nv50: flatten simple IF/ELSE/ENDIF constructs

Less branching means less instructions and less thread divergence.

4 files changed, 123 insertions, 24 deletions

diff --git a/src/gallium/drivers/nv50/nv50_pc.c b/src/gallium/drivers/nv50/nv50_pc.c
index 7601049126f..5041fc75059 100644
--- a/src/gallium/drivers/nv50/nv50_pc.c
+++ b/src/gallium/drivers/nv50/nv50_pc.c
@@ -125,6 +125,20 @@ nv50_nvi_can_load(struct nv_instruction *nvi, int s, struct nv_value *value)
    }
 }
 
+/* Return whether this instruction can be executed conditionally. */
+boolean
+nv50_nvi_can_predicate(struct nv_instruction *nvi)
+{
+   int i;
+
+   if (nvi->flags_src)
+      return FALSE;
+   for (i = 0; i < 4 && nvi->src[i]; ++i)
+      if (nvi->src[i]->value->reg.file == NV_FILE_IMM)
+         return FALSE;
+   return TRUE;
+}
+
 ubyte
 nv50_supported_src_mods(uint opcode, int s)
 {
diff --git a/src/gallium/drivers/nv50/nv50_pc.h b/src/gallium/drivers/nv50/nv50_pc.h
index b24a3067b8d..28208ad247b 100644
--- a/src/gallium/drivers/nv50/nv50_pc.h
+++ b/src/gallium/drivers/nv50/nv50_pc.h
@@ -432,6 +432,7 @@ void nv_print_program(struct nv_basic_block *b);
 boolean nv_op_commutative(uint opcode);
 int nv50_indirect_opnd(struct nv_instruction *);
 boolean nv50_nvi_can_use_imm(struct nv_instruction *, int s);
+boolean nv50_nvi_can_predicate(struct nv_instruction *);
 boolean nv50_nvi_can_load(struct nv_instruction *, int s, struct nv_value *);
 ubyte nv50_supported_src_mods(uint opcode, int s);
 int nv_nvi_refcount(struct nv_instruction *);
diff --git a/src/gallium/drivers/nv50/nv50_pc_optimize.c b/src/gallium/drivers/nv50/nv50_pc_optimize.c
index daf63a3d205..4cf387257df 100644
--- a/src/gallium/drivers/nv50/nv50_pc_optimize.c
+++ b/src/gallium/drivers/nv50/nv50_pc_optimize.c
@@ -119,6 +119,15 @@ nvi_isnop(struct nv_instruction *nvi)
    return values_equal(nvi->def[0], nvi->src[0]->value);
 }
 
+struct nv_pass {
+   struct nv_pc *pc;
+   int n;
+   void *priv;
+};
+
+static int
+nv_pass_flatten(struct nv_pass *ctx, struct nv_basic_block *b);
+
 static void
 nv_pc_pass_pre_emission(void *priv, struct nv_basic_block *b)
 {
@@ -204,6 +213,13 @@ nv_pc_pass_pre_emission(void *priv, struct nv_basic_block *b)
 int
 nv_pc_exec_pass2(struct nv_pc *pc)
 {
+   struct nv_pass pass;
+
+   pass.pc = pc;
+
+   pc->pass_seq++;
+   nv_pass_flatten(&pass, pc->root);
+
    debug_printf("preparing %u blocks for emission\n", pc->num_blocks);
 
    pc->bb_list = CALLOC(pc->num_blocks, sizeof(struct nv_basic_block *));
@@ -273,12 +289,6 @@ check_swap_src_0_1(struct nv_instruction *nvi)
       nvi->set_cond = cc_swapped[nvi->set_cond];
 }
 
-struct nv_pass {
-   struct nv_pc *pc;
-   int n;
-   void *priv;
-};
-
 static int
 nv_pass_fold_stores(struct nv_pass *ctx, struct nv_basic_block *b)
 {
@@ -863,24 +873,95 @@ nv_pass_dce(struct nv_pass_dce *ctx, struct nv_basic_block *b)
    return 0;
 }
 
