uint64_t
qpu_merge_inst(uint64_t a, uint64_t b)
{
uint64_t merge = a | b;
bool ok = true;
uint32_t a_sig = QPU_GET_FIELD(a, QPU_SIG);
uint32_t b_sig = QPU_GET_FIELD(b, QPU_SIG);
if (QPU_GET_FIELD(a, QPU_OP_ADD) != QPU_A_NOP &&
QPU_GET_FIELD(b, QPU_OP_ADD) != QPU_A_NOP) {
if (QPU_GET_FIELD(a, QPU_OP_MUL) != QPU_M_NOP ||
QPU_GET_FIELD(b, QPU_OP_MUL) != QPU_M_NOP ||
!(convert_mov(&a) || convert_mov(&b))) {
return 0;
} else {
merge = a | b;
}
}
if (QPU_GET_FIELD(a, QPU_OP_MUL) != QPU_M_NOP &&
QPU_GET_FIELD(b, QPU_OP_MUL) != QPU_M_NOP)
return 0;
if (qpu_num_sf_accesses(a) && qpu_num_sf_accesses(b))
return 0;
if (a_sig == QPU_SIG_LOAD_IMM ||
b_sig == QPU_SIG_LOAD_IMM ||
a_sig == QPU_SIG_SMALL_IMM ||
b_sig == QPU_SIG_SMALL_IMM ||
a_sig == QPU_SIG_BRANCH ||
b_sig == QPU_SIG_BRANCH) {
return 0;
}
ok = ok && merge_fields(&merge, a, b, QPU_SIG_MASK,
QPU_SET_FIELD(QPU_SIG_NONE, QPU_SIG));
/* Misc fields that have to match exactly. */
ok = ok && merge_fields(&merge, a, b, QPU_SF, ~0);
if (!merge_fields(&merge, a, b, QPU_RADDR_A_MASK,
QPU_SET_FIELD(QPU_R_NOP, QPU_RADDR_A))) {
/* Since we tend to use regfile A by default both for register
* allocation and for our special values (uniforms and
* varyings), try swapping uniforms and varyings to regfile B
* to resolve raddr A conflicts.
*/
if (!try_swap_ra_file(&merge, &a, &b) &&
!try_swap_ra_file(&merge, &b, &a)) {
return 0;
}
}
ok = ok && merge_fields(&merge, a, b, QPU_RADDR_B_MASK,
QPU_SET_FIELD(QPU_R_NOP, QPU_RADDR_B));
ok = ok && merge_fields(&merge, a, b, QPU_WADDR_ADD_MASK,
QPU_SET_FIELD(QPU_W_NOP, QPU_WADDR_ADD));
ok = ok && merge_fields(&merge, a, b, QPU_WADDR_MUL_MASK,
QPU_SET_FIELD(QPU_W_NOP, QPU_WADDR_MUL));
/* Allow disagreement on WS (swapping A vs B physical reg file as the
* destination for ADD/MUL) if one of the original instructions
* ignores it (probably because it's just writing to accumulators).
*/
if (qpu_waddr_ignores_ws(QPU_GET_FIELD(a, QPU_WADDR_ADD)) &&
qpu_waddr_ignores_ws(QPU_GET_FIELD(a, QPU_WADDR_MUL))) {
merge = (merge & ~QPU_WS) | (b & QPU_WS);
} else if (qpu_waddr_ignores_ws(QPU_GET_FIELD(b, QPU_WADDR_ADD)) &&
qpu_waddr_ignores_ws(QPU_GET_FIELD(b, QPU_WADDR_MUL))) {
merge = (merge & ~QPU_WS) | (a & QPU_WS);
} else {
if ((a & QPU_WS) != (b & QPU_WS))
return 0;
}
if (!merge_fields(&merge, a, b, QPU_PM, ~0)) {
/* If one instruction has PM bit set and the other not, the
* one without PM shouldn't do packing/unpacking, and we
* have to make sure non-NOP packing/unpacking from PM
* instruction aren't added to it.
