提交 a3251186 编写于 作者: R Richard Henderson

target-i386: optimize flags checking after sub using CC_SRCT

After a comparison or subtraction, the original value of the LHS will
currently be reconstructed using an addition.  However, in most cases
it is already available: store it in a temp-local variable and save 1
or 2 TCG ops (2 if the result of the addition needs to be extended).

The temp-local can be declared dead as soon as the cc_op changes again,
or also before the translation block ends because gen_prepare_cc will
always make a copy before returning it.  All this magic, plus copy
propagation and dead-code elimination, ensures that the temp local will
(almost) never be spilled.

Example (cmp $0x21,%rax + jbe):

 Before                                     After
----------------------------------------------------------------------------
 movi_i64 tmp1,$0x21                        movi_i64 tmp1,$0x21
 movi_i64 cc_src,$0x21                      movi_i64 cc_src,$0x21
 sub_i64 cc_dst,rax,tmp1                    sub_i64 cc_dst,rax,tmp1
 add_i64 tmp7,cc_dst,cc_src
 movi_i32 cc_op,$0x11                       movi_i32 cc_op,$0x11
 brcond_i64 tmp7,cc_src,leu,$0x0            discard loc11
                                            brcond_i64 rax,cc_src,leu,$0x0

 Before                                     After
----------------------------------------------------------------------------
  mov    (%r14),%rbp                        mov    (%r14),%rbp
  mov    %rbp,%rbx                          mov    %rbp,%rbx
  sub    $0x21,%rbx                         sub    $0x21,%rbx
  lea    0x21(%rbx),%r12
  movl   $0x11,0xa0(%r14)                   movl   $0x11,0xa0(%r14)
  movq   $0x21,0x90(%r14)                   movq   $0x21,0x90(%r14)
  mov    %rbx,0x98(%r14)                    mov    %rbx,0x98(%r14)
  cmp    $0x21,%r12                     |   cmp    $0x21,%rbp
  jbe    ...                                jbe    ...
Signed-off-by: NPaolo Bonzini <pbonzini@redhat.com>
Signed-off-by: NRichard Henderson <rth@twiddle.net>
上级 891a5133
......@@ -60,7 +60,8 @@
/* global register indexes */
static TCGv_ptr cpu_env;
static TCGv cpu_A0, cpu_cc_src, cpu_cc_dst;
static TCGv cpu_A0;
static TCGv cpu_cc_src, cpu_cc_dst, cpu_cc_srcT;
static TCGv_i32 cpu_cc_op;
static TCGv cpu_regs[CPU_NB_REGS];
/* local temps */
......@@ -185,8 +186,9 @@ enum {
};
enum {
USES_CC_DST = 1,
USES_CC_SRC = 2,
USES_CC_DST = 1,
USES_CC_SRC = 2,
USES_CC_SRCT = 4,
};
/* Bit set if the global variable is live after setting CC_OP to X. */
......@@ -196,7 +198,7 @@ static const uint8_t cc_op_live[CC_OP_NB] = {
