LLVMModuleRef r600_tgsi_llvm(
struct radeon_llvm_context * ctx,
const struct tgsi_token * tokens)
{
struct tgsi_shader_info shader_info;
struct lp_build_tgsi_context * bld_base = &ctx->soa.bld_base;
radeon_llvm_context_init(ctx);
#if HAVE_LLVM >= 0x0304
LLVMTypeRef Arguments[32];
unsigned ArgumentsCount = 0;
for (unsigned i = 0; i < ctx->inputs_count; i++)
Arguments[ArgumentsCount++] = LLVMVectorType(bld_base->base.elem_type, 4);
radeon_llvm_create_func(ctx, Arguments, ArgumentsCount);
for (unsigned i = 0; i < ctx->inputs_count; i++) {
LLVMValueRef P = LLVMGetParam(ctx->main_fn, i);
LLVMAddAttribute(P, LLVMInRegAttribute);
}
#else
radeon_llvm_create_func(ctx, NULL, 0);
#endif
tgsi_scan_shader(tokens, &shader_info);
bld_base->info = &shader_info;
bld_base->userdata = ctx;
bld_base->emit_fetch_funcs[TGSI_FILE_CONSTANT] = llvm_fetch_const;
bld_base->emit_prologue = llvm_emit_prologue;
bld_base->emit_epilogue = llvm_emit_epilogue;
ctx->userdata = ctx;
ctx->load_input = llvm_load_input;
ctx->load_system_value = llvm_load_system_value;
bld_base->op_actions[TGSI_OPCODE_DP2] = dot_action;
bld_base->op_actions[TGSI_OPCODE_DP3] = dot_action;
bld_base->op_actions[TGSI_OPCODE_DP4] = dot_action;
bld_base->op_actions[TGSI_OPCODE_DPH] = dot_action;
bld_base->op_actions[TGSI_OPCODE_DDX].emit = llvm_emit_tex;
bld_base->op_actions[TGSI_OPCODE_DDY].emit = llvm_emit_tex;
bld_base->op_actions[TGSI_OPCODE_TEX].emit = llvm_emit_tex;
bld_base->op_actions[TGSI_OPCODE_TEX2].emit = llvm_emit_tex;
bld_base->op_actions[TGSI_OPCODE_TXB].emit = llvm_emit_tex;
bld_base->op_actions[TGSI_OPCODE_TXB2].emit = llvm_emit_tex;
bld_base->op_actions[TGSI_OPCODE_TXD].emit = llvm_emit_tex;
bld_base->op_actions[TGSI_OPCODE_TXL].emit = llvm_emit_tex;
bld_base->op_actions[TGSI_OPCODE_TXL2].emit = llvm_emit_tex;
bld_base->op_actions[TGSI_OPCODE_TXF].emit = llvm_emit_tex;
bld_base->op_actions[TGSI_OPCODE_TXQ].emit = llvm_emit_tex;
bld_base->op_actions[TGSI_OPCODE_TXP].emit = llvm_emit_tex;
bld_base->op_actions[TGSI_OPCODE_CMP].emit = emit_cndlt;
lp_build_tgsi_llvm(bld_base, tokens);
radeon_llvm_finalize_module(ctx);
return ctx->gallivm.module;
}
void
lp_add_function_attr(LLVMValueRef function_or_call,
int attr_idx, enum lp_func_attr attr)
{
#if HAVE_LLVM < 0x0400
LLVMAttribute llvm_attr = lp_attr_to_llvm_attr(attr);
if (LLVMIsAFunction(function_or_call)) {
if (attr_idx == -1) {
LLVMAddFunctionAttr(function_or_call, llvm_attr);
} else {
LLVMAddAttribute(LLVMGetParam(function_or_call, attr_idx - 1), llvm_attr);
}
} else {
LLVMAddInstrAttribute(function_or_call, attr_idx, llvm_attr);
}
#else
LLVMModuleRef module;
if (LLVMIsAFunction(function_or_call)) {
module = LLVMGetGlobalParent(function_or_call);
} else {
LLVMBasicBlockRef bb = LLVMGetInstructionParent(function_or_call);
LLVMValueRef function = LLVMGetBasicBlockParent(bb);
module = LLVMGetGlobalParent(function);
}
LLVMContextRef ctx = LLVMGetModuleContext(module);
const char *attr_name = attr_to_str(attr);
unsigned kind_id = LLVMGetEnumAttributeKindForName(attr_name,
strlen(attr_name));
LLVMAttributeRef llvm_attr = LLVMCreateEnumAttribute(ctx, kind_id, 0);
if (LLVMIsAFunction(function_or_call))
LLVMAddAttributeAtIndex(function_or_call, attr_idx, llvm_attr);
else
LLVMAddCallSiteAttribute(function_or_call, attr_idx, llvm_attr);
#endif
}
LLVMValueRef codegen_addfun(compile_t* c, const char* name, LLVMTypeRef type)
{
// Add the function and set the calling convention.
LLVMValueRef fun = LLVMAddFunction(c->module, name, type);
if(!c->opt->library)
LLVMSetFunctionCallConv(fun, GEN_CALLCONV);
LLVMValueRef arg = LLVMGetFirstParam(fun);
uint32_t i = 1;
while(arg != NULL)
{
LLVMTypeRef type = LLVMTypeOf(arg);
if(LLVMGetTypeKind(type) == LLVMPointerTypeKind)
{
LLVMTypeRef elem = LLVMGetElementType(type);
if(LLVMGetTypeKind(elem) == LLVMStructTypeKind)
{
uint64_t size = LLVMABISizeOfType(c->target_data, elem);
LLVMSetDereferenceable(fun, i, size);
}
// Set the noalias attribute on all arguments. This is fortran-like
// semantics for parameter aliasing, similar to C restrict.
if(!c->opt->no_restrict)
LLVMAddAttribute(arg, LLVMNoAliasAttribute);
}
arg = LLVMGetNextParam(arg);
i++;
}
return fun;
}
static LLVMValueRef get_stritem_len_fn(struct llvm_ctx *ctx)
{
if(ctx->stritem_len_fn != NULL) return ctx->stritem_len_fn;
/* returns (i32 len, i32 new_tpos)
* params (word *utcbptr, i32 tpos)
*
* when return value "new_tpos" > tmax + 1, the result is invalid. the function
* should also not be called when tpos > tmax + 1.
