static void name_param(compile_t* c, reachable_type_t* t,
reachable_method_t* m, LLVMValueRef func, const char* name, unsigned index,
size_t line, size_t pos)
{
LLVMValueRef value = LLVMGetParam(func, index);
LLVMSetValueName(value, name);
LLVMValueRef alloc = LLVMBuildAlloca(c->builder, t->use_type, name);
LLVMBuildStore(c->builder, value, alloc);
codegen_setlocal(c, name, alloc);
LLVMMetadataRef info;
if(index == 0)
{
info = LLVMDIBuilderCreateArtificialVariable(c->di,
m->di_method, name, index + 1, m->di_file, (unsigned)ast_line(m->r_fun),
t->di_type);
} else {
info = LLVMDIBuilderCreateParameterVariable(c->di,
m->di_method, name, index + 1, m->di_file, (unsigned)ast_line(m->r_fun),
t->di_type);
}
LLVMMetadataRef expr = LLVMDIBuilderCreateExpression(c->di, NULL, 0);
LLVMDIBuilderInsertDeclare(c->di, alloc, info, expr,
(unsigned)line, (unsigned)pos, m->di_method,
LLVMGetInsertBlock(c->builder));
}
LLVMValueRef gen_localdecl(compile_t* c, ast_t* ast)
{
ast_t* id = ast_child(ast);
ast_t* type = ast_type(id);
const char* name = ast_name(id);
// If this local has already been generated, don't create another copy. This
// can happen when the same ast node is generated more than once, such as
// the condition block of a while expression.
LLVMValueRef value = codegen_getlocal(c, name);
if(value != NULL)
return GEN_NOVALUE;
reach_type_t* t = reach_type(c->reach, type);
// All alloca should happen in the entry block of a function.
LLVMBasicBlockRef this_block = LLVMGetInsertBlock(c->builder);
LLVMBasicBlockRef entry_block = LLVMGetEntryBasicBlock(codegen_fun(c));
LLVMValueRef inst = LLVMGetFirstInstruction(entry_block);
if(inst != NULL)
LLVMPositionBuilderBefore(c->builder, inst);
else
LLVMPositionBuilderAtEnd(c->builder, entry_block);
LLVMValueRef alloc = LLVMBuildAlloca(c->builder, t->use_type, name);
// Store the alloca to use when we reference this local.
codegen_setlocal(c, name, alloc);
LLVMMetadataRef file = codegen_difile(c);
LLVMMetadataRef scope = codegen_discope(c);
LLVMMetadataRef info = LLVMDIBuilderCreateAutoVariable(c->di, scope, name,
file, (unsigned)ast_line(ast), t->di_type);
LLVMMetadataRef expr = LLVMDIBuilderCreateExpression(c->di, NULL, 0);
LLVMDIBuilderInsertDeclare(c->di, alloc, info, expr,
(unsigned)ast_line(ast), (unsigned)ast_pos(ast), scope,
LLVMGetInsertBlock(c->builder));
// Put the builder back where it was.
LLVMPositionBuilderAtEnd(c->builder, this_block);
return GEN_NOVALUE;
}
LLVMValueRef gendesc_istrait(compile_t* c, LLVMValueRef desc, ast_t* type)
{
// Get the trait identifier.
reach_type_t* t = reach_type(c->reach, type);
assert(t != NULL);
LLVMValueRef trait_id = LLVMConstInt(c->i32, t->type_id, false);
// Read the count and the trait list from the descriptor.
LLVMValueRef count = desc_field(c, desc, DESC_TRAIT_COUNT);
LLVMValueRef list = desc_field(c, desc, DESC_TRAITS);
LLVMBasicBlockRef entry_block = LLVMGetInsertBlock(c->builder);
LLVMBasicBlockRef cond_block = codegen_block(c, "cond");
LLVMBasicBlockRef body_block = codegen_block(c, "body");
LLVMBasicBlockRef post_block = codegen_block(c, "post");
LLVMBuildBr(c->builder, cond_block);
// While the index is less than the count, check an ID.
LLVMPositionBuilderAtEnd(c->builder, cond_block);
LLVMValueRef phi = LLVMBuildPhi(c->builder, c->i32, "");
LLVMValueRef zero = LLVMConstInt(c->i32, 0, false);
LLVMAddIncoming(phi, &zero, &entry_block, 1);
LLVMValueRef test = LLVMBuildICmp(c->builder, LLVMIntULT, phi, count, "");
LLVMBuildCondBr(c->builder, test, body_block, post_block);
// The phi node is the index. Get ID and compare it.
LLVMPositionBuilderAtEnd(c->builder, body_block);
LLVMValueRef gep[2];
gep[0] = LLVMConstInt(c->i32, 0, false);
gep[1] = phi;
LLVMValueRef id_ptr = LLVMBuildInBoundsGEP(c->builder, list, gep, 2, "");
LLVMValueRef id = LLVMBuildLoad(c->builder, id_ptr, "");
LLVMValueRef test_id = LLVMBuildICmp(c->builder, LLVMIntEQ, id, trait_id,
"");
// Add one to the phi node.
LLVMValueRef one = LLVMConstInt(c->i32, 1, false);
LLVMValueRef inc = LLVMBuildAdd(c->builder, phi, one, "");
LLVMAddIncoming(phi, &inc, &body_block, 1);
// Either to the post block or back to the condition.
LLVMBuildCondBr(c->builder, test_id, post_block, cond_block);
LLVMPositionBuilderAtEnd(c->builder, post_block);
LLVMValueRef result = LLVMBuildPhi(c->builder, c->i1, "");
LLVMAddIncoming(result, &test, &cond_block, 1);
LLVMAddIncoming(result, &test_id, &body_block, 1);
return result;
}
LLVMValueRef
lp_build_intrinsic(LLVMBuilderRef builder,
const char *name,
LLVMTypeRef ret_type,
LLVMValueRef *args,
unsigned num_args,
unsigned attr_mask)
{
LLVMModuleRef module = LLVMGetGlobalParent(LLVMGetBasicBlockParent(LLVMGetInsertBlock(builder)));
LLVMValueRef function, call;
bool set_callsite_attrs = HAVE_LLVM >= 0x0400 &&
!(attr_mask & LP_FUNC_ATTR_LEGACY);
function = LLVMGetNamedFunction(module, name);
if(!function) {
LLVMTypeRef arg_types[LP_MAX_FUNC_ARGS];
unsigned i;
assert(num_args <= LP_MAX_FUNC_ARGS);
for(i = 0; i < num_args; ++i) {
assert(args[i]);
arg_types[i] = LLVMTypeOf(args[i]);
}
function = lp_declare_intrinsic(module, name, ret_type, arg_types, num_args);
/*
* If llvm removes an intrinsic we use, we'll hit this abort (rather
* than a call to address zero in the jited code).
