r600/sfn: Make use of variable length DOT

This frees some alu slots for better group scheduling.

Signed-off-by: Gert Wollny <gert.wollny@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20205>

src/gallium/drivers/r600/sfn/sfn_instr_alu.cpp

@@ -694,7 +694,18 @@ AluInstr::split(ValueFactory& vf)
 
    m_dest->del_parent(this);
 
-   for (int s = 0; s < m_alu_slots; ++s) {
+   int start_slot = 0;
+   bool is_dot = m_opcode == op2_dot || opcode() == op2_dot_ieee;
+   auto last_opcode = m_opcode;
+
+   if (is_dot) {
+      start_slot = m_dest->chan();
+      last_opcode = m_opcode == op2_dot ? op2_mul : op2_mul_ieee;
+   }
+
+
+   for (int k = 0; k < m_alu_slots; ++k) {
+      int s = k + start_slot;
 
       PRegister dst = m_dest->chan() == s ? m_dest : vf.dummy_dest(s);
       if (dst->pin() != pin_chgr) {
@@ -720,7 +731,10 @@ AluInstr::split(ValueFactory& vf)
          src.push_back(old_src);
       }
 
-      auto instr = new AluInstr(m_opcode, dst, src, {}, 1);
+      auto opcode = k < m_alu_slots -1 ? m_opcode : last_opcode;
+
+
+      auto instr = new AluInstr(opcode, dst, src, {}, 1);
       instr->set_blockid(block_id(), index());
 
       if (s == 0 || !m_alu_flags.test(alu_64bit_op)) {
@@ -1239,6 +1253,8 @@ emit_unpack_32_2x16_split_y(const nir_alu_instr& alu, Shader& shader);
 static bool
 emit_dot(const nir_alu_instr& alu, int nelm, Shader& shader);
 static bool
+emit_dot4(const nir_alu_instr& alu, Shader& shader);
+static bool
 emit_create_vec(const nir_alu_instr& instr, unsigned nc, Shader& shader);
 
 static bool
@@ -1524,7 +1540,7 @@ AluInstr::from_nir(nir_alu_instr *alu, Shader& shader)
    case nir_op_fdot3:
       return emit_dot(*alu, 3, shader);
    case nir_op_fdot4:
-      return emit_dot(*alu, 4, shader);
+      return emit_dot4(*alu, shader);
 
    case nir_op_feq32:
    case nir_op_feq:
@@ -2463,18 +2479,49 @@ emit_dot(const nir_alu_instr& alu, int n, Shader& shader)
    const nir_alu_src& src0 = alu.src[0];
    const nir_alu_src& src1 = alu.src[1];
 
-   auto dest = value_factory.dest(alu.dest.dest, 0, pin_free);
+   auto dest = value_factory.dest(alu.dest.dest, 0, pin_chan);
 
-   AluInstr::SrcValues srcs(8);
+   AluInstr::SrcValues srcs(2 * n);
 
    for (int i = 0; i < n; ++i) {
       srcs[2 * i] = value_factory.src(src0, i);
       srcs[2 * i + 1] = value_factory.src(src1, i);
    }
 
-   for (int i = n; i < 4; ++i) {
+   auto op =
-      srcs[2 * i] = value_factory.zero();
+      unlikely(shader.has_flag(Shader::sh_legacy_math_rules)) ? op2_dot : op2_dot_ieee;
-      srcs[2 * i + 1] = value_factory.zero();
+   AluInstr *ir = new AluInstr(op, dest, srcs, AluInstr::last_write, n);
+
+   if (src0.negate)
+      ir->set_alu_flag(alu_src0_neg);
+   if (src0.abs)
+      ir->set_alu_flag(alu_src0_abs);
+   if (src1.negate)
+      ir->set_alu_flag(alu_src1_neg);
+   if (src1.abs)
+      ir->set_alu_flag(alu_src1_abs);
+
+   if (alu.dest.saturate)
+      ir->set_alu_flag(alu_dst_clamp);
+
+   shader.emit_instruction(ir);
+   return true;
+}
+
+static bool
+emit_dot4(const nir_alu_instr& alu, Shader& shader)
+{
+   auto& value_factory = shader.value_factory();
+   const nir_alu_src& src0 = alu.src[0];
+   const nir_alu_src& src1 = alu.src[1];
+
+   auto dest = value_factory.dest(alu.dest.dest, 0, pin_free);
+
+   AluInstr::SrcValues srcs(8);
+
+   for (int i = 0; i < 4; ++i) {
+      srcs[2 * i] = value_factory.src(src0, i);
+      srcs[2 * i + 1] = value_factory.src(src1, i);
    }
 
    auto op =
@@ -2497,6 +2544,7 @@ emit_dot(const nir_alu_instr& alu, int n, Shader& shader)
    return true;
 }
 
+
 static bool
 emit_fdph(const nir_alu_instr& alu, Shader& shader)
 {