int main (void) {
LLVMModuleRef module = LLVMModuleCreateWithName("kal");
LLVMBuilderRef builder = LLVMCreateBuilder();
// LLVMInitializeNativeTarget();
LLVMTypeRef funcType = LLVMFunctionType(LLVMVoidType(), NULL, 0, 0);
LLVMValueRef func = LLVMAddFunction(module, "main", funcType);
LLVMSetLinkage(func, LLVMExternalLinkage);
LLVMBasicBlockRef block = LLVMAppendBasicBlock(func, "entry");
LLVMPositionBuilderAtEnd(builder, block);
LLVMValueRef cond = LLVMBuildICmp(builder, LLVMIntNE, LLVMConstInt(LLVMInt32Type(), 2, 0), LLVMConstInt(LLVMInt32Type(), 1, 0), "ifcond");
LLVMValueRef owning_block = LLVMGetBasicBlockParent(LLVMGetInsertBlock(builder)); //TODO: WRONG??
//LLVMValueRef owning_block = LLVMBasicBlockAsValue(LLVMGetPreviousBasicBlock(LLVMGetInsertBlock(builder)));
// 2. Generate new blocks for cases.
LLVMBasicBlockRef then_ref = LLVMAppendBasicBlock(owning_block, "then");
LLVMBasicBlockRef else_ref = LLVMAppendBasicBlock(owning_block, "else");
LLVMBasicBlockRef merge_ref = LLVMAppendBasicBlock(owning_block, "ifmerge");
// 3. Branch conditionally on then or else.
LLVMBuildCondBr(builder, cond, then_ref, else_ref);
// 4. Build then branch prologue.
LLVMPositionBuilderAtEnd(builder, then_ref);
LLVMValueRef hi1 = LLVMBuildXor(builder, LLVMGetUndef(LLVMInt32Type()), LLVMGetUndef(LLVMInt32Type()), "subtmp");
// 5. Connect then branch to merge block.
LLVMBuildBr(builder, merge_ref);
then_ref = LLVMGetInsertBlock(builder);
// 6. Build else branch prologue.
LLVMPositionBuilderAtEnd(builder, else_ref);
LLVMValueRef hi2 = LLVMBuildXor(builder, LLVMGetUndef(LLVMInt32Type()), LLVMGetUndef(LLVMInt32Type()), "subtmp2");
// 7. Connect else branch to merge block.
LLVMBuildBr(builder, merge_ref);
else_ref = LLVMGetInsertBlock(builder);
// 8. Position ourselves after the merge block.
LLVMPositionBuilderAtEnd(builder, merge_ref);
// 9. Build the phi node.
// LLVMValueRef phi = LLVMBuildPhi(builder, LLVMDoubleType(), "phi");
// 10. Add incoming edges.
// LLVMAddIncoming(phi, &hi1, &then_ref, 1);
// LLVMAddIncoming(phi, &hi2, &else_ref, 1);
LLVMDumpModule(module);
LLVMDisposeBuilder(builder);
LLVMDisposeModule(module);
return 0;
}
static LLVMValueRef gen_digestof_value(compile_t* c, LLVMValueRef value)
{
LLVMTypeRef type = LLVMTypeOf(value);
switch(LLVMGetTypeKind(type))
{
case LLVMFloatTypeKind:
value = LLVMBuildBitCast(c->builder, value, c->i32, "");
return LLVMBuildZExt(c->builder, value, c->i64, "");
case LLVMDoubleTypeKind:
return LLVMBuildBitCast(c->builder, value, c->i64, "");
case LLVMIntegerTypeKind:
{
uint32_t width = LLVMGetIntTypeWidth(type);
if(width < 64)
{
value = LLVMBuildZExt(c->builder, value, c->i64, "");
} else if(width == 128) {
LLVMValueRef shift = LLVMConstInt(c->i128, 64, false);
LLVMValueRef high = LLVMBuildLShr(c->builder, value, shift, "");
high = LLVMBuildTrunc(c->builder, high, c->i64, "");
value = LLVMBuildTrunc(c->builder, value, c->i64, "");
value = LLVMBuildXor(c->builder, value, high, "");
}
return value;
}
case LLVMStructTypeKind:
{
uint32_t count = LLVMCountStructElementTypes(type);
