*/
#include "compiler.h"
+#include "bi_print.h"
#include "panfrost/util/lcra.h"
#include "util/u_memory.h"
if (ins->dest && (ins->dest < l->node_count)) {
for (unsigned i = 1; i < l->node_count; ++i) {
if (live[i])
- lcra_add_node_interference(l, ins->dest, ins->writemask, i, live[i]);
+ lcra_add_node_interference(l, ins->dest, bi_writemask(ins), i, live[i]);
}
}
unsigned node_count = bi_max_temp(ctx);
struct lcra_state *l =
- lcra_alloc_equations(node_count, 1, 8, 16, 1);
+ lcra_alloc_equations(node_count, 1);
l->class_start[BI_REG_CLASS_WORK] = 0;
l->class_size[BI_REG_CLASS_WORK] = 64 * 4; /* R0 - R63, all 32-bit */
continue;
l->class[dest] = BI_REG_CLASS_WORK;
- lcra_set_alignment(l, dest, 2); /* 2^2 = 4 */
+ lcra_set_alignment(l, dest, 2, 16); /* 2^2 = 4 */
lcra_restrict_range(l, dest, 4);
}
if (solution < 0)
return index;
- solution += offset;
-
assert((solution & 0x3) == 0);
unsigned reg = solution / 4;
+ reg += offset;
+
return BIR_INDEX_REGISTER | reg;
}
/* TODO: Do we do anything here? */
} else {
/* Use the swizzle as component select */
+ unsigned components = bi_get_component_count(ins, src);
+
nir_alu_type T = ins->src_types[src];
unsigned size = nir_alu_type_get_type_size(T);
- assert(size <= 32); /* TODO: 64-bit */
- unsigned comps_per_reg = 32 / size;
- unsigned components = bi_get_component_count(ins, src);
+
+ /* TODO: 64-bit? */
+ unsigned components_per_word = MAX2(32 / size, 1);
for (unsigned i = 0; i < components; ++i) {
- unsigned off = ins->swizzle[src][i] / comps_per_reg;
- off *= 4; /* 32-bit registers */
+ unsigned off = ins->swizzle[src][i] / components_per_word;
/* We can't cross register boundaries in a swizzle */
if (i == 0)
else
assert(off == offset);
- ins->swizzle[src][i] %= comps_per_reg;
+ ins->swizzle[src][i] %= components_per_word;
}
}
if (ins->dest >= l->node_count)
return;
- bool vector = (bi_class_props[ins->type] & BI_VECTOR);
- unsigned offset = 0;
-
- if (!vector) {
- /* Look at the writemask to get an offset, specifically the
- * trailing zeros */
-
- unsigned tz = __builtin_ctz(ins->writemask);
-
- /* Recall writemask is one bit per byte, so tz is in eytes */
- unsigned regs = tz / 4;
- offset = regs * 4;
-
- /* Adjust writemask to compensate */
- ins->writemask >>= offset;
- }
-
- ins->dest = bi_reg_from_index(l, ins->dest, offset);
-
+ ins->dest = bi_reg_from_index(l, ins->dest, ins->dest_offset);
+ ins->dest_offset = 0;
}
static void