/**
* @file vc4_opt_dead_code.c
*
- * This is a simmple dead code eliminator for QIR with no control flow.
+ * This is a simple dead code eliminator for SSA values in QIR.
*
- * It walks from the bottom of the instruction list, removing instructions
- * with a destination that is never used, and marking the sources of non-dead
- * instructions as used.
+ * It walks all the instructions finding what temps are used, then walks again
+ * to remove instructions writing unused temps.
+ *
+ * This is an inefficient implementation if you have long chains of
+ * instructions where the entire chain is dead, but we expect those to have
+ * been eliminated at the NIR level, and here we're just cleaning up small
+ * problems produced by NIR->QIR.
*/
#include "vc4_qir.h"
qir_dump_inst(c, inst);
fprintf(stderr, "\n");
}
- qir_remove_instruction(inst);
+ assert(!inst->sf);
+ qir_remove_instruction(c, inst);
+}
+
+static bool
+has_nonremovable_reads(struct vc4_compile *c, struct qinst *inst)
+{
+ for (int i = 0; i < qir_get_nsrc(inst); i++) {
+ if (inst->src[i].file == QFILE_VPM) {
+ uint32_t attr = inst->src[i].index / 4;
+ uint32_t offset = (inst->src[i].index % 4) * 4;
+
+ if (c->vattr_sizes[attr] != offset + 4)
+ return true;
+
+ /* Can't get rid of the last VPM read, or the
+ * simulator (at least) throws an error.
+ */
+ uint32_t total_size = 0;
+ for (uint32_t i = 0; i < ARRAY_SIZE(c->vattr_sizes); i++)
+ total_size += c->vattr_sizes[i];
+ if (total_size == 4)
+ return true;
+ }
+
+ if (inst->src[i].file == QFILE_VARY &&
+ c->input_slots[inst->src[i].index].slot == 0xff) {
+ return true;
+ }
+ }
+
+ return false;
}
bool
{
bool progress = false;
bool *used = calloc(c->num_temps, sizeof(bool));
- bool sf_used = false;
-
- struct simple_node *node, *t;
- for (node = c->instructions.prev, t = node->prev;
- &c->instructions != node;
- node = t, t = t->prev) {
- struct qinst *inst = (struct qinst *)node;
- if (inst->dst.file == QFILE_TEMP &&
- !used[inst->dst.index] &&
- !qir_has_side_effects(inst)) {
- dce(c, inst);
- progress = true;
- continue;
+ qir_for_each_inst_inorder(inst, c) {
+ for (int i = 0; i < qir_get_nsrc(inst); i++) {
+ if (inst->src[i].file == QFILE_TEMP)
+ used[inst->src[i].index] = true;
}
+ }
- if (qir_depends_on_flags(inst))
- sf_used = true;
- if (inst->op == QOP_SF) {
- if (!sf_used) {
- dce(c, inst);
- progress = true;
+ qir_for_each_block(block, c) {
+ qir_for_each_inst_safe(inst, block) {
+ if (inst->dst.file != QFILE_NULL &&
+ !(inst->dst.file == QFILE_TEMP &&
+ !used[inst->dst.index])) {
continue;
}
- sf_used = false;
- }
- for (int i = 0; i < qir_get_op_nsrc(inst->op); i++) {
- if (inst->src[i].file == QFILE_TEMP)
- used[inst->src[i].index] = true;
+ if (qir_has_side_effects(c, inst))
+ continue;
+
+ if (inst->sf ||
+ has_nonremovable_reads(c, inst)) {
+ /* If we can't remove the instruction, but we
+ * don't need its destination value, just
+ * remove the destination. The register
+ * allocator would trivially color it and it
+ * wouldn't cause any register pressure, but
+ * it's nicer to read the QIR code without
+ * unused destination regs.
+ */
+ if (inst->dst.file == QFILE_TEMP) {
+ if (debug) {
+ fprintf(stderr,
+ "Removing dst from: ");
+ qir_dump_inst(c, inst);
+ fprintf(stderr, "\n");
+ }
+ c->defs[inst->dst.index] = NULL;
+ inst->dst.file = QFILE_NULL;
+ progress = true;
+ }
+ continue;
+ }
+
+ for (int i = 0; i < qir_get_nsrc(inst); i++) {
+ if (inst->src[i].file != QFILE_VPM)
+ continue;
+ uint32_t attr = inst->src[i].index / 4;
+ uint32_t offset = (inst->src[i].index % 4) * 4;
+
+ if (c->vattr_sizes[attr] == offset + 4) {
+ c->num_inputs--;
+ c->vattr_sizes[attr] -= 4;
+ }
+ }
+
+ dce(c, inst);
+ progress = true;
+ continue;
}
}
projects / mesa.git / blobdiff