[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
[PATCH v2 7/8] mips: Fill delay slots in jit_bgtr, jit_bgti, jit_bler, j
|
From: |
Paul Cercueil |
|
Subject: |
[PATCH v2 7/8] mips: Fill delay slots in jit_bgtr, jit_bgti, jit_bler, jit_blei |
|
Date: |
Mon, 9 Jan 2023 23:04:10 +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 | 214 +++++++++++++++++----------------------------
lib/jit_mips.c | 16 ++--
2 files changed, 87 insertions(+), 143 deletions(-)
diff --git a/lib/jit_mips-cpu.c b/lib/jit_mips-cpu.c
index 984a6f9..a9266ae 100644
--- a/lib/jit_mips-cpu.c
+++ b/lib/jit_mips-cpu.c
@@ -681,14 +681,10 @@ static jit_word_t
_bltr_u(jit_state_t*,jit_word_t,jit_int32_t,jit_int32_t);
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 bler(i0,r0,r1) _bler(_jit,i0,r0,r1)
-static jit_word_t _bler(jit_state_t*,jit_word_t,jit_int32_t,jit_int32_t);
-#define bler_u(i0,r0,r1) _bler_u(_jit,i0,r0,r1)
-static jit_word_t _bler_u(jit_state_t*,jit_word_t,jit_int32_t,jit_int32_t);
-#define blei(i0,r0,i1) _blei(_jit,i0,r0,i1)
-static jit_word_t _blei(jit_state_t*,jit_word_t,jit_int32_t,jit_word_t);
-#define blei_u(i0,r0,i1) _blei_u(_jit,i0,r0,i1)
-static jit_word_t _blei_u(jit_state_t*,jit_word_t,jit_int32_t,jit_word_t);
+#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)
+#define blei_u(i0,r0,i1,no_flag) _bgti(_jit,i0,r0,i1,1,1,no_flag)
#define beqr(i0,r0,r1,no_flag) _beqr(_jit,i0,r0,r1,no_flag)
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)
@@ -701,14 +697,14 @@ static jit_word_t
_bger_u(jit_state_t*,jit_word_t,jit_int32_t,jit_int32_t);
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 bgtr(i0,r0,r1) _bgtr(_jit,i0,r0,r1)
-static jit_word_t _bgtr(jit_state_t*,jit_word_t,jit_int32_t,jit_int32_t);
-#define bgtr_u(i0,r0,r1) _bgtr_u(_jit,i0,r0,r1)
-static jit_word_t _bgtr_u(jit_state_t*,jit_word_t,jit_int32_t,jit_int32_t);
-#define bgti(i0,r0,i1) _bgti(_jit,i0,r0,i1)
-static jit_word_t _bgti(jit_state_t*,jit_word_t,jit_int32_t,jit_word_t);
-#define bgti_u(i0,r0,i1) _bgti_u(_jit,i0,r0,i1)
-static jit_word_t _bgti_u(jit_state_t*,jit_word_t,jit_int32_t,jit_word_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,
+ jit_bool_t,jit_bool_t,jit_bool_t);
+#define bgti(i0,r0,i1,no_flag) _bgti(_jit,i0,r0,i1,0,0,no_flag)
+#define bgti_u(i0,r0,i1,no_flag) _bgti(_jit,i0,r0,i1,1,0,no_flag)
+static jit_word_t _bgti(jit_state_t*,jit_word_t,jit_int32_t,jit_word_t,
+ jit_bool_t,jit_bool_t,jit_bool_t);
#define bner(i0,r0,r1,no_flag) _bner(_jit,i0,r0,r1,no_flag)
static jit_word_t
_bner(jit_state_t*,jit_word_t,jit_int32_t,jit_int32_t,jit_bool_t);
#define bnei(i0,r0,i1,no_flag) _bnei(_jit,i0,r0,i1,no_flag)
@@ -2238,80 +2234,6 @@ _blti_u(jit_state_t *_jit, jit_word_t i0, jit_int32_t
r0, jit_word_t i1)
return (w);
}
-static jit_word_t
-_bler(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);
- SLT(rn(reg), r1, r0);
- w = _jit->pc.w;
- BEQ(rn(reg), _ZERO_REGNO, ((i0 - w) >> 2) - 1);
- NOP(1);
- jit_unget_reg(reg);
-
- return (w);
-}
-
-static jit_word_t
-_bler_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), r1, r0);
- w = _jit->pc.w;
- BEQ(rn(reg), _ZERO_REGNO, ((i0 - w) >> 2) - 1);
- NOP(1);
- jit_unget_reg(reg);
-
- return (w);
-}
-
-static jit_word_t
-_blei(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 (i1 == 0) {
- w = _jit->pc.w;
- BLEZ(r0, ((i0 - w) >> 2) - 1);
- NOP(1);
- }
- else {
- reg = jit_get_reg(jit_class_gpr|jit_class_nospill);
- movi(rn(reg), i1);
- w = bler(i0, r0, rn(reg));
- jit_unget_reg(reg);
- }
-
- return (w);
-}
-
-static jit_word_t
