[PATCH 6/8] mips: Fill delay slots in jit_bner / jit_bnei

[Paul Cercueil]

When we know that the last generated opcode is not the target of a jump,
and that it does not write to the source registers of the BNE opcode, we
can swap it with the BNE opcode, so that it now becomes the delay slot
of the BNE opcode.

Signed-off-by: Paul Cercueil <paul@crapouillou.net>
---
 lib/jit_mips-cpu.c | 52 +++++++++++++++++++++++++++++++++++-----------
 lib/jit_mips.c     |  4 ++--
 2 files changed, 42 insertions(+), 14 deletions(-)

diff --git a/lib/jit_mips-cpu.c b/lib/jit_mips-cpu.c
index a162d93..748db89 100644
--- a/lib/jit_mips-cpu.c
+++ b/lib/jit_mips-cpu.c
@@ -709,10 +709,10 @@ static jit_word_t 
_bgtr_u(jit_state_t*,jit_word_t,jit_int32_t,jit_int32_t);
 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 bner(i0,r0,r1)                 _bner(_jit,i0,r0,r1)
-static jit_word_t _bner(jit_state_t*,jit_word_t,jit_int32_t,jit_int32_t);
-#define bnei(i0,r0,i1)                 _bnei(_jit,i0,r0,i1)
-static jit_word_t _bnei(jit_state_t*,jit_word_t,jit_int32_t,jit_word_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)
+static jit_word_t 
_bnei(jit_state_t*,jit_word_t,jit_int32_t,jit_word_t,jit_bool_t);
 #  define jmpr(r0,no_flag)             _jmpr(_jit,r0,no_flag)
 static void _jmpr(jit_state_t*,jit_int32_t,jit_bool_t);
 #  define jmpi(i0,no_flag)             _jmpi(_jit,i0,no_flag)
@@ -2512,32 +2512,60 @@ _bgti_u(jit_state_t *_jit, jit_word_t i0, jit_int32_t 
r0, jit_word_t i1)
 }
 
 static jit_word_t
-_bner(jit_state_t *_jit, jit_word_t i0, jit_int32_t r0, jit_int32_t r1)
+_bner(jit_state_t *_jit, jit_word_t i0, jit_int32_t r0, jit_int32_t r1, 
jit_bool_t no_flag)
 {
     jit_word_t         w;
+    jit_int32_t                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);
+
+    if (swap_ds)
+       prev = *--_jit->pc.ui;
 
     w = _jit->pc.w;
     BNE(r0, r1, ((i0 - w) >> 2) - 1);
-    NOP(1);
+    if (swap_ds)
+       ii(prev);
+    else
+       NOP(1);
 
     return (w);
 }
 
 static jit_word_t
-_bnei(jit_state_t *_jit, jit_word_t i0, jit_int32_t r0, jit_word_t i1)
+_bnei(jit_state_t *_jit, jit_word_t i0, jit_int32_t r0, jit_word_t i1, 
jit_bool_t no_flag)
 {
     jit_word_t         w;
-    jit_int32_t                reg;
+    jit_int32_t                reg, prev, offset;
+    jit_bool_t         swap_ds;
 
     if (i1 == 0) {
-       w = _jit->pc.w;
-       BNE(r0, _ZERO_REGNO, ((i0 - w) >> 2) - 1);
-       NOP(1);
+       w = bner(i0, r0, _ZERO_REGNO, no_flag);
     }
     else {
+       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);
+
        reg = jit_get_reg(jit_class_gpr|jit_class_nospill);
+
+       if (swap_ds)
+           prev = *--_jit->pc.ui;
+
        movi(rn(reg), i1);
-       w = bner(i0, r0, rn(reg));
+       w = _jit->pc.w;
+       BNE(r0, rn(reg), ((i0 - w) >> 2) - 1);
+       if (swap_ds)
+           ii(prev);
+       else
+           NOP(1);
+
        jit_unget_reg(reg);
     }
 
diff --git a/lib/jit_mips.c b/lib/jit_mips.c
index c3898aa..7683aeb 100644
--- a/lib/jit_mips.c
+++ b/lib/jit_mips.c
@@ -1529,8 +1529,8 @@ _emit_code(jit_state_t *_jit)
                case_brw(bgt,);
                case_brr(bgt, _u);
                case_brw(bgt, _u);
-               case_brr(bne,);
-               case_brw(bne,);
+               case_brrn(bne,, no_flag);
+               case_brwn(bne,, no_flag);
                case_brr(boadd,);
                case_brw(boadd,);
                case_brr(boadd, _u);
-- 
2.39.0