@@ -3954,6 +3954,100 @@ static void tcg_reg_alloc_op(TCGContext *s, const TCGOp *op)
}
}
+static void tcg_reg_alloc_dup2(TCGContext *s, const TCGOp *op)
+{
+ const TCGLifeData arg_life = op->life;
+ TCGTemp *ots, *itsl, *itsh;
+ TCGType vtype = TCGOP_VECL(op) + TCG_TYPE_V64;
+
+ /* This opcode is only valid for 32-bit hosts, for 64-bit elements. */
+ tcg_debug_assert(TCG_TARGET_REG_BITS == 32);
+ tcg_debug_assert(TCGOP_VECE(op) == MO_64);
+
+ ots = arg_temp(op->args[0]);
+ itsl = arg_temp(op->args[1]);
+ itsh = arg_temp(op->args[2]);
+
+ /* ENV should not be modified. */
+ tcg_debug_assert(!temp_readonly(ots));
+
+ /* Allocate the output register now. */
+ if (ots->val_type != TEMP_VAL_REG) {
+ TCGRegSet allocated_regs = s->reserved_regs;
+ TCGRegSet dup_out_regs =
+ tcg_op_defs[INDEX_op_dup_vec].args_ct[0].regs;
+
+ /* Make sure to not spill the input registers. */
+ if (!IS_DEAD_ARG(1) && itsl->val_type == TEMP_VAL_REG) {
+ tcg_regset_set_reg(allocated_regs, itsl->reg);
+ }
+ if (!IS_DEAD_ARG(2) && itsh->val_type == TEMP_VAL_REG) {
+ tcg_regset_set_reg(allocated_regs, itsh->reg);
+ }
+
+ ots->reg = tcg_reg_alloc(s, dup_out_regs, allocated_regs,
+ op->output_pref[0], ots->indirect_base);
+ ots->val_type = TEMP_VAL_REG;
+ ots->mem_coherent = 0;
+ s->reg_to_temp[ots->reg] = ots;
+ }
+
+ /* Promote dup2 of immediates to dupi_vec. */
+ if (itsl->val_type == TEMP_VAL_CONST && itsh->val_type == TEMP_VAL_CONST) {
+ uint64_t val = deposit64(itsl->val, 32, 32, itsh->val);
+ MemOp vece = MO_64;
+
+ if (val == dup_const(MO_8, val)) {
+ vece = MO_8;
+ } else if (val == dup_const(MO_16, val)) {
+ vece = MO_16;
+ } else if (val == dup_const(MO_32, val)) {
+ vece = MO_32;
+ }
+
+ tcg_out_dupi_vec(s, vtype, vece, ots->reg, val);
+ goto done;
+ }
+
+ /* If the two inputs form one 64-bit value, try dupm_vec. */
+ if (itsl + 1 == itsh &&
+ itsl->base_type == TCG_TYPE_I64 &&
+ itsh->base_type == TCG_TYPE_I64) {
+ if (!itsl->mem_coherent) {
+ temp_sync(s, itsl, s->reserved_regs, 0, 0);
+ }
+ if (!itsl->mem_coherent) {
+ temp_sync(s, itsl, s->reserved_regs, 0, 0);
+ }
+#ifdef HOST_WORDS_BIGENDIAN
+ TCGTemp *its = itsh;
+#else
+ TCGTemp *its = itsl;
+#endif
+ if (tcg_out_dupm_vec(s, vtype, MO_64, ots->reg,
+ its->mem_base->reg, its->mem_offset)) {
+ goto done;
+ }
+ }
+
+ /* Fall back to generic expansion. */
+ tcg_reg_alloc_op(s, op);
+ return;
+
+ done:
+ if (IS_DEAD_ARG(1)) {
+ temp_dead(s, itsl);
+ }
+ if (IS_DEAD_ARG(2)) {
+ temp_dead(s, itsh);
+ }
+ if (NEED_SYNC_ARG(0)) {
+ temp_sync(s, ots, s->reserved_regs, 0, IS_DEAD_ARG(0));
+ } else if (IS_DEAD_ARG(0)) {
+ temp_dead(s, ots);
+ }
+}
+
#ifdef TCG_TARGET_STACK_GROWSUP
#define STACK_DIR(x) (-(x))
#else
@@ -4345,6 +4439,9 @@ int tcg_gen_code(TCGContext *s, TranslationBlock *tb)
case INDEX_op_dup_vec:
tcg_reg_alloc_dup(s, op);
break;
+ case INDEX_op_dup2_vec:
+ tcg_reg_alloc_dup2(s, op);
+ break;
case INDEX_op_insn_start:
if (num_insns >= 0) {
size_t off = tcg_current_code_size(s);
There are several ways we can expand a vector dup of a 64-bit element on a 32-bit host. Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- tcg/tcg.c | 97 +++++++++++++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 97 insertions(+) -- 2.25.1