LLVMValueRef result = LLVMConstInt(c->i64, 0, false);
for(uint32_t i = 0; i < count; i++)
{
LLVMValueRef elem = LLVMBuildExtractValue(c->builder, value, i, "");
elem = gen_digestof_value(c, elem);
result = LLVMBuildXor(c->builder, result, elem, "");
}
return result;
}
case LLVMPointerTypeKind:
return LLVMBuildPtrToInt(c->builder, value, c->i64, "");
default: {}
}
assert(0);
return NULL;
}
LLVMValueRef gen_xor(compile_t* c, ast_t* left, ast_t* right)
{
LLVMValueRef l_value = gen_expr(c, left);
LLVMValueRef r_value = gen_expr(c, right);
if((l_value == NULL) || (r_value == NULL))
return NULL;
if(LLVMIsConstant(l_value) && LLVMIsConstant(r_value))
return LLVMConstXor(l_value, r_value);
if(is_always_true(c, l_value))
return LLVMBuildNot(c->builder, r_value, "");
if(is_always_false(c, l_value))
return r_value;
if(is_always_true(c, r_value))
return LLVMBuildNot(c->builder, l_value, "");
if(is_always_false(c, r_value))
return l_value;
return LLVMBuildXor(c->builder, l_value, r_value, "");
}
static void emit_xor(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] = LLVMBuildXor(builder,
emit_data->args[0], emit_data->args[1], "");
}
static void emit_bfi(const struct lp_build_tgsi_action *action,
struct lp_build_tgsi_context *bld_base,
struct lp_build_emit_data *emit_data)
{
struct gallivm_state *gallivm = bld_base->base.gallivm;
LLVMBuilderRef builder = gallivm->builder;
LLVMValueRef bfi_args[3];
LLVMValueRef bfi_sm5;
LLVMValueRef cond;
// Calculate the bitmask: (((1 << src3) - 1) << src2
bfi_args[0] = LLVMBuildShl(builder,
LLVMBuildSub(builder,
LLVMBuildShl(builder,
bld_base->int_bld.one,
emit_data->args[3], ""),
bld_base->int_bld.one, ""),
emit_data->args[2], "");
bfi_args[1] = LLVMBuildShl(builder, emit_data->args[1],
emit_data->args[2], "");
bfi_args[2] = emit_data->args[0];
/* Calculate:
* (arg0 & arg1) | (~arg0 & arg2) = arg2 ^ (arg0 & (arg1 ^ arg2)
* Use the right-hand side, which the LLVM backend can convert to V_BFI.
*/
bfi_sm5 =
LLVMBuildXor(builder, bfi_args[2],
LLVMBuildAnd(builder, bfi_args[0],
LLVMBuildXor(builder, bfi_args[1], bfi_args[2],
""), ""), "");
/* Since shifts of >= 32 bits are undefined in LLVM IR, the backend
* uses the convenient V_BFI lowering for the above, which follows SM5
* and disagrees with GLSL semantics when bits (src3) is 32.
*/
cond = LLVMBuildICmp(builder, LLVMIntUGE, emit_data->args[3],
lp_build_const_int32(gallivm, 32), "");
emit_data->output[emit_data->chan] =
LLVMBuildSelect(builder, cond, emit_data->args[1], bfi_sm5, "");
}
static LLVMValueRef gen_digestof_int64(compile_t* c, LLVMValueRef value)
{
pony_assert(LLVMTypeOf(value) == c->i64);
if(target_is_ilp32(c->opt->triple))
{
LLVMValueRef shift = LLVMConstInt(c->i64, 32, false);
LLVMValueRef high = LLVMBuildLShr(c->builder, value, shift, "");
high = LLVMBuildTrunc(c->builder, high, c->i32, "");
value = LLVMBuildTrunc(c->builder, value, c->i32, "");
value = LLVMBuildXor(c->builder, value, high, "");
}
return value;
}
/**
* Converts int16 half-float to float32
* Note this can be performed in 1 instruction if vcvtph2ps exists (sse5 i think?)