*/
LLVMTypeRef ret_types[2] = { ctx->i32t, ctx->i32t },
parm_types[2] = { LLVMPointerType(ctx->wordt, 0), ctx->i32t },
ret_type = LLVMStructTypeInContext(ctx->ctx, ret_types, 2, 0),
fn_type = LLVMFunctionType(ret_type, parm_types, 2, 0);
LLVMValueRef fn = LLVMAddFunction(ctx->module, "__muidl_get_stritem_len",
fn_type);
LLVMSetVisibility(fn, LLVMHiddenVisibility);
LLVMSetLinkage(fn, LLVMInternalLinkage);
V fn_args[2];
LLVMGetParams(fn, fn_args);
LLVMAddAttribute(fn_args[0], LLVMNoCaptureAttribute);
for(int i=0; i<2; i++) {
LLVMAddAttribute(fn_args[i], LLVMInRegAttribute);
}
ctx->stritem_len_fn = fn;
LLVMBuilderRef old_builder = ctx->builder;
ctx->builder = LLVMCreateBuilderInContext(ctx->ctx);
LLVMBasicBlockRef entry_bb = LLVMAppendBasicBlockInContext(ctx->ctx, fn,
"EntryBlock"),
loop_bb = LLVMAppendBasicBlockInContext(ctx->ctx, fn, "loop"),
valid_bb = LLVMAppendBasicBlockInContext(ctx->ctx, fn, "valid"),
exit_bb = LLVMAppendBasicBlockInContext(ctx->ctx, fn, "exit");
LLVMPositionBuilderAtEnd(ctx->builder, entry_bb);
LLVMValueRef old_utcb = ctx->utcb, old_tpos = ctx->tpos;
ctx->utcb = fn_args[0];
ctx->tpos = fn_args[1];
LLVMBuildBr(ctx->builder, loop_bb);
LLVMPositionBuilderAtEnd(ctx->builder, exit_bb);
LLVMValueRef exit_len_phi = LLVMBuildPhi(ctx->builder, ctx->i32t,
"exit.len.phi"),
exit_tpos_phi = LLVMBuildPhi(ctx->builder, ctx->i32t,
"exit.tpos.phi");
LLVMValueRef rvals[2] = { exit_len_phi, exit_tpos_phi };
LLVMBuildAggregateRet(ctx->builder, rvals, 2);
LLVMPositionBuilderAtEnd(ctx->builder, loop_bb);
LLVMValueRef len_phi = LLVMBuildPhi(ctx->builder, ctx->i32t, "len.phi"),
tpos_phi = LLVMBuildPhi(ctx->builder, ctx->i32t, "tpos.phi");
LLVMAddIncoming(len_phi, &ctx->zero, &entry_bb, 1);
LLVMAddIncoming(tpos_phi, &ctx->tpos, &entry_bb, 1);
ctx->tpos = tpos_phi;
/* test: if *tpos doesn't look like a string item, conk out. */
LLVMValueRef infoword = build_utcb_load(ctx, ctx->tpos, "si.info");
LLVMValueRef is_cond = LLVMBuildICmp(ctx->builder, LLVMIntEQ,
ctx->zero, LLVMBuildAnd(ctx->builder, infoword,
CONST_WORD(1 << 4), "infoword.si.mask"),
"infoword.si.cond");
/* anything + 100 is sure to be > tmax + 1. */
LLVMValueRef fucked_tpos = LLVMBuildAdd(ctx->builder, tpos_phi,
CONST_INT(100), "f****d.tpos");
branch_set_phi(ctx, exit_len_phi, len_phi);
branch_set_phi(ctx, exit_tpos_phi, fucked_tpos);
LLVMBuildCondBr(ctx->builder, is_cond, valid_bb, exit_bb);
LLVMPositionBuilderAtEnd(ctx->builder, valid_bb);
LLVMValueRef string_length = LLVMBuildTruncOrBitCast(ctx->builder,
LLVMBuildLShr(ctx->builder, infoword,
CONST_INT(10), "si.info.len"),
ctx->i32t, "si.info.len.int"),
string_j = LLVMBuildTruncOrBitCast(ctx->builder,
LLVMBuildAnd(ctx->builder, CONST_WORD(0x1f),
LLVMBuildLShr(ctx->builder, infoword, CONST_WORD(4),
"si.info.j.shift"),
"si.info.j.masked"),
ctx->i32t, "si.info.j"),
string_c = LLVMBuildTruncOrBitCast(ctx->builder,
LLVMBuildAnd(ctx->builder, CONST_WORD(1 << 9),
infoword, "si.info.c.masked"),
ctx->i32t, "si.info.c.masked.int"),
c_cond = LLVMBuildICmp(ctx->builder, LLVMIntNE,
string_c, CONST_WORD(0), "si.info.c.cond"),
new_len = LLVMBuildAdd(ctx->builder, len_phi,
LLVMBuildMul(ctx->builder, string_length,
LLVMBuildAdd(ctx->builder, string_j,
CONST_INT(1), "j.plus.one"),
"len.incr"),
"len.new"),
new_tpos = LLVMBuildAdd(ctx->builder, ctx->tpos,
LLVMBuildSelect(ctx->builder, c_cond,
LLVMBuildAdd(ctx->builder, CONST_INT(2),
string_j, "cont.tpos.bump"),
CONST_INT(2), "tpos.bump"),
"tpos.new");
LLVMAddIncoming(len_phi, &new_len, &valid_bb, 1);
LLVMAddIncoming(tpos_phi, &new_tpos, &valid_bb, 1);
LLVMAddIncoming(exit_len_phi, &new_len, &valid_bb, 1);
LLVMAddIncoming(exit_tpos_phi, &new_tpos, &valid_bb, 1);
LLVMBuildCondBr(ctx->builder, c_cond, loop_bb, exit_bb);
LLVMDisposeBuilder(ctx->builder);
ctx->builder = old_builder;
ctx->utcb = old_utcb;
ctx->tpos = old_tpos;
return ctx->stritem_len_fn;
}
static void init_runtime(compile_t* c)
{
c->str_builtin = stringtab("$0");
c->str_Bool = stringtab("Bool");
c->str_I8 = stringtab("I8");
c->str_I16 = stringtab("I16");
c->str_I32 = stringtab("I32");
c->str_I64 = stringtab("I64");
c->str_I128 = stringtab("I128");
c->str_ILong = stringtab("ILong");
c->str_ISize = stringtab("ISize");
c->str_U8 = stringtab("U8");
c->str_U16 = stringtab("U16");
c->str_U32 = stringtab("U32");
c->str_U64 = stringtab("U64");
c->str_U128 = stringtab("U128");
c->str_ULong = stringtab("ULong");
c->str_USize = stringtab("USize");
c->str_F32 = stringtab("F32");
c->str_F64 = stringtab("F64");
c->str_Pointer = stringtab("Pointer");
c->str_Maybe = stringtab("MaybePointer");
c->str_DoNotOptimise = stringtab("DoNotOptimise");
c->str_Array = stringtab("Array");
c->str_String = stringtab("String");
c->str_Platform = stringtab("Platform");
c->str_Main = stringtab("Main");
c->str_Env = stringtab("Env");
c->str_add = stringtab("add");
c->str_sub = stringtab("sub");
c->str_mul = stringtab("mul");
c->str_div = stringtab("div");
c->str_mod = stringtab("mod");
c->str_neg = stringtab("neg");
c->str_add_unsafe = stringtab("add_unsafe");
c->str_sub_unsafe = stringtab("sub_unsafe");
c->str_mul_unsafe = stringtab("mul_unsafe");
c->str_div_unsafe = stringtab("div_unsafe");
c->str_mod_unsafe = stringtab("mod_unsafe");
c->str_neg_unsafe = stringtab("neg_unsafe");
c->str_and = stringtab("op_and");
c->str_or = stringtab("op_or");
c->str_xor = stringtab("op_xor");
c->str_not = stringtab("op_not");
c->str_shl = stringtab("shl");
c->str_shr = stringtab("shr");
c->str_shl_unsafe = stringtab("shl_unsafe");
c->str_shr_unsafe = stringtab("shr_unsafe");
c->str_eq = stringtab("eq");
c->str_ne = stringtab("ne");
c->str_lt = stringtab("lt");
c->str_le = stringtab("le");
c->str_ge = stringtab("ge");
c->str_gt = stringtab("gt");
c->str_eq_unsafe = stringtab("eq_unsafe");
c->str_ne_unsafe = stringtab("ne_unsafe");
c->str_lt_unsafe = stringtab("lt_unsafe");
c->str_le_unsafe = stringtab("le_unsafe");
c->str_ge_unsafe = stringtab("ge_unsafe");
c->str_gt_unsafe = stringtab("gt_unsafe");
c->str_this = stringtab("this");
c->str_create = stringtab("create");
c->str__create = stringtab("_create");
c->str__init = stringtab("_init");
c->str__final = stringtab("_final");
c->str__event_notify = stringtab("_event_notify");
c->str__serialise_space = stringtab("_serialise_space");
c->str__serialise = stringtab("_serialise");
c->str__deserialise = stringtab("_deserialise");
LLVMTypeRef type;
LLVMTypeRef params[5];
LLVMValueRef value;
c->void_type = LLVMVoidTypeInContext(c->context);
c->i1 = LLVMInt1TypeInContext(c->context);
c->i8 = LLVMInt8TypeInContext(c->context);
c->i16 = LLVMInt16TypeInContext(c->context);
c->i32 = LLVMInt32TypeInContext(c->context);
c->i64 = LLVMInt64TypeInContext(c->context);
c->i128 = LLVMIntTypeInContext(c->context, 128);
c->f32 = LLVMFloatTypeInContext(c->context);
c->f64 = LLVMDoubleTypeInContext(c->context);
c->intptr = LLVMIntPtrTypeInContext(c->context, c->target_data);
// i8*
c->void_ptr = LLVMPointerType(c->i8, 0);
// forward declare object
c->object_type = LLVMStructCreateNamed(c->context, "__object");
c->object_ptr = LLVMPointerType(c->object_type, 0);
// padding required in an actor between the descriptor and fields
c->actor_pad = LLVMArrayType(c->i8, PONY_ACTOR_PAD_SIZE);
// message
params[0] = c->i32; // size
params[1] = c->i32; // id
c->msg_type = LLVMStructCreateNamed(c->context, "__message");
c->msg_ptr = LLVMPointerType(c->msg_type, 0);
LLVMStructSetBody(c->msg_type, params, 2, false);
// trace
// void (*)(i8*, __object*)
params[0] = c->void_ptr;
params[1] = c->object_ptr;
c->trace_type = LLVMFunctionType(c->void_type, params, 2, false);
c->trace_fn = LLVMPointerType(c->trace_type, 0);
// serialise
// void (*)(i8*, __object*, i8*, intptr, i32)
params[0] = c->void_ptr;
params[1] = c->object_ptr;
params[2] = c->void_ptr;
params[3] = c->intptr;
params[4] = c->i32;
c->serialise_type = LLVMFunctionType(c->void_type, params, 5, false);
c->serialise_fn = LLVMPointerType(c->serialise_type, 0);
// serialise_space
// i64 (__object*)
params[0] = c->object_ptr;
c->custom_serialise_space_fn = LLVMPointerType(
LLVMFunctionType(c->i64, params, 1, false), 0);
// custom_deserialise
// void (*)(__object*, void*)
params[0] = c->object_ptr;
params[1] = c->void_ptr;
c->custom_deserialise_fn = LLVMPointerType(
LLVMFunctionType(c->void_type, params, 2, false), 0);
// dispatch
// void (*)(i8*, __object*, $message*)
params[0] = c->void_ptr;
params[1] = c->object_ptr;
params[2] = c->msg_ptr;
c->dispatch_type = LLVMFunctionType(c->void_type, params, 3, false);
c->dispatch_fn = LLVMPointerType(c->dispatch_type, 0);
// void (*)(__object*)
params[0] = c->object_ptr;
c->final_fn = LLVMPointerType(
LLVMFunctionType(c->void_type, params, 1, false), 0);
// descriptor, opaque version
// We need this in order to build our own structure.
