type, b));
if(a == b)
return a;
if (type.length == 1) {
mask = LLVMBuildTrunc(builder, mask, LLVMInt1TypeInContext(lc), "");
res = LLVMBuildSelect(builder, mask, a, b, "");
}
else if (0) {
/* Generate a vector select.
*
* XXX: Using vector selects would avoid emitting intrinsics, but they aren't
* properly supported yet.
*
* LLVM 3.0 includes experimental support provided the -promote-elements
* options is passed to LLVM's command line (e.g., via
* llvm::cl::ParseCommandLineOptions), but resulting code quality is much
* worse, probably because some optimization passes don't know how to
* handle vector selects.
*
* See also:
* - http://lists.cs.uiuc.edu/pipermail/llvmdev/2011-October/043659.html
*/
/* Convert the mask to a vector of booleans.
* XXX: There are two ways to do this. Decide what's best.
*/
if (1) {
LLVMTypeRef bool_vec_type = LLVMVectorType(LLVMInt1TypeInContext(lc), type.length);
mask = LLVMBuildTrunc(builder, mask, bool_vec_type, "");
} else {
mask = LLVMBuildICmp(builder, LLVMIntNE, mask, LLVMConstNull(bld->int_vec_type), "");
}
res = LLVMBuildSelect(builder, mask, a, b, "");
}
else if (((util_cpu_caps.has_sse4_1 &&
type.width * type.length == 128) ||
(util_cpu_caps.has_avx &&
type.width * type.length == 256 && type.width >= 32)) &&
!LLVMIsConstant(a) &&
!LLVMIsConstant(b) &&
!LLVMIsConstant(mask)) {
const char *intrinsic;
LLVMTypeRef arg_type;
LLVMValueRef args[3];
/*
* There's only float blend in AVX but can just cast i32/i64
* to float.
*/
if (type.width * type.length == 256) {
if (type.width == 64) {
intrinsic = "llvm.x86.avx.blendv.pd.256";
arg_type = LLVMVectorType(LLVMDoubleTypeInContext(lc), 4);
}
else {
intrinsic = "llvm.x86.avx.blendv.ps.256";
arg_type = LLVMVectorType(LLVMFloatTypeInContext(lc), 8);
}
}
else if (type.floating &&
type.width == 64) {
intrinsic = "llvm.x86.sse41.blendvpd";
arg_type = LLVMVectorType(LLVMDoubleTypeInContext(lc), 2);
} else if (type.floating &&
type.width == 32) {
intrinsic = "llvm.x86.sse41.blendvps";
arg_type = LLVMVectorType(LLVMFloatTypeInContext(lc), 4);
} else {
intrinsic = "llvm.x86.sse41.pblendvb";
arg_type = LLVMVectorType(LLVMInt8TypeInContext(lc), 16);
}
if (arg_type != bld->int_vec_type) {
mask = LLVMBuildBitCast(builder, mask, arg_type, "");
}
if (arg_type != bld->vec_type) {
a = LLVMBuildBitCast(builder, a, arg_type, "");
b = LLVMBuildBitCast(builder, b, arg_type, "");
}
args[0] = b;
args[1] = a;
args[2] = mask;
res = lp_build_intrinsic(builder, intrinsic,
arg_type, args, Elements(args));
if (arg_type != bld->vec_type) {
res = LLVMBuildBitCast(builder, res, bld->vec_type, "");
}
}
else {
res = lp_build_select_bitwise(bld, mask, a, b);
}
return res;
}
static LLVMValueRef
lp_build_extract_soa_chan(struct lp_build_context *bld,
unsigned blockbits,
boolean srgb_chan,
struct util_format_channel_description chan_desc,
LLVMValueRef packed)
{
struct gallivm_state *gallivm = bld->gallivm;
LLVMBuilderRef builder = gallivm->builder;
struct lp_type type = bld->type;
LLVMValueRef input = packed;
const unsigned width = chan_desc.size;
const unsigned start = chan_desc.shift;
const unsigned stop = start + width;
/* Decode the input vector component */
switch(chan_desc.type) {
case UTIL_FORMAT_TYPE_VOID:
input = bld->undef;
break;
case UTIL_FORMAT_TYPE_UNSIGNED:
/*
* Align the LSB
*/
if (start) {
input = LLVMBuildLShr(builder, input,
lp_build_const_int_vec(gallivm, type, start), "");
}
/*
* Zero the MSBs
*/
if (stop < blockbits) {
unsigned mask = ((unsigned long long)1 << width) - 1;
input = LLVMBuildAnd(builder, input,
lp_build_const_int_vec(gallivm, type, mask), "");
}
/*
* Type conversion
*/
if (type.floating) {
if (srgb_chan) {
struct lp_type conv_type = lp_uint_type(type);
input = lp_build_srgb_to_linear(gallivm, conv_type, width, input);
}
else {
if(chan_desc.normalized)
input = lp_build_unsigned_norm_to_float(gallivm, width, type, input);
else
input = LLVMBuildSIToFP(builder, input, bld->vec_type, "");
}
}
else if (chan_desc.pure_integer) {
/* Nothing to do */
} else {
/* FIXME */
assert(0);
}
break;
case UTIL_FORMAT_TYPE_SIGNED:
/*
* Align the sign bit first.
