[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
[PATCH v2 8/8] mips: Fill delay slots in jit_bger, jit_bgei, jit_bltr, j
From: |
Paul Cercueil |
Subject: |
[PATCH v2 8/8] mips: Fill delay slots in jit_bger, jit_bgei, jit_bltr, jit_blti |
Date: |
Mon, 9 Jan 2023 23:04:11 +0000 |
Fill the delay slots with the opcode that precedes the branch opcode, if
possible.
To simplify things, the code has also been factorized into a single
function.
Signed-off-by: Paul Cercueil <paul@crapouillou.net>
---
lib/jit_mips-cpu.c | 233 ++++++++++++++++-----------------------------
lib/jit_mips.c | 16 ++--
2 files changed, 89 insertions(+), 160 deletions(-)
diff --git a/lib/jit_mips-cpu.c b/lib/jit_mips-cpu.c
index a9266ae..9a62776 100644
--- a/lib/jit_mips-cpu.c
+++ b/lib/jit_mips-cpu.c
@@ -673,14 +673,10 @@ static void
_gti_u(jit_state_t*,jit_int32_t,jit_int32_t,jit_word_t);
static void _ner(jit_state_t*,jit_int32_t,jit_int32_t,jit_int32_t);
#define nei(r0,r1,i0) _nei(_jit,r0,r1,i0)
static void _nei(jit_state_t*,jit_int32_t,jit_int32_t,jit_word_t);
-#define bltr(i0,r0,r1) _bltr(_jit,i0,r0,r1)
-static jit_word_t _bltr(jit_state_t*,jit_word_t,jit_int32_t,jit_int32_t);
-#define bltr_u(i0,r0,r1) _bltr_u(_jit,i0,r0,r1)
-static jit_word_t _bltr_u(jit_state_t*,jit_word_t,jit_int32_t,jit_int32_t);
-#define blti(i0,r0,i1) _blti(_jit,i0,r0,i1)
-static jit_word_t _blti(jit_state_t*,jit_word_t,jit_int32_t,jit_word_t);
-#define blti_u(i0,r0,i1) _blti_u(_jit,i0,r0,i1)
-static jit_word_t _blti_u(jit_state_t*,jit_word_t,jit_int32_t,jit_word_t);
+#define bltr(i0,r0,r1,no_flag) bger(i0,r1,r0,no_flag)
+#define bltr_u(i0,r0,r1,no_flag) bger_u(i0,r1,r0,no_flag)
+#define blti(i0,r0,i1,no_flag) _bgei(_jit,i0,r0,i1,0,1,no_flag)
+#define blti_u(i0,r0,i1,no_flag) _bgei(_jit,i0,r0,i1,1,1,no_flag)
#define bler(i0,r0,r1,no_flag) _bgtr(_jit,i0,r1,r0,0,1,no_flag)
#define bler_u(i0,r0,r1,no_flag) _bgtr(_jit,i0,r1,r0,1,1,no_flag)
#define blei(i0,r0,i1,no_flag) _bgti(_jit,i0,r0,i1,0,1,no_flag)
@@ -689,14 +685,13 @@ static jit_word_t
_blti_u(jit_state_t*,jit_word_t,jit_int32_t,jit_word_t);
static jit_word_t
_beqr(jit_state_t*,jit_word_t,jit_int32_t,jit_int32_t,jit_bool_t);
#define beqi(i0,r0,i1,no_flag) _beqi(_jit,i0,r0,i1,no_flag)
static jit_word_t
_beqi(jit_state_t*,jit_word_t,jit_int32_t,jit_word_t,jit_bool_t);
-#define bger(i0,r0,r1) _bger(_jit,i0,r0,r1)
-static jit_word_t _bger(jit_state_t*,jit_word_t,jit_int32_t,jit_int32_t);
-#define bger_u(i0,r0,r1) _bger_u(_jit,i0,r0,r1)
-static jit_word_t _bger_u(jit_state_t*,jit_word_t,jit_int32_t,jit_int32_t);
-#define bgei(i0,r0,i1) _bgei(_jit,i0,r0,i1)
-static jit_word_t _bgei(jit_state_t*,jit_word_t,jit_int32_t,jit_word_t);