*/
uint64_t temp;
/* Let a be the one with PM bit */
if (!(a & QPU_PM)) {
temp = a;
a = b;
b = temp;
}
if ((b & (QPU_PACK_MASK | QPU_UNPACK_MASK)) != 0)
return 0;
if ((a & QPU_PACK_MASK) != 0 &&
QPU_GET_FIELD(b, QPU_OP_MUL) != QPU_M_NOP)
return 0;
if ((a & QPU_UNPACK_MASK) != 0 && reads_r4(b))
return 0;
} else {
/* packing: Make sure that non-NOP packs agree, then deal with
* special-case failing of adding a non-NOP pack to something
* with a NOP pack.
*/
if (!merge_fields(&merge, a, b, QPU_PACK_MASK, 0))
return 0;
bool new_a_pack = (QPU_GET_FIELD(a, QPU_PACK) !=
QPU_GET_FIELD(merge, QPU_PACK));
bool new_b_pack = (QPU_GET_FIELD(b, QPU_PACK) !=
QPU_GET_FIELD(merge, QPU_PACK));
if (!(merge & QPU_PM)) {
/* Make sure we're not going to be putting a new
* a-file packing on either half.
*/
if (new_a_pack && writes_a_file(a))
return 0;
if (new_b_pack && writes_a_file(b))
return 0;
} else {
/* Make sure we're not going to be putting new MUL
* packing oneither half.
*/
if (new_a_pack &&
QPU_GET_FIELD(a, QPU_OP_MUL) != QPU_M_NOP)
return 0;
if (new_b_pack &&
QPU_GET_FIELD(b, QPU_OP_MUL) != QPU_M_NOP)
return 0;
}
/* unpacking: Make sure that non-NOP unpacks agree, then deal
* with special-case failing of adding a non-NOP unpack to
* something with a NOP unpack.
*/
if (!merge_fields(&merge, a, b, QPU_UNPACK_MASK, 0))
return 0;
bool new_a_unpack = (QPU_GET_FIELD(a, QPU_UNPACK) !=
QPU_GET_FIELD(merge, QPU_UNPACK));
bool new_b_unpack = (QPU_GET_FIELD(b, QPU_UNPACK) !=
QPU_GET_FIELD(merge, QPU_UNPACK));
if (!(merge & QPU_PM)) {
/* Make sure we're not going to be putting a new
* a-file packing on either half.
*/
if (new_a_unpack &&
QPU_GET_FIELD(a, QPU_RADDR_A) != QPU_R_NOP)
return 0;
if (new_b_unpack &&
QPU_GET_FIELD(b, QPU_RADDR_A) != QPU_R_NOP)
return 0;
} else {
/* Make sure we're not going to be putting new r4
* unpack on either half.
*/
if (new_a_unpack && reads_r4(a))
return 0;
if (new_b_unpack && reads_r4(b))
return 0;
}
}
if (ok)
return merge;
else
return 0;
}
/**
* Checks for the instruction restrictions from page 37 ("Summary of
* Instruction Restrictions").
*/
void
vc4_qpu_validate(uint64_t *insts, uint32_t num_inst)
{
bool scoreboard_locked = false;
/* We don't want to do validation in release builds, but we want to
* keep compiling the validation code to make sure it doesn't get
* broken.
*/
#ifndef DEBUG
return;
#endif
for (int i = 0; i < num_inst; i++) {
uint64_t inst = insts[i];
if (QPU_GET_FIELD(inst, QPU_SIG) != QPU_SIG_PROG_END) {
if (qpu_inst_is_tlb(inst))
scoreboard_locked = true;
continue;
}
/* "The Thread End instruction must not write to either physical
* regfile A or B."
*/
if (QPU_GET_FIELD(inst, QPU_WADDR_ADD) < 32 ||
QPU_GET_FIELD(inst, QPU_WADDR_MUL) < 32) {
fail_instr(inst, "write to phys reg in thread end");
}
/* Can't trigger an implicit wait on scoreboard in the program
* end instruction.