*/
if (stop < type.width) {
unsigned bits = type.width - stop;
LLVMValueRef bits_val = lp_build_const_int_vec(gallivm, type, bits);
input = LLVMBuildShl(builder, input, bits_val, "");
}
/*
* Align the LSB (with an arithmetic shift to preserve the sign)
*/
if (chan_desc.size < type.width) {
unsigned bits = type.width - chan_desc.size;
LLVMValueRef bits_val = lp_build_const_int_vec(gallivm, type, bits);
input = LLVMBuildAShr(builder, input, bits_val, "");
}
/*
* Type conversion
*/
if (type.floating) {
input = LLVMBuildSIToFP(builder, input, bld->vec_type, "");
if (chan_desc.normalized) {
double scale = 1.0 / ((1 << (chan_desc.size - 1)) - 1);
LLVMValueRef scale_val = lp_build_const_vec(gallivm, type, scale);
input = LLVMBuildFMul(builder, input, scale_val, "");
/*
* The formula above will produce value below -1.0 for most negative
* value but everything seems happy with that hence disable for now.
*/
if (0)
input = lp_build_max(bld, input,
lp_build_const_vec(gallivm, type, -1.0f));
}
}
else if (chan_desc.pure_integer) {
/* Nothing to do */
} else {
/* FIXME */
assert(0);
}
break;
case UTIL_FORMAT_TYPE_FLOAT:
if (type.floating) {
if (chan_desc.size == 16) {
struct lp_type f16i_type = type;
f16i_type.width /= 2;
f16i_type.floating = 0;
if (start) {
input = LLVMBuildLShr(builder, input,
lp_build_const_int_vec(gallivm, type, start), "");
}
input = LLVMBuildTrunc(builder, input,
lp_build_vec_type(gallivm, f16i_type), "");
input = lp_build_half_to_float(gallivm, input);
} else {
assert(start == 0);
assert(stop == 32);
assert(type.width == 32);
}
input = LLVMBuildBitCast(builder, input, bld->vec_type, "");
}
else {
/* FIXME */
assert(0);
input = bld->undef;
}
break;
case UTIL_FORMAT_TYPE_FIXED:
if (type.floating) {
double scale = 1.0 / ((1 << (chan_desc.size/2)) - 1);
LLVMValueRef scale_val = lp_build_const_vec(gallivm, type, scale);
input = LLVMBuildSIToFP(builder, input, bld->vec_type, "");
input = LLVMBuildFMul(builder, input, scale_val, "");
}
else {
/* FIXME */
assert(0);
input = bld->undef;
}
break;
default:
assert(0);
input = bld->undef;
break;
}
return input;
}
/**
* Store depth/stencil values.
* Incoming values are swizzled (typically n 2x2 quads), stored linear.
* If there's a mask it will do select/store otherwise just store.