* [llvm.x86.vcvtph2ps / _mm_cvtph_ps]
*
* @param src_type <vector> type of int16
* @param src value to convert
*
* ref http://fgiesen.wordpress.com/2012/03/28/half-to-float-done-quic/
*/
LLVMValueRef
lp_build_half_to_float(struct gallivm_state *gallivm,
struct lp_type src_type,
LLVMValueRef src)
{
struct lp_type f32_type = lp_type_float_vec(32, 32 * src_type.length);
struct lp_type i32_type = lp_type_int_vec(32, 32 * src_type.length);
LLVMBuilderRef builder = gallivm->builder;
LLVMTypeRef int_vec_type = lp_build_vec_type(gallivm, i32_type);
LLVMTypeRef float_vec_type = lp_build_vec_type(gallivm, f32_type);
/* Constants */
LLVMValueRef i32_13 = lp_build_const_int_vec(gallivm, i32_type, 13);
LLVMValueRef i32_16 = lp_build_const_int_vec(gallivm, i32_type, 16);
LLVMValueRef i32_mask_nosign = lp_build_const_int_vec(gallivm, i32_type, 0x7fff);
LLVMValueRef i32_was_infnan = lp_build_const_int_vec(gallivm, i32_type, 0x7bff);
LLVMValueRef i32_exp_infnan = lp_build_const_int_vec(gallivm, i32_type, 0xff << 23);
LLVMValueRef f32_magic = LLVMBuildBitCast(builder,
lp_build_const_int_vec(gallivm, i32_type, (254 - 15) << 23),
float_vec_type, "");
/* Convert int16 vector to int32 vector by zero ext */
LLVMValueRef h = LLVMBuildZExt(builder, src, int_vec_type, "");
/* Exponent / mantissa bits */
LLVMValueRef expmant = LLVMBuildAnd(builder, i32_mask_nosign, h, "");
LLVMValueRef shifted = LLVMBuildBitCast(builder, LLVMBuildShl(builder, expmant, i32_13, ""), float_vec_type, "");
/* Exponent adjust */
LLVMValueRef scaled = LLVMBuildBitCast(builder, LLVMBuildFMul(builder, shifted, f32_magic, ""), int_vec_type, "");
/* Make sure Inf/NaN survive */
LLVMValueRef b_wasinfnan = lp_build_compare(gallivm, i32_type, PIPE_FUNC_GREATER, expmant, i32_was_infnan);
LLVMValueRef infnanexp = LLVMBuildAnd(builder, b_wasinfnan, i32_exp_infnan, "");
/* Sign bit */
LLVMValueRef justsign = LLVMBuildXor(builder, h, expmant, "");
LLVMValueRef sign = LLVMBuildShl(builder, justsign, i32_16, "");
/* Combine result */
LLVMValueRef sign_inf = LLVMBuildOr(builder, sign, infnanexp, "");
LLVMValueRef final = LLVMBuildOr(builder, scaled, sign_inf, "");
/* Cast from int32 vector to float32 vector */
return LLVMBuildBitCast(builder, final, float_vec_type, "");
}
static void ac_analyze_position_w(struct ac_llvm_context *ctx,
LLVMValueRef pos[3][4],
struct ac_position_w_info *w)
{
LLVMBuilderRef builder = ctx->builder;
LLVMValueRef all_w_negative = ctx->i1true;
w->w_reflection = ctx->i1false;
w->any_w_negative = ctx->i1false;
for (unsigned i = 0; i < 3; i++) {
LLVMValueRef neg_w;
neg_w = LLVMBuildFCmp(builder, LLVMRealOLT, pos[i][3], ctx->f32_0, "");
/* If neg_w is true, negate w_reflection. */
w->w_reflection = LLVMBuildXor(builder, w->w_reflection, neg_w, "");
w->any_w_negative = LLVMBuildOr(builder, w->any_w_negative, neg_w, "");
all_w_negative = LLVMBuildAnd(builder, all_w_negative, neg_w, "");
}
w->all_w_positive = LLVMBuildNot(builder, w->any_w_negative, "");
w->w_accepted = LLVMBuildNot(builder, all_w_negative, "");
}
static LLVMValueRef gen_digestof_value(compile_t* c, ast_t* type,
LLVMValueRef value)
{
LLVMTypeRef impl_type = LLVMTypeOf(value);
switch(LLVMGetTypeKind(impl_type))
{
case LLVMFloatTypeKind:
value = LLVMBuildBitCast(c->builder, value, c->i32, "");
return LLVMBuildZExt(c->builder, value, c->intptr, "");
case LLVMDoubleTypeKind:
value = LLVMBuildBitCast(c->builder, value, c->i64, "");
return gen_digestof_int64(c, value);
case LLVMIntegerTypeKind:
{
uint32_t width = LLVMGetIntTypeWidth(impl_type);
if(width < 64)
{
return LLVMBuildZExt(c->builder, value, c->intptr, "");
} else if(width == 64) {
return gen_digestof_int64(c, value);
} else if(width == 128) {
LLVMValueRef shift = LLVMConstInt(c->i128, 64, false);
LLVMValueRef high = LLVMBuildLShr(c->builder, value, shift, "");
high = LLVMBuildTrunc(c->builder, high, c->i64, "");
value = LLVMBuildTrunc(c->builder, value, c->i64, "");
high = gen_digestof_int64(c, high);
value = gen_digestof_int64(c, value);
return LLVMBuildXor(c->builder, value, high, "");
}
break;
}
case LLVMStructTypeKind:
{
uint32_t count = LLVMCountStructElementTypes(impl_type);
LLVMValueRef result = LLVMConstInt(c->intptr, 0, false);
ast_t* child = ast_child(type);
for(uint32_t i = 0; i < count; i++)
{
LLVMValueRef elem = LLVMBuildExtractValue(c->builder, value, i, "");
elem = gen_digestof_value(c, child, elem);
result = LLVMBuildXor(c->builder, result, elem, "");
child = ast_sibling(child);
}
pony_assert(child == NULL);
return result;
}
case LLVMPointerTypeKind:
if(!is_known(type))
{
reach_type_t* t = reach_type(c->reach, type);
int sub_kind = subtype_kind(t);
if((sub_kind & SUBTYPE_KIND_BOXED) != 0)
return gen_digestof_box(c, t, value, sub_kind);
}
return LLVMBuildPtrToInt(c->builder, value, c->intptr, "");
default: {}
}
pony_assert(0);
return NULL;
}
/*
* gen_operator_expression
*
* Code generation for operator expressions. Most of them have straightforward
* translations into LLVM instructions and are handled directly here.