const char* desc_name = genname_descriptor(NULL);
c->descriptor_type = LLVMStructCreateNamed(c->context, desc_name);
c->descriptor_ptr = LLVMPointerType(c->descriptor_type, 0);
// field descriptor
// Also needed to build a descriptor structure.
params[0] = c->i32;
params[1] = c->descriptor_ptr;
c->field_descriptor = LLVMStructTypeInContext(c->context, params, 2, false);
// descriptor, filled in
gendesc_basetype(c, c->descriptor_type);
// define object
params[0] = c->descriptor_ptr;
LLVMStructSetBody(c->object_type, params, 1, false);
#if PONY_LLVM >= 309
LLVM_DECLARE_ATTRIBUTEREF(nounwind_attr, nounwind, 0);
LLVM_DECLARE_ATTRIBUTEREF(readnone_attr, readnone, 0);
LLVM_DECLARE_ATTRIBUTEREF(readonly_attr, readonly, 0);
LLVM_DECLARE_ATTRIBUTEREF(inacc_or_arg_mem_attr,
inaccessiblemem_or_argmemonly, 0);
LLVM_DECLARE_ATTRIBUTEREF(noalias_attr, noalias, 0);
LLVM_DECLARE_ATTRIBUTEREF(noreturn_attr, noreturn, 0);
LLVM_DECLARE_ATTRIBUTEREF(deref_actor_attr, dereferenceable,
PONY_ACTOR_PAD_SIZE + (target_is_ilp32(c->opt->triple) ? 4 : 8));
LLVM_DECLARE_ATTRIBUTEREF(align_pool_attr, align, ponyint_pool_size(0));
LLVM_DECLARE_ATTRIBUTEREF(align_heap_attr, align, HEAP_MIN);
LLVM_DECLARE_ATTRIBUTEREF(deref_or_null_alloc_attr, dereferenceable_or_null,
HEAP_MIN);
LLVM_DECLARE_ATTRIBUTEREF(deref_alloc_small_attr, dereferenceable, HEAP_MIN);
LLVM_DECLARE_ATTRIBUTEREF(deref_alloc_large_attr, dereferenceable,
HEAP_MAX << 1);
#endif
// i8* pony_ctx()
type = LLVMFunctionType(c->void_ptr, NULL, 0, false);
value = LLVMAddFunction(c->module, "pony_ctx", type);
#if PONY_LLVM >= 309
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex, nounwind_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex, readnone_attr);
#else
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
LLVMAddFunctionAttr(value, LLVMReadNoneAttribute);
#endif
// __object* pony_create(i8*, __Desc*)
params[0] = c->void_ptr;
params[1] = c->descriptor_ptr;
type = LLVMFunctionType(c->object_ptr, params, 2, false);
value = LLVMAddFunction(c->module, "pony_create", type);
#if PONY_LLVM >= 309
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex, nounwind_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex,
inacc_or_arg_mem_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeReturnIndex, noalias_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeReturnIndex, deref_actor_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeReturnIndex, align_pool_attr);
#else
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
# if PONY_LLVM >= 308
LLVMSetInaccessibleMemOrArgMemOnly(value);
# endif
LLVMSetReturnNoAlias(value);
LLVMSetDereferenceable(value, 0, PONY_ACTOR_PAD_SIZE +
(target_is_ilp32(c->opt->triple) ? 4 : 8));
#endif
// void ponyint_destroy(__object*)
params[0] = c->object_ptr;
type = LLVMFunctionType(c->void_type, params, 1, false);
value = LLVMAddFunction(c->module, "ponyint_destroy", type);
#if PONY_LLVM >= 309
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex, nounwind_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex,
inacc_or_arg_mem_attr);
#else
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
# if PONY_LLVM >= 308
LLVMSetInaccessibleMemOrArgMemOnly(value);
# endif
#endif
// void pony_sendv(i8*, __object*, $message*, $message*)
params[0] = c->void_ptr;
params[1] = c->object_ptr;
params[2] = c->msg_ptr;
params[3] = c->msg_ptr;
type = LLVMFunctionType(c->void_type, params, 4, false);
value = LLVMAddFunction(c->module, "pony_sendv", type);
#if PONY_LLVM >= 309
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex, nounwind_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex,
inacc_or_arg_mem_attr);
#else
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
# if PONY_LLVM >= 308
LLVMSetInaccessibleMemOrArgMemOnly(value);
# endif
#endif
// void pony_sendv_single(i8*, __object*, $message*, $message*)
params[0] = c->void_ptr;
params[1] = c->object_ptr;
params[2] = c->msg_ptr;
params[3] = c->msg_ptr;
type = LLVMFunctionType(c->void_type, params, 4, false);
value = LLVMAddFunction(c->module, "pony_sendv_single", type);
#if PONY_LLVM >= 309
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex, nounwind_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex,
inacc_or_arg_mem_attr);
#else
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
# if PONY_LLVM >= 308
LLVMSetInaccessibleMemOrArgMemOnly(value);
# endif
#endif
// i8* pony_alloc(i8*, intptr)
params[0] = c->void_ptr;
params[1] = c->intptr;
type = LLVMFunctionType(c->void_ptr, params, 2, false);
value = LLVMAddFunction(c->module, "pony_alloc", type);
#if PONY_LLVM >= 309
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex, nounwind_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex,
inacc_or_arg_mem_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeReturnIndex, noalias_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeReturnIndex,
deref_or_null_alloc_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeReturnIndex, align_heap_attr);
#else
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
# if PONY_LLVM >= 308
LLVMSetInaccessibleMemOrArgMemOnly(value);
# endif
LLVMSetReturnNoAlias(value);
LLVMSetDereferenceableOrNull(value, 0, HEAP_MIN);
#endif
// i8* pony_alloc_small(i8*, i32)
params[0] = c->void_ptr;
params[1] = c->i32;
type = LLVMFunctionType(c->void_ptr, params, 2, false);
value = LLVMAddFunction(c->module, "pony_alloc_small", type);
#if PONY_LLVM >= 309
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex, nounwind_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex,
inacc_or_arg_mem_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeReturnIndex, noalias_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeReturnIndex,
deref_alloc_small_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeReturnIndex, align_heap_attr);
#else
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
# if PONY_LLVM >= 308
LLVMSetInaccessibleMemOrArgMemOnly(value);
# endif
LLVMSetReturnNoAlias(value);
LLVMSetDereferenceable(value, 0, HEAP_MIN);
#endif
// i8* pony_alloc_large(i8*, intptr)
params[0] = c->void_ptr;
params[1] = c->intptr;
type = LLVMFunctionType(c->void_ptr, params, 2, false);
value = LLVMAddFunction(c->module, "pony_alloc_large", type);
#if PONY_LLVM >= 309
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex, nounwind_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex,
inacc_or_arg_mem_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeReturnIndex, noalias_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeReturnIndex,
deref_alloc_large_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeReturnIndex, align_heap_attr);
#else
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
# if PONY_LLVM >= 308
LLVMSetInaccessibleMemOrArgMemOnly(value);
# endif
LLVMSetReturnNoAlias(value);
LLVMSetDereferenceable(value, 0, HEAP_MAX << 1);
#endif
// i8* pony_realloc(i8*, i8*, intptr)
params[0] = c->void_ptr;
params[1] = c->void_ptr;
params[2] = c->intptr;
type = LLVMFunctionType(c->void_ptr, params, 3, false);
value = LLVMAddFunction(c->module, "pony_realloc", type);
#if PONY_LLVM >= 309
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex, nounwind_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex,
inacc_or_arg_mem_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeReturnIndex, noalias_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeReturnIndex,
deref_or_null_alloc_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeReturnIndex, align_heap_attr);
#else
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
# if PONY_LLVM >= 308
LLVMSetInaccessibleMemOrArgMemOnly(value);
# endif
LLVMSetReturnNoAlias(value);
LLVMSetDereferenceableOrNull(value, 0, HEAP_MIN);
#endif
// i8* pony_alloc_final(i8*, intptr)
params[0] = c->void_ptr;
params[1] = c->intptr;
type = LLVMFunctionType(c->void_ptr, params, 2, false);
value = LLVMAddFunction(c->module, "pony_alloc_final", type);
#if PONY_LLVM >= 309
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex, nounwind_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex,
inacc_or_arg_mem_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeReturnIndex, noalias_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeReturnIndex,
deref_or_null_alloc_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeReturnIndex, align_heap_attr);
#else
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
# if PONY_LLVM >= 308
LLVMSetInaccessibleMemOrArgMemOnly(value);
# endif
LLVMSetReturnNoAlias(value);
LLVMSetDereferenceableOrNull(value, 0, HEAP_MIN);
#endif
// i8* pony_alloc_small_final(i8*, i32)
params[0] = c->void_ptr;
params[1] = c->i32;
type = LLVMFunctionType(c->void_ptr, params, 2, false);
value = LLVMAddFunction(c->module, "pony_alloc_small_final", type);
#if PONY_LLVM >= 309
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex, nounwind_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex,
inacc_or_arg_mem_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeReturnIndex, noalias_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeReturnIndex,
deref_alloc_small_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeReturnIndex, align_heap_attr);
#else
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
# if PONY_LLVM >= 308
LLVMSetInaccessibleMemOrArgMemOnly(value);
# endif
LLVMSetReturnNoAlias(value);
LLVMSetDereferenceable(value, 0, HEAP_MIN);
#endif
// i8* pony_alloc_large_final(i8*, intptr)
params[0] = c->void_ptr;
params[1] = c->intptr;
type = LLVMFunctionType(c->void_ptr, params, 2, false);