*/
if (LLVMGetIntrinsicID(function) == 0) {
_debug_printf("llvm (version 0x%x) found no intrinsic for %s, going to crash...\n",
HAVE_LLVM, name);
abort();
}
if (!set_callsite_attrs)
lp_add_func_attributes(function, attr_mask);
if (gallivm_debug & GALLIVM_DEBUG_IR) {
lp_debug_dump_value(function);
}
}
call = LLVMBuildCall(builder, function, args, num_args, "");
if (set_callsite_attrs)
lp_add_func_attributes(call, attr_mask);
return call;
}
static LLVMValueRef
translateArrayExpr(SymbolTable *TyTable, SymbolTable *ValTable, ASTNode *Node) {
PtrVector *V = &(Node->Child);
ASTNode *SizeNode = (ASTNode*) ptrVectorGet(V, 0),
*InitNode = (ASTNode*) ptrVectorGet(V, 1);
Type *ThisType = createType(IdTy, Node->Value);
LLVMTypeRef ArrayType = getLLVMTypeFromType(TyTable, ThisType);
LLVMValueRef SizeVal = translateExpr(TyTable, ValTable, SizeNode),
InitVal = translateExpr(TyTable, ValTable, InitNode);
LLVMValueRef ArrayVal = LLVMBuildArrayMalloc(Builder, ArrayType, SizeVal, "");
// This BasicBlock and ThisFunction
LLVMBasicBlockRef ThisBB = LLVMGetInsertBlock(Builder);
LLVMValueRef ThisFn = LLVMGetBasicBlockParent(ThisBB);
LLVMValueRef Counter = LLVMBuildAlloca(Builder, LLVMInt32Type(), "");
LLVMBuildStore(Builder, LLVMConstInt(LLVMInt32Type(), 0, 1), Counter);
LLVMTargetDataRef DataRef = LLVMCreateTargetData(LLVMGetDataLayout(Module));
unsigned long long Size = LLVMStoreSizeOfType(DataRef, ArrayType);
LLVMBasicBlockRef InitBB, MidBB, EndBB;
InitBB = LLVMAppendBasicBlock(ThisFn, "for.init");
EndBB = LLVMAppendBasicBlock(ThisFn, "for.end");
MidBB = LLVMAppendBasicBlock(ThisFn, "for.mid");
LLVMBuildBr(Builder, InitBB);
LLVMPositionBuilderAtEnd(Builder, InitBB);
LLVMValueRef CurrentCounter = LLVMBuildLoad(Builder, Counter, "");
LLVMValueRef Comparation = LLVMBuildICmp(Builder, LLVMIntSLT, CurrentCounter, SizeVal, "");
LLVMBuildCondBr(Builder, Comparation, MidBB, EndBB);
LLVMPositionBuilderAtEnd(Builder, MidBB);
CurrentCounter = LLVMBuildLoad(Builder, Counter, "");
LLVMValueRef TheValue = LLVMBuildLoad(Builder, InitVal, "");
LLVMValueRef ElemIdx[] = { LLVMConstInt(LLVMInt32Type(), 0, 1), CurrentCounter };
LLVMValueRef Elem = LLVMBuildInBoundsGEP(Builder, ArrayVal, ElemIdx, 2, "");
copyMemory(Elem, TheValue, getSConstInt(Size));
LLVMBuildBr(Builder, InitBB);
LLVMPositionBuilderAtEnd(Builder, EndBB);
return ArrayVal;
}
void JITImpl::ensureEarlyReturnBB(LLVMTypeRef phiType)
{
if (earlyReturnBB)
return;
// Save off current insert point.
LLVMBasicBlockRef savedBB = LLVMGetInsertBlock(builder);
LLVMValueRef f = LLVMGetBasicBlockParent(savedBB);
earlyReturnBB = LLVMAppendBasicBlock(f, "early_return");
LLVMPositionBuilderAtEnd(builder, earlyReturnBB);
// Create phi (incoming values will be filled in later).
earlyReturnPhi = LLVMBuildPhi(builder, phiType, "");
LLVMBuildRet(builder, earlyReturnPhi);
// Restore insert point.
LLVMPositionBuilderAtEnd(builder, savedBB);
}
LLVMValueRef gen_break(compile_t* c, ast_t* ast)
{
ast_t* body = ast_child(ast);
LLVMBasicBlockRef target;
if(ast_id(body) == TK_NONE)
{
target = c->frame->break_novalue_target;
} else {
ast_t* body_type = ast_type(body);
// Get the break target.
target = c->frame->break_target;
// Get the phi node.
LLVMValueRef post_phi = LLVMGetFirstInstruction(target);
bool needed = (post_phi != NULL) && LLVMIsAPHINode(post_phi);
// Build the break expression.
LLVMValueRef value = gen_expr(c, body);
if(needed)
{
// Cast it to the phi type if we need to.
LLVMTypeRef phi_type = LLVMTypeOf(post_phi);
value = gen_assign_cast(c, phi_type, value, body_type);
}
if(value == NULL)
return NULL;
// Add break value to the post block phi node.
if(needed)
{
LLVMBasicBlockRef insert_block = LLVMGetInsertBlock(c->builder);
LLVMAddIncoming(post_phi, &value, &insert_block, 1);
}
}
// Jump to the break target.
codegen_scope_lifetime_end(c);
codegen_debugloc(c, ast);
LLVMBuildBr(c->builder, target);
codegen_debugloc(c, NULL);
return GEN_NOVALUE;
}
LLVMValueRef gen_label(struct node *ast)
{
LLVMBasicBlockRef label_block, prev_block;
prev_block = LLVMGetInsertBlock(builder);
label_block = symtab_find(ast->one->val);
LLVMMoveBasicBlockAfter(label_block, prev_block);
LLVMPositionBuilderAtEnd(builder, prev_block);
LLVMBuildBr(builder, label_block);
LLVMPositionBuilderAtEnd(builder, label_block);
codegen(ast->two);
return NULL;
}
static LLVMBuilderRef
create_builder_at_entry(struct gallivm_state *gallivm)
{
LLVMBuilderRef builder = gallivm->builder;
LLVMBasicBlockRef current_block = LLVMGetInsertBlock(builder);
LLVMValueRef function = LLVMGetBasicBlockParent(current_block);
LLVMBasicBlockRef first_block = LLVMGetEntryBasicBlock(function);
LLVMValueRef first_instr = LLVMGetFirstInstruction(first_block);
LLVMBuilderRef first_builder = LLVMCreateBuilderInContext(gallivm->context);
if (first_instr) {
LLVMPositionBuilderBefore(first_builder, first_instr);
} else {
LLVMPositionBuilderAtEnd(first_builder, first_block);
}
return first_builder;
}
void Define::visit_module(ast::Module& x) {
// Get the existing module-item
auto item = _ctx.modules_by_context[&x];
if (!item) return;
// Add the module initializer basic block
auto last_block = LLVMGetInsertBlock(_ctx.irb);
auto block = LLVMAppendBasicBlock(item->initializer, "");
LLVMPositionBuilderAtEnd(_ctx.irb, block);
// Define any items that need forward declarations.