*/
static LLVMValueRef
gen_operator_expression (gencodectx_t gctx, expr_node_t *exp, LLVMTypeRef neededtype)
{
expr_node_t *lhs = expr_op_lhs(exp);
expr_node_t *rhs = expr_op_rhs(exp);
optype_t op = expr_op_type(exp);
LLVMBuilderRef builder = gctx->curfn->builder;
LLVMTypeRef inttype;
LLVMValueRef lval, rval, result;
if (op == OPER_FETCH) {
return gen_fetch(gctx, rhs, neededtype);
}
if (op == OPER_ASSIGN) {
LLVMValueRef val = llvmgen_assignment(gctx, lhs, rhs);
return llvmgen_adjustval(gctx, val, neededtype, 0);
}
if (op == OPER_SHIFT) {
return gen_shift(gctx, lhs, rhs, neededtype);
}
inttype = LLVMIntTypeInContext(gctx->llvmctx, machine_scalar_bits(gctx->mach));
lval = (lhs == 0 ? 0 : llvmgen_expression(gctx, lhs, inttype));
rval = llvmgen_expression(gctx, rhs, inttype);
switch (op) {
case OPER_UNARY_PLUS:
result = rval;
break;
case OPER_UNARY_MINUS:
result = LLVMBuildNeg(builder, rval, llvmgen_temp(gctx));
break;
case OPER_ADD:
result = LLVMBuildAdd(builder, lval, rval, llvmgen_temp(gctx));
break;
case OPER_SUBTRACT:
result = LLVMBuildSub(builder, lval, rval, llvmgen_temp(gctx));
break;
case OPER_MULT:
result = LLVMBuildMul(builder, lval, rval, llvmgen_temp(gctx));
break;
case OPER_DIV:
result = LLVMBuildUDiv(builder, lval, rval, llvmgen_temp(gctx));
break;
case OPER_MODULO:
result = LLVMBuildURem(builder, lval, rval, llvmgen_temp(gctx));
break;
case OPER_AND:
result = LLVMBuildAnd(builder, lval, rval, llvmgen_temp(gctx));
break;
case OPER_OR:
result = LLVMBuildOr(builder, lval, rval, llvmgen_temp(gctx));
break;
case OPER_NOT:
result = LLVMBuildNot(builder, rval, llvmgen_temp(gctx));
break;
case OPER_XOR:
result = LLVMBuildXor(builder, lval, rval, llvmgen_temp(gctx));
break;
case OPER_EQV:
result = LLVMBuildXor(builder, lval, rval, llvmgen_temp(gctx));
result = LLVMBuildNot(builder, result, llvmgen_temp(gctx));
break;
default:
if (op >= OPER_CMP_EQL && op <= OPER_CMP_GEQA) {
result = LLVMBuildICmp(builder,
llvmgen_predfromop(op, machine_addr_signed(gctx->mach)),
lval, rval, llvmgen_temp(gctx));
} else {
// Everything should be covered
expr_signal(gctx->ectx, STC__INTCMPERR, "gen_operator_expression");
result = LLVMConstNull(inttype);
}
break;
}
return llvmgen_adjustval(gctx, result, neededtype, 0);
} /* gen_operator_expression */
/**
* Build code to compare two values 'a' and 'b' of 'type' using the given func.
* \param func one of PIPE_FUNC_x
* The result values will be 0 for false or ~0 for true.