value = LLVMAddFunction(c->module, "pony_alloc_large_final", type);
#if PONY_LLVM >= 309
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex, nounwind_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex,
inacc_or_arg_mem_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeReturnIndex, noalias_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeReturnIndex,
deref_alloc_large_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeReturnIndex, align_heap_attr);
#else
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
# if PONY_LLVM >= 308
LLVMSetInaccessibleMemOrArgMemOnly(value);
# endif
LLVMSetReturnNoAlias(value);
LLVMSetDereferenceable(value, 0, HEAP_MAX << 1);
#endif
// $message* pony_alloc_msg(i32, i32)
params[0] = c->i32;
params[1] = c->i32;
type = LLVMFunctionType(c->msg_ptr, params, 2, false);
value = LLVMAddFunction(c->module, "pony_alloc_msg", type);
#if PONY_LLVM >= 309
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex, nounwind_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex,
inacc_or_arg_mem_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeReturnIndex, noalias_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeReturnIndex, align_pool_attr);
#else
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
# if PONY_LLVM >= 308
LLVMSetInaccessibleMemOrArgMemOnly(value);
# endif
LLVMSetReturnNoAlias(value);
#endif
// void pony_trace(i8*, i8*)
params[0] = c->void_ptr;
params[1] = c->void_ptr;
type = LLVMFunctionType(c->void_type, params, 2, false);
value = LLVMAddFunction(c->module, "pony_trace", type);
#if PONY_LLVM >= 309
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex, nounwind_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex,
inacc_or_arg_mem_attr);
LLVMAddAttributeAtIndex(value, 2, readnone_attr);
#else
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
# if PONY_LLVM >= 308
LLVMSetInaccessibleMemOrArgMemOnly(value);
# endif
value = LLVMGetParam(value, 1);
LLVMAddAttribute(value, LLVMReadNoneAttribute);
#endif
// void pony_traceknown(i8*, __object*, __Desc*, i32)
params[0] = c->void_ptr;
params[1] = c->object_ptr;
params[2] = c->descriptor_ptr;
params[3] = c->i32;
type = LLVMFunctionType(c->void_type, params, 4, false);
value = LLVMAddFunction(c->module, "pony_traceknown", type);
#if PONY_LLVM >= 309
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex, nounwind_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex,
inacc_or_arg_mem_attr);
LLVMAddAttributeAtIndex(value, 2, readonly_attr);
#else
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
# if PONY_LLVM >= 308
LLVMSetInaccessibleMemOrArgMemOnly(value);
# endif
value = LLVMGetParam(value, 1);
LLVMAddAttribute(value, LLVMReadOnlyAttribute);
#endif
// void pony_traceunknown(i8*, __object*, i32)
params[0] = c->void_ptr;
params[1] = c->object_ptr;
params[2] = c->i32;
type = LLVMFunctionType(c->void_type, params, 3, false);
value = LLVMAddFunction(c->module, "pony_traceunknown", type);
#if PONY_LLVM >= 309
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex, nounwind_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex,
inacc_or_arg_mem_attr);
LLVMAddAttributeAtIndex(value, 2, readonly_attr);
#else
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
# if PONY_LLVM >= 308
LLVMSetInaccessibleMemOrArgMemOnly(value);
# endif
value = LLVMGetParam(value, 1);
LLVMAddAttribute(value, LLVMReadOnlyAttribute);
#endif
// void pony_gc_send(i8*)
params[0] = c->void_ptr;
type = LLVMFunctionType(c->void_type, params, 1, false);
value = LLVMAddFunction(c->module, "pony_gc_send", type);
#if PONY_LLVM >= 309
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex, nounwind_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex,
inacc_or_arg_mem_attr);
#else
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
# if PONY_LLVM >= 308
LLVMSetInaccessibleMemOrArgMemOnly(value);
# endif
#endif
// void pony_gc_recv(i8*)
params[0] = c->void_ptr;
type = LLVMFunctionType(c->void_type, params, 1, false);
value = LLVMAddFunction(c->module, "pony_gc_recv", type);
#if PONY_LLVM >= 309
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex, nounwind_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex,
inacc_or_arg_mem_attr);
#else
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
# if PONY_LLVM >= 308
LLVMSetInaccessibleMemOrArgMemOnly(value);
# endif
#endif
// void pony_send_done(i8*)
params[0] = c->void_ptr;
type = LLVMFunctionType(c->void_type, params, 1, false);
value = LLVMAddFunction(c->module, "pony_send_done", type);
#if PONY_LLVM >= 309
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex, nounwind_attr);
#else
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
#endif
// void pony_recv_done(i8*)
params[0] = c->void_ptr;
type = LLVMFunctionType(c->void_type, params, 1, false);
value = LLVMAddFunction(c->module, "pony_recv_done", type);
#if PONY_LLVM >= 309
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex, nounwind_attr);
#else
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
#endif
// void pony_serialise_reserve(i8*, i8*, intptr)
params[0] = c->void_ptr;
params[1] = c->void_ptr;
params[2] = c->intptr;
type = LLVMFunctionType(c->void_type, params, 3, false);
value = LLVMAddFunction(c->module, "pony_serialise_reserve", type);
#if PONY_LLVM >= 309
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex, nounwind_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex,
inacc_or_arg_mem_attr);
LLVMAddAttributeAtIndex(value, 2, readnone_attr);
#else
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
# if PONY_LLVM >= 308
LLVMSetInaccessibleMemOrArgMemOnly(value);
# endif
value = LLVMGetParam(value, 1);
LLVMAddAttribute(value, LLVMReadNoneAttribute);
#endif
// intptr pony_serialise_offset(i8*, i8*)
params[0] = c->void_ptr;
params[1] = c->void_ptr;
type = LLVMFunctionType(c->intptr, params, 2, false);
value = LLVMAddFunction(c->module, "pony_serialise_offset", type);
#if PONY_LLVM >= 309
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex, nounwind_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex,
inacc_or_arg_mem_attr);
LLVMAddAttributeAtIndex(value, 2, readonly_attr);
#else
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
# if PONY_LLVM >= 308
LLVMSetInaccessibleMemOrArgMemOnly(value);
# endif
value = LLVMGetParam(value, 1);
LLVMAddAttribute(value, LLVMReadOnlyAttribute);
#endif
// i8* pony_deserialise_offset(i8*, __desc*, intptr)
params[0] = c->void_ptr;
params[1] = c->descriptor_ptr;
params[2] = c->intptr;
type = LLVMFunctionType(c->void_ptr, params, 3, false);
value = LLVMAddFunction(c->module, "pony_deserialise_offset", type);
#if PONY_LLVM >= 309
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex,
inacc_or_arg_mem_attr);
#elif PONY_LLVM == 308
LLVMSetInaccessibleMemOrArgMemOnly(value);
#endif
// i8* pony_deserialise_block(i8*, intptr, intptr)
params[0] = c->void_ptr;
params[1] = c->intptr;
params[2] = c->intptr;
type = LLVMFunctionType(c->void_ptr, params, 3, false);
value = LLVMAddFunction(c->module, "pony_deserialise_block", type);
#if PONY_LLVM >= 309
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex,
inacc_or_arg_mem_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeReturnIndex, noalias_attr);
#else
# if PONY_LLVM >= 308
LLVMSetInaccessibleMemOrArgMemOnly(value);
# endif
LLVMSetReturnNoAlias(value);
#endif
// i32 pony_init(i32, i8**)
params[0] = c->i32;
params[1] = LLVMPointerType(c->void_ptr, 0);
type = LLVMFunctionType(c->i32, params, 2, false);
value = LLVMAddFunction(c->module, "pony_init", type);
#if PONY_LLVM >= 309
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex, nounwind_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex,
inacc_or_arg_mem_attr);
#else
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
# if PONY_LLVM >= 308
LLVMSetInaccessibleMemOrArgMemOnly(value);
# endif
#endif
// void pony_become(i8*, __object*)
params[0] = c->void_ptr;
params[1] = c->object_ptr;
type = LLVMFunctionType(c->void_type, params, 2, false);
value = LLVMAddFunction(c->module, "pony_become", type);
#if PONY_LLVM >= 309
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex, nounwind_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex,
inacc_or_arg_mem_attr);
#else
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
# if PONY_LLVM >= 308
LLVMSetInaccessibleMemOrArgMemOnly(value);
# endif
#endif
// i32 pony_start(i32, i32)
params[0] = c->i32;
params[1] = c->i32;
type = LLVMFunctionType(c->i32, params, 2, false);
value = LLVMAddFunction(c->module, "pony_start", type);
#if PONY_LLVM >= 309
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex, nounwind_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex,
inacc_or_arg_mem_attr);
#else
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
# if PONY_LLVM >= 308
LLVMSetInaccessibleMemOrArgMemOnly(value);
# endif
#endif
// i32 pony_get_exitcode()
type = LLVMFunctionType(c->i32, NULL, 0, false);
value = LLVMAddFunction(c->module, "pony_get_exitcode", type);
#if PONY_LLVM >= 309
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex, nounwind_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex, readonly_attr);
#else
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
LLVMAddFunctionAttr(value, LLVMReadOnlyAttribute);
#endif
// void pony_throw()
type = LLVMFunctionType(c->void_type, NULL, 0, false);
value = LLVMAddFunction(c->module, "pony_throw", type);
#if PONY_LLVM >= 309
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex, noreturn_attr);
#else
LLVMAddFunctionAttr(value, LLVMNoReturnAttribute);
#endif
// i32 pony_personality_v0(...)
