broadcom/compiler: try to use ldunif(a) instead of ldunif(a)rf in v71

The rf variants need to encode the destination in the cond bits, which
prevents these to be merged with any other instruction that need them.

In 4.x, ldunif(a) write to r5 which is a special register that only
ldunif(a) and ldvary can write so we have a special register class for
it and only allow it for them. Then when we need to choose a register
for a node, if this register is available we always use it.

In 7.x these instructions write to rf0, which can be used by any
instruction, so instead of restricting rf0, we track the temps that
are used as ldunif(a) destinations and use that information to favor
rf0 for them.

Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/25450>

diff --git a/src/broadcom/compiler/v3d_compiler.h b/src/broadcom/compiler/v3d_compiler.h
index d6267a3..13a0dad 100644 (file)
--- a/src/broadcom/compiler/v3d_compiler.h
+++ b/src/broadcom/compiler/v3d_compiler.h
@@ -605,6 +605,9 @@ struct v3d_ra_node_info {
         struct {
                 uint32_t priority;
                 uint8_t class_bits;
+
+                /* V3D 7.x */
+                bool is_ldunif_dst;
         } *info;
         uint32_t alloc_count;
 };

diff --git a/src/broadcom/compiler/vir_register_allocate.c b/src/broadcom/compiler/vir_register_allocate.c
index e0adc1d..1be091f 100644 (file)
--- a/src/broadcom/compiler/vir_register_allocate.c
+++ b/src/broadcom/compiler/vir_register_allocate.c
@@ -384,6 +384,7 @@ add_node(struct v3d_compile *c, uint32_t temp, uint8_t class_bits)
         /* We fill the node priority after we are done inserting spills */
         c->nodes.info[node].class_bits = class_bits;
         c->nodes.info[node].priority = 0;
+        c->nodes.info[node].is_ldunif_dst = false;
 }
 
 /* The spill offset for this thread takes a bit of setup, so do it once at
@@ -899,9 +900,22 @@ v3d_ra_select_accum(struct v3d_ra_select_callback_data *v3d_ra,
 
 static bool
 v3d_ra_select_rf(struct v3d_ra_select_callback_data *v3d_ra,
+                 unsigned int node,
                  BITSET_WORD *regs,
                  unsigned int *out)
 {
+        /* In V3D 7.x, try to assign rf0 to temps used as ldunif's dst
+         * so we can avoid turning them into ldunifrf (which uses the
+         * cond field to encode the dst and would prevent merge with
+         * instructions that use cond flags).
+         */
+        if (v3d_ra->nodes->info[node].is_ldunif_dst &&
+            BITSET_TEST(regs, v3d_ra->phys_index)) {
+                assert(v3d_ra->devinfo->ver >= 71);
+                *out = v3d_ra->phys_index;
+                return true;
+        }
+
         for (int i = 0; i < PHYS_COUNT; i++) {
                 int phys_off = (v3d_ra->next_phys + i) % PHYS_COUNT;
                 int phys = v3d_ra->phys_index + phys_off;
@@ -927,7 +941,7 @@ v3d_ra_select_callback(unsigned int n, BITSET_WORD *regs, void *data)
                 return reg;
         }
 
-        if (v3d_ra_select_rf(v3d_ra, regs, &reg))
+        if (v3d_ra_select_rf(v3d_ra, n, regs, &reg))
                 return reg;
 
         /* If we ran out of physical registers try to assign an accumulator
@@ -1139,15 +1153,24 @@ update_graph_and_reg_classes_for_inst(struct v3d_compile *c,
                                 }
                         }
                 } else {
-                        /* If the instruction has an implicit write
-                         * we can't allocate its dest to the same
-                         * register.
+                        /* Make sure we don't allocate the ldvary's
+                         * destination to rf0, since it would clash
+                         * with its implicit write to that register.
                          */
-                        if (v3d_qpu_writes_rf0_implicitly(c->devinfo, &inst->qpu)) {
+                        if (inst->qpu.sig.ldvary) {
                                 ra_add_node_interference(c->g,
                                                          temp_to_node(c, inst->dst.index),
                                                          implicit_rf_nodes[0]);
                         }
+                        /* Flag dst temps from ldunif(a) instructions
+                         * so we can try to assign rf0 to them and avoid
+                         * converting these to ldunif(a)rf.
+                         */
+                        if (inst->qpu.sig.ldunif || inst->qpu.sig.ldunifa) {
+                                const uint32_t dst_n =
+                                        temp_to_node(c, inst->dst.index);
+                                c->nodes.info[dst_n].is_ldunif_dst = true;
+                        }
                 }
         }
 
@@ -1222,6 +1245,7 @@ v3d_register_allocate(struct v3d_compile *c)
          * without accumulators that can have implicit writes to phys regs.
          */
         for (uint32_t i = 0; i < num_ra_nodes; i++) {
+                c->nodes.info[i].is_ldunif_dst = false;
                 if (c->devinfo->has_accumulators && i < ACC_COUNT) {
                         acc_nodes[i] = i;
                         ra_set_node_reg(c->g, acc_nodes[i], ACC_INDEX + i);

diff --git a/src/broadcom/compiler/vir_to_qpu.c b/src/broadcom/compiler/vir_to_qpu.c
index afc4941..cbbb495 100644 (file)
--- a/src/broadcom/compiler/vir_to_qpu.c
+++ b/src/broadcom/compiler/vir_to_qpu.c
@@ -345,8 +345,15 @@ v3d_generate_code_block(struct v3d_compile *c,
                                 assert(qinst->qpu.alu.add.op == V3D_QPU_A_NOP);
                                 assert(qinst->qpu.alu.mul.op == V3D_QPU_M_NOP);
 
-                                if (!dst.magic ||
-                                    dst.index != V3D_QPU_WADDR_R5) {
+                                bool use_rf;
+                                if (c->devinfo->has_accumulators) {
+                                        use_rf = !dst.magic ||
+                                                 dst.index != V3D_QPU_WADDR_R5;
+                                } else {
+                                        use_rf = dst.magic || dst.index != 0;
+                                }
+
+                                if (use_rf) {
                                         assert(c->devinfo->ver >= 40);
 
                                         if (qinst->qpu.sig.ldunif) {