[CC_OP_MULB ... CC_OP_MULQ] = USES_CC_DST | USES_CC_SRC,
[CC_OP_ADDB ... CC_OP_ADDQ] = USES_CC_DST | USES_CC_SRC,
[CC_OP_ADCB ... CC_OP_ADCQ] = USES_CC_DST | USES_CC_SRC,
[CC_OP_SUBB ... CC_OP_SUBQ] = USES_CC_DST | USES_CC_SRC,
[CC_OP_SUBB ... CC_OP_SUBQ] = USES_CC_DST | USES_CC_SRC | USES_CC_SRCT,
[CC_OP_SBBB ... CC_OP_SBBQ] = USES_CC_DST | USES_CC_SRC,
[CC_OP_LOGICB ... CC_OP_LOGICQ] = USES_CC_DST,
[CC_OP_INCB ... CC_OP_INCQ] = USES_CC_DST | USES_CC_SRC,
......@@ -221,6 +223,9 @@ static void set_cc_op(DisasContext *s, CCOp op)
if (dead & USES_CC_SRC) {
tcg_gen_discard_tl(cpu_cc_src);
}
if (dead & USES_CC_SRCT) {
tcg_gen_discard_tl(cpu_cc_srcT);
}
s->cc_op = op;
/* The DYNAMIC setting is translator only, and should never be
......@@ -869,8 +874,9 @@ static inline void gen_op_testl_T0_T1_cc(void)
static void gen_op_update_neg_cc(void)
{
tcg_gen_neg_tl(cpu_cc_src, cpu_T[0]);
tcg_gen_mov_tl(cpu_cc_dst, cpu_T[0]);
tcg_gen_neg_tl(cpu_cc_src, cpu_T[0]);
tcg_gen_movi_tl(cpu_cc_srcT, 0);
}
/* compute all eflags to cc_src */
......@@ -903,12 +909,12 @@ static CCPrepare gen_prepare_eflags_c(DisasContext *s, TCGv reg)
switch (s->cc_op) {
case CC_OP_SUBB ... CC_OP_SUBQ:
/* (DATA_TYPE)(CC_DST + CC_SRC) < (DATA_TYPE)CC_SRC */
/* (DATA_TYPE)CC_SRCT < (DATA_TYPE)CC_SRC */
size = s->cc_op - CC_OP_SUBB;
t1 = gen_ext_tl(cpu_tmp0, cpu_cc_src, size, false);
/* If no temporary was used, be careful not to alias t1 and t0. */
t0 = TCGV_EQUAL(t1, cpu_cc_src) ? cpu_tmp0 : reg;
tcg_gen_add_tl(t0, cpu_cc_dst, cpu_cc_src);
tcg_gen_mov_tl(t0, cpu_cc_srcT);
gen_extu(size, t0);
goto add_sub;
......@@ -1052,7 +1058,7 @@ static CCPrepare gen_prepare_cc(DisasContext *s, int b, TCGv reg)
size = s->cc_op - CC_OP_SUBB;
switch (jcc_op) {
case JCC_BE:
tcg_gen_add_tl(cpu_tmp4, cpu_cc_dst, cpu_cc_src);
tcg_gen_mov_tl(cpu_tmp4, cpu_cc_srcT);
gen_extu(size, cpu_tmp4);
t0 = gen_ext_tl(cpu_tmp0, cpu_cc_src, size, false);
cc = (CCPrepare) { .cond = TCG_COND_LEU, .reg = cpu_tmp4,
......@@ -1065,7 +1071,7 @@ static CCPrepare gen_prepare_cc(DisasContext *s, int b, TCGv reg)
case JCC_LE:
cond = TCG_COND_LE;
fast_jcc_l:
tcg_gen_add_tl(cpu_tmp4, cpu_cc_dst, cpu_cc_src);
tcg_gen_mov_tl(cpu_tmp4, cpu_cc_srcT);
gen_exts(size, cpu_tmp4);
t0 = gen_ext_tl(cpu_tmp0, cpu_cc_src, size, true);
cc = (CCPrepare) { .cond = cond, .reg = cpu_tmp4,
......@@ -1421,6 +1427,10 @@ static void gen_op(DisasContext *s1, int op, int ot, int d)
set_cc_op(s1, CC_OP_DYNAMIC);
break;
case OP_SBBL:
/*
* No need to store cpu_cc_srcT, because it is used only
* when the cc_op is known.
*/
gen_compute_eflags_c(s1, cpu_tmp4);
