pan/bi: Interpret register allocation results

Once LCRA has run, we have a map from IR indices to byte offsets into
the register file, so we need to "install" these results, rewriting the
IR to use native registers and fixing up writemasks/swizzles to
substitute vectorization for adjacent registers (for LCRA, we're
modeling in terms of real vectors).

Signed-off-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Tested-by: Marge Bot <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4158>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4158>

diff --git a/src/panfrost/bifrost/bi_ra.c b/src/panfrost/bifrost/bi_ra.c
index b30f0e7aceb3180ed5c930e0f717686a546492c8..8c0fa92a12e7434bec97e63b41b5296c3aa8dead 100644 (file)
--- a/src/panfrost/bifrost/bi_ra.c
+++ b/src/panfrost/bifrost/bi_ra.c
@@ -89,6 +89,99 @@ bi_allocate_registers(bi_context *ctx, bool *success)
         return l;
 }
 
+static unsigned
+bi_reg_from_index(struct lcra_state *l, unsigned index, unsigned offset)
+{
+        /* Did we run RA for this index at all */
+        if (index >= l->node_count)
+                return index;
+
+        /* LCRA didn't bother solving this index (how lazy!) */
+        signed solution = l->solutions[index];
+        if (solution < 0)
+                return index;
+
+        solution += offset;
+
+        assert((solution & 0x3) == 0);
+        unsigned reg = solution / 4;
+        return BIR_INDEX_REGISTER | reg;
+}
+
+static void
+bi_adjust_src_ra(bi_instruction *ins, struct lcra_state *l, unsigned src)
+{
+        if (ins->src[src] >= l->node_count)
+                return;
+
+        bool vector = (bi_class_props[ins->type] & BI_VECTOR);
+        unsigned offset = 0;
+
+        if (vector) {
+                /* TODO: Do we do anything here? */
+        } else {
+                /* Use the swizzle as component select */
+                nir_alu_type T = ins->src_types[src];
+                unsigned size = nir_alu_type_get_type_size(T);
+                unsigned bytes = (MAX2(size, 8) / 8);
+                unsigned comps_per_reg = 4 / bytes;
+                unsigned components = bi_get_component_count(ins);
+
+                for (unsigned i = 0; i < components; ++i) {
+                        unsigned off = ins->swizzle[src][i] / comps_per_reg;
+                        off *= 4; /* 32-bit registers */
+
+                        /* We can't cross register boundaries in a swizzle */
+                        if (i == 0)
+                                offset = off;
+                        else
+                                assert(off == offset);
+
+                        ins->swizzle[src][i] %= comps_per_reg;
+                }
+        }
+
+        ins->src[src] = bi_reg_from_index(l, ins->src[src], offset);
+}
+
+static void
+bi_adjust_dest_ra(bi_instruction *ins, struct lcra_state *l)
+{
+        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 bytes */
+                unsigned regs = tz / 4;
+                offset = regs * 4;
+
+                /* Adjust writemask to compensate */
+                ins->writemask >>= offset;
+        }
+
+        ins->dest = bi_reg_from_index(l, ins->dest, offset);
+
+}
+
+static void
+bi_install_registers(bi_context *ctx, struct lcra_state *l)
+{
+        bi_foreach_instr_global(ctx, ins) {
+                bi_adjust_dest_ra(ins, l);
+
+                bi_foreach_src(ins, s)
+                        bi_adjust_src_ra(ins, l, s);
+        }
+}
+
 void
 bi_register_allocate(bi_context *ctx)
 {
@@ -108,5 +201,7 @@ bi_register_allocate(bi_context *ctx)
                 assert(success);
         } while(!success);
 
+        bi_install_registers(ctx, l);
+
         lcra_free(l);
 }

diff --git a/src/panfrost/bifrost/bifrost_compile.c b/src/panfrost/bifrost/bifrost_compile.c
index b9cb89d961bc850c0a8ef921a3f4ad2a65f76d32..f25f742440997b451fa9d607a4f884cfead2c2ff 100644 (file)
--- a/src/panfrost/bifrost/bifrost_compile.c
+++ b/src/panfrost/bifrost/bifrost_compile.c
@@ -864,6 +864,7 @@ bifrost_compile_shader_nir(nir_shader *nir, panfrost_program *program, unsigned
         bi_print_shader(ctx, stdout);
         bi_schedule(ctx);
         bi_register_allocate(ctx);
+        bi_print_shader(ctx, stdout);
 
         ralloc_free(ctx);
 }