Define(_ctx, item->scope).run(*x.block);
// Move instruction ptr back to where it was (if it was somewhere)
if (last_block) {
LLVMPositionBuilderAtEnd(_ctx.irb, last_block);
}
}
/**
* 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;
}
void codegen_startfun(compile_t* c, LLVMValueRef fun, bool has_source)
{
compile_frame_t* frame = push_frame(c);
frame->fun = fun;
frame->restore_builder = LLVMGetInsertBlock(c->builder);
frame->has_source = has_source;
frame->is_function = true;
c->dwarf.has_source = has_source;
// Reset debug locations
dwarf_location(&c->dwarf, NULL);
if(LLVMCountBasicBlocks(fun) == 0)
{
LLVMBasicBlockRef block = codegen_block(c, "entry");
LLVMPositionBuilderAtEnd(c->builder, block);
}
}
LLVMValueRef gen_return(struct node *ast)
{
LLVMValueRef func, retval;
LLVMBasicBlockRef next_block, ret_block;
ret_block = symtab_find(".return");
retval = symtab_find(".retval");
if (ast->one)
LLVMBuildStore(builder, codegen(ast->one), retval);
LLVMBuildBr(builder, ret_block);
func = LLVMGetBasicBlockParent(LLVMGetInsertBlock(builder));
next_block = LLVMAppendBasicBlock(func, "");
LLVMPositionBuilderAtEnd(builder, next_block);;
return NULL;
}
LLVMValueRef gen_while(struct node *ast)
{
LLVMValueRef condition, func;
LLVMBasicBlockRef do_block, end;
condition = LLVMBuildICmp(builder, LLVMIntNE, codegen(ast->one), CONST(0), "");
func = LLVMGetBasicBlockParent(LLVMGetInsertBlock(builder));
do_block = LLVMAppendBasicBlock(func, "");
end = LLVMAppendBasicBlock(func, "");
LLVMBuildCondBr(builder, condition, do_block, end);
LLVMPositionBuilderAtEnd(builder, do_block);
codegen(ast->two);
condition = LLVMBuildICmp(builder, LLVMIntNE, codegen(ast->one), CONST(0), "");
LLVMBuildCondBr(builder, condition, do_block, end);
LLVMPositionBuilderAtEnd(builder, end);
return NULL;
}
LLVMValueRef gen_names(struct node *ast)
{
LLVMValueRef func, pam, var;
char *name;
func = LLVMGetBasicBlockParent(LLVMGetInsertBlock(builder));
/* Pam param, pam pam param... */
for (pam = LLVMGetLastParam(func); pam; pam = LLVMGetPreviousParam(pam))
{
name = ast->one->val;
var = LLVMBuildAlloca(builder, TYPE_INT, name);
symtab_enter(name, var);
LLVMBuildStore(builder, pam, var);
ast = ast->two;
}
return NULL;
}
static LLVMValueRef invoke_fun(compile_t* c, LLVMValueRef fun,
LLVMValueRef* args, int count, const char* ret, bool setcc)
{
if(fun == NULL)
return NULL;
LLVMBasicBlockRef this_block = LLVMGetInsertBlock(c->builder);
LLVMBasicBlockRef then_block = LLVMInsertBasicBlockInContext(c->context,
this_block, "invoke");
LLVMMoveBasicBlockAfter(then_block, this_block);
LLVMBasicBlockRef else_block = c->frame->invoke_target;
LLVMValueRef invoke = LLVMBuildInvoke(c->builder, fun, args, count,
then_block, else_block, ret);
if(setcc)
LLVMSetInstructionCallConv(invoke, c->callconv);
LLVMPositionBuilderAtEnd(c->builder, then_block);
return invoke;
}
LLVMValueRef gen_goto(struct node *ast)
{
LLVMValueRef branch, func;
LLVMBasicBlockRef next_block;
int i;
branch = LLVMBuildIndirectBr(builder,
LLVMBuildIntToPtr(builder,
codegen(ast->one),
TYPE_LABEL,
""),
label_count);
for (i = 0; i < label_count; i++)
LLVMAddDestination(branch, label_blocks[i]);
func = LLVMGetBasicBlockParent(LLVMGetInsertBlock(builder));
next_block = LLVMAppendBasicBlock(func, "");
LLVMPositionBuilderAtEnd(builder, next_block);
return NULL;
}
static void createPopHeapFunction() {
// Saving last BasicBlock;
LLVMBasicBlockRef OldBB = LLVMGetInsertBlock(Builder);
LLVMTypeRef FunctionType = LLVMFunctionType(LLVMVoidType(), NULL, 0, 0);
LLVMValueRef Function = LLVMAddFunction(Module, "pop.heap", FunctionType);
LLVMBasicBlockRef Entry = LLVMAppendBasicBlock(Function, "entry");
LLVMPositionBuilderAtEnd(Builder, Entry);
// Function Body
LLVMValueRef HeapHdPtr = LLVMBuildLoad(Builder, HeapHead, "");
LLVMValueRef LastPtrIdx[] = { getSConstInt(0), getSConstInt(1) };
LLVMValueRef LastPtr = LLVMBuildInBoundsGEP(Builder, HeapHdPtr, LastPtrIdx, 2, "heap.last");
LLVMValueRef LastPtrLd = LLVMBuildLoad(Builder, LastPtr, "ld.heap.last");
LLVMBuildStore(Builder, LastPtrLd, HeapHead);
LLVMBuildRetVoid(Builder);
// Restoring last BasicBlock
LLVMPositionBuilderAtEnd(Builder, OldBB);
}
LLVMValueRef codegen_ctx(compile_t* c)
{
compile_frame_t* frame = c->frame;
while(!frame->is_function)
frame = frame->prev;
if(frame->ctx == NULL)
{
LLVMBasicBlockRef this_block = LLVMGetInsertBlock(c->builder);
LLVMBasicBlockRef entry_block = LLVMGetEntryBasicBlock(frame->fun);
LLVMValueRef inst = LLVMGetFirstInstruction(entry_block);
if(inst != NULL)
LLVMPositionBuilderBefore(c->builder, inst);
else
LLVMPositionBuilderAtEnd(c->builder, entry_block);
frame->ctx = gencall_runtime(c, "pony_ctx", NULL, 0, "");
LLVMPositionBuilderAtEnd(c->builder, this_block);
}
return frame->ctx;
}
/**
* Insert a new block, right where builder is pointing to.
*
* This is useful important not only for aesthetic reasons, but also for
* performance reasons, as frequently run blocks should be laid out next to
* each other and fall-throughs maximized.
*
* See also llvm/lib/Transforms/Scalar/BasicBlockPlacement.cpp.
*
* Note: this function has no dependencies on the flow code and could
* be used elsewhere.
*/
LLVMBasicBlockRef
lp_build_insert_new_block(struct gallivm_state *gallivm, const char *name)
{
LLVMBasicBlockRef current_block;
LLVMBasicBlockRef next_block;
LLVMBasicBlockRef new_block;
/* get current basic block */
current_block = LLVMGetInsertBlock(gallivm->builder);
/* check if there's another block after this one */
next_block = LLVMGetNextBasicBlock(current_block);
if (next_block) {
/* insert the new block before the next block */
new_block = LLVMInsertBasicBlockInContext(gallivm->context, next_block, name);
}
else {
/* append new block after current block */
LLVMValueRef function = LLVMGetBasicBlockParent(current_block);
new_block = LLVMAppendBasicBlockInContext(gallivm->context, function, name);
}
return new_block;
}
/**
* Begin an if/else/endif construct.