*
* \param type the data type of the fragment depth/stencil values
* \param format_desc description of the depth/stencil surface
* \param mask the alive/dead pixel mask for the quad (vector)
* \param z_fb z values read from fb (with padding)
* \param s_fb s values read from fb (with padding)
* \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_value the depth values to store (with padding)
* \param s_value the stencil values to store (with padding)
*/
void
lp_build_depth_stencil_write_swizzled(struct gallivm_state *gallivm,
struct lp_type z_src_type,
const struct util_format_description *format_desc,
struct lp_build_mask_context *mask,
LLVMValueRef z_fb,
LLVMValueRef s_fb,
LLVMValueRef loop_counter,
LLVMValueRef depth_ptr,
LLVMValueRef depth_stride,
LLVMValueRef z_value,
LLVMValueRef s_value)
{
struct lp_build_context z_bld;
LLVMValueRef shuffles[LP_MAX_VECTOR_LENGTH / 4];
LLVMBuilderRef builder = gallivm->builder;
LLVMValueRef mask_value = NULL;
LLVMValueRef zs_dst1, zs_dst2;
LLVMValueRef zs_dst_ptr1, zs_dst_ptr2;
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 z_type = zs_type;
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);
z_type.width = z_src_type.width;
lp_build_context_init(&z_bld, gallivm, z_type);
/*
* This is far from ideal, at least for late depth write we should do this
* outside the fs loop to avoid all the swizzle stuff.
*/
if (z_src_type.length == 4) {
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, "");
}
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, "");
zs_dst_ptr1 = LLVMBuildGEP(builder, depth_ptr, &depth_offset1, 1, "");
zs_dst_ptr1 = LLVMBuildBitCast(builder, zs_dst_ptr1, load_ptr_type, "");
zs_dst_ptr2 = LLVMBuildGEP(builder, depth_ptr, &depth_offset2, 1, "");
zs_dst_ptr2 = LLVMBuildBitCast(builder, zs_dst_ptr2, load_ptr_type, "");
if (format_desc->block.bits > 32) {
s_value = LLVMBuildBitCast(builder, s_value, z_bld.vec_type, "");
}
if (mask) {
mask_value = lp_build_mask_value(mask);
z_value = lp_build_select(&z_bld, mask_value, z_value, z_fb);
if (format_desc->block.bits > 32) {
s_fb = LLVMBuildBitCast(builder, s_fb, z_bld.vec_type, "");
s_value = lp_build_select(&z_bld, mask_value, s_value, s_fb);
}
}
if (zs_type.width < z_src_type.width) {
/* Truncate ZS values (e.g., when writing to Z16_UNORM) */
z_value = LLVMBuildTrunc(builder, z_value,
lp_build_int_vec_type(gallivm, zs_type), "");
}
if (format_desc->block.bits <= 32) {
if (z_src_type.length == 4) {
zs_dst1 = lp_build_extract_range(gallivm, z_value, 0, 2);
zs_dst2 = lp_build_extract_range(gallivm, z_value, 2, 2);
}
else {
assert(z_src_type.length == 8);
zs_dst1 = LLVMBuildShuffleVector(builder, z_value, z_value,
LLVMConstVector(&shuffles[0],
zs_load_type.length), "");
zs_dst2 = LLVMBuildShuffleVector(builder, z_value, z_value,
LLVMConstVector(&shuffles[4],
zs_load_type.length), "");
}
}
else {
if (z_src_type.length == 4) {
zs_dst1 = lp_build_interleave2(gallivm, z_type,
z_value, s_value, 0);
zs_dst2 = lp_build_interleave2(gallivm, z_type,
z_value, s_value, 1);
}
else {
unsigned i;
LLVMValueRef shuffles[LP_MAX_VECTOR_LENGTH / 2];
assert(z_src_type.length == 8);
for (i = 0; i < 8; i++) {
shuffles[i*2] = lp_build_const_int32(gallivm, (i&1) + (i&2) * 2 + (i&4) / 2);
shuffles[i*2+1] = lp_build_const_int32(gallivm, (i&1) + (i&2) * 2 + (i&4) / 2 +
z_src_type.length);
}
zs_dst1 = LLVMBuildShuffleVector(builder, z_value, s_value,
LLVMConstVector(&shuffles[0],
z_src_type.length), "");
zs_dst2 = LLVMBuildShuffleVector(builder, z_value, s_value,
LLVMConstVector(&shuffles[8],
z_src_type.length), "");
}
zs_dst1 = LLVMBuildBitCast(builder, zs_dst1,
lp_build_vec_type(gallivm, zs_load_type), "");
zs_dst2 = LLVMBuildBitCast(builder, zs_dst2,
lp_build_vec_type(gallivm, zs_load_type), "");
}
LLVMBuildStore(builder, zs_dst1, zs_dst_ptr1);
LLVMBuildStore(builder, zs_dst2, zs_dst_ptr2);
}
/**
* Perform the occlusion test and increase the counter.
