/**
* Normalized 8bit multiplication.
*
* - alpha plus one
*
* makes the following approximation to the division (Sree)
*
* a*b/255 ~= (a*(b + 1)) >> 256
*
* which is the fastest method that satisfies the following OpenGL criteria
*
* 0*0 = 0 and 255*255 = 255
*
* - geometric series
*
* takes the geometric series approximation to the division
*
* t/255 = (t >> 8) + (t >> 16) + (t >> 24) ..
*
* in this case just the first two terms to fit in 16bit arithmetic
*
* t/255 ~= (t + (t >> 8)) >> 8
*
* note that just by itself it doesn't satisfies the OpenGL criteria, as
* 255*255 = 254, so the special case b = 255 must be accounted or roundoff
* must be used
*
* - geometric series plus rounding
*
* when using a geometric series division instead of truncating the result
* use roundoff in the approximation (Jim Blinn)
*
* t/255 ~= (t + (t >> 8) + 0x80) >> 8
*
* achieving the exact results
*
* @sa Alvy Ray Smith, Image Compositing Fundamentals, Tech Memo 4, Aug 15, 1995,
* ftp://ftp.alvyray.com/Acrobat/4_Comp.pdf
* @sa Michael Herf, The "double blend trick", May 2000,
* http://www.stereopsis.com/doubleblend.html
*/
static LLVMValueRef
lp_build_mul_u8n(LLVMBuilderRef builder,
struct lp_type i16_type,
LLVMValueRef a, LLVMValueRef b)
{
LLVMValueRef c8;
LLVMValueRef ab;
c8 = lp_build_int_const_scalar(i16_type, 8);
#if 0
/* a*b/255 ~= (a*(b + 1)) >> 256 */
b = LLVMBuildAdd(builder, b, lp_build_int_const_scalar(i16_type, 1), "");
ab = LLVMBuildMul(builder, a, b, "");
#else
/* ab/255 ~= (ab + (ab >> 8) + 0x80) >> 8 */
ab = LLVMBuildMul(builder, a, b, "");
ab = LLVMBuildAdd(builder, ab, LLVMBuildLShr(builder, ab, c8, ""), "");
ab = LLVMBuildAdd(builder, ab, lp_build_int_const_scalar(i16_type, 0x80), "");
#endif
ab = LLVMBuildLShr(builder, ab, c8, "");
return ab;
}
/*
* Derive from the quad's upper left scalar coordinates the coordinates for
* all other quad pixels
*/
static void
generate_pos0(LLVMBuilderRef builder,
LLVMValueRef x,
LLVMValueRef y,
LLVMValueRef *x0,
LLVMValueRef *y0)
{
LLVMTypeRef int_elem_type = LLVMInt32Type();
LLVMTypeRef int_vec_type = LLVMVectorType(int_elem_type, QUAD_SIZE);
LLVMTypeRef elem_type = LLVMFloatType();
LLVMTypeRef vec_type = LLVMVectorType(elem_type, QUAD_SIZE);
LLVMValueRef x_offsets[QUAD_SIZE];
LLVMValueRef y_offsets[QUAD_SIZE];
unsigned i;
x = lp_build_broadcast(builder, int_vec_type, x);
y = lp_build_broadcast(builder, int_vec_type, y);
for(i = 0; i < QUAD_SIZE; ++i) {
x_offsets[i] = LLVMConstInt(int_elem_type, quad_offset_x[i], 0);
y_offsets[i] = LLVMConstInt(int_elem_type, quad_offset_y[i], 0);
}
x = LLVMBuildAdd(builder, x, LLVMConstVector(x_offsets, QUAD_SIZE), "");
y = LLVMBuildAdd(builder, y, LLVMConstVector(y_offsets, QUAD_SIZE), "");
*x0 = LLVMBuildSIToFP(builder, x, vec_type, "");
*y0 = LLVMBuildSIToFP(builder, y, vec_type, "");
}
LLVMValueRef gen_add(struct node *ast)
{
return LLVMBuildAdd(builder,
codegen(ast->one),
codegen(ast->two),
"");
}
int main(int argc, char const *argv[]) {
LLVMModuleRef mod = LLVMModuleCreateWithName("sum");
LLVMTypeRef param_types[] = { LLVMInt32Type(), LLVMInt32Type() };
LLVMTypeRef ret_type = LLVMFunctionType(LLVMInt32Type(), /* ret type */
param_types, /* arg types */
2, /* arg count */
0 /* is variadic */);
LLVMValueRef sum = LLVMAddFunction(mod, "sum", ret_type);
LLVMBasicBlockRef entry = LLVMAppendBasicBlock(sum, "entry");