+/* Register allocation inserted ELSE blocks for all IF/ENDIF without ELSE.
+ * Returns TRUE if @bb initiates an IF/ELSE/ENDIF clause, or is an IF with
+ * BREAK and dummy ELSE block.
+ */
 static INLINE boolean
-bb_simple_if_endif(struct nv_basic_block *bb)
+bb_is_if_else_endif(struct nv_basic_block *bb)
+{
+   if (!bb->out[0] || !bb->out[1])
+      return FALSE;
+
+   if (bb->out[0]->out_kind[0] == CFG_EDGE_LOOP_LEAVE) {
+      return (bb->out[0]->out[1] == bb->out[1]->out[0] &&
+              !bb->out[1]->out[1]);
+   } else {
+      return (bb->out[0]->out[0] == bb->out[1]->out[0] &&
+              !bb->out[0]->out[1] &&
+              !bb->out[1]->out[1]);
+   }
+}
+
+/* predicate instructions and remove branch at the end */
+static void
+predicate_instructions(struct nv_pc *pc, struct nv_basic_block *b,
+                       struct nv_value *p, ubyte cc)
 {
-   return (bb->out[0] && bb->out[1] &&
-           bb->out[0]->out[0] == bb->out[1] &&
-           !bb->out[0]->out[1]);
+   struct nv_instruction *nvi;
+
+   if (!b->entry)
+      return;
+   for (nvi = b->entry; nvi->next; nvi = nvi->next) {
+      if (!nvi_isnop(nvi)) {
+         nvi->cc = cc;
+         nv_reference(pc, &nvi->flags_src, p);
+      }
+   }
+
+   if (nvi->opcode == NV_OP_BRA)
+      nv_nvi_delete(nvi);
+   else
+   if (!nvi_isnop(nvi)) {
+      nvi->cc = cc;
+      nv_reference(pc, &nvi->flags_src, p);
+   }
 }
 
+/* NOTE: Run this after register allocation, we can just cut out the cflow
+ * instructions and hook the predicates to the conditional OPs if they are
+ * not using immediates; better than inserting SELECT to join definitions.
+ *
+ * NOTE: Should adapt prior optimization to make this possible more often.
+ */
 static int
 nv_pass_flatten(struct nv_pass *ctx, struct nv_basic_block *b)
 {
-   int j;
+   struct nv_instruction *nvi;
+   struct nv_value *pred;
+   int i;
+   int n0 = 0, n1 = 0;
+
+   if (bb_is_if_else_endif(b)) {
+
+      debug_printf("nv_pass_flatten: IF/ELSE/ENDIF construct at BB:%i\n", b->id);
 
-   if (bb_simple_if_endif(b)) {
-      ++ctx->n;
-      debug_printf("nv_pass_flatten: total IF/ENDIF constructs: %i\n", ctx->n);
+      for (n0 = 0, nvi = b->out[0]->entry; nvi; nvi = nvi->next, ++n0)
+         if (!nv50_nvi_can_predicate(nvi))
+            break;
+      if (!nvi) {
+         for (n1 = 0, nvi = b->out[1]->entry; nvi; nvi = nvi->next, ++n1)
+            if (!nv50_nvi_can_predicate(nvi))
+               break;
+         if (nvi) {
+            debug_printf("cannot predicate: "); nv_print_instruction(nvi);
+         }
+      } else {
+         debug_printf("cannot predicate: "); nv_print_instruction(nvi);
+      }
+
+      if (!nvi && n0 < 12 && n1 < 12) { /* 12 as arbitrary limit */
+         assert(b->exit && b->exit->flags_src);
+         pred = b->exit->flags_src->value;
+
+         predicate_instructions(ctx->pc, b->out[0], pred, NV_CC_NE | NV_CC_U);
+         predicate_instructions(ctx->pc, b->out[1], pred, NV_CC_EQ);
+
+         assert(b->exit && b->exit->opcode == NV_OP_BRA);
+         nv_nvi_delete(b->exit);
+      }
    }
-   DESCEND_ARBITRARY(j, nv_pass_flatten);
+   DESCEND_ARBITRARY(i, nv_pass_flatten);
 