* Test the depth mask. Add the number of channel which has none zero mask
* into the occlusion counter. e.g. maskvalue is {-1, -1, -1, -1}.
* The counter will add 4.
*
* \param type holds element type of the mask vector.
* \param maskvalue is the depth test mask.
* \param counter is a pointer of the uint32 counter.
*/
void
lp_build_occlusion_count(struct gallivm_state *gallivm,
struct lp_type type,
LLVMValueRef maskvalue,
LLVMValueRef counter)
{
LLVMBuilderRef builder = gallivm->builder;
LLVMContextRef context = gallivm->context;
LLVMValueRef countmask = lp_build_const_int_vec(gallivm, type, 1);
LLVMValueRef count, newcount;
assert(type.length <= 16);
assert(type.floating);
if(util_cpu_caps.has_sse && type.length == 4) {
const char *movmskintr = "llvm.x86.sse.movmsk.ps";
const char *popcntintr = "llvm.ctpop.i32";
LLVMValueRef bits = LLVMBuildBitCast(builder, maskvalue,
lp_build_vec_type(gallivm, type), "");
bits = lp_build_intrinsic_unary(builder, movmskintr,
LLVMInt32TypeInContext(context), bits);
count = lp_build_intrinsic_unary(builder, popcntintr,
LLVMInt32TypeInContext(context), bits);
}
else if(util_cpu_caps.has_avx && type.length == 8) {
const char *movmskintr = "llvm.x86.avx.movmsk.ps.256";
const char *popcntintr = "llvm.ctpop.i32";
LLVMValueRef bits = LLVMBuildBitCast(builder, maskvalue,
lp_build_vec_type(gallivm, type), "");
bits = lp_build_intrinsic_unary(builder, movmskintr,
LLVMInt32TypeInContext(context), bits);
count = lp_build_intrinsic_unary(builder, popcntintr,
LLVMInt32TypeInContext(context), bits);
}
else {
unsigned i;
LLVMValueRef countv = LLVMBuildAnd(builder, maskvalue, countmask, "countv");
LLVMTypeRef counttype = LLVMIntTypeInContext(context, type.length * 8);
LLVMTypeRef i8vntype = LLVMVectorType(LLVMInt8TypeInContext(context), type.length * 4);
LLVMValueRef shufflev, countd;
LLVMValueRef shuffles[16];
const char *popcntintr = NULL;
countv = LLVMBuildBitCast(builder, countv, i8vntype, "");
for (i = 0; i < type.length; i++) {
shuffles[i] = lp_build_const_int32(gallivm, 4*i);
}
shufflev = LLVMConstVector(shuffles, type.length);
countd = LLVMBuildShuffleVector(builder, countv, LLVMGetUndef(i8vntype), shufflev, "");
countd = LLVMBuildBitCast(builder, countd, counttype, "countd");
/*
* XXX FIXME
* this is bad on cpus without popcount (on x86 supported by intel
* nehalem, amd barcelona, and up - not tied to sse42).
* Would be much faster to just sum the 4 elements of the vector with
* some horizontal add (shuffle/add/shuffle/add after the initial and).
*/
switch (type.length) {
case 4:
popcntintr = "llvm.ctpop.i32";
break;
case 8:
popcntintr = "llvm.ctpop.i64";
break;
case 16:
popcntintr = "llvm.ctpop.i128";
break;
default:
assert(0);
}
count = lp_build_intrinsic_unary(builder, popcntintr, counttype, countd);
if (type.length > 4) {
count = LLVMBuildTrunc(builder, count, LLVMIntTypeInContext(context, 32), "");
}
}
newcount = LLVMBuildLoad(builder, counter, "origcount");
newcount = LLVMBuildAdd(builder, newcount, count, "newcount");
LLVMBuildStore(builder, newcount, counter);
}
/**
* Sample the texture/mipmap using given image filter and mip filter.
* data0_ptr and data1_ptr point to the two mipmap levels to sample
* from. width0/1_vec, height0/1_vec, depth0/1_vec indicate their sizes.
