diff options
Diffstat (limited to 'tcg/tcg-op-vec.c')
-rw-r--r-- | tcg/tcg-op-vec.c | 102 |
1 files changed, 102 insertions, 0 deletions
diff --git a/tcg/tcg-op-vec.c b/tcg/tcg-op-vec.c index cfb18682b1..914fe42b1e 100644 --- a/tcg/tcg-op-vec.c +++ b/tcg/tcg-op-vec.c @@ -34,6 +34,90 @@ extern TCGv_i32 TCGV_HIGH_link_error(TCGv_i64); #define TCGV_HIGH TCGV_HIGH_link_error #endif +/* + * Vector optional opcode tracking. + * Except for the basic logical operations (and, or, xor), and + * data movement (mov, ld, st, dupi), many vector opcodes are + * optional and may not be supported on the host. Thank Intel + * for the irregularity in their instruction set. + * + * The gvec expanders allow custom vector operations to be composed, + * generally via the .fniv callback in the GVecGen* structures. At + * the same time, in deciding whether to use this hook we need to + * know if the host supports the required operations. This is + * presented as an array of opcodes, terminated by 0. Each opcode + * is assumed to be expanded with the given VECE. + * + * For debugging, we want to validate this array. Therefore, when + * tcg_ctx->vec_opt_opc is non-NULL, the tcg_gen_*_vec expanders + * will validate that their opcode is present in the list. + */ +#ifdef CONFIG_DEBUG_TCG +void tcg_assert_listed_vecop(TCGOpcode op) +{ + const TCGOpcode *p = tcg_ctx->vecop_list; + if (p) { + for (; *p; ++p) { + if (*p == op) { + return; + } + } + g_assert_not_reached(); + } +} +#endif + +bool tcg_can_emit_vecop_list(const TCGOpcode *list, + TCGType type, unsigned vece) +{ + if (list == NULL) { + return true; + } + + for (; *list; ++list) { + TCGOpcode opc = *list; + +#ifdef CONFIG_DEBUG_TCG + switch (opc) { + case INDEX_op_and_vec: + case INDEX_op_or_vec: + case INDEX_op_xor_vec: + case INDEX_op_mov_vec: + case INDEX_op_dup_vec: + case INDEX_op_dupi_vec: + case INDEX_op_dup2_vec: + case INDEX_op_ld_vec: + case INDEX_op_st_vec: + /* These opcodes are mandatory and should not be listed. */ + g_assert_not_reached(); + default: + break; + } +#endif + + if (tcg_can_emit_vec_op(opc, type, vece)) { + continue; + } + + /* + * The opcode list is created by front ends based on what they + * actually invoke. We must mirror the logic in the routines + * below for generic expansions using other opcodes. + */ + switch (opc) { + case INDEX_op_neg_vec: + if (tcg_can_emit_vec_op(INDEX_op_sub_vec, type, vece)) { + continue; + } + break; + default: + break; + } + return false; + } + return true; +} + void vec_gen_2(TCGOpcode opc, TCGType type, unsigned vece, TCGArg r, TCGArg a) { TCGOp *op = tcg_emit_op(opc); @@ -296,11 +380,14 @@ static bool do_op2(unsigned vece, TCGv_vec r, TCGv_vec a, TCGOpcode opc) int can; tcg_debug_assert(at->base_type >= type); + tcg_assert_listed_vecop(opc); can = tcg_can_emit_vec_op(opc, type, vece); if (can > 0) { vec_gen_2(opc, type, vece, ri, ai); } else if (can < 0) { + const TCGOpcode *hold_list = tcg_swap_vecop_list(NULL); tcg_expand_vec_op(opc, type, vece, ri, ai); + tcg_swap_vecop_list(hold_list); } else { return false; } @@ -318,11 +405,17 @@ void tcg_gen_not_vec(unsigned vece, TCGv_vec r, TCGv_vec a) void tcg_gen_neg_vec(unsigned vece, TCGv_vec r, TCGv_vec a) { + const TCGOpcode *hold_list; + + tcg_assert_listed_vecop(INDEX_op_neg_vec); + hold_list = tcg_swap_vecop_list(NULL); + if (!TCG_TARGET_HAS_neg_vec || !do_op2(vece, r, a, INDEX_op_neg_vec)) { TCGv_vec t = tcg_const_zeros_vec_matching(r); tcg_gen_sub_vec(vece, r, t, a); tcg_temp_free_vec(t); } + tcg_swap_vecop_list(hold_list); } static void do_shifti(TCGOpcode opc, unsigned vece, @@ -337,6 +430,7 @@ static void do_shifti(TCGOpcode opc, unsigned vece, tcg_debug_assert(at->base_type == type); tcg_debug_assert(i >= 0 && i < (8 << vece)); + tcg_assert_listed_vecop(opc); if (i == 0) { tcg_gen_mov_vec(r, a); @@ -350,8 +444,10 @@ static void do_shifti(TCGOpcode opc, unsigned vece, /* We leave the choice of expansion via scalar or vector shift to the target. Often, but not always, dupi can feed a vector shift easier than a scalar. */ + const TCGOpcode *hold_list = tcg_swap_vecop_list(NULL); tcg_debug_assert(can < 0); tcg_expand_vec_op(opc, type, vece, ri, ai, i); + tcg_swap_vecop_list(hold_list); } } @@ -384,12 +480,15 @@ void tcg_gen_cmp_vec(TCGCond cond, unsigned vece, tcg_debug_assert(at->base_type >= type); tcg_debug_assert(bt->base_type >= type); + tcg_assert_listed_vecop(INDEX_op_cmp_vec); can = tcg_can_emit_vec_op(INDEX_op_cmp_vec, type, vece); if (can > 0) { vec_gen_4(INDEX_op_cmp_vec, type, vece, ri, ai, bi, cond); } else { + const TCGOpcode *hold_list = tcg_swap_vecop_list(NULL); tcg_debug_assert(can < 0); tcg_expand_vec_op(INDEX_op_cmp_vec, type, vece, ri, ai, bi, cond); + tcg_swap_vecop_list(hold_list); } } @@ -407,12 +506,15 @@ static void do_op3(unsigned vece, TCGv_vec r, TCGv_vec a, tcg_debug_assert(at->base_type >= type); tcg_debug_assert(bt->base_type >= type); + tcg_assert_listed_vecop(opc); can = tcg_can_emit_vec_op(opc, type, vece); if (can > 0) { vec_gen_3(opc, type, vece, ri, ai, bi); } else { + const TCGOpcode *hold_list = tcg_swap_vecop_list(NULL); tcg_debug_assert(can < 0); tcg_expand_vec_op(opc, type, vece, ri, ai, bi); + tcg_swap_vecop_list(hold_list); } } |