LLVMBuilderRef builder = LLVMCreateBuilder();
LLVMPositionBuilderAtEnd(builder, entry);
LLVMValueRef tmp = LLVMBuildAdd(builder,
LLVMGetParam(sum, 0),
LLVMGetParam(sum, 1), "tmp");
LLVMBuildRet(builder, tmp);
char *error = NULL;
LLVMVerifyModule(mod, LLVMAbortProcessAction, &error);
LLVMDisposeMessage(error);
LLVMExecutionEngineRef engine;
error = NULL;
LLVMLinkInJIT();
LLVMInitializeNativeTarget();
if (LLVMCreateExecutionEngineForModule(&engine, mod, &error) != 0) {
fprintf(stderr, "failed to create execution engine\n");
abort();
}
if (error) {
fprintf(stderr, "error: %s\n", error);
LLVMDisposeMessage(error);
exit(EXIT_FAILURE);
}
if (argc < 3) {
fprintf(stderr, "usage: %s x y\n", argv[0]);
exit(EXIT_FAILURE);
}
long long x = strtoll(argv[1], NULL, 10);
long long y = strtoll(argv[2], NULL, 10);
LLVMGenericValueRef args[] = {
LLVMCreateGenericValueOfInt(LLVMInt32Type(), x, 0),
LLVMCreateGenericValueOfInt(LLVMInt32Type(), y, 0),
};
LLVMGenericValueRef res = LLVMRunFunction(engine, sum, 2, args);
printf("%d\n", (int)LLVMGenericValueToInt(res, 0));
// write bitcode to file
if (LLVMWriteBitcodeToFile(mod, "sum.bc") != 0) {
fprintf(stderr, "error writing bitcode to file\n");
}
LLVMDisposeBuilder(builder);
LLVMDisposeExecutionEngine(engine);
}
void
lp_build_loop_end_cond(struct lp_build_loop_state *state,
LLVMValueRef end,
LLVMValueRef step,
LLVMIntPredicate llvm_cond)
{
LLVMBuilderRef builder = state->gallivm->builder;
LLVMValueRef next;
LLVMValueRef cond;
LLVMBasicBlockRef after_block;
if (!step)
step = LLVMConstInt(LLVMTypeOf(end), 1, 0);
next = LLVMBuildAdd(builder, state->counter, step, "");
LLVMBuildStore(builder, next, state->counter_var);
cond = LLVMBuildICmp(builder, llvm_cond, next, end, "");
after_block = lp_build_insert_new_block(state->gallivm, "loop_end");
LLVMBuildCondBr(builder, cond, after_block, state->block);
LLVMPositionBuilderAtEnd(builder, after_block);
state->counter = LLVMBuildLoad(builder, state->counter_var, "");
}
static void pointer_offset(compile_t* c, reach_type_t* t, reach_type_t* t_elem)
{
FIND_METHOD("_offset");
LLVMTypeRef params[3];
params[0] = t->use_type;
params[1] = c->intptr;
start_function(c, m, t->use_type, params, 2);
// Set up a constant integer for the allocation size.
size_t size = (size_t)LLVMABISizeOfType(c->target_data, t_elem->use_type);
LLVMValueRef l_size = LLVMConstInt(c->intptr, size, false);
LLVMValueRef ptr = LLVMGetParam(m->func, 0);
LLVMValueRef n = LLVMGetParam(m->func, 1);
// Return ptr + (n * sizeof(len)).
LLVMValueRef src = LLVMBuildPtrToInt(c->builder, ptr, c->intptr, "");
LLVMValueRef offset = LLVMBuildMul(c->builder, n, l_size, "");
LLVMValueRef result = LLVMBuildAdd(c->builder, src, offset, "");
result = LLVMBuildIntToPtr(c->builder, result, t->use_type, "");
LLVMBuildRet(c->builder, result);
codegen_finishfun(c);
BOX_FUNCTION();
}
void genprim_string_serialise_trace(compile_t* c, reach_type_t* t)
{
// Generate the serialise_trace function.
t->serialise_trace_fn = codegen_addfun(c, genname_serialise_trace(t->name),
c->serialise_type);
codegen_startfun(c, t->serialise_trace_fn, NULL, NULL);
LLVMSetFunctionCallConv(t->serialise_trace_fn, LLVMCCallConv);
LLVMValueRef ctx = LLVMGetParam(t->serialise_trace_fn, 0);
LLVMValueRef arg = LLVMGetParam(t->serialise_trace_fn, 1);
LLVMValueRef object = LLVMBuildBitCast(c->builder, arg, t->use_type, "");
// Read the size.