tcg_gen_sub_tl(cpu_T[0], cpu_T[0], cpu_T[1]);
tcg_gen_sub_tl(cpu_T[0], cpu_T[0], cpu_tmp4);
......@@ -1445,6 +1455,7 @@ static void gen_op(DisasContext *s1, int op, int ot, int d)
set_cc_op(s1, CC_OP_ADDB + ot);
break;
case OP_SUBL:
tcg_gen_mov_tl(cpu_cc_srcT, cpu_T[0]);
tcg_gen_sub_tl(cpu_T[0], cpu_T[0], cpu_T[1]);
if (d != OR_TMP0)
gen_op_mov_reg_T0(ot, d);
......@@ -1483,6 +1494,7 @@ static void gen_op(DisasContext *s1, int op, int ot, int d)
break;
case OP_CMPL:
tcg_gen_mov_tl(cpu_cc_src, cpu_T[1]);
tcg_gen_mov_tl(cpu_cc_srcT, cpu_T[0]);
tcg_gen_sub_tl(cpu_cc_dst, cpu_T[0], cpu_T[1]);
set_cc_op(s1, CC_OP_SUBB + ot);
break;
......@@ -2799,8 +2811,9 @@ static void gen_eob(DisasContext *s)
direct call to the next block may occur */
static void gen_jmp_tb(DisasContext *s, target_ulong eip, int tb_num)
{
gen_update_cc_op(s);
set_cc_op(s, CC_OP_DYNAMIC);
if (s->jmp_opt) {
gen_update_cc_op(s);
gen_goto_tb(s, tb_num, eip);
s->is_jmp = DISAS_TB_JUMP;
} else {
......@@ -5017,9 +5030,10 @@ static target_ulong disas_insn(CPUX86State *env, DisasContext *s,
rm = 0; /* avoid warning */
}
label1 = gen_new_label();
tcg_gen_sub_tl(t2, cpu_regs[R_EAX], t0);
tcg_gen_mov_tl(t2, cpu_regs[R_EAX]);
gen_extu(ot, t0);
gen_extu(ot, t2);
tcg_gen_brcondi_tl(TCG_COND_EQ, t2, 0, label1);
tcg_gen_brcond_tl(TCG_COND_EQ, t2, t0, label1);
label2 = gen_new_label();
if (mod == 3) {
gen_op_mov_reg_v(ot, R_EAX, t0);
......@@ -5038,7 +5052,8 @@ static target_ulong disas_insn(CPUX86State *env, DisasContext *s,
}
gen_set_label(label2);
tcg_gen_mov_tl(cpu_cc_src, t0);
tcg_gen_mov_tl(cpu_cc_dst, t2);
tcg_gen_mov_tl(cpu_cc_srcT, t2);
tcg_gen_sub_tl(cpu_cc_dst, t2, t0);
set_cc_op(s, CC_OP_SUBB + ot);
tcg_temp_free(t0);
tcg_temp_free(t1);
......@@ -7746,10 +7761,10 @@ void optimize_flags_init(void)
cpu_env = tcg_global_reg_new_ptr(TCG_AREG0, "env");
cpu_cc_op = tcg_global_mem_new_i32(TCG_AREG0,
offsetof(CPUX86State, cc_op), "cc_op");
cpu_cc_src = tcg_global_mem_new(TCG_AREG0, offsetof(CPUX86State, cc_src),
"cc_src");
cpu_cc_dst = tcg_global_mem_new(TCG_AREG0, offsetof(CPUX86State, cc_dst),
"cc_dst");
cpu_cc_src = tcg_global_mem_new(TCG_AREG0, offsetof(CPUX86State, cc_src),
"cc_src");
#ifdef TARGET_X86_64
cpu_regs[R_EAX] = tcg_global_mem_new_i64(TCG_AREG0,
......@@ -7885,6 +7900,7 @@ static inline void gen_intermediate_code_internal(CPUX86State *env,
cpu_tmp5 = tcg_temp_new();
cpu_ptr0 = tcg_temp_new_ptr();
cpu_ptr1 = tcg_temp_new_ptr();
cpu_cc_srcT = tcg_temp_local_new();
gen_opc_end = tcg_ctx.gen_opc_buf + OPC_MAX_SIZE;
......
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册