nvc0: fix branching ops

- bra is PC relative
- jump to else condition was inverted
- handle integer comparisons

4 files changed, 26 insertions, 13 deletions

diff --git a/src/gallium/drivers/nvc0/nvc0_pc.c b/src/gallium/drivers/nvc0/nvc0_pc.c
index e38f6ced24b..cf7b8e347fb 100644
--- a/src/gallium/drivers/nvc0/nvc0_pc.c
+++ b/src/gallium/drivers/nvc0/nvc0_pc.c
@@ -328,6 +328,7 @@ nvc0_emit_program(struct nv_pc *pc)
 
    NOUVEAU_DBG("emitting program: size = %u\n", pc->emit_size);
 
+   pc->emit_pos = 0;
    for (n = 0; n < pc->num_blocks; ++n) {
       struct nv_instruction *i;
       struct nv_basic_block *b = pc->bb_list[n];
@@ -335,7 +336,7 @@ nvc0_emit_program(struct nv_pc *pc)
       for (i = b->entry; i; i = i->next) {
          nvc0_emit_instruction(pc, i);
          pc->emit += 2;
-         pc->emit_pos += 2;
+         pc->emit_pos += 8;
       }
    }
    assert(pc->emit == &code[pc->emit_size / 4]);
diff --git a/src/gallium/drivers/nvc0/nvc0_pc.h b/src/gallium/drivers/nvc0/nvc0_pc.h
index b48b0b1fbad..df0314965a3 100644
--- a/src/gallium/drivers/nvc0/nvc0_pc.h
+++ b/src/gallium/drivers/nvc0/nvc0_pc.h
@@ -425,7 +425,7 @@ struct nv_basic_block {
    uint priv; /* reset to 0 after you're done */
    uint pass_seq;
 
-   uint32_t emit_pos; /* position, size in emitted code */
+   uint32_t emit_pos; /* position, size in emitted code (in bytes) */
    uint32_t emit_size;
 
    uint32_t live_set[NV_PC_MAX_VALUES / 32];
@@ -457,8 +457,8 @@ struct nv_pc {
    unsigned immd_count;
 
    uint32_t *emit;
-   unsigned emit_size;
-   unsigned emit_pos;
+   uint32_t emit_size;
+   uint32_t emit_pos;
 
    void *reloc_entries;
    unsigned num_relocs;
diff --git a/src/gallium/drivers/nvc0/nvc0_pc_emit.c b/src/gallium/drivers/nvc0/nvc0_pc_emit.c
index 6735f93fd3c..cd1ad03b00b 100644
--- a/src/gallium/drivers/nvc0/nvc0_pc_emit.c
+++ b/src/gallium/drivers/nvc0/nvc0_pc_emit.c
@@ -73,7 +73,7 @@ create_fixup(struct nv_pc *pc, uint8_t ty,
 
    f = (struct nvc0_fixup *)pc->reloc_entries;
 
-   f[n].ofst = (pc->emit_pos + w) * 4;
+   f[n].ofst = pc->emit_pos + w * 4;
    f[n].type = ty;
    f[n].data = data;
    f[n].mask = m;
@@ -217,19 +217,26 @@ const_space_index(struct nv_instruction *i, int s)
 static void
 emit_flow(struct nv_pc *pc, struct nv_instruction *i, uint8_t op)
 {
-   pc->emit[0] = 0x000001e7;
+   pc->emit[0] = 0x00000007;
    pc->emit[1] = op << 24;
 
-   set_pred(pc, i);
+   if (op == 0x40 || (op >= 0x80 && op <= 0x98)) {
+      /* bra, exit, ret or kil */
+      pc->emit[0] |= 0x1e0;
+      set_pred(pc, i);
+   }
 
    if (i->target) {
-      uint32_t pos = i->target->emit_pos;
+      int32_t pcrel = i->target->emit_pos - (pc->emit_pos + 8);
 
+      /* we will need relocations only for global functions */
+      /*
       create_fixup(pc, NVC0_FIXUP_CODE_RELOC, 0, pos, 26, 0xfc000000);
       create_fixup(pc, NVC0_FIXUP_CODE_RELOC, 1, pos, -6, 0x0001ffff);
+      */
 
-      pc->emit[0] |= (pos & 0x3f) << 26;
-      pc->emit[1] |= (pos >> 6) & 0x1ffff;
+      pc->emit[0] |= (pcrel & 0x3f) << 26;
+      pc->emit[1] |= (pcrel >> 6) & 0x1ffff;
    }
 }
 
@@ -893,6 +900,11 @@ nvc0_emit_instruction(struct nv_pc *pc, struct nv_instruction *i)
       emit_mul_f32(pc, i);
       break;
    case NV_OP_SET_F32:
+   case NV_OP_SET_F32_AND:
+   case NV_OP_SET_F32_OR:
+   case NV_OP_SET_F32_XOR:
+   case NV_OP_SET_S32:
+   case NV_OP_SET_U32:
    case NV_OP_FSET_F32:
       emit_set(pc, i);
       break;
@@ -926,8 +938,8 @@ nvc0_emit_instruction(struct nv_pc *pc, struct nv_instruction *i)
       break;
    case NV_OP_JOIN:
    case NV_OP_NOP:
-      pc->emit[0] = 0x00003c00;
-      pc->emit[1] = 0x00000000;
+      pc->emit[0] = 0x00003de4;
+      pc->emit[1] = 0x40000000;
       break;
    case NV_OP_SELP:
       emit_selp(pc, i);
diff --git a/src/gallium/drivers/nvc0/nvc0_tgsi_to_nc.c b/src/gallium/drivers/nvc0/nvc0_tgsi_to_nc.c
index a6797db9c54..26f9e735fb2 100644
--- a/src/gallium/drivers/nvc0/nvc0_tgsi_to_nc.c
+++ b/src/gallium/drivers/nvc0/nvc0_tgsi_to_nc.c
@@ -1494,7 +1494,7 @@ 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_setp(bld, NV_OP_SET_U32, NV_CC_NE,
+      src1 = bld_setp(bld, NV_OP_SET_U32, NV_CC_EQ,
                       emit_fetch(bld, insn, 0, 0), bld->zero);
 
       bld_flow(bld, NV_OP_BRA, src1, NULL, (bld->cond_lvl == 0));