*/
LLVMValueRef
lp_build_compare(struct gallivm_state *gallivm,
const struct lp_type type,
unsigned func,
LLVMValueRef a,
LLVMValueRef b)
{
LLVMBuilderRef builder = gallivm->builder;
LLVMTypeRef int_vec_type = lp_build_int_vec_type(gallivm, type);
LLVMValueRef zeros = LLVMConstNull(int_vec_type);
LLVMValueRef ones = LLVMConstAllOnes(int_vec_type);
LLVMValueRef cond;
LLVMValueRef res;
assert(func >= PIPE_FUNC_NEVER);
assert(func <= PIPE_FUNC_ALWAYS);
assert(lp_check_value(type, a));
assert(lp_check_value(type, b));
if(func == PIPE_FUNC_NEVER)
return zeros;
if(func == PIPE_FUNC_ALWAYS)
return ones;
#if defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64)
/*
* There are no unsigned integer comparison instructions in SSE.
*/
if (!type.floating && !type.sign &&
type.width * type.length == 128 &&
util_cpu_caps.has_sse2 &&
(func == PIPE_FUNC_LESS ||
func == PIPE_FUNC_LEQUAL ||
func == PIPE_FUNC_GREATER ||
func == PIPE_FUNC_GEQUAL) &&
(gallivm_debug & GALLIVM_DEBUG_PERF)) {
debug_printf("%s: inefficient <%u x i%u> unsigned comparison\n",
__FUNCTION__, type.length, type.width);
}
#endif
#if HAVE_LLVM < 0x0207
#if defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64)
if(type.width * type.length == 128) {
if(type.floating && util_cpu_caps.has_sse) {
/* float[4] comparison */
LLVMTypeRef vec_type = lp_build_vec_type(gallivm, type);
LLVMValueRef args[3];
unsigned cc;
boolean swap;
swap = FALSE;
switch(func) {
case PIPE_FUNC_EQUAL:
cc = 0;
break;
case PIPE_FUNC_NOTEQUAL:
cc = 4;
break;
case PIPE_FUNC_LESS:
cc = 1;
break;
case PIPE_FUNC_LEQUAL:
cc = 2;
break;
case PIPE_FUNC_GREATER:
cc = 1;
swap = TRUE;
break;
case PIPE_FUNC_GEQUAL:
cc = 2;
swap = TRUE;
break;
default:
assert(0);
return lp_build_undef(gallivm, type);
}
if(swap) {
args[0] = b;
args[1] = a;
}
else {
args[0] = a;
args[1] = b;
}
args[2] = LLVMConstInt(LLVMInt8TypeInContext(gallivm->context), cc, 0);
res = lp_build_intrinsic(builder,
"llvm.x86.sse.cmp.ps",
vec_type,
args, 3);
res = LLVMBuildBitCast(builder, res, int_vec_type, "");
return res;
}
else if(util_cpu_caps.has_sse2) {
/* int[4] comparison */
static const struct {
unsigned swap:1;
unsigned eq:1;
unsigned gt:1;
unsigned not:1;
} table[] = {
{0, 0, 0, 1}, /* PIPE_FUNC_NEVER */
{1, 0, 1, 0}, /* PIPE_FUNC_LESS */
{0, 1, 0, 0}, /* PIPE_FUNC_EQUAL */
{0, 0, 1, 1}, /* PIPE_FUNC_LEQUAL */
{0, 0, 1, 0}, /* PIPE_FUNC_GREATER */
{0, 1, 0, 1}, /* PIPE_FUNC_NOTEQUAL */
{1, 0, 1, 1}, /* PIPE_FUNC_GEQUAL */
{0, 0, 0, 0} /* PIPE_FUNC_ALWAYS */
};
const char *pcmpeq;
const char *pcmpgt;
LLVMValueRef args[2];
LLVMValueRef res;
LLVMTypeRef vec_type = lp_build_vec_type(gallivm, type);
switch (type.width) {
case 8:
pcmpeq = "llvm.x86.sse2.pcmpeq.b";
pcmpgt = "llvm.x86.sse2.pcmpgt.b";
break;
case 16:
pcmpeq = "llvm.x86.sse2.pcmpeq.w";
pcmpgt = "llvm.x86.sse2.pcmpgt.w";
break;
case 32:
pcmpeq = "llvm.x86.sse2.pcmpeq.d";
pcmpgt = "llvm.x86.sse2.pcmpgt.d";
break;
default:
assert(0);
return lp_build_undef(gallivm, type);
}
/* There are no unsigned comparison instructions. So flip the sign bit
* so that the results match.
