nir: add legal bit_sizes to intrinsics

With OpenCL some system values match the address bits, but in GLSL we also
have some system values being 64 bit like subgroup masks.

With this it is possible to adjust the builder functions so that depending
on the bit_sizes the correct bit_size is used or an additional argument is
added in case of multiple possible values.

v2: validate dest bit_size
v3: generate hex values in python code
    remove useless imports
    rename and move bit_sizes
v4: add 1 to legal bit_sizes for front_face

Signed-off-by: Karol Herbst <kherbst@redhat.com>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>

src/compiler/nir/nir.h

@@ -1313,6 +1313,9 @@ typedef struct {
     */
    unsigned dest_components;
 
+   /** bitfield of legal bit sizes */
+   unsigned dest_bit_sizes;
+
    /** the number of constant indices used by the intrinsic */
    unsigned num_indices;
 

src/compiler/nir/nir_intrinsics.py

@@ -32,7 +32,7 @@ class Intrinsic(object):
    NOTE: this must be kept in sync with nir_intrinsic_info.
    """
    def __init__(self, name, src_components, dest_components,
-                indices, flags, sysval):
+                indices, flags, sysval, bit_sizes):
        """Parameters:
 
        - name: the intrinsic name
@@ -45,6 +45,7 @@ class Intrinsic(object):
        - indices: list of constant indicies
        - flags: list of semantic flags
        - sysval: is this a system-value intrinsic
+       - bit_sizes: allowed dest bit_sizes
        """
        assert isinstance(name, str)
        assert isinstance(src_components, list)
@@ -58,6 +59,8 @@ class Intrinsic(object):
        if flags:
            assert isinstance(flags[0], str)
        assert isinstance(sysval, bool)
+       if bit_sizes:
+           assert isinstance(bit_sizes[0], int)
 
        self.name = name
        self.num_srcs = len(src_components)
@@ -68,6 +71,7 @@ class Intrinsic(object):
        self.indices = indices
        self.flags = flags
        self.sysval = sysval
+       self.bit_sizes = bit_sizes
 
 #
 # Possible indices:
@@ -125,10 +129,10 @@ CAN_REORDER   = "NIR_INTRINSIC_CAN_REORDER"
 INTR_OPCODES = {}
 
 def intrinsic(name, src_comp=[], dest_comp=-1, indices=[],
-              flags=[], sysval=False):
+              flags=[], sysval=False, bit_sizes=[]):
     assert name not in INTR_OPCODES
     INTR_OPCODES[name] = Intrinsic(name, src_comp, dest_comp,
-                                   indices, flags, sysval)
+                                   indices, flags, sysval, bit_sizes)
 
 intrinsic("nop", flags=[CAN_ELIMINATE])
 
@@ -454,12 +458,13 @@ intrinsic("shared_atomic_fmin",  src_comp=[1, 1], dest_comp=1, indices=[BASE])
 intrinsic("shared_atomic_fmax",  src_comp=[1, 1], dest_comp=1, indices=[BASE])
 intrinsic("shared_atomic_fcomp_swap", src_comp=[1, 1, 1], dest_comp=1, indices=[BASE])
 
-def system_value(name, dest_comp, indices=[]):
+def system_value(name, dest_comp, indices=[], bit_sizes=[32]):
     intrinsic("load_" + name, [], dest_comp, indices,
-              flags=[CAN_ELIMINATE, CAN_REORDER], sysval=True)
+              flags=[CAN_ELIMINATE, CAN_REORDER], sysval=True,
+              bit_sizes=bit_sizes)
 
 system_value("frag_coord", 4)
-system_value("front_face", 1)
+system_value("front_face", 1, bit_sizes=[1, 32])
 system_value("vertex_id", 1)
 system_value("vertex_id_zero_base", 1)
 system_value("first_vertex", 1)
@@ -485,17 +490,17 @@ system_value("local_invocation_index", 1)
 system_value("work_group_id", 3)
 system_value("user_clip_plane", 4, indices=[UCP_ID])
 system_value("num_work_groups", 3)
-system_value("helper_invocation", 1)
+system_value("helper_invocation", 1, bit_sizes=[1, 32])
 system_value("alpha_ref_float", 1)
 system_value("layer_id", 1)
 system_value("view_index", 1)
 system_value("subgroup_size", 1)
 system_value("subgroup_invocation", 1)
-system_value("subgroup_eq_mask", 0)
-system_value("subgroup_ge_mask", 0)
-system_value("subgroup_gt_mask", 0)
-system_value("subgroup_le_mask", 0)
-system_value("subgroup_lt_mask", 0)
+system_value("subgroup_eq_mask", 0, bit_sizes=[32, 64])
+system_value("subgroup_ge_mask", 0, bit_sizes=[32, 64])
+system_value("subgroup_gt_mask", 0, bit_sizes=[32, 64])
+system_value("subgroup_le_mask", 0, bit_sizes=[32, 64])
+system_value("subgroup_lt_mask", 0, bit_sizes=[32, 64])
 system_value("num_subgroups", 1)
 system_value("subgroup_id", 1)
 system_value("local_group_size", 3)

src/compiler/nir/nir_intrinsics_c.py

@@ -1,3 +1,5 @@
+from functools import reduce
+import operator
 
 template = """\
 /* Copyright (C) 2018 Red Hat
@@ -36,6 +38,7 @@ const nir_intrinsic_info nir_intrinsic_infos[nir_num_intrinsics] = {
 % endif
    .has_dest = ${"true" if opcode.has_dest else "false"},
    .dest_components = ${max(opcode.dest_components, 0)},
+   .dest_bit_sizes = ${hex(reduce(operator.or_, opcode.bit_sizes, 0))},
    .num_indices = ${opcode.num_indices},
 % if opcode.indices:
    .index_map = {
@@ -64,7 +67,7 @@ def main():
 
     path = os.path.join(args.outdir, 'nir_intrinsics.c')
     with open(path, 'wb') as f:
-        f.write(Template(template, output_encoding='utf-8').render(INTR_OPCODES=INTR_OPCODES))
+        f.write(Template(template, output_encoding='utf-8').render(INTR_OPCODES=INTR_OPCODES, reduce=reduce, operator=operator))
 
 if __name__ == '__main__':
     main()

src/compiler/nir/nir_validate.c

@@ -563,9 +563,15 @@ validate_intrinsic_instr(nir_intrinsic_instr *instr, validate_state *state)
 
    if (nir_intrinsic_infos[instr->intrinsic].has_dest) {
       unsigned components_written = nir_intrinsic_dest_components(instr);
+      unsigned bit_sizes = nir_intrinsic_infos[instr->intrinsic].dest_bit_sizes;
 
       validate_assert(state, components_written > 0);
 
+      if (dest_bit_size && bit_sizes)
+         validate_assert(state, dest_bit_size & bit_sizes);
+      else
+         dest_bit_size = dest_bit_size ? dest_bit_size : bit_sizes;
+
       validate_dest(&instr->dest, state, dest_bit_size, components_written);
    }
 }