type = LLVMFunctionType(c->i32, NULL, 0, true);
c->personality = LLVMAddFunction(c->module, "pony_personality_v0", type);
// i32 memcmp(i8*, i8*, intptr)
params[0] = c->void_ptr;
params[1] = c->void_ptr;
params[2] = c->intptr;
type = LLVMFunctionType(c->i32, params, 3, false);
value = LLVMAddFunction(c->module, "memcmp", type);
#if PONY_LLVM >= 309
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex, nounwind_attr);
LLVMAddAttributeAtIndex(value, LLVMAttributeFunctionIndex, readonly_attr);
LLVMAddAttributeAtIndex(value, 1, readonly_attr);
LLVMAddAttributeAtIndex(value, 2, readonly_attr);
#else
LLVMAddFunctionAttr(value, LLVMNoUnwindAttribute);
LLVMAddFunctionAttr(value, LLVMReadOnlyAttribute);
LLVMValueRef param = LLVMGetParam(value, 0);
LLVMAddAttribute(param, LLVMReadOnlyAttribute);
param = LLVMGetParam(value, 1);
LLVMAddAttribute(param, LLVMReadOnlyAttribute);
#endif
}
LLVMValueRef get_struct_fn(
struct llvm_ctx *ctx,
IDL_tree ctyp,
bool for_encode)
{
const char *s_id = IDL_IDENT(IDL_TYPE_STRUCT(ctyp).ident).repo_id;
char *lookup_name = talloc_asprintf(ctx, "%c%s",
for_encode ? 'e' : 'd', s_id);
LLVMValueRef fn = strmap_get(&ctx->struct_decoder_fns, lookup_name);
if(fn != NULL) {
talloc_free(lookup_name);
return fn;
}
const struct packed_format *fmt = packed_format_of(ctyp);
assert(fmt != NULL); /* only sane for packable structs */
int namelen = strlen(s_id);
char flatname[namelen + 1];
/* FIXME: make this proper, i.e. use a name mangler that works */
for(int i=0; i < namelen; i++) {
flatname[i] = isalnum(s_id[i]) ? s_id[i] : '_';
}
flatname[namelen] = '\0';
T types[3], rettyp = LLVMVoidTypeInContext(ctx->ctx);
types[0] = LLVMPointerType(llvm_rigid_type(ctx, ctyp), 0);
int nparms;
if(!for_encode) {
/* decoder */
types[1] = ctx->i32t;
types[2] = ctx->i32t;
nparms = fmt->num_bits < BITS_PER_WORD ? 3 : 2;
} else if(fmt->num_bits < BITS_PER_WORD) {
/* subword encoder */
rettyp = ctx->wordt;
types[1] = ctx->wordt;
types[2] = ctx->i32t;
nparms = 3;
} else {
/* non-subword encoder */
types[1] = ctx->i32t;
nparms = 2;
}
T fntype = LLVMFunctionType(rettyp, types, nparms, 0);
char *fnname = g_strdup_printf("__muidl_idl_%scode__%s",
for_encode ? "en" : "de", flatname);
fn = LLVMAddFunction(ctx->module, fnname, fntype);
LLVMSetLinkage(fn, LLVMExternalLinkage);
V params[nparms];
assert(LLVMCountParams(fn) == nparms);
LLVMGetParams(fn, params);
LLVMAddAttribute(params[0], LLVMNoAliasAttribute);
LLVMAddAttribute(params[0], LLVMNoCaptureAttribute);
for(int i=0; i<nparms; i++) {
LLVMAddAttribute(params[i], LLVMInRegAttribute);
}
g_free(fnname);
bool ok = strmap_add(&ctx->struct_decoder_fns, lookup_name, fn);
assert(ok || errno != EEXIST);
return fn;
}
/**
* Generate the runtime callable function for the whole fragment pipeline.
* Note that the function which we generate operates on a block of 16
* pixels at at time. The block contains 2x2 quads. Each quad contains
* 2x2 pixels.
*/
static void
generate_fragment(struct llvmpipe_context *lp,
struct lp_fragment_shader *shader,
struct lp_fragment_shader_variant *variant,
unsigned do_tri_test)
{
struct llvmpipe_screen *screen = llvmpipe_screen(lp->pipe.screen);
const struct lp_fragment_shader_variant_key *key = &variant->key;
struct lp_type fs_type;
struct lp_type blend_type;
LLVMTypeRef fs_elem_type;
LLVMTypeRef fs_vec_type;
LLVMTypeRef fs_int_vec_type;
LLVMTypeRef blend_vec_type;
LLVMTypeRef blend_int_vec_type;
LLVMTypeRef arg_types[14];
LLVMTypeRef func_type;
LLVMTypeRef int32_vec4_type = lp_build_int32_vec4_type();
LLVMValueRef context_ptr;
LLVMValueRef x;
LLVMValueRef y;
LLVMValueRef a0_ptr;
LLVMValueRef dadx_ptr;
LLVMValueRef dady_ptr;
LLVMValueRef color_ptr_ptr;
LLVMValueRef depth_ptr;
LLVMValueRef c0, c1, c2, step0_ptr, step1_ptr, step2_ptr;
LLVMBasicBlockRef block;
LLVMBuilderRef builder;
LLVMValueRef x0;
LLVMValueRef y0;
struct lp_build_sampler_soa *sampler;
struct lp_build_interp_soa_context interp;
LLVMValueRef fs_mask[LP_MAX_VECTOR_LENGTH];
LLVMValueRef fs_out_color[PIPE_MAX_COLOR_BUFS][NUM_CHANNELS][LP_MAX_VECTOR_LENGTH];
LLVMValueRef blend_mask;
LLVMValueRef blend_in_color[NUM_CHANNELS];
LLVMValueRef function;
unsigned num_fs;
unsigned i;
unsigned chan;
unsigned cbuf;
/* TODO: actually pick these based on the fs and color buffer
* characteristics. */
memset(&fs_type, 0, sizeof fs_type);
fs_type.floating = TRUE; /* floating point values */
fs_type.sign = TRUE; /* values are signed */
fs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
fs_type.width = 32; /* 32-bit float */
fs_type.length = 4; /* 4 elements per vector */
num_fs = 4; /* number of quads per block */
memset(&blend_type, 0, sizeof blend_type);
blend_type.floating = FALSE; /* values are integers */
blend_type.sign = FALSE; /* values are unsigned */
blend_type.norm = TRUE; /* values are in [0,1] or [-1,1] */
blend_type.width = 8; /* 8-bit ubyte values */
blend_type.length = 16; /* 16 elements per vector */
/*
* Generate the function prototype. Any change here must be reflected in
* lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
*/
fs_elem_type = lp_build_elem_type(fs_type);
fs_vec_type = lp_build_vec_type(fs_type);
fs_int_vec_type = lp_build_int_vec_type(fs_type);
blend_vec_type = lp_build_vec_type(blend_type);
blend_int_vec_type = lp_build_int_vec_type(blend_type);
arg_types[0] = screen->context_ptr_type; /* context */
arg_types[1] = LLVMInt32Type(); /* x */
arg_types[2] = LLVMInt32Type(); /* y */
arg_types[3] = LLVMPointerType(fs_elem_type, 0); /* a0 */
arg_types[4] = LLVMPointerType(fs_elem_type, 0); /* dadx */
arg_types[5] = LLVMPointerType(fs_elem_type, 0); /* dady */
arg_types[6] = LLVMPointerType(LLVMPointerType(blend_vec_type, 0), 0); /* color */
arg_types[7] = LLVMPointerType(fs_int_vec_type, 0); /* depth */
arg_types[8] = LLVMInt32Type(); /* c0 */
arg_types[9] = LLVMInt32Type(); /* c1 */
arg_types[10] = LLVMInt32Type(); /* c2 */
/* Note: the step arrays are built as int32[16] but we interpret
* them here as int32_vec4[4].