*/
if (table[func].gt && !type.sign) {
LLVMValueRef msb = lp_build_const_int_vec(gallivm, type, (unsigned long long)1 << (type.width - 1));
a = LLVMBuildXor(builder, a, msb, "");
b = LLVMBuildXor(builder, b, msb, "");
}
if(table[func].swap) {
args[0] = b;
args[1] = a;
}
else {
args[0] = a;
args[1] = b;
}
if(table[func].eq)
res = lp_build_intrinsic(builder, pcmpeq, vec_type, args, 2);
else if (table[func].gt)
res = lp_build_intrinsic(builder, pcmpgt, vec_type, args, 2);
else
res = LLVMConstNull(vec_type);
if(table[func].not)
res = LLVMBuildNot(builder, res, "");
return res;
}
} /* if (type.width * type.length == 128) */
#endif
#endif /* HAVE_LLVM < 0x0207 */
/* XXX: It is not clear if we should use the ordered or unordered operators */
if(type.floating) {
LLVMRealPredicate op;
switch(func) {
case PIPE_FUNC_NEVER:
op = LLVMRealPredicateFalse;
break;
case PIPE_FUNC_ALWAYS:
op = LLVMRealPredicateTrue;
break;
case PIPE_FUNC_EQUAL:
op = LLVMRealUEQ;
break;
case PIPE_FUNC_NOTEQUAL:
op = LLVMRealUNE;
break;
case PIPE_FUNC_LESS:
op = LLVMRealULT;
break;
case PIPE_FUNC_LEQUAL:
op = LLVMRealULE;
break;
case PIPE_FUNC_GREATER:
op = LLVMRealUGT;
break;
case PIPE_FUNC_GEQUAL:
op = LLVMRealUGE;
break;
default:
assert(0);
return lp_build_undef(gallivm, type);
}
#if HAVE_LLVM >= 0x0207
cond = LLVMBuildFCmp(builder, op, a, b, "");
res = LLVMBuildSExt(builder, cond, int_vec_type, "");
#else
if (type.length == 1) {
cond = LLVMBuildFCmp(builder, op, a, b, "");
res = LLVMBuildSExt(builder, cond, int_vec_type, "");
}
else {
unsigned i;
res = LLVMGetUndef(int_vec_type);
debug_printf("%s: warning: using slow element-wise float"
" vector comparison\n", __FUNCTION__);
for (i = 0; i < type.length; ++i) {
LLVMValueRef index = lp_build_const_int32(gallivm, i);
cond = LLVMBuildFCmp(builder, op,
LLVMBuildExtractElement(builder, a, index, ""),
LLVMBuildExtractElement(builder, b, index, ""),
"");
cond = LLVMBuildSelect(builder, cond,
LLVMConstExtractElement(ones, index),
LLVMConstExtractElement(zeros, index),
"");
res = LLVMBuildInsertElement(builder, res, cond, index, "");
}
}
#endif
}
else {
LLVMIntPredicate op;
switch(func) {
case PIPE_FUNC_EQUAL:
op = LLVMIntEQ;
break;
case PIPE_FUNC_NOTEQUAL:
op = LLVMIntNE;
break;
case PIPE_FUNC_LESS:
op = type.sign ? LLVMIntSLT : LLVMIntULT;
break;
case PIPE_FUNC_LEQUAL:
op = type.sign ? LLVMIntSLE : LLVMIntULE;
break;
case PIPE_FUNC_GREATER:
op = type.sign ? LLVMIntSGT : LLVMIntUGT;
break;
case PIPE_FUNC_GEQUAL:
op = type.sign ? LLVMIntSGE : LLVMIntUGE;
break;
default:
assert(0);
return lp_build_undef(gallivm, type);
}
#if HAVE_LLVM >= 0x0207
cond = LLVMBuildICmp(builder, op, a, b, "");
res = LLVMBuildSExt(builder, cond, int_vec_type, "");
#else
if (type.length == 1) {
cond = LLVMBuildICmp(builder, op, a, b, "");
res = LLVMBuildSExt(builder, cond, int_vec_type, "");
}
else {
unsigned i;
res = LLVMGetUndef(int_vec_type);
if (gallivm_debug & GALLIVM_DEBUG_PERF) {
debug_printf("%s: using slow element-wise int"
" vector comparison\n", __FUNCTION__);
}
for(i = 0; i < type.length; ++i) {
LLVMValueRef index = lp_build_const_int32(gallivm, i);
cond = LLVMBuildICmp(builder, op,
LLVMBuildExtractElement(builder, a, index, ""),
LLVMBuildExtractElement(builder, b, index, ""),
"");
cond = LLVMBuildSelect(builder, cond,
LLVMConstExtractElement(ones, index),
LLVMConstExtractElement(zeros, index),
"");
res = LLVMBuildInsertElement(builder, res, cond, index, "");
}
}
#endif
}
return res;
}
/*
* Do a cached lookup.