*/
void
lp_build_if(struct lp_build_if_state *ifthen,
struct gallivm_state *gallivm,
LLVMValueRef condition)
{
LLVMBasicBlockRef block = LLVMGetInsertBlock(gallivm->builder);
memset(ifthen, 0, sizeof *ifthen);
ifthen->gallivm = gallivm;
ifthen->condition = condition;
ifthen->entry_block = block;
/* create endif/merge basic block for the phi functions */
ifthen->merge_block = lp_build_insert_new_block(gallivm, "endif-block");
/* create/insert true_block before merge_block */
ifthen->true_block =
LLVMInsertBasicBlockInContext(gallivm->context,
ifthen->merge_block,
"if-true-block");
/* successive code goes into the true block */
LLVMPositionBuilderAtEnd(gallivm->builder, ifthen->true_block);
}
static bool case_body(compile_t* c, ast_t* body,
LLVMBasicBlockRef post_block, LLVMValueRef phi, LLVMTypeRef phi_type)
{
LLVMValueRef body_value = gen_expr(c, body);
// If it returns, we don't branch to the post block.
if(body_value == GEN_NOVALUE)
return true;
if(is_result_needed(body))
{
ast_t* body_type = ast_type(body);
body_value = gen_assign_cast(c, phi_type, body_value, body_type);
if(body_value == NULL)
return false;
LLVMBasicBlockRef block = LLVMGetInsertBlock(c->builder);
LLVMAddIncoming(phi, &body_value, &block, 1);
}
LLVMBuildBr(c->builder, post_block);
return true;
}
/**
* Fetch a pixel into a 4 float AoS.
*
* \param format_desc describes format of the image we're fetching from
* \param ptr address of the pixel block (or the texel if uncompressed)
* \param i, j the sub-block pixel coordinates. For non-compressed formats
* these will always be (0, 0).
* \return a 4 element vector with the pixel's RGBA values.
*/
LLVMValueRef
lp_build_fetch_rgba_aos(struct gallivm_state *gallivm,
const struct util_format_description *format_desc,
struct lp_type type,
LLVMValueRef base_ptr,
LLVMValueRef offset,
LLVMValueRef i,
LLVMValueRef j)
{
LLVMBuilderRef builder = gallivm->builder;
unsigned num_pixels = type.length / 4;
struct lp_build_context bld;
assert(type.length <= LP_MAX_VECTOR_LENGTH);
assert(type.length % 4 == 0);
lp_build_context_init(&bld, gallivm, type);
/*
* Trivial case
*
* The format matches the type (apart of a swizzle) so no need for
* scaling or converting.
*/
if (format_matches_type(format_desc, type) &&
format_desc->block.bits <= type.width * 4 &&
util_is_power_of_two(format_desc->block.bits)) {
LLVMValueRef packed;
/*
* The format matches the type (apart of a swizzle) so no need for
* scaling or converting.
*/
packed = lp_build_gather(gallivm, type.length/4,
format_desc->block.bits, type.width*4,
base_ptr, offset);
assert(format_desc->block.bits <= type.width * type.length);
packed = LLVMBuildBitCast(gallivm->builder, packed,
lp_build_vec_type(gallivm, type), "");
return lp_build_format_swizzle_aos(format_desc, &bld, packed);
}
/*
* Bit arithmetic
*/
if (format_desc->layout == UTIL_FORMAT_LAYOUT_PLAIN &&
(format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB ||
format_desc->colorspace == UTIL_FORMAT_COLORSPACE_ZS) &&
format_desc->block.width == 1 &&
format_desc->block.height == 1 &&
util_is_power_of_two(format_desc->block.bits) &&
format_desc->block.bits <= 32 &&
format_desc->is_bitmask &&
!format_desc->is_mixed &&
(format_desc->channel[0].type == UTIL_FORMAT_TYPE_UNSIGNED ||
format_desc->channel[1].type == UTIL_FORMAT_TYPE_UNSIGNED)) {
LLVMValueRef tmps[LP_MAX_VECTOR_LENGTH/4];
LLVMValueRef res;
unsigned k;
/*
* Unpack a pixel at a time into a <4 x float> RGBA vector
*/
for (k = 0; k < num_pixels; ++k) {
LLVMValueRef packed;
packed = lp_build_gather_elem(gallivm, num_pixels,
format_desc->block.bits, 32,
base_ptr, offset, k);
tmps[k] = lp_build_unpack_arith_rgba_aos(gallivm,
format_desc,
packed);
}
/*
* Type conversion.
*
* TODO: We could avoid floating conversion for integer to
* integer conversions.
*/
if (gallivm_debug & GALLIVM_DEBUG_PERF && !type.floating) {
debug_printf("%s: unpacking %s with floating point\n",
__FUNCTION__, format_desc->short_name);
}
lp_build_conv(gallivm,
lp_float32_vec4_type(),
type,
tmps, num_pixels, &res, 1);
return lp_build_format_swizzle_aos(format_desc, &bld, res);
}
/*
* YUV / subsampled formats
*/
if (format_desc->layout == UTIL_FORMAT_LAYOUT_SUBSAMPLED) {
struct lp_type tmp_type;
LLVMValueRef tmp;
memset(&tmp_type, 0, sizeof tmp_type);
tmp_type.width = 8;
tmp_type.length = num_pixels * 4;
tmp_type.norm = TRUE;
tmp = lp_build_fetch_subsampled_rgba_aos(gallivm,
format_desc,
num_pixels,
base_ptr,
offset,
i, j);
lp_build_conv(gallivm,
tmp_type, type,
&tmp, 1, &tmp, 1);
return tmp;
}
/*
* Fallback to util_format_description::fetch_rgba_8unorm().
*/
if (format_desc->fetch_rgba_8unorm &&
!type.floating && type.width == 8 && !type.sign && type.norm) {
/*
* Fallback to calling util_format_description::fetch_rgba_8unorm.
*
* This is definitely not the most efficient way of fetching pixels, as
* we miss the opportunity to do vectorization, but this it is a
* convenient for formats or scenarios for which there was no opportunity
* or incentive to optimize.
*/
LLVMModuleRef module = LLVMGetGlobalParent(LLVMGetBasicBlockParent(LLVMGetInsertBlock(gallivm->builder)));
char name[256];
LLVMTypeRef i8t = LLVMInt8TypeInContext(gallivm->context);
LLVMTypeRef pi8t = LLVMPointerType(i8t, 0);
LLVMTypeRef i32t = LLVMInt32TypeInContext(gallivm->context);
LLVMValueRef function;
LLVMValueRef tmp_ptr;
LLVMValueRef tmp;
LLVMValueRef res;
LLVMValueRef callee;
unsigned k;
util_snprintf(name, sizeof name, "util_format_%s_fetch_rgba_8unorm",
format_desc->short_name);
if (gallivm_debug & GALLIVM_DEBUG_PERF) {
debug_printf("%s: falling back to %s\n", __FUNCTION__, name);
}
/*
* Declare and bind format_desc->fetch_rgba_8unorm().