LLVMValueRef size = field_value(c, object, 1);
LLVMValueRef alloc = LLVMBuildAdd(c->builder, size,
LLVMConstInt(c->intptr, 1, false), "");
// Reserve space for the contents.
LLVMValueRef ptr = field_value(c, object, 3);
LLVMValueRef args[3];
args[0] = ctx;
args[1] = ptr;
args[2] = alloc;
gencall_runtime(c, "pony_serialise_reserve", args, 3, "");
LLVMBuildRetVoid(c->builder);
codegen_finishfun(c);
}
static LLVMValueRef
CreateFibFunction(LLVMModuleRef M, LLVMContextRef Context)
{
LLVMBuilderRef B = LLVMCreateBuilderInContext(Context);
// Create the fib function and insert it into module M. This function is said
// to return an int and take an int parameter.
LLVMTypeRef ParamTypes[] = {LLVMInt32TypeInContext(Context)};
LLVMTypeRef ReturnType = LLVMInt32TypeInContext(Context);
LLVMTypeRef FunctionTy = LLVMFunctionType(ReturnType, ParamTypes, 1, 0);
LLVMValueRef FibF = LLVMAddFunction(M, "fib", FunctionTy);
// Add a basic block to the function.
LLVMBasicBlockRef BB = LLVMAppendBasicBlockInContext(Context, FibF, "EntryBlock");
// Get pointers to the constants.
LLVMValueRef One = LLVMConstInt(LLVMInt32TypeInContext(Context), 1, 0);
LLVMValueRef Two = LLVMConstInt(LLVMInt32TypeInContext(Context), 2, 0);
// Get pointer to the integer argument of the add1 function...
LLVMValueRef ArgX = LLVMGetFirstParam(FibF); // Get the arg.
LLVMSetValueName(ArgX, "AnArg"); // Give it a nice symbolic name for fun.
// Create the true_block.
LLVMBasicBlockRef RetBB = LLVMAppendBasicBlockInContext(Context, FibF, "return");
// Create an exit block.
LLVMBasicBlockRef RecurseBB = LLVMAppendBasicBlockInContext(Context, FibF, "recurse");
// Create the "if (arg <= 2) goto exitbb"
LLVMPositionBuilderAtEnd(B, BB);
LLVMValueRef CondInst = LLVMBuildICmp(B, LLVMIntSLE, ArgX, Two, "cond");
LLVMBuildCondBr(B, CondInst, RetBB, RecurseBB);
// Create: ret int 1
LLVMPositionBuilderAtEnd(B, RetBB);
LLVMBuildRet(B, One);
// create fib(x-1)
LLVMPositionBuilderAtEnd(B, RecurseBB);
LLVMValueRef Sub = LLVMBuildSub(B, ArgX, One, "arg");
LLVMValueRef CallFibX1 = LLVMBuildCall(B, FibF, &Sub, 1, "fibx1");
LLVMSetTailCall(CallFibX1, 1);
// create fib(x-2)
LLVMPositionBuilderAtEnd(B, RecurseBB);
Sub = LLVMBuildSub(B, ArgX, Two, "arg");
LLVMValueRef CallFibX2 = LLVMBuildCall(B, FibF, &Sub, 1, "fibx2");
LLVMSetTailCall(CallFibX2, 1);
// fib(x-1)+fib(x-2)
LLVMPositionBuilderAtEnd(B, RecurseBB);
LLVMValueRef Sum = LLVMBuildAdd(B, CallFibX1, CallFibX2, "addresult");
// Create the return instruction and add it to the basic block
LLVMPositionBuilderAtEnd(B, RecurseBB);
LLVMBuildRet(B, Sum);
return FibF;
}
LLVMValueRef gendesc_fieldptr(compile_t* c, LLVMValueRef ptr,
LLVMValueRef field_info)
{
LLVMValueRef offset = LLVMBuildExtractValue(c->builder, field_info, 0, "");
offset = LLVMBuildZExt(c->builder, offset, c->intptr, "");
return LLVMBuildAdd(c->builder, ptr, offset, "");
}
struct LLVMOpaqueValue *bllvm_compile_rirbop(const struct rir_expression *expr,
struct llvm_traversal_ctx *ctx)
{
LLVMValueRef ret;
LLVMValueRef left = bllvm_value_from_rir_value_or_die(expr->binaryop.a, ctx);
LLVMValueRef right = bllvm_value_from_rir_value_or_die(expr->binaryop.b, ctx);
switch(expr->type) {
case RIR_EXPRESSION_ADD:
ret = LLVMBuildAdd(ctx->builder, left, right, "");
break;
case RIR_EXPRESSION_SUB:
ret = LLVMBuildSub(ctx->builder, left, right, "");
break;
case RIR_EXPRESSION_MUL:
ret = LLVMBuildMul(ctx->builder, left, right, "");
break;
case RIR_EXPRESSION_DIV:
ret = LLVMBuildUDiv(ctx->builder, left, right, "");
break;
default:
RF_CRITICAL_FAIL("Should never get anything other than binaryop here");
break;
}
return ret;
}
static void
store_cached_block(struct gallivm_state *gallivm,
LLVMValueRef *col,