* If we're using nearest miplevel sampling the '1' values will be null/unused.
*/
static void
lp_build_sample_mipmap(struct lp_build_sample_context *bld,
unsigned img_filter,
unsigned mip_filter,
LLVMValueRef s,
LLVMValueRef t,
LLVMValueRef r,
LLVMValueRef ilevel0,
LLVMValueRef ilevel1,
LLVMValueRef lod_fpart,
LLVMValueRef colors_lo_var,
LLVMValueRef colors_hi_var)
{
LLVMBuilderRef builder = bld->gallivm->builder;
LLVMValueRef size0;
LLVMValueRef size1;
LLVMValueRef row_stride0_vec;
LLVMValueRef row_stride1_vec;
LLVMValueRef img_stride0_vec;
LLVMValueRef img_stride1_vec;
LLVMValueRef data_ptr0;
LLVMValueRef data_ptr1;
LLVMValueRef colors0_lo, colors0_hi;
LLVMValueRef colors1_lo, colors1_hi;
/* sample the first mipmap level */
lp_build_mipmap_level_sizes(bld, ilevel0,
&size0,
&row_stride0_vec, &img_stride0_vec);
data_ptr0 = lp_build_get_mipmap_level(bld, ilevel0);
if (img_filter == PIPE_TEX_FILTER_NEAREST) {
lp_build_sample_image_nearest(bld,
size0,
row_stride0_vec, img_stride0_vec,
data_ptr0, s, t, r,
&colors0_lo, &colors0_hi);
}
else {
assert(img_filter == PIPE_TEX_FILTER_LINEAR);
lp_build_sample_image_linear(bld,
size0,
row_stride0_vec, img_stride0_vec,
data_ptr0, s, t, r,
&colors0_lo, &colors0_hi);
}
/* Store the first level's colors in the output variables */
LLVMBuildStore(builder, colors0_lo, colors_lo_var);
LLVMBuildStore(builder, colors0_hi, colors_hi_var);
if (mip_filter == PIPE_TEX_MIPFILTER_LINEAR) {
LLVMValueRef h16_scale = lp_build_const_float(bld->gallivm, 256.0);
LLVMTypeRef i32_type = LLVMIntTypeInContext(bld->gallivm->context, 32);
struct lp_build_if_state if_ctx;
LLVMValueRef need_lerp;
lod_fpart = LLVMBuildFMul(builder, lod_fpart, h16_scale, "");
lod_fpart = LLVMBuildFPToSI(builder, lod_fpart, i32_type, "lod_fpart.fixed16");
/* need_lerp = lod_fpart > 0 */
need_lerp = LLVMBuildICmp(builder, LLVMIntSGT,
lod_fpart, LLVMConstNull(i32_type),
"need_lerp");
lp_build_if(&if_ctx, bld->gallivm, need_lerp);
{
struct lp_build_context h16_bld;
lp_build_context_init(&h16_bld, bld->gallivm, lp_type_ufixed(16));
/* sample the second mipmap level */
lp_build_mipmap_level_sizes(bld, ilevel1,
&size1,
&row_stride1_vec, &img_stride1_vec);
data_ptr1 = lp_build_get_mipmap_level(bld, ilevel1);
if (img_filter == PIPE_TEX_FILTER_NEAREST) {
lp_build_sample_image_nearest(bld,
size1,
row_stride1_vec, img_stride1_vec,
data_ptr1, s, t, r,
&colors1_lo, &colors1_hi);
}
else {
lp_build_sample_image_linear(bld,
size1,
row_stride1_vec, img_stride1_vec,
data_ptr1, s, t, r,
&colors1_lo, &colors1_hi);
}
/* interpolate samples from the two mipmap levels */
lod_fpart = LLVMBuildTrunc(builder, lod_fpart, h16_bld.elem_type, "");
lod_fpart = lp_build_broadcast_scalar(&h16_bld, lod_fpart);
#if HAVE_LLVM == 0x208
/* This is a work-around for a bug in LLVM 2.8.
* Evidently, something goes wrong in the construction of the
* lod_fpart short[8] vector. Adding this no-effect shuffle seems
* to force the vector to be properly constructed.