*/
if (qpu_inst_is_tlb(inst) && !scoreboard_locked)
fail_instr(inst, "implicit sb wait in program end");
/* Two delay slots will be executed. */
assert(i + 2 <= num_inst);
for (int j = i; j < i + 2; j++) {
/* "The last three instructions of any program
* (Thread End plus the following two delay-slot
* instructions) must not do varyings read, uniforms
* read or any kind of VPM, VDR, or VDW read or
* write."
*/
if (writes_reg(insts[j], QPU_W_VPM) ||
reads_reg(insts[j], QPU_R_VARY) ||
reads_reg(insts[j], QPU_R_UNIF) ||
reads_reg(insts[j], QPU_R_VPM)) {
fail_instr(insts[j], "last 3 instructions "
"using fixed functions");
}
/* "The Thread End instruction and the following two
* delay slot instructions must not write or read
* address 14 in either regfile A or B."
*/
if (writes_reg(insts[j], 14) ||
reads_reg(insts[j], 14)) {
fail_instr(insts[j], "last 3 instructions "
"must not use r14");
}
}
/* "The final program instruction (the second delay slot
* instruction) must not do a TLB Z write."
*/
if (writes_reg(insts[i + 2], QPU_W_TLB_Z)) {
fail_instr(insts[i + 2], "final instruction doing "
"Z write");
}
}
/* "A scoreboard wait must not occur in the first two instructions of
* a fragment shader. This is either the explicit Wait for Scoreboard
* signal or an implicit wait with the first tile-buffer read or
* write instruction."
*/
for (int i = 0; i < 2; i++) {
uint64_t inst = insts[i];
if (qpu_inst_is_tlb(inst))
fail_instr(inst, "sb wait in first two insts");
}
/* "If TMU_NOSWAP is written, the write must be three instructions
* before the first TMU write instruction. For example, if
* TMU_NOSWAP is written in the first shader instruction, the first
* TMU write cannot occur before the 4th shader instruction."
*/
int last_tmu_noswap = -10;
for (int i = 0; i < num_inst; i++) {
uint64_t inst = insts[i];
if ((i - last_tmu_noswap) <= 3 &&
(writes_reg(inst, QPU_W_TMU0_S) ||
writes_reg(inst, QPU_W_TMU1_S))) {
fail_instr(inst, "TMU write too soon after TMU_NOSWAP");
}
if (writes_reg(inst, QPU_W_TMU_NOSWAP))
last_tmu_noswap = i;
}
/* "An instruction must not read from a location in physical regfile A
* or B that was written to by the previous instruction."
*/
for (int i = 0; i < num_inst - 1; i++) {
uint64_t inst = insts[i];
uint32_t add_waddr = QPU_GET_FIELD(inst, QPU_WADDR_ADD);
uint32_t mul_waddr = QPU_GET_FIELD(inst, QPU_WADDR_MUL);
uint32_t waddr_a, waddr_b;
if (inst & QPU_WS) {
waddr_b = add_waddr;
waddr_a = mul_waddr;
} else {
waddr_a = add_waddr;
waddr_b = mul_waddr;
}
if ((waddr_a < 32 && reads_a_reg(insts[i + 1], waddr_a)) ||
(waddr_b < 32 && reads_b_reg(insts[i + 1], waddr_b))) {
fail_instr(insts[i + 1],
"Reads physical reg too soon after write");
}
}
/* "After an SFU lookup instruction, accumulator r4 must not be read
* in the following two instructions. Any other instruction that
* results in r4 being written (that is, TMU read, TLB read, SFU
* lookup) cannot occur in the two instructions following an SFU
* lookup."