LLVMValueRef tag_value,
LLVMValueRef hash_index,
LLVMValueRef cache)
{
LLVMBuilderRef builder = gallivm->builder;
LLVMValueRef ptr, indices[3];
LLVMTypeRef type_ptr4x32;
unsigned count;
type_ptr4x32 = LLVMPointerType(LLVMVectorType(LLVMInt32TypeInContext(gallivm->context), 4), 0);
indices[0] = lp_build_const_int32(gallivm, 0);
indices[1] = lp_build_const_int32(gallivm, LP_BUILD_FORMAT_CACHE_MEMBER_TAGS);
indices[2] = hash_index;
ptr = LLVMBuildGEP(builder, cache, indices, ARRAY_SIZE(indices), "");
LLVMBuildStore(builder, tag_value, ptr);
indices[1] = lp_build_const_int32(gallivm, LP_BUILD_FORMAT_CACHE_MEMBER_DATA);
hash_index = LLVMBuildMul(builder, hash_index,
lp_build_const_int32(gallivm, 16), "");
for (count = 0; count < 4; count++) {
indices[2] = hash_index;
ptr = LLVMBuildGEP(builder, cache, indices, ARRAY_SIZE(indices), "");
ptr = LLVMBuildBitCast(builder, ptr, type_ptr4x32, "");
LLVMBuildStore(builder, col[count], ptr);
hash_index = LLVMBuildAdd(builder, hash_index,
lp_build_const_int32(gallivm, 4), "");
}
}
/**
* Special case for converting clamped IEEE-754 floats to unsigned norms.
*
* The mathematical voodoo below may seem excessive but it is actually
* paramount we do it this way for several reasons. First, there is no single
* precision FP to unsigned integer conversion Intel SSE instruction. Second,
* secondly, even if there was, since the FP's mantissa takes only a fraction
* of register bits the typically scale and cast approach would require double
* precision for accurate results, and therefore half the throughput
*
* Although the result values can be scaled to an arbitrary bit width specified
* by dst_width, the actual result type will have the same width.
*/
LLVMValueRef
lp_build_clamped_float_to_unsigned_norm(LLVMBuilderRef builder,
struct lp_type src_type,
unsigned dst_width,
LLVMValueRef src)
{
LLVMTypeRef int_vec_type = lp_build_int_vec_type(src_type);
LLVMValueRef res;
unsigned mantissa;
unsigned n;
unsigned long long ubound;
unsigned long long mask;
double scale;
double bias;
assert(src_type.floating);
mantissa = lp_mantissa(src_type);
/* We cannot carry more bits than the mantissa */
n = MIN2(mantissa, dst_width);
/* This magic coefficients will make the desired result to appear in the
* lowest significant bits of the mantissa.
*/
ubound = ((unsigned long long)1 << n);
mask = ubound - 1;
scale = (double)mask/ubound;
bias = (double)((unsigned long long)1 << (mantissa - n));
res = LLVMBuildMul(builder, src, lp_build_const_scalar(src_type, scale), "");
res = LLVMBuildAdd(builder, res, lp_build_const_scalar(src_type, bias), "");
res = LLVMBuildBitCast(builder, res, int_vec_type, "");
if(dst_width > n) {
int shift = dst_width - n;
res = LLVMBuildShl(builder, res, lp_build_int_const_scalar(src_type, shift), "");
/* TODO: Fill in the empty lower bits for additional precision? */
#if 0
{
LLVMValueRef msb;
msb = LLVMBuildLShr(builder, res, lp_build_int_const_scalar(src_type, dst_width - 1), "");
msb = LLVMBuildShl(builder, msb, lp_build_int_const_scalar(src_type, shift), "");
msb = LLVMBuildSub(builder, msb, lp_build_int_const_scalar(src_type, 1), "");
res = LLVMBuildOr(builder, res, msb, "");
}
#elif 0
while(shift > 0) {
res = LLVMBuildOr(builder, res, LLVMBuildLShr(builder, res, lp_build_int_const_scalar(src_type, n), ""), "");
shift -= n;
n *= 2;
}
#endif
}
else
res = LLVMBuildAnd(builder, res, lp_build_int_const_scalar(src_type, mask), "");
return res;
}
void
lp_build_exp2_approx(struct lp_build_context *bld,
LLVMValueRef x,
LLVMValueRef *p_exp2_int_part,
LLVMValueRef *p_frac_part,
LLVMValueRef *p_exp2)
{
const struct lp_type type = bld->type;
LLVMTypeRef vec_type = lp_build_vec_type(type);