-#define bgei_u(i0,r0,i1) _bgei_u(_jit,i0,r0,i1)
-static jit_word_t _bgei_u(jit_state_t*,jit_word_t,jit_int32_t,jit_word_t);
+#define bger(i0,r0,r1,no_flag) _bger(_jit,i0,r0,r1,0,no_flag)
+#define bger_u(i0,r0,r1,no_flag) _bger(_jit,i0,r0,r1,1,no_flag)
+static jit_word_t
_bger(jit_state_t*,jit_word_t,jit_int32_t,jit_int32_t,jit_bool_t,jit_bool_t);
+#define bgei(i0,r0,i1,no_flag) _bgei(_jit,i0,r0,i1,0,0,no_flag)
+#define bgei_u(i0,r0,i1,no_flag) _bgei(_jit,i0,r0,i1,1,0,no_flag)
+static jit_word_t _bgei(jit_state_t*,jit_word_t,jit_int32_t,jit_word_t,
+ jit_bool_t,jit_bool_t,jit_bool_t);
#define bgtr(i0,r0,r1,no_flag) _bgtr(_jit,i0,r0,r1,0,0,no_flag)
#define bgtr_u(i0,r0,r1,no_flag) _bgtr(_jit,i0,r0,r1,1,0,no_flag)
static jit_word_t _bgtr(jit_state_t*,jit_word_t,jit_int32_t,jit_int32_t,
@@ -2149,91 +2144,6 @@ _nei(jit_state_t *_jit, jit_int32_t r0, jit_int32_t r1,
jit_word_t i0)
SLTU(r0, _ZERO_REGNO, r1);
}
-static jit_word_t
-_bltr(jit_state_t *_jit, jit_word_t i0, jit_int32_t r0, jit_int32_t r1)
-{
- jit_word_t w;
- jit_int32_t reg;
-
- reg = jit_get_reg(jit_class_gpr);
- SLT(rn(reg), r0, r1);
- w = _jit->pc.w;
- BNE(rn(reg), _ZERO_REGNO, ((i0 - w) >> 2) - 1);
- NOP(1);
- jit_unget_reg(reg);
-
- return (w);
-}
-
-static jit_word_t
-_bltr_u(jit_state_t *_jit, jit_word_t i0, jit_int32_t r0, jit_int32_t r1)
-{
- jit_word_t w;
- jit_int32_t reg;
-
- reg = jit_get_reg(jit_class_gpr|jit_class_nospill);
- SLTU(rn(reg), r0, r1);
- w = _jit->pc.w;
- BNE(rn(reg), _ZERO_REGNO, ((i0 - w) >> 2) - 1);
- NOP(1);
- jit_unget_reg(reg);
-
- return (w);
-}
-
-static jit_word_t
-_blti(jit_state_t *_jit, jit_word_t i0, jit_int32_t r0, jit_word_t i1)
-{
- jit_word_t w;
- jit_word_t d;
- jit_int32_t reg;
- jit_bool_t zero_p;
-
- if (!(zero_p = i1 == 0))
- reg = jit_get_reg(jit_class_gpr|jit_class_nospill);
- if (can_sign_extend_short_p(i1)) {
- if (!zero_p)
- SLTI(rn(reg), r0, i1);
- w = _jit->pc.w;
- d = ((i0 - w) >> 2) - 1;
- if (!zero_p)
- BNE(rn(reg), _ZERO_REGNO, d);
- else
- BLTZ(r0, d);
- NOP(1);
- }
- else {
- movi(rn(reg), i1);
- w = bltr(i0, r0, rn(reg));
- }
- if (!zero_p)
- jit_unget_reg(reg);
-
- return (w);
-}
-
-static jit_word_t
-_blti_u(jit_state_t *_jit, jit_word_t i0, jit_int32_t r0, jit_word_t i1)
-{
- jit_word_t w;
- jit_int32_t reg;
-
- reg = jit_get_reg(jit_class_gpr|jit_class_nospill);
- if (can_sign_extend_short_p(i1)) {
- SLTIU(rn(reg), r0, i1);
- w = _jit->pc.w;
- BNE(rn(reg), _ZERO_REGNO, ((i0 - w) >> 2) - 1);
- NOP(1);
- }
- else {
- movi(rn(reg), i1);
- w = bltr_u(i0, r0, rn(reg));
- }
- jit_unget_reg(reg);
-
- return (w);
-}
-
static jit_word_t