*/
arg_types[11] = LLVMPointerType(int32_vec4_type, 0);/* step0 */
arg_types[12] = LLVMPointerType(int32_vec4_type, 0);/* step1 */
arg_types[13] = LLVMPointerType(int32_vec4_type, 0);/* step2 */
func_type = LLVMFunctionType(LLVMVoidType(), arg_types, Elements(arg_types), 0);
function = LLVMAddFunction(screen->module, "shader", func_type);
LLVMSetFunctionCallConv(function, LLVMCCallConv);
variant->function[do_tri_test] = function;
/* XXX: need to propagate noalias down into color param now we are
* passing a pointer-to-pointer?
*/
for(i = 0; i < Elements(arg_types); ++i)
if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
LLVMAddAttribute(LLVMGetParam(function, i), LLVMNoAliasAttribute);
context_ptr = LLVMGetParam(function, 0);
x = LLVMGetParam(function, 1);
y = LLVMGetParam(function, 2);
a0_ptr = LLVMGetParam(function, 3);
dadx_ptr = LLVMGetParam(function, 4);
dady_ptr = LLVMGetParam(function, 5);
color_ptr_ptr = LLVMGetParam(function, 6);
depth_ptr = LLVMGetParam(function, 7);
c0 = LLVMGetParam(function, 8);
c1 = LLVMGetParam(function, 9);
c2 = LLVMGetParam(function, 10);
step0_ptr = LLVMGetParam(function, 11);
step1_ptr = LLVMGetParam(function, 12);
step2_ptr = LLVMGetParam(function, 13);
lp_build_name(context_ptr, "context");
lp_build_name(x, "x");
lp_build_name(y, "y");
lp_build_name(a0_ptr, "a0");
lp_build_name(dadx_ptr, "dadx");
lp_build_name(dady_ptr, "dady");
lp_build_name(color_ptr_ptr, "color_ptr");
lp_build_name(depth_ptr, "depth");
lp_build_name(c0, "c0");
lp_build_name(c1, "c1");
lp_build_name(c2, "c2");
lp_build_name(step0_ptr, "step0");
lp_build_name(step1_ptr, "step1");
lp_build_name(step2_ptr, "step2");
/*
* Function body
*/
block = LLVMAppendBasicBlock(function, "entry");
builder = LLVMCreateBuilder();
LLVMPositionBuilderAtEnd(builder, block);
generate_pos0(builder, x, y, &x0, &y0);
lp_build_interp_soa_init(&interp,
shader->base.tokens,
key->flatshade,
builder, fs_type,
a0_ptr, dadx_ptr, dady_ptr,
x0, y0);
/* code generated texture sampling */
sampler = lp_llvm_sampler_soa_create(key->sampler, context_ptr);
/* loop over quads in the block */
for(i = 0; i < num_fs; ++i) {
LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0);
LLVMValueRef out_color[PIPE_MAX_COLOR_BUFS][NUM_CHANNELS];
LLVMValueRef depth_ptr_i;
int cbuf;
if(i != 0)
lp_build_interp_soa_update(&interp, i);
depth_ptr_i = LLVMBuildGEP(builder, depth_ptr, &index, 1, "");
generate_fs(lp, shader, key,
builder,
fs_type,
context_ptr,
i,
&interp,
sampler,
&fs_mask[i], /* output */
out_color,
depth_ptr_i,
do_tri_test,
c0, c1, c2,
step0_ptr, step1_ptr, step2_ptr);
for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++)
for(chan = 0; chan < NUM_CHANNELS; ++chan)
fs_out_color[cbuf][chan][i] = out_color[cbuf][chan];
}
sampler->destroy(sampler);
/* Loop over color outputs / color buffers to do blending.
*/
for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
LLVMValueRef color_ptr;
LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), cbuf, 0);
/*
* Convert the fs's output color and mask to fit to the blending type.
*/
for(chan = 0; chan < NUM_CHANNELS; ++chan) {
lp_build_conv(builder, fs_type, blend_type,
fs_out_color[cbuf][chan], num_fs,
&blend_in_color[chan], 1);
lp_build_name(blend_in_color[chan], "color%d.%c", cbuf, "rgba"[chan]);
}
lp_build_conv_mask(builder, fs_type, blend_type,
fs_mask, num_fs,
&blend_mask, 1);
color_ptr = LLVMBuildLoad(builder,
LLVMBuildGEP(builder, color_ptr_ptr, &index, 1, ""),
"");
lp_build_name(color_ptr, "color_ptr%d", cbuf);
/*
* Blending.
*/
generate_blend(&key->blend,
builder,
blend_type,
context_ptr,
blend_mask,
blend_in_color,
color_ptr);
}
LLVMBuildRetVoid(builder);
LLVMDisposeBuilder(builder);
/* Verify the LLVM IR. If invalid, dump and abort */
#ifdef DEBUG
if(LLVMVerifyFunction(function, LLVMPrintMessageAction)) {
if (1)
LLVMDumpValue(function);
abort();
}
#endif
/* Apply optimizations to LLVM IR */
if (1)
LLVMRunFunctionPassManager(screen->pass, function);
if (LP_DEBUG & DEBUG_JIT) {
/* Print the LLVM IR to stderr */
LLVMDumpValue(function);
debug_printf("\n");
}
/*
* Translate the LLVM IR into machine code.
*/
variant->jit_function[do_tri_test] = (lp_jit_frag_func)LLVMGetPointerToGlobal(screen->engine, function);
if (LP_DEBUG & DEBUG_ASM)
lp_disassemble(variant->jit_function[do_tri_test]);
}
/**
* Generate the runtime callable function for the whole fragment pipeline.