LLVMTypeRef int_vec_type = lp_build_int_vec_type(type);
LLVMValueRef ipart = NULL;
LLVMValueRef fpart = NULL;
LLVMValueRef expipart = NULL;
LLVMValueRef expfpart = NULL;
LLVMValueRef res = NULL;
if(p_exp2_int_part || p_frac_part || p_exp2) {
/* TODO: optimize the constant case */
if(LLVMIsConstant(x))
debug_printf("%s: inefficient/imprecise constant arithmetic\n",
__FUNCTION__);
assert(type.floating && type.width == 32);
x = lp_build_min(bld, x, lp_build_const_scalar(type, 129.0));
x = lp_build_max(bld, x, lp_build_const_scalar(type, -126.99999));
/* ipart = int(x - 0.5) */
ipart = LLVMBuildSub(bld->builder, x, lp_build_const_scalar(type, 0.5f), "");
ipart = LLVMBuildFPToSI(bld->builder, ipart, int_vec_type, "");
/* fpart = x - ipart */
fpart = LLVMBuildSIToFP(bld->builder, ipart, vec_type, "");
fpart = LLVMBuildSub(bld->builder, x, fpart, "");
}
if(p_exp2_int_part || p_exp2) {
/* expipart = (float) (1 << ipart) */
expipart = LLVMBuildAdd(bld->builder, ipart, lp_build_int_const_scalar(type, 127), "");
expipart = LLVMBuildShl(bld->builder, expipart, lp_build_int_const_scalar(type, 23), "");
expipart = LLVMBuildBitCast(bld->builder, expipart, vec_type, "");
}
if(p_exp2) {
expfpart = lp_build_polynomial(bld, fpart, lp_build_exp2_polynomial,
Elements(lp_build_exp2_polynomial));
res = LLVMBuildMul(bld->builder, expipart, expfpart, "");
}
if(p_exp2_int_part)
*p_exp2_int_part = expipart;
if(p_frac_part)
*p_frac_part = fpart;
if(p_exp2)
*p_exp2 = res;
}
static LLVMValueRef lvalue_index(struct node *ast)
{
LLVMValueRef ptr;
ptr = LLVMBuildAdd(builder, codegen(ast->one), codegen(ast->two), "");
return rvalue_to_lvalue(ptr);
}
static void emit_uadd(const struct lp_build_tgsi_action *action,
struct lp_build_tgsi_context *bld_base,
struct lp_build_emit_data *emit_data)
{
LLVMBuilderRef builder = bld_base->base.gallivm->builder;
emit_data->output[emit_data->chan] = LLVMBuildAdd(builder,
emit_data->args[0], emit_data->args[1], "");
}
static LLVMValueRef field_loc(compile_t* c, LLVMValueRef offset,
LLVMTypeRef structure, LLVMTypeRef ftype, int index)
{
LLVMValueRef f_offset = LLVMBuildAdd(c->builder, offset,
LLVMConstInt(c->intptr,
LLVMOffsetOfElement(c->target_data, structure, index), false), "");
return LLVMBuildIntToPtr(c->builder, f_offset,
LLVMPointerType(ftype, 0), "");
}
static LLVMValueRef
emit_array_index(
struct lp_build_tgsi_soa_context *bld,
const struct tgsi_full_src_register *reg,
unsigned swizzle)
{
struct gallivm_state * gallivm = bld->bld_base.base.gallivm;
LLVMValueRef addr = LLVMBuildLoad(gallivm->builder,
bld->addr[reg->Indirect.Index][swizzle], "");
LLVMValueRef offset = lp_build_const_int32(gallivm, reg->Register.Index);
LLVMValueRef hw_index = LLVMBuildAdd(gallivm->builder, addr, offset, "");
LLVMValueRef soa_index = LLVMBuildMul(gallivm->builder, hw_index,
lp_build_const_int32(gallivm, 4), "");
LLVMValueRef array_index = LLVMBuildAdd(gallivm->builder, soa_index,
lp_build_const_int32(gallivm, swizzle), "");
return array_index;
}
static LLVMValueRef
emit_array_index(
struct lp_build_tgsi_soa_context *bld,
const struct tgsi_ind_register *reg,
unsigned offset)
{
struct gallivm_state * gallivm = bld->bld_base.base.gallivm;
LLVMValueRef addr = LLVMBuildLoad(gallivm->builder, bld->addr[reg->Index][reg->Swizzle], "");
return LLVMBuildAdd(gallivm->builder, addr, lp_build_const_int32(gallivm, offset), "");
}
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;
}
/*
* SI implements derivatives using the local data store (LDS)
* All writes to the LDS happen in all executing threads at
* the same time. TID is the Thread ID for the current
* thread and is a value between 0 and 63, representing
* the thread's position in the wavefront.
