broadcom/compiler: convert mul to add when needed to allow merge

V3D 7.x added 'mov' opcodes to the ADD alu, so now it is possible to
move these to the ADD alu to facilitate merging them with other MUL
instructions.

Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/25450>

diff --git a/src/broadcom/compiler/qpu_schedule.c b/src/broadcom/compiler/qpu_schedule.c
index f0d59973e66893b53210489b0149a5c40ca9ca68..a4983d880ed8f583a1f628b02ee7827cf66c164d 100644 (file)
--- a/src/broadcom/compiler/qpu_schedule.c
+++ b/src/broadcom/compiler/qpu_schedule.c
@@ -1086,6 +1086,57 @@ qpu_convert_add_to_mul(struct v3d_qpu_instr *inst)
         inst->alu.add.b.unpack = V3D_QPU_UNPACK_NONE;
 }
 
+static bool
+can_do_mul_as_add(const struct v3d_device_info *devinfo, enum v3d_qpu_mul_op op)
+{
+        switch (op) {
+        case V3D_QPU_M_MOV:
+        case V3D_QPU_M_FMOV:
+                return devinfo->ver >= 71;
+        default:
+                return false;
+        }
+}
+
+static enum v3d_qpu_mul_op
+mul_op_as_add_op(enum v3d_qpu_mul_op op)
+{
+        switch (op) {
+        case V3D_QPU_M_MOV:
+                return V3D_QPU_A_MOV;
+        case V3D_QPU_M_FMOV:
+                return V3D_QPU_A_FMOV;
+        default:
+                unreachable("unexpected mov opcode");
+        }
+}
+
+static void
+qpu_convert_mul_to_add(struct v3d_qpu_instr *inst)
+{
+        STATIC_ASSERT(sizeof(inst->alu.add) == sizeof(inst->alu.mul));
+        assert(inst->alu.mul.op != V3D_QPU_M_NOP);
+        assert(inst->alu.add.op == V3D_QPU_A_NOP);
+
+        memcpy(&inst->alu.add, &inst->alu.mul, sizeof(inst->alu.add));
+        inst->alu.add.op = mul_op_as_add_op(inst->alu.mul.op);
+        inst->alu.mul.op = V3D_QPU_M_NOP;
+
+        inst->flags.ac = inst->flags.mc;
+        inst->flags.apf = inst->flags.mpf;
+        inst->flags.auf = inst->flags.muf;
+        inst->flags.mc = V3D_QPU_COND_NONE;
+        inst->flags.mpf = V3D_QPU_PF_NONE;
+        inst->flags.muf = V3D_QPU_UF_NONE;
+
+        inst->alu.add.output_pack = inst->alu.mul.output_pack;
+        inst->alu.add.a.unpack = inst->alu.mul.a.unpack;
+        inst->alu.add.b.unpack = inst->alu.mul.b.unpack;
+        inst->alu.mul.output_pack = V3D_QPU_PACK_NONE;
+        inst->alu.mul.a.unpack = V3D_QPU_UNPACK_NONE;
+        inst->alu.mul.b.unpack = V3D_QPU_UNPACK_NONE;
+}
+
 static bool
 qpu_merge_inst(const struct v3d_device_info *devinfo,
                struct v3d_qpu_instr *result,
@@ -1151,17 +1202,52 @@ qpu_merge_inst(const struct v3d_device_info *devinfo,
                 }
         }
 
+        struct v3d_qpu_instr add_inst;
         if (b->alu.mul.op != V3D_QPU_M_NOP) {
-                if (a->alu.mul.op != V3D_QPU_M_NOP)
-                        return false;
-                merge.alu.mul = b->alu.mul;
+                if (a->alu.mul.op == V3D_QPU_M_NOP) {
+                        merge.alu.mul = b->alu.mul;
+
+                        merge.flags.mc = b->flags.mc;
+                        merge.flags.mpf = b->flags.mpf;
+                        merge.flags.muf = b->flags.muf;
+
+                        mul_instr = b;
+                        add_instr = a;
+                }
+                /* If a's mul op is used but its add op is not, then see if we
+                 * can convert either a's mul op or b's mul op to an add op
+                 * so we can merge.
+                 */
+                else if (a->alu.add.op == V3D_QPU_A_NOP &&
+                         can_do_mul_as_add(devinfo, b->alu.mul.op)) {
+                        add_inst = *b;
+                        qpu_convert_mul_to_add(&add_inst);
 
-                merge.flags.mc = b->flags.mc;
-                merge.flags.mpf = b->flags.mpf;
-                merge.flags.muf = b->flags.muf;
+                        merge.alu.add = add_inst.alu.add;
 
-                mul_instr = b;
-                add_instr = a;
+                        merge.flags.ac = b->flags.mc;
+                        merge.flags.apf = b->flags.mpf;
+                        merge.flags.auf = b->flags.muf;
+
+                        mul_instr = a;
+                        add_instr = &add_inst;
+                } else if (a->alu.add.op == V3D_QPU_A_NOP &&
+                           can_do_mul_as_add(devinfo, a->alu.mul.op)) {
+                        add_inst = *a;
+                        qpu_convert_mul_to_add(&add_inst);
+
+                        merge = add_inst;
+                        merge.alu.mul = b->alu.mul;
+
+                        merge.flags.mc = b->flags.mc;
+                        merge.flags.mpf = b->flags.mpf;
+                        merge.flags.muf = b->flags.muf;
+
+                        mul_instr = b;
+                        add_instr = &add_inst;
+                } else {
+                        return false;
+                }
         }
 
         /* V3D 4.x and earlier use muxes to select the inputs for the ALUs and