mir_foreach_instr_global(ctx, ins) {
mir_rewrite_index_dst_single(ins, old, new);
}
+
+ /* Implicitly written before the shader */
+ if (ctx->blend_input == old)
+ ctx->blend_input = new;
+
+ if (ctx->blend_src1 == old)
+ ctx->blend_src1 = new;
}
void
return mir_use_count(ctx, value) <= 1;
}
-static bool
-mir_nontrivial_raw_mod(midgard_vector_alu_src src, bool is_int)
+bool
+mir_nontrivial_mod(midgard_instruction *ins, unsigned i, bool check_swizzle)
{
- if (is_int)
- return src.mod == midgard_int_shift;
- else
- return src.mod;
-}
+ bool is_int = midgard_is_integer_op(ins->op);
-static bool
-mir_nontrivial_mod(midgard_vector_alu_src src, bool is_int, unsigned mask, unsigned *swizzle)
-{
- if (mir_nontrivial_raw_mod(src, is_int)) return true;
+ if (is_int) {
+ if (ins->src_shift[i]) return true;
+ } else {
+ if (ins->src_neg[i]) return true;
+ if (ins->src_abs[i]) return true;
+ }
- /* size-conversion */
- if (src.half) return true;
+ if (ins->dest_type != ins->src_types[i]) return true;
- for (unsigned c = 0; c < 16; ++c) {
- if (!(mask & (1 << c))) continue;
- if (swizzle[c] != c) return true;
+ if (check_swizzle) {
+ for (unsigned c = 0; c < 16; ++c) {
+ if (!(ins->mask & (1 << c))) continue;
+ if (ins->swizzle[i][c] != c) return true;
+ }
}
return false;
}
-bool
-mir_nontrivial_source2_mod(midgard_instruction *ins)
-{
- bool is_int = midgard_is_integer_op(ins->alu.op);
-
- midgard_vector_alu_src src2 =
- vector_alu_from_unsigned(ins->alu.src2);
-
- return mir_nontrivial_mod(src2, is_int, ins->mask, ins->swizzle[1]);
-}
-
-bool
-mir_nontrivial_source2_mod_simple(midgard_instruction *ins)
-{
- bool is_int = midgard_is_integer_op(ins->alu.op);
-
- midgard_vector_alu_src src2 =
- vector_alu_from_unsigned(ins->alu.src2);
-
- return mir_nontrivial_raw_mod(src2, is_int) || src2.half;
-}
-
bool
mir_nontrivial_outmod(midgard_instruction *ins)
{
- bool is_int = midgard_is_integer_op(ins->alu.op);
- unsigned mod = ins->alu.outmod;
+ bool is_int = midgard_is_integer_op(ins->op);
+ unsigned mod = ins->outmod;
- /* Type conversion is a sort of outmod */
- if (ins->alu.dest_override != midgard_dest_override_none)
+ if (ins->dest_type != ins->src_types[1])
return true;
if (is_int)
return mod != midgard_outmod_none;
}
+/* 128 / sz = exp2(log2(128 / sz))
+ * = exp2(log2(128) - log2(sz))
+ * = exp2(7 - log2(sz))
+ * = 1 << (7 - log2(sz))
+ */
+
+static unsigned
+mir_components_for_bits(unsigned bits)
+{
+ return 1 << (7 - util_logbase2(bits));
+}
+
+unsigned
+mir_components_for_type(nir_alu_type T)
+{
+ unsigned sz = nir_alu_type_get_type_size(T);
+ return mir_components_for_bits(sz);
+}
+
uint16_t
mir_from_bytemask(uint16_t bytemask, unsigned bits)
{
{
unsigned bytes = bits / 8;
unsigned maxmask = mask_of(bytes);
- unsigned channels = 16 / bytes;
+ unsigned channels = mir_components_for_bits(bits);
for (unsigned c = 0; c < channels; ++c) {
unsigned submask = maxmask << (c * bytes);
/* Checks if we should use an upper destination override, rather than the lower
* one in the IR. Returns zero if no, returns the bytes to shift otherwise */
-unsigned