_beqr(jit_state_t *_jit, jit_word_t i0, jit_int32_t r0, jit_int32_t r1,
jit_bool_t no_flag)
{
@@ -2296,86 +2206,105 @@ _beqi(jit_state_t *_jit, jit_word_t i0, jit_int32_t
r0, jit_word_t i1, jit_bool_
}
static jit_word_t
-_bger(jit_state_t *_jit, jit_word_t i0, jit_int32_t r0, jit_int32_t r1)
+_bger(jit_state_t *_jit, jit_word_t i0, jit_int32_t r0, jit_int32_t r1,
+ jit_bool_t sltu, jit_bool_t no_flag)
{
jit_word_t w;
- jit_int32_t reg;
+ jit_int32_t reg, prev, offset;
+ jit_bool_t swap_ds;
+
+ offset = ((jit_word_t)_jit->pc.ui - (jit_word_t)_jit->code.ptr) /
sizeof(jit_instr_t);
+ swap_ds = no_flag
+ && (offset < 2 || !has_delay_slot((jit_instr_t)*(_jit->pc.ui - 2)))
+ && !op_writes_register((jit_instr_t)*(_jit->pc.ui - 1), r0)
+ && !op_writes_register((jit_instr_t)*(_jit->pc.ui - 1), r1);
reg = jit_get_reg(jit_class_gpr|jit_class_nospill);
- SLT(rn(reg), r0, r1);
- w = _jit->pc.w;
- BEQ(rn(reg), _ZERO_REGNO, ((i0 - w) >> 2) - 1);
- NOP(1);
- jit_unget_reg(reg);
- return (w);
-}
+ if (swap_ds)
+ prev = *--_jit->pc.ui;
-static jit_word_t
-_bger_u(jit_state_t *_jit, jit_word_t i0, jit_int32_t r0, jit_int32_t r1)
-{
- jit_word_t w;
- jit_int32_t reg;
+ if (sltu)
+ SLTU(rn(reg), r0, r1);
+ else
+ SLT(rn(reg), r0, r1);
- reg = jit_get_reg(jit_class_gpr|jit_class_nospill);
- SLTU(rn(reg), r0, r1);
w = _jit->pc.w;
BEQ(rn(reg), _ZERO_REGNO, ((i0 - w) >> 2) - 1);
- NOP(1);
+
+ if (swap_ds)
+ ii(prev);
+ else
+ NOP(1);
+
jit_unget_reg(reg);
return (w);
}
static jit_word_t
-_bgei(jit_state_t *_jit, jit_word_t i0, jit_int32_t r0, jit_word_t i1)
+_bgei(jit_state_t *_jit, jit_word_t i0, jit_int32_t r0, jit_word_t i1,
+ jit_bool_t sltiu, jit_bool_t bne, jit_bool_t no_flag)
{
jit_word_t w;
jit_word_t d;
- jit_int32_t reg;
- jit_bool_t zero_p;
+ jit_int32_t reg, prev, offset;
+ jit_bool_t zero_p, swap_ds;
+
+ offset = ((jit_word_t)_jit->pc.ui - (jit_word_t)_jit->code.ptr) /
sizeof(jit_instr_t);
+ swap_ds = no_flag
+ && (offset < 2 || !has_delay_slot((jit_instr_t)*(_jit->pc.ui - 2)))
+ && !op_writes_register((jit_instr_t)*(_jit->pc.ui - 1), r0);
+ zero_p = !sltiu && i1 == 0;
- if (!(zero_p = i1 == 0))
+ if (!zero_p)
reg = jit_get_reg(jit_class_gpr|jit_class_nospill);
+ if (swap_ds)
+ prev = *--_jit->pc.ui;
+
if (can_sign_extend_short_p(i1)) {
- if (!zero_p)
- SLTI(rn(reg), r0, i1);
+ if (!zero_p) {
+ if (sltiu)
+ SLTIU(rn(reg), r0, i1);
+ else
+ SLTI(rn(reg), r0, i1);
+ }
+
w = _jit->pc.w;
d = ((i0 - w) >> 2) - 1;
- if (!zero_p)
- BEQ(rn(reg), _ZERO_REGNO, d);
- else
- BGEZ(r0, d);
- NOP(1);
+ if (bne) {
+ if (!zero_p)
+ BNE(rn(reg), _ZERO_REGNO, d);
+ else
+ BLTZ(r0, d);
+ } else {
+ if (!zero_p)