* Tested with mesa-demos/src/tests/mipmap_limits.c (press t, f).
*/
{
LLVMValueRef shuffles[8], shuffle;
int i;
assert(h16_bld.type.length <= Elements(shuffles));
for (i = 0; i < h16_bld.type.length; i++)
shuffles[i] = lp_build_const_int32(bld->gallivm, 2 * (i & 1));
shuffle = LLVMConstVector(shuffles, h16_bld.type.length);
lod_fpart = LLVMBuildShuffleVector(builder,
lod_fpart, lod_fpart,
shuffle, "");
}
#endif
colors0_lo = lp_build_lerp(&h16_bld, lod_fpart,
colors0_lo, colors1_lo);
colors0_hi = lp_build_lerp(&h16_bld, lod_fpart,
colors0_hi, colors1_hi);
LLVMBuildStore(builder, colors0_lo, colors_lo_var);
LLVMBuildStore(builder, colors0_hi, colors_hi_var);
}
lp_build_endif(&if_ctx);
}
}
LLVMValueRef gen_while(compile_t* c, ast_t* ast)
{
bool needed = is_result_needed(ast);
AST_GET_CHILDREN(ast, cond, body, else_clause);
ast_t* type = ast_type(ast);
ast_t* body_type = ast_type(body);
ast_t* else_type = ast_type(else_clause);
reach_type_t* phi_type = NULL;
if(needed && !is_control_type(type))
phi_type = reach_type(c->reach, type);
LLVMBasicBlockRef init_block = codegen_block(c, "while_init");
LLVMBasicBlockRef body_block = codegen_block(c, "while_body");
LLVMBasicBlockRef else_block = codegen_block(c, "while_else");
LLVMBasicBlockRef post_block = NULL;
LLVMBuildBr(c->builder, init_block);
// start the post block so that a break can modify the phi node
LLVMValueRef phi = GEN_NOTNEEDED;
if(!is_control_type(type))
{
// Start the post block so that a break can modify the phi node.
post_block = codegen_block(c, "while_post");
LLVMPositionBuilderAtEnd(c->builder, post_block);
if(needed)
phi = LLVMBuildPhi(c->builder, phi_type->use_type, "");
}
// Push the loop status.
codegen_pushloop(c, init_block, post_block, else_block);
// init
// This jumps either to the body or the else clause. This is not evaluated
// on each loop iteration: only on the first entry or after a continue.
LLVMPositionBuilderAtEnd(c->builder, init_block);
LLVMValueRef i_value = gen_expr(c, cond);
if(i_value == NULL)
return NULL;
LLVMValueRef test = LLVMBuildTrunc(c->builder, i_value, c->i1, "");
LLVMBuildCondBr(c->builder, test, body_block, else_block);
// Body.
LLVMPositionBuilderAtEnd(c->builder, body_block);
LLVMValueRef l_value = gen_expr(c, body);
if(needed)
l_value = gen_assign_cast(c, phi_type->use_type, l_value, body_type);
if(l_value == NULL)
return NULL;
LLVMBasicBlockRef body_from = NULL;
// If the body can't result in a value, don't generate the conditional
// evaluation. This basic block for the body already has a terminator.
if(l_value != GEN_NOVALUE)
{
// The body evaluates the condition itself, jumping either back to the body
// or directly to the post block.
LLVMValueRef c_value = gen_expr(c, cond);
if(c_value == NULL)
return NULL;
body_from = LLVMGetInsertBlock(c->builder);
LLVMValueRef test = LLVMBuildTrunc(c->builder, c_value, c->i1, "");
LLVMBuildCondBr(c->builder, test, body_block, post_block);
}
// Don't need loop status for the else block.
codegen_poploop(c);
// else
// If the loop doesn't generate a value (doesn't execute, or continues on the
// last iteration), the else clause generates the value.
