Import Mesa 13.0.2

5 files changed, 80 insertions, 69 deletions

diff --git a/lib/mesa/src/gallium/drivers/vc4/vc4_opt_peephole_sf.c b/lib/mesa/src/gallium/drivers/vc4/vc4_opt_peephole_sf.c
index 577290b1f..f4856673b 100644
--- a/lib/mesa/src/gallium/drivers/vc4/vc4_opt_peephole_sf.c
+++ b/lib/mesa/src/gallium/drivers/vc4/vc4_opt_peephole_sf.c
@@ -62,7 +62,7 @@ inst_srcs_updated(struct qinst *inst, struct qinst *writer)
          */
         switch (writer->dst.file) {
         case QFILE_TEMP:
-                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_TEMP &&
                             inst->src[i].index == writer->dst.index) {
                                 return true;
@@ -95,7 +95,7 @@ inst_result_equals(struct qinst *a, struct qinst *b)
                 return false;
         }
 
-        for (int i = 0; i < qir_get_nsrc(a); i++) {
+        for (int i = 0; i < qir_get_op_nsrc(a->op); i++) {
                 if (!qir_reg_equals(a->src[i], b->src[i]) ||
                     src_file_varies_on_reread(a->src[i]) ||
                     src_file_varies_on_reread(b->src[i])) {
diff --git a/lib/mesa/src/gallium/drivers/vc4/vc4_opt_vpm.c b/lib/mesa/src/gallium/drivers/vc4/vc4_opt_vpm.c
index 6f196e7d1..83ba11b81 100644
--- a/lib/mesa/src/gallium/drivers/vc4/vc4_opt_vpm.c
+++ b/lib/mesa/src/gallium/drivers/vc4/vc4_opt_vpm.c
@@ -24,8 +24,10 @@
 /**
  * @file vc4_opt_vpm.c
  *
- * This modifies instructions that exclusively consume a value read from the
- * VPM to directly read the VPM if other operands allow it.
+ * This modifies instructions that:
+ * 1. exclusively consume a value read from the VPM to directly read the VPM if
+ *    other operands allow it.
+ * 2. generate the value consumed by a VPM write to write directly into the VPM.
  */
 
 #include "vc4_qir.h"
@@ -42,11 +44,21 @@ qir_opt_vpm(struct vc4_compile *c)
                 return false;
 
         bool progress = false;
+        struct qinst *vpm_writes[64] = { 0 };
         uint32_t use_count[c->num_temps];
+        uint32_t vpm_write_count = 0;
         memset(&use_count, 0, sizeof(use_count));
 
         qir_for_each_inst_inorder(inst, c) {
-                for (int i = 0; i < qir_get_nsrc(inst); i++) {
+                switch (inst->dst.file) {
+                case QFILE_VPM:
+                        vpm_writes[vpm_write_count++] = inst;
+                        break;
+                default:
+                        break;
+                }
+
+                for (int i = 0; i < qir_get_op_nsrc(inst->op); i++) {
                         if (inst->src[i].file == QFILE_TEMP) {
                                 uint32_t temp = inst->src[i].index;
                                 use_count[temp]++;
@@ -69,7 +81,7 @@ qir_opt_vpm(struct vc4_compile *c)
                     qir_is_tex(inst))
                         continue;
 
-                for (int j = 0; j < qir_get_nsrc(inst); j++) {
+                for (int j = 0; j < qir_get_op_nsrc(inst->op); j++) {
                         if (inst->src[j].file != QFILE_TEMP ||
                             inst->src[j].pack)
                                 continue;
@@ -94,7 +106,7 @@ qir_opt_vpm(struct vc4_compile *c)
                         }
 
                         uint32_t temps = 0;
-                        for (int k = 0; k < qir_get_nsrc(inst); k++) {
+                        for (int k = 0; k < qir_get_op_nsrc(inst->op); k++) {
                                 if (inst->src[k].file == QFILE_TEMP)
                                         temps++;
                         }
@@ -115,5 +127,42 @@ qir_opt_vpm(struct vc4_compile *c)
                 }
         }
 
