unsigned
ac_count_scratch_private_memory(LLVMValueRef function)
{
unsigned private_mem_vgprs = 0;
/* Process all LLVM instructions. */
LLVMBasicBlockRef bb = LLVMGetFirstBasicBlock(function);
while (bb) {
LLVMValueRef next = LLVMGetFirstInstruction(bb);
while (next) {
LLVMValueRef inst = next;
next = LLVMGetNextInstruction(next);
if (LLVMGetInstructionOpcode(inst) != LLVMAlloca)
continue;
LLVMTypeRef type = LLVMGetElementType(LLVMTypeOf(inst));
/* No idea why LLVM aligns allocas to 4 elements. */
unsigned alignment = LLVMGetAlignment(inst);
unsigned dw_size = align(ac_get_type_size(type) / 4, alignment);
private_mem_vgprs += dw_size;
}
bb = LLVMGetNextBasicBlock(bb);
}
return private_mem_vgprs;
}
void ac_optimize_vs_outputs(struct ac_llvm_context *ctx,
LLVMValueRef main_fn,
uint8_t *vs_output_param_offset,
uint32_t num_outputs,
uint8_t *num_param_exports)
{
LLVMBasicBlockRef bb;
bool removed_any = false;
struct ac_vs_exports exports;
exports.num = 0;
/* Process all LLVM instructions. */
bb = LLVMGetFirstBasicBlock(main_fn);
while (bb) {
LLVMValueRef inst = LLVMGetFirstInstruction(bb);
while (inst) {
LLVMValueRef cur = inst;
inst = LLVMGetNextInstruction(inst);
struct ac_vs_exp_inst exp;
if (LLVMGetInstructionOpcode(cur) != LLVMCall)
continue;
LLVMValueRef callee = ac_llvm_get_called_value(cur);
if (!ac_llvm_is_function(callee))
continue;
const char *name = LLVMGetValueName(callee);
unsigned num_args = LLVMCountParams(callee);
/* Check if this is an export instruction. */
if ((num_args != 9 && num_args != 8) ||
(strcmp(name, "llvm.SI.export") &&
strcmp(name, "llvm.amdgcn.exp.f32")))
continue;
LLVMValueRef arg = LLVMGetOperand(cur, AC_EXP_TARGET);
unsigned target = LLVMConstIntGetZExtValue(arg);
if (target < V_008DFC_SQ_EXP_PARAM)
continue;
target -= V_008DFC_SQ_EXP_PARAM;
/* Parse the instruction. */
memset(&exp, 0, sizeof(exp));
exp.offset = target;
exp.inst = cur;
for (unsigned i = 0; i < 4; i++) {
LLVMValueRef v = LLVMGetOperand(cur, AC_EXP_OUT0 + i);
exp.chan[i].value = v;
if (LLVMIsUndef(v)) {
exp.chan[i].type = AC_IR_UNDEF;
} else if (LLVMIsAConstantFP(v)) {
LLVMBool loses_info;
exp.chan[i].type = AC_IR_CONST;
exp.chan[i].const_float =
LLVMConstRealGetDouble(v, &loses_info);
} else {
exp.chan[i].type = AC_IR_VALUE;
}
}
/* Eliminate constant and duplicated PARAM exports. */
if (ac_eliminate_const_output(vs_output_param_offset,
num_outputs, &exp) ||
ac_eliminate_duplicated_output(vs_output_param_offset,
num_outputs, &exports,
&exp)) {
removed_any = true;
} else {
exports.exp[exports.num++] = exp;
}
}
bb = LLVMGetNextBasicBlock(bb);
}
/* Remove holes in export memory due to removed PARAM exports.
* This is done by renumbering all PARAM exports.
*/
if (removed_any) {
uint8_t old_offset[VARYING_SLOT_MAX];
unsigned out, i;
/* Make a copy of the offsets. We need the old version while
* we are modifying some of them. */
memcpy(old_offset, vs_output_param_offset,
sizeof(old_offset));
for (i = 0; i < exports.num; i++) {
unsigned offset = exports.exp[i].offset;
/* Update vs_output_param_offset. Multiple outputs can
* have the same offset.
*/
for (out = 0; out < num_outputs; out++) {
if (old_offset[out] == offset)
vs_output_param_offset[out] = i;
}
/* Change the PARAM offset in the instruction. */
LLVMSetOperand(exports.exp[i].inst, AC_EXP_TARGET,
LLVMConstInt(ctx->i32,
V_008DFC_SQ_EXP_PARAM + i, 0));
}
*num_param_exports = exports.num;
}
}