LLVMPositionBuilderAtEnd(c->builder, else_block);
LLVMValueRef r_value = gen_expr(c, else_clause);
LLVMBasicBlockRef else_from = NULL;
if(r_value != GEN_NOVALUE)
{
if(r_value == NULL)
return NULL;
if(needed)
r_value = gen_assign_cast(c, phi_type->use_type, r_value, else_type);
else_from = LLVMGetInsertBlock(c->builder);
LLVMBuildBr(c->builder, post_block);
}
if(is_control_type(type))
return GEN_NOVALUE;
// post
LLVMPositionBuilderAtEnd(c->builder, post_block);
if(needed)
{
if(l_value != GEN_NOVALUE)
LLVMAddIncoming(phi, &l_value, &body_from, 1);
if(r_value != GEN_NOVALUE)
LLVMAddIncoming(phi, &r_value, &else_from, 1);
return phi;
}
return GEN_NOTNEEDED;
}
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;
}
/**
* Converts float32 to int16 half-float
* Note this can be performed in 1 instruction if vcvtps2ph exists (f16c/cvt16)
* [llvm.x86.vcvtps2ph / _mm_cvtps_ph]
*
* @param src value to convert
*
* Convert float32 to half floats, preserving Infs and NaNs,
* with rounding towards zero (trunc).
*/
LLVMValueRef
lp_build_float_to_half(struct gallivm_state *gallivm,
LLVMValueRef src)
{
LLVMBuilderRef builder = gallivm->builder;
LLVMTypeRef f32_vec_type = LLVMTypeOf(src);
unsigned length = LLVMGetTypeKind(f32_vec_type) == LLVMVectorTypeKind
? LLVMGetVectorSize(f32_vec_type) : 1;
struct lp_type i32_type = lp_type_int_vec(32, 32 * length);
struct lp_type i16_type = lp_type_int_vec(16, 16 * length);
LLVMValueRef result;
if (util_cpu_caps.has_f16c && HAVE_LLVM >= 0x0301 &&
(length == 4 || length == 8)) {
struct lp_type i168_type = lp_type_int_vec(16, 16 * 8);
unsigned mode = 3; /* same as LP_BUILD_ROUND_TRUNCATE */
LLVMTypeRef i32t = LLVMInt32TypeInContext(gallivm->context);
const char *intrinsic = NULL;
if (length == 4) {
intrinsic = "llvm.x86.vcvtps2ph.128";
}
else {
intrinsic = "llvm.x86.vcvtps2ph.256";
}
result = lp_build_intrinsic_binary(builder, intrinsic,
lp_build_vec_type(gallivm, i168_type),
src, LLVMConstInt(i32t, mode, 0));
if (length == 4) {
result = lp_build_extract_range(gallivm, result, 0, 4);
}
}
else {
result = lp_build_float_to_smallfloat(gallivm, i32_type, src, 10, 5, 0, true);
/* Convert int32 vector to int16 vector by trunc (might generate bad code) */
result = LLVMBuildTrunc(builder, result, lp_build_vec_type(gallivm, i16_type), "");
}
/*
* Debugging code.
*/
if (0) {
LLVMTypeRef i32t = LLVMInt32TypeInContext(gallivm->context);
LLVMTypeRef i16t = LLVMInt16TypeInContext(gallivm->context);
LLVMTypeRef f32t = LLVMFloatTypeInContext(gallivm->context);
LLVMValueRef ref_result = LLVMGetUndef(LLVMVectorType(i16t, length));
unsigned i;
LLVMTypeRef func_type = LLVMFunctionType(i16t, &f32t, 1, 0);
LLVMValueRef func = lp_build_const_int_pointer(gallivm, func_to_pointer((func_pointer)util_float_to_half));
func = LLVMBuildBitCast(builder, func, LLVMPointerType(func_type, 0), "util_float_to_half");
for (i = 0; i < length; ++i) {
LLVMValueRef index = LLVMConstInt(i32t, i, 0);
LLVMValueRef f32 = LLVMBuildExtractElement(builder, src, index, "");
#if 0
/* XXX: not really supported by backends */
LLVMValueRef f16 = lp_build_intrinsic_unary(builder, "llvm.convert.to.fp16", i16t, f32);
#else
LLVMValueRef f16 = LLVMBuildCall(builder, func, &f32, 1, "");
#endif
ref_result = LLVMBuildInsertElement(builder, ref_result, f16, index, "");
}
lp_build_print_value(gallivm, "src = ", src);
lp_build_print_value(gallivm, "llvm = ", result);
lp_build_print_value(gallivm, "util = ", ref_result);
lp_build_printf(gallivm, "\n");
}
return result;
}
/**
* 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;
}
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;
}
