X-Git-Url: https://git.libre-soc.org/?a=blobdiff_plain;f=src%2Fgallium%2Fdrivers%2Fvc4%2Fvc4_register_allocate.c;h=1d860ea058ee208c271c94950acc62f9bbe3e2d7;hb=e0c6135625e1a2e2dc21a4e1472acb331af043cc;hp=d6bf71c97249509b3e0fef1ad2d337c3266b492d;hpb=755037173d19b65777a97f55455c1f64bf618264;p=mesa.git

diff --git a/src/gallium/drivers/vc4/vc4_register_allocate.c b/src/gallium/drivers/vc4/vc4_register_allocate.c
index d6bf71c9724..1d860ea058e 100644
--- a/src/gallium/drivers/vc4/vc4_register_allocate.c
+++ b/src/gallium/drivers/vc4/vc4_register_allocate.c
@@ -101,7 +101,9 @@ static const struct qpu_reg vc4_regs[] = {
         QPU_R(B, 31),
 };
 #define ACC_INDEX     0
-#define AB_INDEX      (ACC_INDEX + 5)
+#define ACC_COUNT     5
+#define AB_INDEX      (ACC_INDEX + ACC_COUNT)
+#define AB_COUNT      64
 
 static void
 vc4_alloc_reg_set(struct vc4_context *vc4)
@@ -141,14 +143,16 @@ vc4_alloc_reg_set(struct vc4_context *vc4)
         for (int i = 0; i < 2; i++) {
                 ra_class_add_reg(vc4->regs, vc4->reg_class_r4_or_a[i],
                                  ACC_INDEX + 4);
+                ra_class_add_reg(vc4->regs, vc4->reg_class_any[i],
+                                 ACC_INDEX + 4);
         }
 
         /* A/B */
         for (uint32_t i = AB_INDEX; i < AB_INDEX + 64; i ++) {
-                /* Reserve ra31/rb31 for spilling fixup_raddr_conflict() in
+                /* Reserve ra14/rb14 for spilling fixup_raddr_conflict() in
                  * vc4_qpu_emit.c
                  */
-                if (vc4_regs[i].addr == 31)
+                if (vc4_regs[i].addr == 14)
                         continue;
 
                 ra_class_add_reg(vc4->regs, vc4->reg_class_any[0], i);
@@ -198,6 +202,49 @@ node_to_temp_priority(const void *in_a, const void *in_b)
 #define CLASS_BIT_R4			(1 << 2)
 #define CLASS_BIT_R0_R3			(1 << 4)
 
+struct vc4_ra_select_callback_data {
+        uint32_t next_acc;
+        uint32_t next_ab;
+};
+
+static unsigned int
+vc4_ra_select_callback(struct ra_graph *g, BITSET_WORD *regs, void *data)
+{
+        struct vc4_ra_select_callback_data *vc4_ra = data;
+
+        /* If r4 is available, always choose it -- few other things can go
+         * there, and choosing anything else means inserting a mov.
+         */
+        if (BITSET_TEST(regs, ACC_INDEX + 4))
+                return ACC_INDEX + 4;
+
+        /* Choose an accumulator if possible (no delay between write and
+         * read), but round-robin through them to give post-RA instruction
+         * selection more options.
+         */
+        for (int i = 0; i < ACC_COUNT; i++) {
+                int acc_off = (vc4_ra->next_acc + i) % ACC_COUNT;
+                int acc = ACC_INDEX + acc_off;
+
+                if (BITSET_TEST(regs, acc)) {
+                        vc4_ra->next_acc = acc_off + 1;
+                        return acc;
+                }
+        }
+
+        for (int i = 0; i < AB_COUNT; i++) {
+                int ab_off = (vc4_ra->next_ab + i) % AB_COUNT;
+                int ab = AB_INDEX + ab_off;
+
+                if (BITSET_TEST(regs, ab)) {
+                        vc4_ra->next_ab = ab_off + 1;
+                        return ab;
+                }
+        }
+
+        unreachable("RA must pass us at least one possible reg.");
+}
+
 /**
  * Returns a mapping from QFILE_TEMP indices to struct qpu_regs.
  *
@@ -211,6 +258,10 @@ vc4_register_allocate(struct vc4_context *vc4, struct vc4_compile *c)
         uint8_t class_bits[c->num_temps];
         struct qpu_reg *temp_registers = calloc(c->num_temps,
                                                 sizeof(*temp_registers));
+        struct vc4_ra_select_callback_data callback_data = {
+                .next_acc = 0,
+                .next_ab = 0,
+        };
 
         /* If things aren't ever written (undefined values), just read from
          * r0.
@@ -226,6 +277,8 @@ vc4_register_allocate(struct vc4_context *vc4, struct vc4_compile *c)
         /* Compute the live ranges so we can figure out interference. */
         qir_calculate_live_intervals(c);
 
+        ra_set_select_reg_callback(g, vc4_ra_select_callback, &callback_data);
+
         for (uint32_t i = 0; i < c->num_temps; i++) {
                 map[i].temp = i;
                 map[i].priority = c->temp_end[i] - c->temp_start[i];
@@ -254,6 +307,14 @@ vc4_register_allocate(struct vc4_context *vc4, struct vc4_compile *c)
                                 if (c->temp_start[i] < ip && c->temp_end[i] > ip)
                                         class_bits[i] &= ~CLASS_BIT_R4;
                         }
+
+                        /* If we're doing a conditional write of something
+                         * writing R4 (math, tex results), then make sure that
+                         * we store in a temp so that we actually
+                         * conditionally move the result.
+                         */
+                        if (inst->cond != QPU_COND_ALWAYS)
+                                class_bits[inst->dst.index] &= ~CLASS_BIT_R4;
                 } else {
                         /* R4 can't be written as a general purpose
                          * register. (it's TMU_NOSWAP as a write address).
@@ -304,7 +365,7 @@ vc4_register_allocate(struct vc4_context *vc4, struct vc4_compile *c)
                  * can only be done from regfile A, while float unpacks can be
                  * either A or R4.
                  */
-                for (int i = 0; i < qir_get_op_nsrc(inst->op); i++) {
+                for (int i = 0; i < qir_get_nsrc(inst); i++) {
                         if (inst->src[i].file == QFILE_TEMP &&
                             inst->src[i].pack) {
                                 if (qir_is_float_input(inst)) {
@@ -354,6 +415,7 @@ vc4_register_allocate(struct vc4_context *vc4, struct vc4_compile *c)
                          */
                         if (c->fs_threaded) {
                                 c->failed = true;
+                                free(temp_registers);
                                 return NULL;
                         }
 
@@ -383,6 +445,7 @@ vc4_register_allocate(struct vc4_context *vc4, struct vc4_compile *c)
                 }
 
                 c->failed = true;
+                free(temp_registers);
                 return NULL;
         }