*
* For the pixel shader threads are grouped into quads of four pixels.
* The TIDs of the pixels of a quad are:
*
* +------+------+
* |4n + 0|4n + 1|
* +------+------+
* |4n + 2|4n + 3|
* +------+------+
*
* So, masking the TID with 0xfffffffc yields the TID of the top left pixel
* of the quad, masking with 0xfffffffd yields the TID of the top pixel of
* the current pixel's column, and masking with 0xfffffffe yields the TID
* of the left pixel of the current pixel's row.
*
* Adding 1 yields the TID of the pixel to the right of the left pixel, and
* adding 2 yields the TID of the pixel below the top pixel.
*/
LLVMValueRef
ac_build_ddxy(struct ac_llvm_context *ctx,
bool has_ds_bpermute,
uint32_t mask,
int idx,
LLVMValueRef lds,
LLVMValueRef val)
{
LLVMValueRef thread_id, tl, trbl, tl_tid, trbl_tid, args[2];
LLVMValueRef result;
thread_id = ac_get_thread_id(ctx);
tl_tid = LLVMBuildAnd(ctx->builder, thread_id,
LLVMConstInt(ctx->i32, mask, false), "");
trbl_tid = LLVMBuildAdd(ctx->builder, tl_tid,
LLVMConstInt(ctx->i32, idx, false), "");
if (has_ds_bpermute) {
args[0] = LLVMBuildMul(ctx->builder, tl_tid,
LLVMConstInt(ctx->i32, 4, false), "");
args[1] = val;
tl = ac_build_intrinsic(ctx,
"llvm.amdgcn.ds.bpermute", ctx->i32,
args, 2,
AC_FUNC_ATTR_READNONE |
AC_FUNC_ATTR_CONVERGENT);
args[0] = LLVMBuildMul(ctx->builder, trbl_tid,
LLVMConstInt(ctx->i32, 4, false), "");
trbl = ac_build_intrinsic(ctx,
"llvm.amdgcn.ds.bpermute", ctx->i32,
args, 2,
AC_FUNC_ATTR_READNONE |
AC_FUNC_ATTR_CONVERGENT);
} else {
LLVMValueRef store_ptr, load_ptr0, load_ptr1;
store_ptr = ac_build_gep0(ctx, lds, thread_id);
load_ptr0 = ac_build_gep0(ctx, lds, tl_tid);
load_ptr1 = ac_build_gep0(ctx, lds, trbl_tid);
LLVMBuildStore(ctx->builder, val, store_ptr);
tl = LLVMBuildLoad(ctx->builder, load_ptr0, "");
trbl = LLVMBuildLoad(ctx->builder, load_ptr1, "");
}
tl = LLVMBuildBitCast(ctx->builder, tl, ctx->f32, "");
trbl = LLVMBuildBitCast(ctx->builder, trbl, ctx->f32, "");
result = LLVMBuildFSub(ctx->builder, trbl, tl, "");
return result;
}
LLVMValueRef build_store_mrs(
struct llvm_ctx *ctx,
LLVMValueRef pos,
const LLVMValueRef *words,
int num_words)
{
for(int i=0; i<num_words; i++) {
V t_addr = UTCB_ADDR_VAL(ctx, pos, "store.mr.ptr");
LLVMBuildStore(ctx->builder, words[i], t_addr);
pos = LLVMBuildAdd(ctx->builder, pos, CONST_INT(1), "t.pos");
}
return pos;
}
LLVMValueRef gendesc_ptr_to_fields(compile_t* c, LLVMValueRef object,
LLVMValueRef desc)
{
// Skip the descriptor.
LLVMValueRef offset = desc_field(c, desc, DESC_FIELD_OFFSET);
offset = LLVMBuildZExt(c->builder, offset, c->intptr, "");
LLVMValueRef base = LLVMBuildPtrToInt(c->builder, object, c->intptr, "");
LLVMValueRef result = LLVMBuildAdd(c->builder, base, offset, "");
// Return as a c->intptr.
return result;
}
LLVMValueRef gen_preinc(struct node *ast)
{
LLVMValueRef result;
result = LLVMBuildAdd(builder,
codegen(ast->one),
CONST(1),
"");
check_store(result, lvalue(ast->one));
return result;
}
LLVMValueRef gen_postinc(struct node *ast)
{
LLVMValueRef orig, result;
orig = codegen(ast->one);
result = LLVMBuildAdd(builder,
orig,
CONST(1),
"");
check_store(result, lvalue(ast->one));
return orig;
}
/**
* Small vector x scale multiplication optimization.