*
* Returns (vectors of) 4x8 rgba aos value
*/
LLVMValueRef
lp_build_fetch_cached_texels(struct gallivm_state *gallivm,
const struct util_format_description *format_desc,
unsigned n,
LLVMValueRef base_ptr,
LLVMValueRef offset,
LLVMValueRef i,
LLVMValueRef j,
LLVMValueRef cache)
{
LLVMBuilderRef builder = gallivm->builder;
unsigned count, low_bit, log2size;
LLVMValueRef color, offset_stored, addr, ptr_addrtrunc, tmp;
LLVMValueRef ij_index, hash_index, hash_mask, block_index;
LLVMTypeRef i8t = LLVMInt8TypeInContext(gallivm->context);
LLVMTypeRef i32t = LLVMInt32TypeInContext(gallivm->context);
LLVMTypeRef i64t = LLVMInt64TypeInContext(gallivm->context);
struct lp_type type;
struct lp_build_context bld32;
memset(&type, 0, sizeof type);
type.width = 32;
type.length = n;
assert(format_desc->block.width == 4);
assert(format_desc->block.height == 4);
lp_build_context_init(&bld32, gallivm, type);
/*
* compute hash - we use direct mapped cache, the hash function could
* be better but it needs to be simple
* per-element:
* compare offset with offset stored at tag (hash)
* if not equal decode/store block, update tag
* extract color from cache
* assemble result vector
*/
/* TODO: not ideal with 32bit pointers... */
low_bit = util_logbase2(format_desc->block.bits / 8);
log2size = util_logbase2(LP_BUILD_FORMAT_CACHE_SIZE);
addr = LLVMBuildPtrToInt(builder, base_ptr, i64t, "");
ptr_addrtrunc = LLVMBuildPtrToInt(builder, base_ptr, i32t, "");
ptr_addrtrunc = lp_build_broadcast_scalar(&bld32, ptr_addrtrunc);
/* For the hash function, first mask off the unused lowest bits. Then just
do some xor with address bits - only use lower 32bits */
ptr_addrtrunc = LLVMBuildAdd(builder, offset, ptr_addrtrunc, "");
ptr_addrtrunc = LLVMBuildLShr(builder, ptr_addrtrunc,
lp_build_const_int_vec(gallivm, type, low_bit), "");
/* This only really makes sense for size 64,128,256 */
hash_index = ptr_addrtrunc;
ptr_addrtrunc = LLVMBuildLShr(builder, ptr_addrtrunc,
lp_build_const_int_vec(gallivm, type, 2*log2size), "");
hash_index = LLVMBuildXor(builder, ptr_addrtrunc, hash_index, "");
tmp = LLVMBuildLShr(builder, hash_index,
lp_build_const_int_vec(gallivm, type, log2size), "");
hash_index = LLVMBuildXor(builder, hash_index, tmp, "");
hash_mask = lp_build_const_int_vec(gallivm, type, LP_BUILD_FORMAT_CACHE_SIZE - 1);
hash_index = LLVMBuildAnd(builder, hash_index, hash_mask, "");
ij_index = LLVMBuildShl(builder, i, lp_build_const_int_vec(gallivm, type, 2), "");
ij_index = LLVMBuildAdd(builder, ij_index, j, "");
block_index = LLVMBuildShl(builder, hash_index,
lp_build_const_int_vec(gallivm, type, 4), "");
block_index = LLVMBuildAdd(builder, ij_index, block_index, "");
if (n > 1) {
color = LLVMGetUndef(LLVMVectorType(i32t, n));
for (count = 0; count < n; count++) {
LLVMValueRef index, cond, colorx;
LLVMValueRef block_indexx, hash_indexx, addrx, offsetx, ptr_addrx;
struct lp_build_if_state if_ctx;
index = lp_build_const_int32(gallivm, count);
offsetx = LLVMBuildExtractElement(builder, offset, index, "");
addrx = LLVMBuildZExt(builder, offsetx, i64t, "");
addrx = LLVMBuildAdd(builder, addrx, addr, "");
block_indexx = LLVMBuildExtractElement(builder, block_index, index, "");