*/
function = LLVMGetNamedFunction(module, name);
if (!function) {
/*
* Function to call looks like:
* fetch(uint8_t *dst, const uint8_t *src, unsigned i, unsigned j)
*/
LLVMTypeRef ret_type;
LLVMTypeRef arg_types[4];
LLVMTypeRef function_type;
ret_type = LLVMVoidTypeInContext(gallivm->context);
arg_types[0] = pi8t;
arg_types[1] = pi8t;
arg_types[2] = i32t;
arg_types[3] = i32t;
function_type = LLVMFunctionType(ret_type, arg_types,
Elements(arg_types), 0);
function = LLVMAddFunction(module, name, function_type);
LLVMSetFunctionCallConv(function, LLVMCCallConv);
LLVMSetLinkage(function, LLVMExternalLinkage);
assert(LLVMIsDeclaration(function));
}
/* make const pointer for the C fetch_rgba_float function */
callee = lp_build_const_int_pointer(gallivm,
func_to_pointer((func_pointer) format_desc->fetch_rgba_8unorm));
/* cast the callee pointer to the function's type */
function = LLVMBuildBitCast(builder, callee,
LLVMTypeOf(function), "cast callee");
tmp_ptr = lp_build_alloca(gallivm, i32t, "");
res = LLVMGetUndef(LLVMVectorType(i32t, num_pixels));
/*
* Invoke format_desc->fetch_rgba_8unorm() for each pixel and insert the result
* in the SoA vectors.
*/
for (k = 0; k < num_pixels; ++k) {
LLVMValueRef index = lp_build_const_int32(gallivm, k);
LLVMValueRef args[4];
args[0] = LLVMBuildBitCast(builder, tmp_ptr, pi8t, "");
args[1] = lp_build_gather_elem_ptr(gallivm, num_pixels,
base_ptr, offset, k);
if (num_pixels == 1) {
args[2] = i;
args[3] = j;
}
else {
args[2] = LLVMBuildExtractElement(builder, i, index, "");
args[3] = LLVMBuildExtractElement(builder, j, index, "");
}
LLVMBuildCall(builder, function, args, Elements(args), "");
tmp = LLVMBuildLoad(builder, tmp_ptr, "");
if (num_pixels == 1) {
res = tmp;
}
else {
res = LLVMBuildInsertElement(builder, res, tmp, index, "");
}
}
/* Bitcast from <n x i32> to <4n x i8> */
res = LLVMBuildBitCast(builder, res, bld.vec_type, "");
return res;
}
/*
* Fallback to util_format_description::fetch_rgba_float().
*/
if (format_desc->fetch_rgba_float) {
/*
* Fallback to calling util_format_description::fetch_rgba_float.
*
* This is definitely not the most efficient way of fetching pixels, as
* we miss the opportunity to do vectorization, but this it is a
* convenient for formats or scenarios for which there was no opportunity
* or incentive to optimize.
*/
LLVMModuleRef module = LLVMGetGlobalParent(LLVMGetBasicBlockParent(LLVMGetInsertBlock(builder)));
char name[256];
LLVMTypeRef f32t = LLVMFloatTypeInContext(gallivm->context);
LLVMTypeRef f32x4t = LLVMVectorType(f32t, 4);
LLVMTypeRef pf32t = LLVMPointerType(f32t, 0);
LLVMTypeRef pi8t = LLVMPointerType(LLVMInt8TypeInContext(gallivm->context), 0);
LLVMTypeRef i32t = LLVMInt32TypeInContext(gallivm->context);
LLVMValueRef function;
LLVMValueRef tmp_ptr;
LLVMValueRef tmps[LP_MAX_VECTOR_LENGTH/4];
LLVMValueRef res;
LLVMValueRef callee;
unsigned k;
util_snprintf(name, sizeof name, "util_format_%s_fetch_rgba_float",
format_desc->short_name);
if (gallivm_debug & GALLIVM_DEBUG_PERF) {
debug_printf("%s: falling back to %s\n", __FUNCTION__, name);
}
/*
* Declare and bind format_desc->fetch_rgba_float().
*/
function = LLVMGetNamedFunction(module, name);
if (!function) {
/*
* Function to call looks like:
* fetch(float *dst, const uint8_t *src, unsigned i, unsigned j)
*/
LLVMTypeRef ret_type;
LLVMTypeRef arg_types[4];
LLVMTypeRef function_type;
ret_type = LLVMVoidTypeInContext(gallivm->context);
arg_types[0] = pf32t;
arg_types[1] = pi8t;
arg_types[2] = i32t;
arg_types[3] = i32t;
function_type = LLVMFunctionType(ret_type, arg_types,
Elements(arg_types), 0);
function = LLVMAddFunction(module, name, function_type);
LLVMSetFunctionCallConv(function, LLVMCCallConv);
LLVMSetLinkage(function, LLVMExternalLinkage);
assert(LLVMIsDeclaration(function));
}
/* Note: we're using this casting here instead of LLVMAddGlobalMapping()
* to work around a bug in LLVM 2.6.
*/
/* make const pointer for the C fetch_rgba_float function */
callee = lp_build_const_int_pointer(gallivm,
func_to_pointer((func_pointer) format_desc->fetch_rgba_float));
/* cast the callee pointer to the function's type */
function = LLVMBuildBitCast(builder, callee,
LLVMTypeOf(function), "cast callee");
tmp_ptr = lp_build_alloca(gallivm, f32x4t, "");
/*
* Invoke format_desc->fetch_rgba_float() for each pixel and insert the result
* in the SoA vectors.