*/
LLVMValueRef
lp_build_mul_imm(struct lp_build_context *bld,
LLVMValueRef a,
int b)
{
LLVMValueRef factor;
if(b == 0)
return bld->zero;
if(b == 1)
return a;
if(b == -1)
return LLVMBuildNeg(bld->builder, a, "");
if(b == 2 && bld->type.floating)
return lp_build_add(bld, a, a);
if(util_is_pot(b)) {
unsigned shift = ffs(b) - 1;
if(bld->type.floating) {
#if 0
/*
* Power of two multiplication by directly manipulating the mantissa.
*
* XXX: This might not be always faster, it will introduce a small error
* for multiplication by zero, and it will produce wrong results
* for Inf and NaN.
*/
unsigned mantissa = lp_mantissa(bld->type);
factor = lp_build_int_const_scalar(bld->type, (unsigned long long)shift << mantissa);
a = LLVMBuildBitCast(bld->builder, a, lp_build_int_vec_type(bld->type), "");
a = LLVMBuildAdd(bld->builder, a, factor, "");
a = LLVMBuildBitCast(bld->builder, a, lp_build_vec_type(bld->type), "");
return a;
#endif
}
else {
factor = lp_build_const_scalar(bld->type, shift);
return LLVMBuildShl(bld->builder, a, factor, "");
}
}
factor = lp_build_const_scalar(bld->type, (double)b);
return lp_build_mul(bld, a, factor);
}
static void trace_array_elements(compile_t* c, reach_type_t* t,
LLVMValueRef ctx, LLVMValueRef object, LLVMValueRef pointer)
{
// Get the type argument for the array. This will be used to generate the
// per-element trace call.
ast_t* typeargs = ast_childidx(t->ast, 2);
ast_t* typearg = ast_child(typeargs);
if(!gentrace_needed(typearg))
return;
reach_type_t* t_elem = reach_type(c->reach, typearg);
pointer = LLVMBuildBitCast(c->builder, pointer,
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");
// Read the size.
LLVMValueRef size = field_value(c, object, 1);
LLVMBuildBr(c->builder, cond_block);
// While the index is less than the size, trace 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 trace it.
LLVMPositionBuilderAtEnd(c->builder, body_block);
LLVMValueRef elem_ptr = LLVMBuildGEP(c->builder, pointer, &phi, 1, "elem");
LLVMValueRef elem = LLVMBuildLoad(c->builder, elem_ptr, "");
gentrace(c, ctx, elem, typearg);
// 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);
}
static LLVMValueRef get_instance_index(
struct radeon_llvm_context * radeon_bld,
unsigned divisor)
{
struct gallivm_state * gallivm = radeon_bld->soa.bld_base.base.gallivm;
LLVMValueRef result = LLVMGetParam(radeon_bld->main_fn, SI_PARAM_INSTANCE_ID);
result = LLVMBuildAdd(gallivm->builder, result, LLVMGetParam(
radeon_bld->main_fn, SI_PARAM_START_INSTANCE), "");
if (divisor > 1)
result = LLVMBuildUDiv(gallivm->builder, result,
lp_build_const_int32(gallivm, divisor), "");
return result;
}
/**
* Build LLVM code for texture coord wrapping, for nearest filtering,
* for scaled integer texcoords.
