import Mesa 11.0.6

1 files changed, 60 insertions, 52 deletions

diff --git a/lib/mesa/src/gallium/drivers/vc4/vc4_opt_dead_code.c b/lib/mesa/src/gallium/drivers/vc4/vc4_opt_dead_code.c
index f04d0ff97..ffd42422d 100644
--- a/lib/mesa/src/gallium/drivers/vc4/vc4_opt_dead_code.c
+++ b/lib/mesa/src/gallium/drivers/vc4/vc4_opt_dead_code.c
@@ -24,15 +24,11 @@
 /**
  * @file vc4_opt_dead_code.c
  *
- * This is a simple dead code eliminator for SSA values in QIR.
+ * This is a simmple dead code eliminator for QIR with no control flow.
  *
- * 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.
+ * 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.
  */
 
 #include "vc4_qir.h"
@@ -54,7 +50,7 @@ dce(struct vc4_compile *c, struct qinst *inst)
 static bool
 has_nonremovable_reads(struct vc4_compile *c, struct qinst *inst)
 {
-        for (int i = 0; i < qir_get_nsrc(inst); i++) {
+        for (int i = 0; i < qir_get_op_nsrc(inst->op); 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;
@@ -73,7 +69,7 @@ has_nonremovable_reads(struct vc4_compile *c, struct qinst *inst)
                 }
 
                 if (inst->src[i].file == QFILE_VARY &&
-                    c->input_slots[inst->src[i].index].slot == 0xff) {
+                    c->input_semantics[inst->src[i].index].semantic == 0xff) {
                         return true;
                 }
         }
@@ -86,50 +82,28 @@ qir_opt_dead_code(struct vc4_compile *c)
 {
         bool progress = false;
         bool *used = calloc(c->num_temps, sizeof(bool));
-
-        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;
-                }
-        }
-
-        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;
+        bool sf_used = false;
+        /* Whether we're eliminating texture setup currently. */
+        bool dce_tex = false;
+
+        struct list_head *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] &&
+                    !inst->sf &&
+                    (!qir_has_side_effects(c, inst) ||
+                     inst->op == QOP_TEX_RESULT) &&
+                    !has_nonremovable_reads(c, inst)) {
+                        if (inst->op == QOP_TEX_RESULT) {
+                                dce_tex = true;
+                                c->num_texture_samples--;
                         }
 
-                        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++) {
+                        for (int i = 0; i < qir_get_op_nsrc(inst->op); i++) {
                                 if (inst->src[i].file != QFILE_VPM)
                                         continue;
                                 uint32_t attr = inst->src[i].index / 4;
@@ -145,6 +119,40 @@ qir_opt_dead_code(struct vc4_compile *c)
                         progress = true;
                         continue;
                 }
+
+                if (qir_depends_on_flags(inst))
+                        sf_used = true;
+                if (inst->sf) {
+                        if (!sf_used) {
+                                if (debug) {
+                                        fprintf(stderr, "Removing SF on: ");
+                                        qir_dump_inst(c, inst);
+                                        fprintf(stderr, "\n");
+                                }
+
+                                inst->sf = false;
+                                progress = true;
+                        }
+                        sf_used = false;
+                }
+
+                if (inst->op == QOP_TEX_RESULT)
+                        dce_tex = false;
+
+                if (dce_tex && (inst->op == QOP_TEX_S ||
+                                inst->op == QOP_TEX_T ||
+                                inst->op == QOP_TEX_R ||
+                                inst->op == QOP_TEX_B ||
+                                inst->op == QOP_TEX_DIRECT)) {
+                        dce(c, inst);
+                        progress = true;
+                        continue;
+                }
+
+                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;
+                }
         }
 
         free(used);