From a1b3c48d2b23d6eaeb4529d3e1183d2648731bf8 Mon Sep 17 00:00:00 2001 From: Richard Henderson Date: Wed, 22 Jun 2016 15:46:09 -0700 Subject: [PATCH] tcg: Compress liveness data to 16 bits This reduces both memory usage and per-insn cacheline usage during code generation. Reviewed-by: Aurelien Jarno Signed-off-by: Richard Henderson --- tcg/tcg.c | 58 ++++++++++++++++++++++------------------------------------ tcg/tcg.h | 16 ++++++++++------ 2 files changed, 32 insertions(+), 42 deletions(-) diff --git a/tcg/tcg.c b/tcg/tcg.c index 0c46c43cfa..4aa1933a3e 100644 --- a/tcg/tcg.c +++ b/tcg/tcg.c @@ -1341,7 +1341,7 @@ static inline void tcg_la_bb_end(TCGContext *s, uint8_t *dead_temps, } } -/* Liveness analysis : update the opc_dead_args array to tell if a +/* Liveness analysis : update the opc_arg_life array to tell if a given input arguments is dead. Instructions updating dead temporaries are removed. */ static void tcg_liveness_analysis(TCGContext *s) @@ -1350,9 +1350,8 @@ static void tcg_liveness_analysis(TCGContext *s) int oi, oi_prev, nb_ops; nb_ops = s->gen_next_op_idx; - s->op_dead_args = tcg_malloc(nb_ops * sizeof(uint16_t)); - s->op_sync_args = tcg_malloc(nb_ops * sizeof(uint8_t)); - + s->op_arg_life = tcg_malloc(nb_ops * sizeof(TCGLifeData)); + dead_temps = tcg_malloc(s->nb_temps); mem_temps = tcg_malloc(s->nb_temps); tcg_la_func_end(s, dead_temps, mem_temps); @@ -1361,8 +1360,7 @@ static void tcg_liveness_analysis(TCGContext *s) int i, nb_iargs, nb_oargs; TCGOpcode opc_new, opc_new2; bool have_opc_new2; - uint16_t dead_args; - uint8_t sync_args; + TCGLifeData arg_life = 0; TCGArg arg; TCGOp * const op = &s->gen_op_buf[oi]; @@ -1394,15 +1392,13 @@ static void tcg_liveness_analysis(TCGContext *s) do_not_remove_call: /* output args are dead */ - dead_args = 0; - sync_args = 0; for (i = 0; i < nb_oargs; i++) { arg = args[i]; if (dead_temps[arg]) { - dead_args |= (1 << i); + arg_life |= DEAD_ARG << i; } if (mem_temps[arg]) { - sync_args |= (1 << i); + arg_life |= SYNC_ARG << i; } dead_temps[arg] = 1; mem_temps[arg] = 0; @@ -1423,7 +1419,7 @@ static void tcg_liveness_analysis(TCGContext *s) arg = args[i]; if (arg != TCG_CALL_DUMMY_ARG) { if (dead_temps[arg]) { - dead_args |= (1 << i); + arg_life |= DEAD_ARG << i; } } } @@ -1432,8 +1428,6 @@ static void tcg_liveness_analysis(TCGContext *s) arg = args[i]; dead_temps[arg] = 0; } - s->op_dead_args[oi] = dead_args; - s->op_sync_args[oi] = sync_args; } } break; @@ -1544,15 +1538,13 @@ static void tcg_liveness_analysis(TCGContext *s) } else { do_not_remove: /* output args are dead */ - dead_args = 0; - sync_args = 0; for (i = 0; i < nb_oargs; i++) { arg = args[i]; if (dead_temps[arg]) { - dead_args |= (1 << i); + arg_life |= DEAD_ARG << i; } if (mem_temps[arg]) { - sync_args |= (1 << i); + arg_life |= SYNC_ARG << i; } dead_temps[arg] = 1; mem_temps[arg] = 0; @@ -1570,7 +1562,7 @@ static void tcg_liveness_analysis(TCGContext *s) for (i = nb_oargs; i < nb_oargs + nb_iargs; i++) { arg = args[i]; if (dead_temps[arg]) { - dead_args |= (1 << i); + arg_life |= DEAD_ARG << i; } } /* input arguments are live for preceding opcodes */ @@ -1578,11 +1570,10 @@ static void tcg_liveness_analysis(TCGContext *s) arg = args[i]; dead_temps[arg] = 0; } - s->op_dead_args[oi] = dead_args; - s->op_sync_args[oi] = sync_args; } break; } + s->op_arg_life[oi] = arg_life; } } #else @@ -1921,11 +1912,11 @@ static void tcg_reg_alloc_bb_end(TCGContext *s, TCGRegSet allocated_regs) save_globals(s, allocated_regs); } -#define IS_DEAD_ARG(n) ((dead_args >> (n)) & 1) -#define NEED_SYNC_ARG(n) ((sync_args >> (n)) & 1) +#define IS_DEAD_ARG(n) (arg_life & (DEAD_ARG << (n))) +#define NEED_SYNC_ARG(n) (arg_life & (SYNC_ARG << (n))) static void tcg_reg_alloc_movi(TCGContext *s, const TCGArg *args, - uint16_t dead_args, uint8_t sync_args) + TCGLifeData arg_life) { TCGTemp *ots; tcg_target_ulong val; @@ -1954,8 +1945,7 @@ static void tcg_reg_alloc_movi(TCGContext *s, const TCGArg *args, } static void tcg_reg_alloc_mov(TCGContext *s, const TCGOpDef *def, - const TCGArg *args, uint16_t dead_args, - uint8_t sync_args) + const TCGArg *args, TCGLifeData arg_life) { TCGRegSet allocated_regs; TCGTemp *ts, *ots; @@ -2040,8 +2030,7 @@ static void tcg_reg_alloc_mov(TCGContext *s, const TCGOpDef *def, static void tcg_reg_alloc_op(TCGContext *s, const TCGOpDef *def, TCGOpcode opc, - const TCGArg *args, uint16_t dead_args, - uint8_t sync_args) + const TCGArg *args, TCGLifeData arg_life) { TCGRegSet allocated_regs; int i, k, nb_iargs, nb_oargs; @@ -2206,8 +2195,7 @@ static void tcg_reg_alloc_op(TCGContext *s, #endif static void tcg_reg_alloc_call(TCGContext *s, int nb_oargs, int nb_iargs, - const TCGArg * const args, uint16_t dead_args, - uint8_t sync_args) + const TCGArg * const args, TCGLifeData arg_life) { int flags, nb_regs, i; TCGReg reg; @@ -2427,8 +2415,7 @@ int tcg_gen_code(TCGContext *s, TranslationBlock *tb) TCGArg * const args = &s->gen_opparam_buf[op->args]; TCGOpcode opc = op->opc; const TCGOpDef *def = &tcg_op_defs[opc]; - uint16_t dead_args = s->op_dead_args[oi]; - uint8_t sync_args = s->op_sync_args[oi]; + TCGLifeData arg_life = s->op_arg_life[oi]; oi_next = op->next; #ifdef CONFIG_PROFILER @@ -2438,11 +2425,11 @@ int tcg_gen_code(TCGContext *s, TranslationBlock *tb) switch (opc) { case INDEX_op_mov_i32: case INDEX_op_mov_i64: - tcg_reg_alloc_mov(s, def, args, dead_args, sync_args); + tcg_reg_alloc_mov(s, def, args, arg_life); break; case INDEX_op_movi_i32: case INDEX_op_movi_i64: - tcg_reg_alloc_movi(s, args, dead_args, sync_args); + tcg_reg_alloc_movi(s, args, arg_life); break; case INDEX_op_insn_start: if (num_insns >= 0) { @@ -2467,8 +2454,7 @@ int tcg_gen_code(TCGContext *s, TranslationBlock *tb) tcg_out_label(s, arg_label(args[0]), s->code_ptr); break; case INDEX_op_call: - tcg_reg_alloc_call(s, op->callo, op->calli, args, - dead_args, sync_args); + tcg_reg_alloc_call(s, op->callo, op->calli, args, arg_life); break; default: /* Sanity check that we've not introduced any unhandled opcodes. */ @@ -2478,7 +2464,7 @@ int tcg_gen_code(TCGContext *s, TranslationBlock *tb) /* Note: in order to speed up the code, it would be much faster to have specialized register allocator functions for some common argument patterns */ - tcg_reg_alloc_op(s, def, opc, args, dead_args, sync_args); + tcg_reg_alloc_op(s, def, opc, args, arg_life); break; } #ifdef CONFIG_DEBUG_TCG diff --git a/tcg/tcg.h b/tcg/tcg.h index 6046dcdc89..7c0a138152 100644 --- a/tcg/tcg.h +++ b/tcg/tcg.h @@ -575,6 +575,14 @@ typedef struct TCGTempSet { unsigned long l[BITS_TO_LONGS(TCG_MAX_TEMPS)]; } TCGTempSet; +/* While we limit helpers to 6 arguments, for 32-bit hosts, with padding, + this imples a max of 6*2 (64-bit in) + 2 (64-bit out) = 14 operands. + There are never more than 2 outputs, which means that we can store all + dead + sync data within 16 bits. */ +#define DEAD_ARG 4 +#define SYNC_ARG 1 +typedef uint16_t TCGLifeData; + typedef struct TCGOp { TCGOpcode opc : 8; @@ -608,12 +616,8 @@ struct TCGContext { uintptr_t *tb_jmp_target_addr; /* tb->jmp_target_addr if !USE_DIRECT_JUMP */ /* liveness analysis */ - uint16_t *op_dead_args; /* for each operation, each bit tells if the - corresponding argument is dead */ - uint8_t *op_sync_args; /* for each operation, each bit tells if the - corresponding output argument needs to be - sync to memory. */ - + TCGLifeData *op_arg_life; + TCGRegSet reserved_regs; intptr_t current_frame_offset; intptr_t frame_start; -- 2.11.0