* \param block_length is the length of the pixel block along the
* coordinate axis
* \param coord the incoming texcoord (s,t,r or q) scaled to the texture size
* \param length the texture size along one dimension
* \param stride pixel stride along the coordinate axis (in bytes)
* \param is_pot if TRUE, length is a power of two
* \param wrap_mode one of PIPE_TEX_WRAP_x
* \param out_offset byte offset for the wrapped coordinate
* \param out_i resulting sub-block pixel coordinate for coord0
*/
static void
lp_build_sample_wrap_nearest_int(struct lp_build_sample_context *bld,
unsigned block_length,
LLVMValueRef coord,
LLVMValueRef length,
LLVMValueRef stride,
boolean is_pot,
unsigned wrap_mode,
LLVMValueRef *out_offset,
LLVMValueRef *out_i)
{
struct lp_build_context *int_coord_bld = &bld->int_coord_bld;
LLVMBuilderRef builder = bld->gallivm->builder;
LLVMValueRef length_minus_one;
length_minus_one = lp_build_sub(int_coord_bld, length, int_coord_bld->one);
switch(wrap_mode) {
case PIPE_TEX_WRAP_REPEAT:
if(is_pot)
coord = LLVMBuildAnd(builder, coord, length_minus_one, "");
else {
/* Add a bias to the texcoord to handle negative coords */
LLVMValueRef bias = lp_build_mul_imm(int_coord_bld, length, 1024);
coord = LLVMBuildAdd(builder, coord, bias, "");
coord = LLVMBuildURem(builder, coord, length, "");
}
break;
case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
coord = lp_build_max(int_coord_bld, coord, int_coord_bld->zero);
coord = lp_build_min(int_coord_bld, coord, length_minus_one);
break;
case PIPE_TEX_WRAP_CLAMP:
case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
case PIPE_TEX_WRAP_MIRROR_REPEAT:
case PIPE_TEX_WRAP_MIRROR_CLAMP:
case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE:
case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER:
default:
assert(0);
}
lp_build_sample_partial_offset(int_coord_bld, block_length, coord, stride,
out_offset, out_i);
}
/**
* Generate a + b
*/
LLVMValueRef
lp_build_add(struct lp_build_context *bld,
LLVMValueRef a,
LLVMValueRef b)
{
const struct lp_type type = bld->type;
LLVMValueRef res;
if(a == bld->zero)
return b;
if(b == bld->zero)
return a;
if(a == bld->undef || b == bld->undef)
return bld->undef;
if(bld->type.norm) {
const char *intrinsic = NULL;
if(a == bld->one || b == bld->one)
return bld->one;
if(util_cpu_caps.has_sse2 &&
type.width * type.length == 128 &&
!type.floating && !type.fixed) {
if(type.width == 8)
intrinsic = type.sign ? "llvm.x86.sse2.padds.b" : "llvm.x86.sse2.paddus.b";
if(type.width == 16)
intrinsic = type.sign ? "llvm.x86.sse2.padds.w" : "llvm.x86.sse2.paddus.w";
}
if(intrinsic)
return lp_build_intrinsic_binary(bld->builder, intrinsic, lp_build_vec_type(bld->type), a, b);
}
if(LLVMIsConstant(a) && LLVMIsConstant(b))
res = LLVMConstAdd(a, b);
else
res = LLVMBuildAdd(bld->builder, a, b, "");
/* clamp to ceiling of 1.0 */
if(bld->type.norm && (bld->type.floating || bld->type.fixed))
res = lp_build_min_simple(bld, res, bld->one);
/* XXX clamp to floor of -1 or 0??? */
return res;
}
static void pointer_delete(compile_t* c, reach_type_t* t, reach_type_t* t_elem)
{
FIND_METHOD("_delete");
LLVMTypeRef params[3];
params[0] = t->use_type;
params[1] = c->intptr;
params[2] = c->intptr;
start_function(c, m, t_elem->use_type, params, 3);
// Set up a constant integer for the allocation size.
size_t size = (size_t)LLVMABISizeOfType(c->target_data, t_elem->use_type);
LLVMValueRef l_size = LLVMConstInt(c->intptr, size, false);
LLVMValueRef ptr = LLVMGetParam(m->func, 0);
LLVMValueRef n = LLVMGetParam(m->func, 1);
LLVMValueRef len = LLVMGetParam(m->func, 2);
LLVMValueRef elem_ptr = LLVMBuildBitCast(c->builder, ptr,
LLVMPointerType(t_elem->use_type, 0), "");
LLVMValueRef result = LLVMBuildLoad(c->builder, elem_ptr, "");
LLVMValueRef dst = LLVMBuildPtrToInt(c->builder, elem_ptr, c->intptr, "");
LLVMValueRef offset = LLVMBuildMul(c->builder, n, l_size, "");
LLVMValueRef src = LLVMBuildAdd(c->builder, dst, offset, "");
LLVMValueRef elen = LLVMBuildMul(c->builder, len, l_size, "");
LLVMValueRef args[5];
args[0] = LLVMBuildIntToPtr(c->builder, dst, c->void_ptr, "");
args[1] = LLVMBuildIntToPtr(c->builder, src, c->void_ptr, "");
args[2] = elen;
args[3] = LLVMConstInt(c->i32, 1, false);
args[4] = LLVMConstInt(c->i1, 0, false);
// llvm.memmove.*(ptr, ptr + (n * sizeof(elem)), len * sizeof(elem))
if(target_is_ilp32(c->opt->triple))
{
gencall_runtime(c, "llvm.memmove.p0i8.p0i8.i32", args, 5, "");
} else {
gencall_runtime(c, "llvm.memmove.p0i8.p0i8.i64", args, 5, "");
}
// Return ptr[0].
LLVMBuildRet(c->builder, result);
codegen_finishfun(c);
}