-_blei_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 (i1 == 0) {
- w = _jit->pc.w;
- BEQ(r0, _ZERO_REGNO, ((i0 - w) >> 2) - 1);
- NOP(1);
- }
- else {
- reg = jit_get_reg(jit_class_gpr|jit_class_nospill);
- movi(rn(reg), i1);
- w = bler_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)
{
@@ -2459,76 +2381,98 @@ _bgei_u(jit_state_t *_jit, jit_word_t i0, jit_int32_t
r0, jit_word_t i1)
}
static jit_word_t
-_bgtr(jit_state_t *_jit, jit_word_t i0, jit_int32_t r0, jit_int32_t r1)
+_bgtr(jit_state_t *_jit, jit_word_t i0, jit_int32_t r0, jit_int32_t r1,
+ jit_bool_t sltu, jit_bool_t inv, 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), r1, r0);
- w = _jit->pc.w;
- BNE(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
-_bgtr_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), r1, r0);
+ else
+ SLT(rn(reg), r1, r0);
- reg = jit_get_reg(jit_class_gpr|jit_class_nospill);
- SLTU(rn(reg), r1, r0);
w = _jit->pc.w;
- BNE(rn(reg), _ZERO_REGNO, ((i0 - w) >> 2) - 1);
- NOP(1);
+ if (inv)
+ BEQ(rn(reg), _ZERO_REGNO, ((i0 - w) >> 2) - 1);
+ else
+ BNE(rn(reg), _ZERO_REGNO, ((i0 - w) >> 2) - 1);
+
+ if (swap_ds)
+ ii(prev);
+ else
+ NOP(1);
+
jit_unget_reg(reg);
return (w);
}
static jit_word_t
-_bgti(jit_state_t *_jit, jit_word_t i0, jit_int32_t r0, jit_word_t i1)
+_bgti(jit_state_t *_jit, jit_word_t i0, jit_int32_t r0, jit_word_t i1,
+ jit_bool_t sltiu, jit_bool_t inv, 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);
+
+ if (swap_ds)
+ prev = *--_jit->pc.ui;
if (i1 == 0) {
- w = _jit->pc.w;
- BGTZ(r0, ((i0 - w) >> 2) - 1);
- NOP(1);
+ w = _jit->pc.w;
+ if (inv) {
+ if (sltiu)
+ BEQ(r0, _ZERO_REGNO, ((i0 - w) >> 2) - 1);
+ else
+ BLEZ(r0, ((i0 - w) >> 2) - 1);
+ } else {
+ if (sltiu)
+ BNE(r0, _ZERO_REGNO, ((i0 - w) >> 2) - 1);
+ else
+ BGTZ(r0, ((i0 - w) >> 2) - 1);
+ }
}
else {
- reg = jit_get_reg(jit_class_gpr|jit_class_nospill);
- movi(rn(reg), i1);
- w = bgtr(i0, r0, rn(reg));
- jit_unget_reg(reg);
- }
+ reg = jit_get_reg(jit_class_gpr|jit_class_nospill);
- return (w);
-}
+ movi(rn(reg), i1);
+ if (sltiu)
+ SLTU(rn(reg), rn(reg), r0);
+ else
+ SLT(rn(reg), rn(reg), r0);
-static jit_word_t
-_bgti_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;
+ w = _jit->pc.w;
+ if (inv)
+ BEQ(rn(reg), _ZERO_REGNO, ((i0 - w) >> 2) - 1);
+ else
+ BNE(rn(reg), _ZERO_REGNO, ((i0 - w) >> 2) - 1);
- if (i1 == 0) {
- w = _jit->pc.w;
- BNE(r0, _ZERO_REGNO, ((i0 - w) >> 2) - 1);
- NOP(1);
- }
- else {
- reg = jit_get_reg(jit_class_gpr|jit_class_nospill);
- movi(rn(reg), i1);
- w = bgtr_u(i0, r0, rn(reg));
- jit_unget_reg(reg);
+ jit_unget_reg(reg);
}
+ if (swap_ds)
+ ii(prev);
+ else
+ NOP(1);
+
return (w);
}
diff --git a/lib/jit_mips.c b/lib/jit_mips.c
index 67df648..0b7f544 100644
--- a/lib/jit_mips.c
+++ b/lib/jit_mips.c
@@ -1515,20 +1515,20 @@ _emit_code(jit_state_t *_jit)
case_brw(blt,);
case_brr(blt, _u);
case_brw(blt, _u);
- case_brr(ble,);
- case_brw(ble,);
- case_brr(ble, _u);
- case_brw(ble, _u);
+ 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_brr(bgt,);
- case_brw(bgt,);
- case_brr(bgt, _u);
- case_brw(bgt, _u);
+ case_brrn(bgt,, no_flag);
+ case_brwn(bgt,, no_flag);
+ case_brrn(bgt, _u, no_flag);
+ case_brwn(bgt, _u, no_flag);
case_brrn(bne,, no_flag);
case_brwn(bne,, no_flag);
case_brr(boadd,);
--
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 <=
- [PATCH v2 8/8] mips: Fill delay slots in jit_bger, jit_bgei, jit_bltr, jit_blti, Paul Cercueil, 2023/01/09
- Re: [PATCH v2 0/9] mips: Fill delay slots v2, Paulo César Pereira de Andrade, 2023/01/10