    return 0;
 }
@@ -960,11 +1041,6 @@ nv_pc_exec_pass0(struct nv_pc *pc)
    pass.n = 0;
    pass.pc = pc;
 
-   pc->pass_seq++;
-   ret = nv_pass_flatten(&pass, pc->root);
-   if (ret)
-      return ret;
-
    /* Do this first, so we don't have to pay attention
     * to whether sources are supported memory loads.
     */
diff --git a/src/gallium/drivers/nv50/nv50_tgsi_to_nc.c b/src/gallium/drivers/nv50/nv50_tgsi_to_nc.c
index 7e77ed6ef6e..b23c285dc12 100644
--- a/src/gallium/drivers/nv50/nv50_tgsi_to_nc.c
+++ b/src/gallium/drivers/nv50/nv50_tgsi_to_nc.c
@@ -591,7 +591,7 @@ bld_get_address(struct bld_context *bld, int id, struct nv_value *indirect)
 
 
 static struct nv_value *
-bld_predicate(struct bld_context *bld, struct nv_value *src)
+bld_predicate(struct bld_context *bld, struct nv_value *src, boolean bool_only)
 {
    struct nv_instruction *nvi = src->insn;
 
@@ -600,6 +600,14 @@ bld_predicate(struct bld_context *bld, struct nv_value *src)
        nvi->bb != bld->pc->current_block) {
       nvi = new_instruction(bld->pc, NV_OP_CVT);
       nv_reference(bld->pc, &nvi->src[0], src);
+   } else
+   if (bool_only) {
+      while (nvi->opcode == NV_OP_ABS || nvi->opcode == NV_OP_CVT ||
+             nvi->opcode == NV_OP_NEG) {
+         /* TGSI SET gets conversion to f32, we only need source 0/~0 */
+         if (!nvi->def[0]->insn->flags_src)
+            nvi = nvi->src[0]->value->insn;
+      }
    }
 
    if (!nvi->flags_def) {
@@ -614,7 +622,7 @@ bld_kil(struct bld_context *bld, struct nv_value *src)
 {
    struct nv_instruction *nvi;
 
-   src = bld_predicate(bld, src);
+   src = bld_predicate(bld, src, FALSE);
    nvi = new_instruction(bld->pc, NV_OP_KIL);
    nvi->fixed = 1;
    nvi->flags_src = new_ref(bld->pc, src);
@@ -1223,7 +1231,7 @@ bld_instruction(struct bld_context *bld,
          src0 = emit_fetch(bld, insn, 0, c);
          src1 = emit_fetch(bld, insn, 1, c);
          src2 = emit_fetch(bld, insn, 2, c);
-         src0 = bld_predicate(bld, src0);
+         src0 = bld_predicate(bld, src0, FALSE);
 
          src1 = bld_insn_1(bld, NV_OP_MOV, src1);
          src1->insn->flags_src = new_ref(bld->pc, src0);
@@ -1304,7 +1312,7 @@ bld_instruction(struct bld_context *bld,
       bld->join_bb[bld->cond_lvl] = bld->pc->current_block;
       bld->cond_bb[bld->cond_lvl] = bld->pc->current_block;
 
-      src1 = bld_predicate(bld, emit_fetch(bld, insn, 0, 0));
+      src1 = bld_predicate(bld, emit_fetch(bld, insn, 0, 0), TRUE);
 
       bld_flow(bld, NV_OP_BRA, NV_CC_EQ, src1, NULL, FALSE);