nv50: SSG

3 files changed, 18 insertions, 10 deletions

diff --git a/src/gallium/drivers/nv50/nv50_pc_emit.c b/src/gallium/drivers/nv50/nv50_pc_emit.c
index 7808335e50b..e1d7bc64593 100644
--- a/src/gallium/drivers/nv50/nv50_pc_emit.c
+++ b/src/gallium/drivers/nv50/nv50_pc_emit.c
@@ -729,7 +729,7 @@ emit_bitop2(struct nv_pc *pc, struct nv_instruction *i)
 {
    pc->emit[0] = 0xd0000000;
 
-   if (SFILE(i, 0) == NV_FILE_IMM) {
+   if (SFILE(i, 1) == NV_FILE_IMM) {
       emit_form_IMM(pc, i, 0);
 
       if (i->opcode == NV_OP_OR)
@@ -761,7 +761,7 @@ emit_arl(struct nv_pc *pc, struct nv_instruction *i)
    pc->emit[0] = 0x00000001;
    pc->emit[1] = 0xc0000000;
 
-   set_dst(pc, i->def[0]);
+   pc->emit[0] |= (i->def[0]->reg.id + 1) << 2;
    set_pred(pc, i);
    set_src_0(pc, i->src[0]);
    pc->emit[0] |= (get_immd_u32(i->src[1]) & 0x3f) << 16;
diff --git a/src/gallium/drivers/nv50/nv50_pc_optimize.c b/src/gallium/drivers/nv50/nv50_pc_optimize.c
index 1d2710a8aca..4a3a51512e2 100644
--- a/src/gallium/drivers/nv50/nv50_pc_optimize.c
+++ b/src/gallium/drivers/nv50/nv50_pc_optimize.c
@@ -264,11 +264,8 @@ check_swap_src_0_1(struct nv_instruction *nvi)
       return;
    assert(src0 && src1);
 
-   if (src1->value->reg.file == NV_FILE_IMM) {
-      /* should only be present from folding a constant MUL part of a MAD */
-      assert(nvi->opcode == NV_OP_ADD);
+   if (src1->value->reg.file == NV_FILE_IMM)
       return;
-   }
 
    if (is_cmem_load(src0->value->insn)) {
       if (!is_cmem_load(src1->value->insn)) {
@@ -305,7 +302,7 @@ nv_pass_fold_stores(struct nv_pass *ctx, struct nv_basic_block *b)
          continue;
 
       nvi = sti->src[0]->value->insn;
-      if (!nvi || nvi->opcode == NV_OP_PHI)
+      if (!nvi || nvi->opcode == NV_OP_PHI || nv_is_vector_op(nvi->opcode))
          continue;
       assert(nvi->def[0] == sti->src[0]->value);
 
@@ -536,9 +533,9 @@ constant_expression(struct nv_pc *pc, struct nv_instruction *nvi,
       break;
    case NV_OP_SUB:
       switch (type) {
-      case NV_TYPE_F32: u.f32 = u0.f32 - u1.f32;
-      case NV_TYPE_U32: u.u32 = u0.u32 - u1.u32;
-      case NV_TYPE_S32: u.s32 = u0.s32 - u1.s32;
+      case NV_TYPE_F32: u.f32 = u0.f32 - u1.f32; break;
+      case NV_TYPE_U32: u.u32 = u0.u32 - u1.u32; break;
+      case NV_TYPE_S32: u.s32 = u0.s32 - u1.s32; break;
       default:
          assert(0);
          break;
diff --git a/src/gallium/drivers/nv50/nv50_tgsi_to_nc.c b/src/gallium/drivers/nv50/nv50_tgsi_to_nc.c
index 5ac61f108e4..0a4c88c8179 100644
--- a/src/gallium/drivers/nv50/nv50_tgsi_to_nc.c
+++ b/src/gallium/drivers/nv50/nv50_tgsi_to_nc.c
@@ -1595,6 +1595,17 @@ bld_instruction(struct bld_context *bld,
       if (insn->Dst[0].Register.WriteMask & 0x8)
          dst0[3] = bld_imm_f32(bld, 1.0f);
       break;
+   case TGSI_OPCODE_SSG:
+      FOR_EACH_DST0_ENABLED_CHANNEL(c, insn) {
+         src0 = emit_fetch(bld, insn, 0, c);
+         src1 = bld_predicate(bld, src0, FALSE);
+         temp = bld_insn_2(bld, NV_OP_AND, src0, bld_imm_u32(bld, 0x80000000));
+         temp = bld_insn_2(bld, NV_OP_OR,  temp, bld_imm_f32(bld, 1.0f));
+         dst0[c] = bld_insn_2(bld, NV_OP_XOR, temp, temp);
+         dst0[c]->insn->cc = NV_CC_EQ;
+         nv_reference(bld->pc, &dst0[c]->insn->flags_src, src1);
+      }
+      break;
    case TGSI_OPCODE_SUB:
       FOR_EACH_DST0_ENABLED_CHANNEL(c, insn) {
          src0 = emit_fetch(bld, insn, 0, c);