*/
static struct lp_fragment_shader_variant *
generate_fragment(struct llvmpipe_context *lp,
struct lp_fragment_shader *shader,
const struct lp_fragment_shader_variant_key *key)
{
struct llvmpipe_screen *screen = llvmpipe_screen(lp->pipe.screen);
struct lp_fragment_shader_variant *variant;
struct lp_type fs_type;
struct lp_type blend_type;
LLVMTypeRef fs_elem_type;
LLVMTypeRef fs_vec_type;
LLVMTypeRef fs_int_vec_type;
LLVMTypeRef blend_vec_type;
LLVMTypeRef blend_int_vec_type;
LLVMTypeRef arg_types[9];
LLVMTypeRef func_type;
LLVMValueRef context_ptr;
LLVMValueRef x;
LLVMValueRef y;
LLVMValueRef a0_ptr;
LLVMValueRef dadx_ptr;
LLVMValueRef dady_ptr;
LLVMValueRef mask_ptr;
LLVMValueRef color_ptr;
LLVMValueRef depth_ptr;
LLVMBasicBlockRef block;
LLVMBuilderRef builder;
LLVMValueRef x0;
LLVMValueRef y0;
struct lp_build_sampler_soa *sampler;
struct lp_build_interp_soa_context interp;
LLVMValueRef fs_mask[LP_MAX_VECTOR_LENGTH];
LLVMValueRef fs_out_color[NUM_CHANNELS][LP_MAX_VECTOR_LENGTH];
LLVMValueRef blend_mask;
LLVMValueRef blend_in_color[NUM_CHANNELS];
unsigned num_fs;
unsigned i;
unsigned chan;
#ifdef DEBUG
tgsi_dump(shader->base.tokens, 0);
if(key->depth.enabled) {
debug_printf("depth.format = %s\n", pf_name(key->zsbuf_format));
debug_printf("depth.func = %s\n", debug_dump_func(key->depth.func, TRUE));
debug_printf("depth.writemask = %u\n", key->depth.writemask);
}
if(key->alpha.enabled) {
debug_printf("alpha.func = %s\n", debug_dump_func(key->alpha.func, TRUE));
debug_printf("alpha.ref_value = %f\n", key->alpha.ref_value);
}
if(key->blend.logicop_enable) {
debug_printf("blend.logicop_func = %u\n", key->blend.logicop_func);
}
else if(key->blend.blend_enable) {
debug_printf("blend.rgb_func = %s\n", debug_dump_blend_func (key->blend.rgb_func, TRUE));
debug_printf("rgb_src_factor = %s\n", debug_dump_blend_factor(key->blend.rgb_src_factor, TRUE));
debug_printf("rgb_dst_factor = %s\n", debug_dump_blend_factor(key->blend.rgb_dst_factor, TRUE));
debug_printf("alpha_func = %s\n", debug_dump_blend_func (key->blend.alpha_func, TRUE));
debug_printf("alpha_src_factor = %s\n", debug_dump_blend_factor(key->blend.alpha_src_factor, TRUE));
debug_printf("alpha_dst_factor = %s\n", debug_dump_blend_factor(key->blend.alpha_dst_factor, TRUE));
}
debug_printf("blend.colormask = 0x%x\n", key->blend.colormask);
for(i = 0; i < PIPE_MAX_SAMPLERS; ++i) {
if(key->sampler[i].format) {
debug_printf("sampler[%u] = \n", i);
debug_printf(" .format = %s\n",
pf_name(key->sampler[i].format));
debug_printf(" .target = %s\n",
debug_dump_tex_target(key->sampler[i].target, TRUE));
debug_printf(" .pot = %u %u %u\n",
key->sampler[i].pot_width,
key->sampler[i].pot_height,
key->sampler[i].pot_depth);
debug_printf(" .wrap = %s %s %s\n",
debug_dump_tex_wrap(key->sampler[i].wrap_s, TRUE),
debug_dump_tex_wrap(key->sampler[i].wrap_t, TRUE),
debug_dump_tex_wrap(key->sampler[i].wrap_r, TRUE));
debug_printf(" .min_img_filter = %s\n",
debug_dump_tex_filter(key->sampler[i].min_img_filter, TRUE));
debug_printf(" .min_mip_filter = %s\n",
debug_dump_tex_mipfilter(key->sampler[i].min_mip_filter, TRUE));
debug_printf(" .mag_img_filter = %s\n",
debug_dump_tex_filter(key->sampler[i].mag_img_filter, TRUE));
if(key->sampler[i].compare_mode)
debug_printf(" .compare_mode = %s\n", debug_dump_func(key->sampler[i].compare_func, TRUE));
debug_printf(" .normalized_coords = %u\n", key->sampler[i].normalized_coords);
debug_printf(" .prefilter = %u\n", key->sampler[i].prefilter);
}
}
#endif
variant = CALLOC_STRUCT(lp_fragment_shader_variant);
if(!variant)
return NULL;
variant->shader = shader;
memcpy(&variant->key, key, sizeof *key);
/* TODO: actually pick these based on the fs and color buffer
* characteristics. */
memset(&fs_type, 0, sizeof fs_type);
fs_type.floating = TRUE; /* floating point values */
fs_type.sign = TRUE; /* values are signed */
fs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
fs_type.width = 32; /* 32-bit float */
fs_type.length = 4; /* 4 element per vector */
num_fs = 4;
memset(&blend_type, 0, sizeof blend_type);
blend_type.floating = FALSE; /* values are integers */
blend_type.sign = FALSE; /* values are unsigned */
blend_type.norm = TRUE; /* values are in [0,1] or [-1,1] */
blend_type.width = 8; /* 8-bit ubyte values */
blend_type.length = 16; /* 16 elements per vector */
/*
* Generate the function prototype. Any change here must be reflected in
* lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
*/
fs_elem_type = lp_build_elem_type(fs_type);
fs_vec_type = lp_build_vec_type(fs_type);
fs_int_vec_type = lp_build_int_vec_type(fs_type);
blend_vec_type = lp_build_vec_type(blend_type);
blend_int_vec_type = lp_build_int_vec_type(blend_type);
arg_types[0] = screen->context_ptr_type; /* context */
arg_types[1] = LLVMInt32Type(); /* x */
arg_types[2] = LLVMInt32Type(); /* y */
arg_types[3] = LLVMPointerType(fs_elem_type, 0); /* a0 */
arg_types[4] = LLVMPointerType(fs_elem_type, 0); /* dadx */
arg_types[5] = LLVMPointerType(fs_elem_type, 0); /* dady */
arg_types[6] = LLVMPointerType(fs_int_vec_type, 0); /* mask */
arg_types[7] = LLVMPointerType(blend_vec_type, 0); /* color */
arg_types[8] = LLVMPointerType(fs_int_vec_type, 0); /* depth */
func_type = LLVMFunctionType(LLVMVoidType(), arg_types, Elements(arg_types), 0);
variant->function = LLVMAddFunction(screen->module, "shader", func_type);
LLVMSetFunctionCallConv(variant->function, LLVMCCallConv);
for(i = 0; i < Elements(arg_types); ++i)
if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
LLVMAddAttribute(LLVMGetParam(variant->function, i), LLVMNoAliasAttribute);
context_ptr = LLVMGetParam(variant->function, 0);
x = LLVMGetParam(variant->function, 1);
y = LLVMGetParam(variant->function, 2);
a0_ptr = LLVMGetParam(variant->function, 3);
dadx_ptr = LLVMGetParam(variant->function, 4);
dady_ptr = LLVMGetParam(variant->function, 5);
mask_ptr = LLVMGetParam(variant->function, 6);
color_ptr = LLVMGetParam(variant->function, 7);
depth_ptr = LLVMGetParam(variant->function, 8);
lp_build_name(context_ptr, "context");
lp_build_name(x, "x");
lp_build_name(y, "y");
lp_build_name(a0_ptr, "a0");
lp_build_name(dadx_ptr, "dadx");
lp_build_name(dady_ptr, "dady");
lp_build_name(mask_ptr, "mask");
lp_build_name(color_ptr, "color");
lp_build_name(depth_ptr, "depth");
/*
* Function body
*/
block = LLVMAppendBasicBlock(variant->function, "entry");
builder = LLVMCreateBuilder();
LLVMPositionBuilderAtEnd(builder, block);
generate_pos0(builder, x, y, &x0, &y0);
lp_build_interp_soa_init(&interp, shader->base.tokens, builder, fs_type,
a0_ptr, dadx_ptr, dady_ptr,
x0, y0, 2, 0);
#if 0
/* C texture sampling */
sampler = lp_c_sampler_soa_create(context_ptr);
#else
/* code generated texture sampling */
sampler = lp_llvm_sampler_soa_create(key->sampler, context_ptr);
#endif
for(i = 0; i < num_fs; ++i) {
LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0);
LLVMValueRef out_color[NUM_CHANNELS];
LLVMValueRef depth_ptr_i;
if(i != 0)
lp_build_interp_soa_update(&interp);
fs_mask[i] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, mask_ptr, &index, 1, ""), "");
depth_ptr_i = LLVMBuildGEP(builder, depth_ptr, &index, 1, "");
generate_fs(lp, shader, key,
builder,
fs_type,
context_ptr,
i,
&interp,
sampler,
&fs_mask[i],
out_color,
depth_ptr_i);
for(chan = 0; chan < NUM_CHANNELS; ++chan)
fs_out_color[chan][i] = out_color[chan];
}
sampler->destroy(sampler);
/*
* Convert the fs's output color and mask to fit to the blending type.
*/
for(chan = 0; chan < NUM_CHANNELS; ++chan) {
lp_build_conv(builder, fs_type, blend_type,
fs_out_color[chan], num_fs,
&blend_in_color[chan], 1);
lp_build_name(blend_in_color[chan], "color.%c", "rgba"[chan]);
}
lp_build_conv_mask(builder, fs_type, blend_type,
fs_mask, num_fs,
&blend_mask, 1);
/*
* Blending.
*/
generate_blend(&key->blend,
builder,
blend_type,
context_ptr,
blend_mask,
blend_in_color,
color_ptr);
LLVMBuildRetVoid(builder);
LLVMDisposeBuilder(builder);
/*
* Translate the LLVM IR into machine code.
*/
if(LLVMVerifyFunction(variant->function, LLVMPrintMessageAction)) {
LLVMDumpValue(variant->function);
abort();
}
LLVMRunFunctionPassManager(screen->pass, variant->function);
#ifdef DEBUG
LLVMDumpValue(variant->function);
debug_printf("\n");
#endif
variant->jit_function = (lp_jit_frag_func)LLVMGetPointerToGlobal(screen->engine, variant->function);
#ifdef DEBUG
lp_disassemble(variant->jit_function);
#endif
variant->next = shader->variants;
shader->variants = variant;
return variant;
}
/**
* Generate the runtime callable function for the whole fragment pipeline.
* Note that the function which we generate operates on a block of 16
* pixels at at time. The block contains 2x2 quads. Each quad contains
* 2x2 pixels.