*/
for (k = 0; k < num_pixels; ++k) {
LLVMValueRef args[4];
args[0] = LLVMBuildBitCast(builder, tmp_ptr, pf32t, "");
args[1] = lp_build_gather_elem_ptr(gallivm, num_pixels,
base_ptr, offset, k);
if (num_pixels == 1) {
args[2] = i;
args[3] = j;
}
else {
LLVMValueRef index = lp_build_const_int32(gallivm, k);
args[2] = LLVMBuildExtractElement(builder, i, index, "");
args[3] = LLVMBuildExtractElement(builder, j, index, "");
}
LLVMBuildCall(builder, function, args, Elements(args), "");
tmps[k] = LLVMBuildLoad(builder, tmp_ptr, "");
}
lp_build_conv(gallivm,
lp_float32_vec4_type(),
type,
tmps, num_pixels, &res, 1);
return res;
}
assert(0);
return lp_build_undef(gallivm, type);
}
LLVMValueRef JITImpl::getCurrentFunction()
{
return LLVMGetBasicBlockParent(LLVMGetInsertBlock(builder));
}
void
lp_build_tgsi_aos(struct gallivm_state *gallivm,
const struct tgsi_token *tokens,
struct lp_type type,
const unsigned char swizzles[4],
LLVMValueRef consts_ptr,
const LLVMValueRef *inputs,
LLVMValueRef *outputs,
struct lp_build_sampler_aos *sampler,
const struct tgsi_shader_info *info)
{
struct lp_build_tgsi_aos_context bld;
struct tgsi_parse_context parse;
uint num_immediates = 0;
unsigned chan;
int pc = 0;
/* Setup build context */
memset(&bld, 0, sizeof bld);
lp_build_context_init(&bld.bld_base.base, gallivm, type);
lp_build_context_init(&bld.bld_base.uint_bld, gallivm, lp_uint_type(type));
lp_build_context_init(&bld.bld_base.int_bld, gallivm, lp_int_type(type));
lp_build_context_init(&bld.int_bld, gallivm, lp_int_type(type));
for (chan = 0; chan < 4; ++chan) {
bld.swizzles[chan] = swizzles[chan];
bld.inv_swizzles[swizzles[chan]] = chan;
}
bld.inputs = inputs;
bld.outputs = outputs;
bld.consts_ptr = consts_ptr;
bld.sampler = sampler;
bld.indirect_files = info->indirect_files;
bld.bld_base.emit_swizzle = swizzle_aos;
bld.bld_base.info = info;
bld.bld_base.emit_fetch_funcs[TGSI_FILE_CONSTANT] = emit_fetch_constant;
bld.bld_base.emit_fetch_funcs[TGSI_FILE_IMMEDIATE] = emit_fetch_immediate;
bld.bld_base.emit_fetch_funcs[TGSI_FILE_INPUT] = emit_fetch_input;
bld.bld_base.emit_fetch_funcs[TGSI_FILE_TEMPORARY] = emit_fetch_temporary;
/* Set opcode actions */
lp_set_default_actions_cpu(&bld.bld_base);
if (!lp_bld_tgsi_list_init(&bld.bld_base)) {
return;
}
tgsi_parse_init(&parse, tokens);
while (!tgsi_parse_end_of_tokens(&parse)) {
tgsi_parse_token(&parse);
switch(parse.FullToken.Token.Type) {
case TGSI_TOKEN_TYPE_DECLARATION:
/* Inputs already interpolated */
lp_emit_declaration_aos(&bld, &parse.FullToken.FullDeclaration);
break;
case TGSI_TOKEN_TYPE_INSTRUCTION:
/* save expanded instruction */
lp_bld_tgsi_add_instruction(&bld.bld_base,
&parse.FullToken.FullInstruction);
break;
case TGSI_TOKEN_TYPE_IMMEDIATE:
/* simply copy the immediate values into the next immediates[] slot */
{
const uint size = parse.FullToken.FullImmediate.Immediate.NrTokens - 1;
float imm[4];
assert(size <= 4);
assert(num_immediates < LP_MAX_TGSI_IMMEDIATES);
for (chan = 0; chan < 4; ++chan) {
imm[chan] = 0.0f;
}
for (chan = 0; chan < size; ++chan) {
unsigned swizzle = bld.swizzles[chan];
imm[swizzle] = parse.FullToken.FullImmediate.u[chan].Float;
}
bld.immediates[num_immediates] =
lp_build_const_aos(gallivm, type,
imm[0], imm[1], imm[2], imm[3],
NULL);
num_immediates++;
}
break;
case TGSI_TOKEN_TYPE_PROPERTY:
break;
default:
assert(0);
}
}
while (pc != -1) {
struct tgsi_full_instruction *instr = bld.bld_base.instructions + pc;
const struct tgsi_opcode_info *opcode_info =
tgsi_get_opcode_info(instr->Instruction.Opcode);
if (!lp_emit_instruction_aos(&bld, instr, opcode_info, &pc))
_debug_printf("warning: failed to translate tgsi opcode %s to LLVM\n",
opcode_info->mnemonic);
}
if (0) {
LLVMBasicBlockRef block = LLVMGetInsertBlock(gallivm->builder);
LLVMValueRef function = LLVMGetBasicBlockParent(block);
debug_printf("11111111111111111111111111111 \n");
tgsi_dump(tokens, 0);
lp_debug_dump_value(function);
debug_printf("2222222222222222222222222222 \n");
}
tgsi_parse_free(&parse);
FREE(bld.bld_base.instructions);
if (0) {
LLVMModuleRef module = LLVMGetGlobalParent(
LLVMGetBasicBlockParent(LLVMGetInsertBlock(gallivm->builder)));
LLVMDumpModule(module);
}
}
void Build::visit_select(ast::Select& x) {
// Get the current LLVM block and function
auto current_block = LLVMGetInsertBlock(_ctx.irb);
auto current_fn = LLVMGetBasicBlockParent(current_block);
std::vector<LLVMBasicBlockRef> blocks;
std::vector<LLVMValueRef> values;
std::vector<LLVMBasicBlockRef> value_blocks;
// Resolve the type of us (in full; only used if we have a value)
auto type = Resolve(_scope).run(x);
// A Select expression has a value IIF it has an else and
// each of its branches have a value
bool has_value = !!type;
auto do_select_branch = [&](ast::Block& block) -> bool {
// Build the block
auto value = Build(_ctx, _scope).run_scalar(block);
auto iblock = LLVMGetInsertBlock(_ctx.irb);
if (has_value && value && !value->type.is<code::TypeNone>()) {
// Cast the value to the type analyzed result
value = util::cast(_ctx, value, block, type, false);
if (!value) return false;
// Append to the value chain
values.push_back(value->get_value(_ctx));
value_blocks.push_back(iblock);
} else if (!LLVMGetBasicBlockTerminator(iblock)) {
// This block wasn't terminated and it has no value
// We no longer have a value
has_value = false;
}
// Append to the block chain
blocks.push_back(iblock);
return true;
};
// Iterate through each branch and build its contained block ..
for (auto& br : x.branches) {
// Build the condition expression
auto cond = Build(_ctx, _scope).run_scalar(*br->condition);
if (!cond) return;
// Create the THEN and NEXT LLVM blocks
auto then_block = LLVMAppendBasicBlock(current_fn, "select-then");
auto next_block = LLVMAppendBasicBlock(current_fn, "select-next");
// Build the conditional branch
LLVMBuildCondBr(_ctx.irb, cond->get_value(_ctx), then_block, next_block);
// Activate the THEN block
LLVMPositionBuilderAtEnd(_ctx.irb, then_block);
// Process the branch ..
if (!do_select_branch(*br->block)) return;
// Insert the `next` block after our current block.
LLVMMoveBasicBlockAfter(next_block, LLVMGetInsertBlock(_ctx.irb));
// Replace the outer-block with our new "merge" block.
LLVMPositionBuilderAtEnd(_ctx.irb, next_block);
}
// Check for and build the else_block
LLVMBasicBlockRef merge_block;
if (x.else_block) {
// Process the branch ..
do_select_branch(*x.else_block);
// Create the final "merge" block
merge_block = LLVMAppendBasicBlock(current_fn, "select-merge");
} else {
// Use the elided "else" block as the "merge" block
merge_block = LLVMGetLastBasicBlock(current_fn);
}
// Iterate through the established branches and have them return to
// the "merge" block (if they are not otherwise terminated).
unsigned term = 0;
for (auto& ib : blocks) {
if (!LLVMGetBasicBlockTerminator(ib)) {
// Insert the non-conditional branch.