+        for (int i = 0; i < vpm_write_count; i++) {
+                if (!qir_is_raw_mov(vpm_writes[i]) ||
+                    vpm_writes[i]->src[0].file != QFILE_TEMP) {
+                        continue;
+                }
+
+                uint32_t temp = vpm_writes[i]->src[0].index;
+                if (use_count[temp] != 1)
+                        continue;
+
+                struct qinst *inst = c->defs[temp];
+                if (!inst)
+                        continue;
+
+                if (qir_depends_on_flags(inst) || inst->sf)
+                        continue;
+
+                if (qir_has_side_effects(c, inst) ||
+                    qir_has_side_effect_reads(c, inst)) {
+                        continue;
+                }
+
+                /* Move the generating instruction to the end of the program
+                 * to maintain the order of the VPM writes.
+                 */
+                assert(!vpm_writes[i]->sf);
+                list_del(&inst->link);
+                list_addtail(&inst->link, &vpm_writes[i]->link);
+                qir_remove_instruction(c, vpm_writes[i]);
+
+                c->defs[inst->dst.index] = NULL;
+                inst->dst.file = QFILE_VPM;
+                inst->dst.index = 0;
+
+                progress = true;
+        }
+
         return progress;
 }
diff --git a/lib/mesa/src/gallium/drivers/vc4/vc4_qir_emit_uniform_stream_resets.c b/lib/mesa/src/gallium/drivers/vc4/vc4_qir_emit_uniform_stream_resets.c
index 443682a46..3fd6358e3 100644
--- a/lib/mesa/src/gallium/drivers/vc4/vc4_qir_emit_uniform_stream_resets.c
+++ b/lib/mesa/src/gallium/drivers/vc4/vc4_qir_emit_uniform_stream_resets.c
@@ -36,10 +36,24 @@
 #include "util/u_math.h"
 
 static bool
+inst_reads_a_uniform(struct qinst *inst)
+{
+        if (qir_is_tex(inst))
+                return true;
+
+        for (int i = 0; i < qir_get_op_nsrc(inst->op); i++) {
+                if (inst->src[i].file == QFILE_UNIF)
+                        return true;
+        }
+
+        return false;
+}
+
+static bool
 block_reads_any_uniform(struct qblock *block)
 {
         qir_for_each_inst(inst, block) {
-                if (qir_has_uniform_read(inst))
+                if (inst_reads_a_uniform(inst))
                         return true;
         }
 
@@ -80,7 +94,7 @@ qir_emit_uniform_stream_resets(struct vc4_compile *c)
                 }
 
                 qir_for_each_inst(inst, block) {
-                        if (qir_has_uniform_read(inst))
+                        if (inst_reads_a_uniform(inst))
                                 uniform_count++;
                 }
         }
diff --git a/lib/mesa/src/gallium/drivers/vc4/vc4_qir_live_variables.c b/lib/mesa/src/gallium/drivers/vc4/vc4_qir_live_variables.c
index 7108b3ee9..beefb0d7f 100644
--- a/lib/mesa/src/gallium/drivers/vc4/vc4_qir_live_variables.c
+++ b/lib/mesa/src/gallium/drivers/vc4/vc4_qir_live_variables.c
@@ -205,7 +205,7 @@ qir_setup_def_use(struct vc4_compile *c)
                 _mesa_hash_table_clear(partial_update_ht, NULL);
 
                 qir_for_each_inst(inst, block) {
-                        for (int i = 0; i < qir_get_nsrc(inst); i++)
+                        for (int i = 0; i < qir_get_op_nsrc(inst->op); i++)
                                 qir_setup_use(c, block, ip, inst->src[i]);
 
                         qir_setup_def(c, block, ip, partial_update_ht, inst);
@@ -301,13 +301,8 @@ qir_calculate_live_intervals(struct vc4_compile *c)
 {
         int bitset_words = BITSET_WORDS(c->num_temps);
 