*/
static void
generate_fragment(struct llvmpipe_context *lp,
struct lp_fragment_shader *shader,
struct lp_fragment_shader_variant *variant,
unsigned partial_mask)
{
struct llvmpipe_screen *screen = llvmpipe_screen(lp->pipe.screen);
const struct lp_fragment_shader_variant_key *key = &variant->key;
char func_name[256];
struct lp_type fs_type;
struct lp_type blend_type;
LLVMTypeRef fs_elem_type;
LLVMTypeRef fs_int_vec_type;
LLVMTypeRef blend_vec_type;
LLVMTypeRef arg_types[11];
LLVMTypeRef func_type;
LLVMValueRef context_ptr;
LLVMValueRef x;
LLVMValueRef y;
LLVMValueRef a0_ptr;
LLVMValueRef dadx_ptr;
LLVMValueRef dady_ptr;
LLVMValueRef color_ptr_ptr;
LLVMValueRef depth_ptr;
LLVMValueRef mask_input;
LLVMValueRef counter = NULL;
LLVMBasicBlockRef block;
LLVMBuilderRef builder;
struct lp_build_sampler_soa *sampler;
struct lp_build_interp_soa_context interp;
LLVMValueRef fs_mask[LP_MAX_VECTOR_LENGTH];
LLVMValueRef fs_out_color[PIPE_MAX_COLOR_BUFS][NUM_CHANNELS][LP_MAX_VECTOR_LENGTH];
LLVMValueRef blend_mask;
LLVMValueRef function;
LLVMValueRef facing;
unsigned num_fs;
unsigned i;
unsigned chan;
unsigned cbuf;
/* TODO: actually pick these based on the fs and color buffer
* characteristics. */
memset(&fs_type, 0, sizeof fs_type);
fs_type.floating = TRUE; /* floating point values */
fs_type.sign = TRUE; /* values are signed */
fs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
fs_type.width = 32; /* 32-bit float */
fs_type.length = 4; /* 4 elements per vector */
num_fs = 4; /* number of quads per block */
memset(&blend_type, 0, sizeof blend_type);
blend_type.floating = FALSE; /* values are integers */
blend_type.sign = FALSE; /* values are unsigned */
blend_type.norm = TRUE; /* values are in [0,1] or [-1,1] */
blend_type.width = 8; /* 8-bit ubyte values */
blend_type.length = 16; /* 16 elements per vector */
/*
* Generate the function prototype. Any change here must be reflected in
* lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
*/
fs_elem_type = lp_build_elem_type(fs_type);
fs_int_vec_type = lp_build_int_vec_type(fs_type);
blend_vec_type = lp_build_vec_type(blend_type);
util_snprintf(func_name, sizeof(func_name), "fs%u_variant%u_%s",
shader->no, variant->no, partial_mask ? "partial" : "whole");
arg_types[0] = screen->context_ptr_type; /* context */
arg_types[1] = LLVMInt32Type(); /* x */
arg_types[2] = LLVMInt32Type(); /* y */
arg_types[3] = LLVMFloatType(); /* facing */
arg_types[4] = LLVMPointerType(fs_elem_type, 0); /* a0 */
arg_types[5] = LLVMPointerType(fs_elem_type, 0); /* dadx */
arg_types[6] = LLVMPointerType(fs_elem_type, 0); /* dady */
arg_types[7] = LLVMPointerType(LLVMPointerType(blend_vec_type, 0), 0); /* color */
arg_types[8] = LLVMPointerType(fs_int_vec_type, 0); /* depth */
arg_types[9] = LLVMInt32Type(); /* mask_input */
arg_types[10] = LLVMPointerType(LLVMInt32Type(), 0);/* counter */
func_type = LLVMFunctionType(LLVMVoidType(), arg_types, Elements(arg_types), 0);
function = LLVMAddFunction(screen->module, func_name, func_type);
LLVMSetFunctionCallConv(function, LLVMCCallConv);
variant->function[partial_mask] = function;
/* XXX: need to propagate noalias down into color param now we are
* passing a pointer-to-pointer?
*/
for(i = 0; i < Elements(arg_types); ++i)
if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
LLVMAddAttribute(LLVMGetParam(function, i), LLVMNoAliasAttribute);
context_ptr = LLVMGetParam(function, 0);
x = LLVMGetParam(function, 1);
y = LLVMGetParam(function, 2);
facing = LLVMGetParam(function, 3);
a0_ptr = LLVMGetParam(function, 4);
dadx_ptr = LLVMGetParam(function, 5);
dady_ptr = LLVMGetParam(function, 6);
color_ptr_ptr = LLVMGetParam(function, 7);
depth_ptr = LLVMGetParam(function, 8);
mask_input = LLVMGetParam(function, 9);
lp_build_name(context_ptr, "context");
lp_build_name(x, "x");
lp_build_name(y, "y");
lp_build_name(a0_ptr, "a0");
lp_build_name(dadx_ptr, "dadx");
lp_build_name(dady_ptr, "dady");
lp_build_name(color_ptr_ptr, "color_ptr_ptr");
lp_build_name(depth_ptr, "depth");
lp_build_name(mask_input, "mask_input");
if (key->occlusion_count) {
counter = LLVMGetParam(function, 10);
lp_build_name(counter, "counter");
}
/*
* Function body
*/
block = LLVMAppendBasicBlock(function, "entry");
builder = LLVMCreateBuilder();
LLVMPositionBuilderAtEnd(builder, block);
/*
* The shader input interpolation info is not explicitely baked in the
* shader key, but everything it derives from (TGSI, and flatshade) is
* already included in the shader key.
*/
lp_build_interp_soa_init(&interp,
lp->num_inputs,
lp->inputs,
builder, fs_type,
a0_ptr, dadx_ptr, dady_ptr,
x, y);
/* code generated texture sampling */
sampler = lp_llvm_sampler_soa_create(key->sampler, context_ptr);
/* loop over quads in the block */
for(i = 0; i < num_fs; ++i) {
LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0);
LLVMValueRef out_color[PIPE_MAX_COLOR_BUFS][NUM_CHANNELS];
LLVMValueRef depth_ptr_i;
if(i != 0)
lp_build_interp_soa_update(&interp, i);
depth_ptr_i = LLVMBuildGEP(builder, depth_ptr, &index, 1, "");
generate_fs(lp, shader, key,
builder,
fs_type,
context_ptr,
i,
&interp,
sampler,
&fs_mask[i], /* output */
out_color,
depth_ptr_i,
facing,
partial_mask,
mask_input,
counter);
for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++)
for(chan = 0; chan < NUM_CHANNELS; ++chan)
fs_out_color[cbuf][chan][i] = out_color[cbuf][chan];
}
sampler->destroy(sampler);
/* Loop over color outputs / color buffers to do blending.
*/
for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
LLVMValueRef color_ptr;
LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), cbuf, 0);
LLVMValueRef blend_in_color[NUM_CHANNELS];
unsigned rt;
/*
* Convert the fs's output color and mask to fit to the blending type.
*/
for(chan = 0; chan < NUM_CHANNELS; ++chan) {
lp_build_conv(builder, fs_type, blend_type,
fs_out_color[cbuf][chan], num_fs,
&blend_in_color[chan], 1);
lp_build_name(blend_in_color[chan], "color%d.%c", cbuf, "rgba"[chan]);
}
if (partial_mask || !variant->opaque) {
lp_build_conv_mask(builder, fs_type, blend_type,
fs_mask, num_fs,
&blend_mask, 1);
} else {
blend_mask = lp_build_const_int_vec(blend_type, ~0);
}
color_ptr = LLVMBuildLoad(builder,
LLVMBuildGEP(builder, color_ptr_ptr, &index, 1, ""),
"");
lp_build_name(color_ptr, "color_ptr%d", cbuf);
/* which blend/colormask state to use */
rt = key->blend.independent_blend_enable ? cbuf : 0;
/*
* Blending.
*/
generate_blend(&key->blend,
rt,
builder,
blend_type,
context_ptr,
blend_mask,
blend_in_color,
color_ptr);
}
#ifdef PIPE_ARCH_X86
/* Avoid corrupting the FPU stack on 32bit OSes. */
lp_build_intrinsic(builder, "llvm.x86.mmx.emms", LLVMVoidType(), NULL, 0);
#endif
LLVMBuildRetVoid(builder);
LLVMDisposeBuilder(builder);
/* Verify the LLVM IR. If invalid, dump and abort */
#ifdef DEBUG
if(LLVMVerifyFunction(function, LLVMPrintMessageAction)) {
if (1)
lp_debug_dump_value(function);
abort();
}
#endif
/* Apply optimizations to LLVM IR */
LLVMRunFunctionPassManager(screen->pass, function);
if (gallivm_debug & GALLIVM_DEBUG_IR) {
/* Print the LLVM IR to stderr */
lp_debug_dump_value(function);
debug_printf("\n");
}
/*
* Translate the LLVM IR into machine code.
*/
{
void *f = LLVMGetPointerToGlobal(screen->engine, function);
variant->jit_function[partial_mask] = (lp_jit_frag_func)pointer_to_func(f);
if (gallivm_debug & GALLIVM_DEBUG_ASM) {
lp_disassemble(f);
}
lp_func_delete_body(function);
}
}