hash_indexx = LLVMBuildLShr(builder, block_indexx,
lp_build_const_int32(gallivm, 4), "");
offset_stored = lookup_tag_data(gallivm, cache, hash_indexx);
cond = LLVMBuildICmp(builder, LLVMIntNE, offset_stored, addrx, "");
lp_build_if(&if_ctx, gallivm, cond);
{
ptr_addrx = LLVMBuildIntToPtr(builder, addrx,
LLVMPointerType(i8t, 0), "");
update_cached_block(gallivm, format_desc, ptr_addrx, hash_indexx, cache);
#if LP_BUILD_FORMAT_CACHE_DEBUG
update_cache_access(gallivm, cache, 1,
LP_BUILD_FORMAT_CACHE_MEMBER_ACCESS_MISS);
#endif
}
lp_build_endif(&if_ctx);
colorx = lookup_cached_pixel(gallivm, cache, block_indexx);
color = LLVMBuildInsertElement(builder, color, colorx,
lp_build_const_int32(gallivm, count), "");
}
}
else {
LLVMValueRef cond;
struct lp_build_if_state if_ctx;
tmp = LLVMBuildZExt(builder, offset, i64t, "");
addr = LLVMBuildAdd(builder, tmp, addr, "");
offset_stored = lookup_tag_data(gallivm, cache, hash_index);
cond = LLVMBuildICmp(builder, LLVMIntNE, offset_stored, addr, "");
lp_build_if(&if_ctx, gallivm, cond);
{
tmp = LLVMBuildIntToPtr(builder, addr, LLVMPointerType(i8t, 0), "");
update_cached_block(gallivm, format_desc, tmp, hash_index, cache);
#if LP_BUILD_FORMAT_CACHE_DEBUG
update_cache_access(gallivm, cache, 1,
LP_BUILD_FORMAT_CACHE_MEMBER_ACCESS_MISS);
#endif
}
lp_build_endif(&if_ctx);
color = lookup_cached_pixel(gallivm, cache, block_index);
}
#if LP_BUILD_FORMAT_CACHE_DEBUG
update_cache_access(gallivm, cache, n,
LP_BUILD_FORMAT_CACHE_MEMBER_ACCESS_TOTAL);
#endif
return LLVMBuildBitCast(builder, color, LLVMVectorType(i8t, n * 4), "");
}
LLVMValueRef
lp_build_logicop(LLVMBuilderRef builder,
unsigned logicop_func,
LLVMValueRef src,
LLVMValueRef dst)
{
LLVMTypeRef type;
LLVMValueRef res;
type = LLVMTypeOf(src);
switch (logicop_func) {
case PIPE_LOGICOP_CLEAR:
res = LLVMConstNull(type);
break;
case PIPE_LOGICOP_NOR:
res = LLVMBuildNot(builder, LLVMBuildOr(builder, src, dst, ""), "");
break;
case PIPE_LOGICOP_AND_INVERTED:
res = LLVMBuildAnd(builder, LLVMBuildNot(builder, src, ""), dst, "");
break;
case PIPE_LOGICOP_COPY_INVERTED:
res = LLVMBuildNot(builder, src, "");
break;
case PIPE_LOGICOP_AND_REVERSE:
res = LLVMBuildAnd(builder, src, LLVMBuildNot(builder, dst, ""), "");
break;
case PIPE_LOGICOP_INVERT:
res = LLVMBuildNot(builder, dst, "");
break;
case PIPE_LOGICOP_XOR:
res = LLVMBuildXor(builder, src, dst, "");
break;
case PIPE_LOGICOP_NAND:
res = LLVMBuildNot(builder, LLVMBuildAnd(builder, src, dst, ""), "");
break;
case PIPE_LOGICOP_AND:
res = LLVMBuildAnd(builder, src, dst, "");
break;
case PIPE_LOGICOP_EQUIV:
res = LLVMBuildNot(builder, LLVMBuildXor(builder, src, dst, ""), "");
break;
case PIPE_LOGICOP_NOOP:
res = dst;
break;
case PIPE_LOGICOP_OR_INVERTED:
res = LLVMBuildOr(builder, LLVMBuildNot(builder, src, ""), dst, "");
break;
case PIPE_LOGICOP_COPY:
res = src;
break;
case PIPE_LOGICOP_OR_REVERSE:
res = LLVMBuildOr(builder, src, LLVMBuildNot(builder, dst, ""), "");
break;
case PIPE_LOGICOP_OR:
res = LLVMBuildOr(builder, src, dst, "");
break;
case PIPE_LOGICOP_SET:
res = LLVMConstAllOnes(type);
break;
default:
assert(0);
res = src;
}
return res;
}