LLVMPositionBuilderAtEnd(_ctx.irb, ib);
LLVMBuildBr(_ctx.irb, merge_block);
} else {
term += 1;
}
}
// If all blocks were terminated and there is an ELSE present;
// remove the merge block
if (term == blocks.size() && x.else_block) {
LLVMDeleteBasicBlock(merge_block);
}
// Re-establish our insertion point.
LLVMPositionBuilderAtEnd(_ctx.irb, merge_block);
// If we still have a value ..
if (has_value && values.size() > 0) {
// Make the PHI value
auto res = LLVMBuildPhi(_ctx.irb, type->handle(), "");
LLVMAddIncoming(res, values.data(), value_blocks.data(), values.size());
Ref<code::Value> res_value = new code::Value(res, type);
_stack.push_front(res_value);
}
}
LLVMValueRef make_divmod(compile_t* c, ast_t* left, ast_t* right,
const_binop const_f, const_binop const_ui, const_binop const_si,
build_binop build_f, build_binop build_ui, build_binop build_si)
{
ast_t* type = ast_type(left);
bool sign = is_signed(c->opt, type);
LLVMValueRef l_value = gen_expr(c, left);
LLVMValueRef r_value = gen_expr(c, right);
if((l_value == NULL) || (r_value == NULL))
return NULL;
if(!is_fp(r_value) &&
LLVMIsConstant(r_value) &&
(LLVMConstIntGetSExtValue(r_value) == 0)
)
{
ast_error(right, "constant divide or mod by zero");
return NULL;
}
if(LLVMIsConstant(l_value) && LLVMIsConstant(r_value))
{
if(is_fp(l_value))
return const_f(l_value, r_value);
if(sign)
return const_si(l_value, r_value);
return const_ui(l_value, r_value);
}
if(is_fp(l_value))
return build_f(c->builder, l_value, r_value, "");
// Setup additional blocks.
LLVMBasicBlockRef insert = LLVMGetInsertBlock(c->builder);
LLVMBasicBlockRef then_block = codegen_block(c, "div_then");
LLVMBasicBlockRef post_block = codegen_block(c, "div_post");
// Check for div by zero.
LLVMTypeRef r_type = LLVMTypeOf(r_value);
LLVMValueRef zero = LLVMConstInt(r_type, 0, false);
LLVMValueRef cmp = LLVMBuildICmp(c->builder, LLVMIntNE, r_value, zero, "");
LLVMBuildCondBr(c->builder, cmp, then_block, post_block);
// Divisor is not zero.
LLVMPositionBuilderAtEnd(c->builder, then_block);
LLVMValueRef result;
if(sign)
result = build_si(c->builder, l_value, r_value, "");
else
result = build_ui(c->builder, l_value, r_value, "");
LLVMBuildBr(c->builder, post_block);
// Phi node.
LLVMPositionBuilderAtEnd(c->builder, post_block);
LLVMValueRef phi = LLVMBuildPhi(c->builder, r_type, "");
LLVMAddIncoming(phi, &zero, &insert, 1);
LLVMAddIncoming(phi, &result, &then_block, 1);
return phi;
}
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(is_constant_i1(c, l_value))
{
LLVMInstructionEraseFromParent(branch);
LLVMDeleteBasicBlock(left_block);
LLVMPositionBuilderAtEnd(c->builder, entry_block);
if(is_and)
{
if(is_always_false(c, l_value))
return gen_expr(c, left);
} else {
if(is_always_true(c, 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");
if(is_and)
LLVMBuildCondBr(c->builder, l_value, right_block, post_block);
else
LLVMBuildCondBr(c->builder, l_value, 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->i1, "");
LLVMAddIncoming(phi, &l_value, &left_exit_block, 1);
LLVMAddIncoming(phi, &r_value, &right_exit_block, 1);
if(is_constant_i1(c, r_value))
{
if(is_and)
{
if(is_always_false(c, r_value))
return r_value;
} else {
if(is_always_true(c, r_value))
return r_value;
}
return l_value;
}
return phi;
}
void genprim_array_deserialise(compile_t* c, reach_type_t* t)
{
// Generate the deserisalise function.
t->deserialise_fn = codegen_addfun(c, genname_serialise(t->name),
c->trace_type);
codegen_startfun(c, t->deserialise_fn, NULL, NULL);
LLVMSetFunctionCallConv(t->deserialise_fn, LLVMCCallConv);
LLVMValueRef ctx = LLVMGetParam(t->deserialise_fn, 0);
LLVMValueRef arg = LLVMGetParam(t->deserialise_fn, 1);
LLVMValueRef object = LLVMBuildBitCast(c->builder, arg, t->structure_ptr,
"");
gendeserialise_typeid(c, t, object);
// Deserialise the array contents.
LLVMValueRef alloc = field_value(c, object, 2);
LLVMValueRef ptr_offset = field_value(c, object, 3);
ptr_offset = LLVMBuildPtrToInt(c->builder, ptr_offset, c->intptr, "");
ast_t* typeargs = ast_childidx(t->ast, 2);
ast_t* typearg = ast_child(typeargs);
reach_type_t* t_elem = reach_type(c->reach, typearg);
size_t abisize = (size_t)LLVMABISizeOfType(c->target_data, t_elem->use_type);
LLVMValueRef l_size = LLVMConstInt(c->intptr, abisize, false);
LLVMValueRef args[3];
args[0] = ctx;
args[1] = ptr_offset;
args[2] = LLVMBuildMul(c->builder, alloc, l_size, "");
LLVMValueRef ptr = gencall_runtime(c, "pony_deserialise_block", args, 3, "");
LLVMValueRef ptr_loc = LLVMBuildStructGEP(c->builder, object, 3, "");
LLVMBuildStore(c->builder, ptr, ptr_loc);
if((t_elem->underlying == TK_PRIMITIVE) && (t_elem->primitive != NULL))
{
// Do nothing. A memcpy is sufficient.
} else {
LLVMValueRef size = field_value(c, object, 1);
ptr = LLVMBuildBitCast(c->builder, ptr,
LLVMPointerType(t_elem->use_type, 0), "");
LLVMBasicBlockRef entry_block = LLVMGetInsertBlock(c->builder);
LLVMBasicBlockRef cond_block = codegen_block(c, "cond");
LLVMBasicBlockRef body_block = codegen_block(c, "body");
LLVMBasicBlockRef post_block = codegen_block(c, "post");
LLVMBuildBr(c->builder, cond_block);
// While the index is less than the size, deserialise an element. The
// initial index when coming from the entry block is zero.
LLVMPositionBuilderAtEnd(c->builder, cond_block);
LLVMValueRef phi = LLVMBuildPhi(c->builder, c->intptr, "");
LLVMValueRef zero = LLVMConstInt(c->intptr, 0, false);
LLVMAddIncoming(phi, &zero, &entry_block, 1);
LLVMValueRef test = LLVMBuildICmp(c->builder, LLVMIntULT, phi, size, "");
LLVMBuildCondBr(c->builder, test, body_block, post_block);
// The phi node is the index. Get the element and deserialise it.
LLVMPositionBuilderAtEnd(c->builder, body_block);
LLVMValueRef elem_ptr = LLVMBuildGEP(c->builder, ptr, &phi, 1, "");
gendeserialise_element(c, t_elem, false, ctx, elem_ptr);
// Add one to the phi node and branch back to the cond block.