+ BEQ(rn(reg), _ZERO_REGNO, d);
+ else
+ BGEZ(r0, d);
+ }
}
else {
movi(rn(reg), i1);
- w = bger(i0, r0, rn(reg));
- }
- if (!zero_p)
- jit_unget_reg(reg);
-
- return (w);
-}
-
-static jit_word_t
-_bgei_u(jit_state_t *_jit, jit_word_t i0, jit_int32_t r0, jit_word_t i1)
-{
- jit_word_t w;
- jit_int32_t reg;
+ if (sltiu)
+ SLTU(rn(reg), r0, rn(reg));
+ else
+ SLT(rn(reg), r0, rn(reg));
- reg = jit_get_reg(jit_class_gpr|jit_class_nospill);
- if (can_sign_extend_short_p(i1)) {
- SLTIU(rn(reg), r0, i1);
w = _jit->pc.w;
- BEQ(rn(reg), _ZERO_REGNO, ((i0 - w) >> 2) - 1);
- NOP(1);
- }
- else {
- movi(rn(reg), i1);
- w = bger_u(i0, r0, rn(reg));
+ if (bne)
+ BNE(rn(reg), _ZERO_REGNO, ((i0 - w) >> 2) - 1);
+ else
+ BEQ(rn(reg), _ZERO_REGNO, ((i0 - w) >> 2) - 1);
}
- jit_unget_reg(reg);
+
+ if (swap_ds)
+ ii(prev);
+ else
+ NOP(1);
+
+ if (!zero_p)
+ jit_unget_reg(reg);
return (w);
}
diff --git a/lib/jit_mips.c b/lib/jit_mips.c
index 0b7f544..4170e57 100644
--- a/lib/jit_mips.c
+++ b/lib/jit_mips.c
@@ -1511,20 +1511,20 @@ _emit_code(jit_state_t *_jit)
case_rrw(gt, _u);
case_rrr(ne,);
case_rrw(ne,);
- case_brr(blt,);
- case_brw(blt,);
- case_brr(blt, _u);
- case_brw(blt, _u);
+ case_brrn(blt,, no_flag);
+ case_brwn(blt,, no_flag);
+ case_brrn(blt, _u, no_flag);
+ case_brwn(blt, _u, no_flag);
case_brrn(ble,, no_flag);
case_brwn(ble,, no_flag);
case_brrn(ble, _u, no_flag);
case_brwn(ble, _u, no_flag);
case_brrn(beq,, no_flag);
case_brwn(beq,, no_flag);
- case_brr(bge,);
- case_brw(bge,);
- case_brr(bge, _u);
- case_brw(bge, _u);
+ case_brrn(bge,, no_flag);
+ case_brwn(bge,, no_flag);
+ case_brrn(bge, _u, no_flag);
+ case_brwn(bge, _u, no_flag);
case_brrn(bgt,, no_flag);
case_brwn(bgt,, no_flag);
case_brrn(bgt, _u, no_flag);
--
2.39.0
- [PATCH v2 0/9] mips: Fill delay slots v2, Paul Cercueil, 2023/01/09
- [PATCH v2 1/8] mips: Optimize jit_eqr / jit_eqi, Paul Cercueil, 2023/01/09
- [PATCH v2 2/8] mips: Fill delay slots of JR opcodes in jit_jmpr, Paul Cercueil, 2023/01/09
- [PATCH v2 3/8] mips: Fill delay slots of JALR opcodes in jit_callr, Paul Cercueil, 2023/01/09
- [PATCH v2 4/8] mips: Fill delay slots of J in jit_jmpi, Paul Cercueil, 2023/01/09
- [PATCH v2 5/8] mips: Fill delay slots in jit_beqr / jit_beqi, Paul Cercueil, 2023/01/09
- [PATCH v2 6/8] mips: Fill delay slots in jit_bner / jit_bnei, Paul Cercueil, 2023/01/09
- [PATCH v2 7/8] mips: Fill delay slots in jit_bgtr, jit_bgti, jit_bler, jit_blei, Paul Cercueil, 2023/01/09
- [PATCH v2 8/8] mips: Fill delay slots in jit_bger, jit_bgei, jit_bltr, jit_blti,
Paul Cercueil <=
- Re: [PATCH v2 0/9] mips: Fill delay slots v2, Paulo César Pereira de Andrade, 2023/01/10