-        /* If we called this function more than once, then we should be
-         * freeing the previous arrays.
-         */
-        assert(!c->temp_start);
-
-        c->temp_start = rzalloc_array(c, int, c->num_temps);
-        c->temp_end = rzalloc_array(c, int, c->num_temps);
+        c->temp_start = reralloc(c, c->temp_start, int, c->num_temps);
+        c->temp_end = reralloc(c, c->temp_end, int, c->num_temps);
 
         for (int i = 0; i < c->num_temps; i++) {
                 c->temp_start[i] = MAX_INSTRUCTION;
@@ -315,10 +310,10 @@ qir_calculate_live_intervals(struct vc4_compile *c)
         }
 
         qir_for_each_block(block, c) {
-                block->def = rzalloc_array(c, BITSET_WORD, bitset_words);
-                block->use = rzalloc_array(c, BITSET_WORD, bitset_words);
-                block->live_in = rzalloc_array(c, BITSET_WORD, bitset_words);
-                block->live_out = rzalloc_array(c, BITSET_WORD, bitset_words);
+                block->def = reralloc(c, block->def, BITSET_WORD, bitset_words);
+                block->use = reralloc(c, block->use, BITSET_WORD, bitset_words);
+                block->live_in = reralloc(c, block->live_in, BITSET_WORD, bitset_words);
+                block->live_out = reralloc(c, block->live_out, BITSET_WORD, bitset_words);
         }
 
         qir_setup_def_use(c);
@@ -327,27 +322,4 @@ qir_calculate_live_intervals(struct vc4_compile *c)
                 ;
 
         qir_compute_start_end(c, c->num_temps);
-
-        if (vc4_debug & VC4_DEBUG_SHADERDB) {
-                int last_ip = 0;
-                for (int i = 0; i < c->num_temps; i++)
-                        last_ip = MAX2(last_ip, c->temp_end[i]);
-
-                int reg_pressure = 0;
-                int max_reg_pressure = 0;
-                for (int i = 0; i < last_ip; i++) {
-                        for (int j = 0; j < c->num_temps; j++) {
-                                if (c->temp_start[j] == i)
-                                        reg_pressure++;
-                                if (c->temp_end[j] == i)
-                                        reg_pressure--;
-                        }
-                        max_reg_pressure = MAX2(max_reg_pressure, reg_pressure);
-                }
-
-                fprintf(stderr, "SHADER-DB: %s prog %d/%d: %d max temps\n",
-                        qir_get_stage_name(c->stage),
-                        c->program_id, c->variant_id,
-                        max_reg_pressure);
-        }
 }
diff --git a/lib/mesa/src/gallium/drivers/vc4/vc4_qir_validate.c b/lib/mesa/src/gallium/drivers/vc4/vc4_qir_validate.c
index 302eb4826..e7cfe5ad2 100644
--- a/lib/mesa/src/gallium/drivers/vc4/vc4_qir_validate.c
+++ b/lib/mesa/src/gallium/drivers/vc4/vc4_qir_validate.c
@@ -84,28 +84,9 @@ void qir_validate(struct vc4_compile *c)
                 case QFILE_LOAD_IMM:
                         fail_instr(c, inst, "Bad dest file");
                         break;
-
-                case QFILE_TEX_S:
-                case QFILE_TEX_T:
-                case QFILE_TEX_R:
-                case QFILE_TEX_B:
-                        if (inst->src[qir_get_tex_uniform_src(inst)].file !=
-                            QFILE_UNIF) {
-                                fail_instr(c, inst,
-                                           "tex op missing implicit uniform");
-                        }
-                        break;
-
-                case QFILE_TEX_S_DIRECT:
-                        if (inst->op != QOP_ADD) {
-                                fail_instr(c, inst,
-                                           "kernel validation requires that "
-                                           "direct texture lookups use an ADD");
-                        }
-                        break;
                 }
 
-                for (int i = 0; i < qir_get_nsrc(inst); i++) {
+                for (int i = 0; i < qir_get_op_nsrc(inst->op); i++) {
                         struct qreg src = inst->src[i];
 
                         switch (src.file) {
@@ -138,11 +119,6 @@ void qir_validate(struct vc4_compile *c)
                         case QFILE_TLB_COLOR_WRITE_MS:
                         case QFILE_TLB_Z_WRITE:
                         case QFILE_TLB_STENCIL_SETUP:
-                        case QFILE_TEX_S_DIRECT:
-                        case QFILE_TEX_S:
-                        case QFILE_TEX_T:
-                        case QFILE_TEX_R:
-                        case QFILE_TEX_B:
                                 fail_instr(c, inst, "Bad src file");
                                 break;
                         }