LLVMValueRef one = LLVMConstInt(c->intptr, 1, false);
LLVMValueRef inc = LLVMBuildAdd(c->builder, phi, one, "");
body_block = LLVMGetInsertBlock(c->builder);
LLVMAddIncoming(phi, &inc, &body_block, 1);
LLVMBuildBr(c->builder, cond_block);
LLVMPositionBuilderAtEnd(c->builder, post_block);
}
LLVMBuildRetVoid(c->builder);
codegen_finishfun(c);
}
void genprim_array_serialise(compile_t* c, reach_type_t* t)
{
// Generate the serialise function.
t->serialise_fn = codegen_addfun(c, genname_serialise(t->name),
c->serialise_type);
codegen_startfun(c, t->serialise_fn, NULL, NULL);
LLVMSetFunctionCallConv(t->serialise_fn, LLVMCCallConv);
LLVMValueRef ctx = LLVMGetParam(t->serialise_fn, 0);
LLVMValueRef arg = LLVMGetParam(t->serialise_fn, 1);
LLVMValueRef addr = LLVMGetParam(t->serialise_fn, 2);
LLVMValueRef offset = LLVMGetParam(t->serialise_fn, 3);
LLVMValueRef mut = LLVMGetParam(t->serialise_fn, 4);
LLVMValueRef object = LLVMBuildBitCast(c->builder, arg, t->structure_ptr,
"");
LLVMValueRef offset_addr = LLVMBuildAdd(c->builder,
LLVMBuildPtrToInt(c->builder, addr, c->intptr, ""), offset, "");
genserialise_typeid(c, t, offset_addr);
// Don't serialise our contents if we are opaque.
LLVMBasicBlockRef body_block = codegen_block(c, "body");
LLVMBasicBlockRef post_block = codegen_block(c, "post");
LLVMValueRef test = LLVMBuildICmp(c->builder, LLVMIntNE, mut,
LLVMConstInt(c->i32, PONY_TRACE_OPAQUE, false), "");
LLVMBuildCondBr(c->builder, test, body_block, post_block);
LLVMPositionBuilderAtEnd(c->builder, body_block);
// Write the size twice, effectively rewriting alloc to be the same as size.
LLVMValueRef size = field_value(c, object, 1);
LLVMValueRef size_loc = field_loc(c, offset_addr, t->structure,
c->intptr, 1);
LLVMBuildStore(c->builder, size, size_loc);
LLVMValueRef alloc_loc = field_loc(c, offset_addr, t->structure,
c->intptr, 2);
LLVMBuildStore(c->builder, size, alloc_loc);
// Write the pointer.
LLVMValueRef ptr = field_value(c, object, 3);
// The resulting offset will only be invalid (i.e. have the high bit set) if
// the size is zero. For an opaque array, we don't serialise the contents,
// so we don't get here, so we don't end up with an invalid offset.
LLVMValueRef args[5];
args[0] = ctx;
args[1] = ptr;
LLVMValueRef ptr_offset = gencall_runtime(c, "pony_serialise_offset",
args, 2, "");
LLVMValueRef ptr_loc = field_loc(c, offset_addr, t->structure, c->intptr, 3);
LLVMBuildStore(c->builder, ptr_offset, ptr_loc);
LLVMValueRef ptr_offset_addr = LLVMBuildAdd(c->builder, ptr_offset,
LLVMBuildPtrToInt(c->builder, addr, c->intptr, ""), "");
// Serialise elements.
ast_t* typeargs = ast_childidx(t->ast, 2);
ast_t* typearg = ast_child(typeargs);
reach_type_t* t_elem = reach_type(c->reach, typearg);
size_t abisize = (size_t)LLVMABISizeOfType(c->target_data, t_elem->use_type);
LLVMValueRef l_size = LLVMConstInt(c->intptr, abisize, false);
if((t_elem->underlying == TK_PRIMITIVE) && (t_elem->primitive != NULL))
{
// memcpy machine words
args[0] = LLVMBuildIntToPtr(c->builder, ptr_offset_addr, c->void_ptr, "");
args[1] = LLVMBuildBitCast(c->builder, ptr, c->void_ptr, "");
args[2] = LLVMBuildMul(c->builder, size, l_size, "");
args[3] = LLVMConstInt(c->i32, 1, false);
args[4] = LLVMConstInt(c->i1, 0, false);
if(target_is_ilp32(c->opt->triple))
{
gencall_runtime(c, "llvm.memcpy.p0i8.p0i8.i32", args, 5, "");
} else {
gencall_runtime(c, "llvm.memcpy.p0i8.p0i8.i64", args, 5, "");
}
} else {
ptr = LLVMBuildBitCast(c->builder, ptr,
LLVMPointerType(t_elem->use_type, 0), "");
LLVMBasicBlockRef entry_block = LLVMGetInsertBlock(c->builder);
LLVMBasicBlockRef cond_block = codegen_block(c, "cond");
LLVMBasicBlockRef body_block = codegen_block(c, "body");
LLVMBasicBlockRef post_block = codegen_block(c, "post");
LLVMValueRef offset_var = LLVMBuildAlloca(c->builder, c->intptr, "");
LLVMBuildStore(c->builder, ptr_offset_addr, offset_var);
LLVMBuildBr(c->builder, cond_block);
// While the index is less than the size, serialise an element. The
// initial index when coming from the entry block is zero.
LLVMPositionBuilderAtEnd(c->builder, cond_block);
LLVMValueRef phi = LLVMBuildPhi(c->builder, c->intptr, "");
LLVMValueRef zero = LLVMConstInt(c->intptr, 0, false);
LLVMAddIncoming(phi, &zero, &entry_block, 1);
LLVMValueRef test = LLVMBuildICmp(c->builder, LLVMIntULT, phi, size, "");
LLVMBuildCondBr(c->builder, test, body_block, post_block);
// The phi node is the index. Get the element and serialise it.
LLVMPositionBuilderAtEnd(c->builder, body_block);
LLVMValueRef elem_ptr = LLVMBuildGEP(c->builder, ptr, &phi, 1, "");
ptr_offset_addr = LLVMBuildLoad(c->builder, offset_var, "");
genserialise_element(c, t_elem, false, ctx, elem_ptr, ptr_offset_addr);
ptr_offset_addr = LLVMBuildAdd(c->builder, ptr_offset_addr, l_size, "");
LLVMBuildStore(c->builder, ptr_offset_addr, offset_var);
// Add one to the phi node and branch back to the cond block.
LLVMValueRef one = LLVMConstInt(c->intptr, 1, false);
LLVMValueRef inc = LLVMBuildAdd(c->builder, phi, one, "");
body_block = LLVMGetInsertBlock(c->builder);
LLVMAddIncoming(phi, &inc, &body_block, 1);
LLVMBuildBr(c->builder, cond_block);
LLVMPositionBuilderAtEnd(c->builder, post_block);
}
LLVMBuildBr(c->builder, post_block);
LLVMPositionBuilderAtEnd(c->builder, post_block);
LLVMBuildRetVoid(c->builder);
codegen_finishfun(c);
}