-mir_upper_override(midgard_instruction *ins)
+signed
+mir_upper_override(midgard_instruction *ins, unsigned inst_size)
{
- /* If there is no override, there is no upper override, tautology */
- if (ins->alu.dest_override == midgard_dest_override_none)
- return 0;
+ unsigned type_size = nir_alu_type_get_type_size(ins->dest_type);
- /* Make sure we didn't already lower somehow */
- assert(ins->alu.dest_override == midgard_dest_override_lower);
+ /* If the sizes are the same, there's nothing to override */
+ if (type_size == inst_size)
+ return -1;
/* There are 16 bytes per vector, so there are (16/bytes)
* components per vector. So the magic half is half of
* (16/bytes), which simplifies to 8/bytes = 8 / (bits / 8) = 64 / bits
* */
- unsigned type_size = nir_alu_type_get_type_size(ins->dest_type);
- unsigned threshold = 64 / type_size;
+ unsigned threshold = mir_components_for_bits(type_size) >> 1;
/* How many components did we shift over? */
unsigned zeroes = __builtin_ctz(ins->mask);
uint16_t
mir_bytemask_of_read_components_index(midgard_instruction *ins, unsigned i)
{
- if (ins->compact_branch && ins->writeout && (i == 0)) {
- /* Non-ZS writeout uses all components */
- if (!ins->writeout_depth && !ins->writeout_stencil)
- return 0xFFFF;
-
- /* For ZS-writeout, if both Z and S are written we need two
- * components, otherwise we only need one.
- */
- if (ins->writeout_depth && ins->writeout_stencil)
- return 0xFF;
- else
- return 0xF;
- }
-
/* Conditional branches read one 32-bit component = 4 bytes (TODO: multi branch??) */
if (ins->compact_branch && ins->branch.conditional && (i == 0))
return 0xF;
/* ALU ops act componentwise so we need to pay attention to
* their mask. Texture/ldst does not so we don't clamp source
* readmasks based on the writemask */
- unsigned qmask = (ins->type == TAG_ALU_4) ? ins->mask : ~0;
+ unsigned qmask = ~0;
/* Handle dot products and things */
if (ins->type == TAG_ALU_4 && !ins->compact_branch) {
- unsigned props = alu_opcode_props[ins->alu.op].props;
+ unsigned props = alu_opcode_props[ins->op].props;
unsigned channel_override = GET_CHANNEL_COUNT(props);
if (channel_override)
qmask = mask_of(channel_override);
+ else
+ qmask = ins->mask;
}
return mir_bytemask_of_read_components_single(ins->swizzle[i], qmask,
};
if (bundle.tag == TAG_ALU_4) {
- assert(OP_IS_MOVE(u->alu.op));
+ assert(OP_IS_MOVE(u->op));
u->unit = UNIT_VMUL;
size_t bytes_emitted = sizeof(uint32_t) + sizeof(midgard_reg_info) + sizeof(midgard_vector_alu);
assert(ins->type == TAG_ALU_4);
- temp = ins->alu.src1;
- ins->alu.src1 = ins->alu.src2;
- ins->alu.src2 = temp;
+ temp = ins->src_types[0];
+ ins->src_types[0] = ins->src_types[1];
+ ins->src_types[1] = temp;
+
+ temp = ins->src_abs[0];
+ ins->src_abs[0] = ins->src_abs[1];
+ ins->src_abs[1] = temp;
+
+ temp = ins->src_neg[0];
+ ins->src_neg[0] = ins->src_neg[1];
+ ins->src_neg[1] = temp;
+
+ temp = ins->src_invert[0];
+ ins->src_invert[0] = ins->src_invert[1];
+ ins->src_invert[1] = temp;
unsigned temp_swizzle[16];
memcpy(temp_swizzle, ins->swizzle[0], sizeof(ins->swizzle[0]));