*/
int last_sfu_inst = -10;
for (int i = 0; i < num_inst - 1; i++) {
uint64_t inst = insts[i];
uint32_t sig = QPU_GET_FIELD(inst, QPU_SIG);
if (i - last_sfu_inst <= 2 &&
(writes_sfu(inst) ||
sig == QPU_SIG_LOAD_TMU0 ||
sig == QPU_SIG_LOAD_TMU1 ||
sig == QPU_SIG_COLOR_LOAD)) {
fail_instr(inst, "R4 write too soon after SFU write");
}
if (writes_sfu(inst))
last_sfu_inst = i;
}
for (int i = 0; i < num_inst - 1; i++) {
uint64_t inst = insts[i];
if (QPU_GET_FIELD(inst, QPU_SIG) == QPU_SIG_SMALL_IMM &&
QPU_GET_FIELD(inst, QPU_SMALL_IMM) >=
QPU_SMALL_IMM_MUL_ROT) {
uint32_t mux_a = QPU_GET_FIELD(inst, QPU_MUL_A);
uint32_t mux_b = QPU_GET_FIELD(inst, QPU_MUL_B);
/* "The full horizontal vector rotate is only
* available when both of the mul ALU input arguments
* are taken from accumulators r0-r3."
*/
if (mux_a > QPU_MUX_R3 || mux_b > QPU_MUX_R3) {
fail_instr(inst,
"MUL rotate using non-accumulator "
"input");
}
if (QPU_GET_FIELD(inst, QPU_SMALL_IMM) ==
QPU_SMALL_IMM_MUL_ROT) {
/* "An instruction that does a vector rotate
* by r5 must not immediately follow an
* instruction that writes to r5."
*/
if (writes_reg(insts[i - 1], QPU_W_ACC5)) {
fail_instr(inst,
"vector rotate by r5 "
"immediately after r5 write");
}
}
/* "An instruction that does a vector rotate must not
* immediately follow an instruction that writes to the
* accumulator that is being rotated."
*/
if (writes_reg(insts[i - 1], QPU_W_ACC0 + mux_a) ||
writes_reg(insts[i - 1], QPU_W_ACC0 + mux_b)) {
fail_instr(inst,
"vector rotate of value "
"written in previous instruction");
}
}
}
/* "An instruction that does a vector rotate must not immediately
* follow an instruction that writes to the accumulator that is being
* rotated.
*
* XXX: TODO.
*/
/* "After an instruction that does a TLB Z write, the multisample mask
* must not be read as an instruction input argument in the following
* two instruction. The TLB Z write instruction can, however, be
* followed immediately by a TLB color write."
*/
for (int i = 0; i < num_inst - 1; i++) {
uint64_t inst = insts[i];
if (writes_reg(inst, QPU_W_TLB_Z) &&
(reads_a_reg(insts[i + 1], QPU_R_MS_REV_FLAGS) ||
reads_a_reg(insts[i + 2], QPU_R_MS_REV_FLAGS))) {
fail_instr(inst, "TLB Z write followed by MS mask read");
}
}
/*
* "A single instruction can only perform a maximum of one of the
* following closely coupled peripheral accesses in a single
* instruction: TMU write, TMU read, TLB write, TLB read, TLB
* combined color read and write, SFU write, Mutex read or Semaphore
* access."
*/
for (int i = 0; i < num_inst - 1; i++) {
uint64_t inst = insts[i];
if (qpu_num_sf_accesses(inst) > 1)
fail_instr(inst, "Single instruction writes SFU twice");
}
/* "The uniform base pointer can be written (from SIMD element 0) by
* the processor to reset the stream, there must be at least two
* nonuniform-accessing instructions following a pointer change
* before uniforms can be accessed once more."
*/
int last_unif_pointer_update = -3;
for (int i = 0; i < num_inst; i++) {
uint64_t inst = insts[i];
uint32_t waddr_add = QPU_GET_FIELD(inst, QPU_WADDR_ADD);
uint32_t waddr_mul = QPU_GET_FIELD(inst, QPU_WADDR_MUL);
if (reads_reg(inst, QPU_R_UNIF) &&
i - last_unif_pointer_update <= 2) {
fail_instr(inst,
"uniform read too soon after pointer update");
}
if (waddr_add == QPU_W_UNIFORMS_ADDRESS ||
waddr_mul == QPU_W_UNIFORMS_ADDRESS)
last_unif_pointer_update = i;
}
}