v3d: Rematerialize MOVs of uniforms instead of spilling them.

If we have a MOV of a uniform value available to spill, that's one of our
best choices.  We can just not spill the value, and emit a new load of the
uniform as the fill.  This saves bothering the TMU and the thrsw, and is
the same cost in uniforms (since the spill offset is a uniform anyway).

This doesn't have a huge impact on shader-db, since there aren't a whole
lot of spills and we usually copy-prop the uniforms at the VIR level such
that the only uniform MOVs are from vir_lower_uniforms:

total instructions in shared programs: 6430292 -> 6430279 (<.01%)
total uniforms in shared programs: 2386023 -> 2385787 (<.01%)
total spills in shared programs: 4961 -> 4960 (-0.02%)
total fills in shared programs: 6352 -> 6350 (-0.03%)

However, I'm interested in dropping the uniforms copy-prop in the backend,
since it would be cheaper to not load repeated uniforms if we have the
registers to spare.  This also saves many spills on
dEQP-GLES31.functional.ubo.random.all_per_block_buffers.20, which is what
motivated a bunch of my recent backend work in the first place:

before: 46 spills, 106 fills, 3062 instructions
after: 0 spills, 0 fills, 2611 instructions

src/broadcom/compiler/v3d_compiler.h

@@ -1152,4 +1152,8 @@ vir_BRANCH(struct v3d_compile *c, enum v3d_qpu_branch_cond cond)
         vir_for_each_block(_block, c)                                   \
                 vir_for_each_inst(inst, _block)
 
+#define vir_for_each_inst_inorder_safe(inst, c)                         \
+        vir_for_each_block(_block, c)                                   \
+                vir_for_each_inst_safe(inst, _block)
+
 #endif /* V3D_COMPILER_H */

src/broadcom/compiler/vir_register_allocate.c

@@ -47,10 +47,21 @@ is_last_ldtmu(struct qinst *inst, struct qblock *block)
         return true;
 }
 
+static bool
+vir_is_mov_uniform(struct v3d_compile *c, int temp)
+{
+        struct qinst *def = c->defs[temp];
+
+        return (def &&
+                vir_is_raw_mov(def) &&
+                def->src[0].file == QFILE_UNIF);
+}
+
 static int
 v3d_choose_spill_node(struct v3d_compile *c, struct ra_graph *g,
                       uint32_t *temp_to_node)
 {
+        const float tmu_scale = 5;
         float block_scale = 1.0;
         float spill_costs[c->num_temps];
         bool in_tmu_operation = false;
@@ -75,22 +86,28 @@ v3d_choose_spill_node(struct v3d_compile *c, struct ra_graph *g,
                                         continue;
 
                                 int temp = inst->src[i].index;
-                                if (no_spilling) {
+                                if (vir_is_mov_uniform(c, temp)) {
-                                        BITSET_CLEAR(c->spillable,
-                                                     temp);
-                                } else {
                                         spill_costs[temp] += block_scale;
+                                } else if (!no_spilling) {
+                                        spill_costs[temp] += (block_scale *
+                                                              tmu_scale);
+                                } else {
+                                        BITSET_CLEAR(c->spillable, temp);
                                 }
                         }
 
                         if (inst->dst.file == QFILE_TEMP) {
                                 int temp = inst->dst.index;
 
-                                if (no_spilling) {
+                                if (vir_is_mov_uniform(c, temp)) {
-                                        BITSET_CLEAR(c->spillable,
+                                        /* We just rematerialize the unform
-                                                     temp);
+                                         * later.
+                                         */
+                                } else if (!no_spilling) {
+                                        spill_costs[temp] += (block_scale *
+                                                              tmu_scale);
                                 } else {
-                                        spill_costs[temp] += block_scale;
+                                        BITSET_CLEAR(c->spillable, temp);
                                 }
                         }
 
@@ -184,18 +201,28 @@ v3d_emit_spill_tmua(struct v3d_compile *c, uint32_t spill_offset)
 static void
 v3d_spill_reg(struct v3d_compile *c, int spill_temp)
 {
+        bool is_uniform = vir_is_mov_uniform(c, spill_temp);
+
+        uint32_t spill_offset = 0;
+
+        if (!is_uniform) {
                 uint32_t spill_offset = c->spill_size;
                 c->spill_size += 16 * sizeof(uint32_t);
 
                 if (spill_offset == 0)
                         v3d_setup_spill_base(c);
+        }
 
         struct qinst *last_thrsw = c->last_thrsw;
         assert(!last_thrsw || last_thrsw->is_last_thrsw);
 
         int start_num_temps = c->num_temps;
 
-        vir_for_each_inst_inorder(inst, c) {
+        struct qreg uniform_src = c->undef;
+        if (is_uniform)
+                uniform_src = c->defs[spill_temp]->src[0];
+
+        vir_for_each_inst_inorder_safe(inst, c) {
                 for (int i = 0; i < vir_get_nsrc(inst); i++) {
                         if (inst->src[i].file != QFILE_TEMP ||
                             inst->src[i].index != spill_temp) {
@@ -204,31 +231,41 @@ v3d_spill_reg(struct v3d_compile *c, int spill_temp)
 
                         c->cursor = vir_before_inst(inst);
 
+                        if (is_uniform) {
+                                inst->src[i] = vir_MOV(c, uniform_src);
+                        } else {
                                 v3d_emit_spill_tmua(c, spill_offset);
                                 vir_emit_thrsw(c);
                                 inst->src[i] = vir_LDTMU(c);
                                 c->fills++;
                         }
+                }
 
                 if (inst->dst.file == QFILE_TEMP &&
                     inst->dst.index == spill_temp) {
+                        if (is_uniform) {
+                                c->cursor.link = NULL;
+                                vir_remove_instruction(c, inst);
+                        } else {
                                 c->cursor = vir_after_inst(inst);
 
                                 inst->dst.index = c->num_temps++;
-                        vir_MOV_dest(c, vir_reg(QFILE_MAGIC, V3D_QPU_WADDR_TMUD),
+                                vir_MOV_dest(c, vir_reg(QFILE_MAGIC,
+                                                        V3D_QPU_WADDR_TMUD),
                                              inst->dst);
                                 v3d_emit_spill_tmua(c, spill_offset);
                                 vir_emit_thrsw(c);
                                 vir_TMUWT(c);
                                 c->spills++;
                         }
+                }
 
                 /* If we didn't have a last-thrsw inserted by nir_to_vir and
                  * we've been inserting thrsws, then insert a new last_thrsw
                  * right before we start the vpm/tlb sequence for the last
                  * thread segment.
                  */
-                if (!last_thrsw && c->last_thrsw &&
+                if (!is_uniform && !last_thrsw && c->last_thrsw &&
                     (v3d_qpu_writes_vpm(&inst->qpu) ||
                      v3d_qpu_uses_tlb(&inst->qpu))) {
                         c->cursor = vir_before_inst(inst);