[PATCH v2 4/8] mips: Fill delay slots of J in jit_jmpi

[Paul Cercueil]

When we know that the last generated opcode is not the target of a jump,
we can swap it with the J opcode, so that it now becomes the delay slot
of the J opcode.

Signed-off-by: Paul Cercueil <paul@crapouillou.net>
---
 lib/jit_mips-cpu.c | 37 +++++++++++++++++++++++++++++++------
 lib/jit_mips.c     |  6 +++---
 2 files changed, 34 insertions(+), 9 deletions(-)

diff --git a/lib/jit_mips-cpu.c b/lib/jit_mips-cpu.c
index 512ec7e..ec2745b 100644
--- a/lib/jit_mips-cpu.c
+++ b/lib/jit_mips-cpu.c
@@ -715,8 +715,8 @@ static jit_word_t 
_bner(jit_state_t*,jit_word_t,jit_int32_t,jit_int32_t);
 static jit_word_t _bnei(jit_state_t*,jit_word_t,jit_int32_t,jit_word_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)                     _jmpi(_jit,i0)
-static jit_word_t _jmpi(jit_state_t*,jit_word_t);
+#  define jmpi(i0,no_flag)             _jmpi(_jit,i0,no_flag)
+static jit_word_t _jmpi(jit_state_t*,jit_word_t,jit_bool_t);
 #  define boaddr(i0,r0,r1)             _boaddr(_jit,i0,r0,r1)
 static jit_word_t _boaddr(jit_state_t*,jit_word_t,jit_int32_t,jit_int32_t);
 #  define boaddi(i0,r0,i1)             _boaddi(_jit,i0,r0,i1)
@@ -2558,20 +2558,45 @@ _jmpr(jit_state_t *_jit, jit_int32_t r0, jit_bool_t 
no_flag)
 }
 
 static jit_word_t
-_jmpi(jit_state_t *_jit, jit_word_t i0)
+_jmpi(jit_state_t *_jit, jit_word_t i0, 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_int32_t);
+    swap_ds = no_flag
+           && (offset < 2 || !has_delay_slot((jit_instr_t)*(_jit->pc.ui - 2)));
 
     w = _jit->pc.w;
     if (((w + sizeof(jit_int32_t)) & 0xf0000000) == (i0 & 0xf0000000)) {
+
+        if (swap_ds) {
+            prev = *--_jit->pc.ui;
+           w -= sizeof(jit_int32_t);
+       }
+
        J((i0 & ~0xf0000000) >> 2);
-       NOP(1);
+       if (swap_ds)
+            ii(prev);
+       else
+            NOP(1);
     }
     else {
        reg = jit_get_reg(jit_class_gpr|jit_class_nospill);
+
+       if (swap_ds) {
+            prev = *--_jit->pc.ui;
+           w -= sizeof(jit_int32_t);
+       }
+
        movi_p(rn(reg), i0);
-       jmpr(rn(reg), 0);
+       JR(rn(reg));
+       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 fc303b2..cf917aa 100644
--- a/lib/jit_mips.c
+++ b/lib/jit_mips.c
@@ -1696,14 +1696,14 @@ _emit_code(jit_state_t *_jit)
                    assert(temp->code == jit_code_label ||
                           temp->code == jit_code_epilog);
                    if (temp->flag & jit_flag_patch)
-                       jmpi(temp->u.w);
+                       jmpi(temp->u.w, no_flag);
                    else {
-                       word = jmpi(_jit->pc.w);
+                       word = jmpi(_jit->pc.w, no_flag);
                        patch(word, node);
                    }
                }
                else
-                   jmpi(node->u.w);
+                   jmpi(node->u.w, no_flag);
                break;
            case jit_code_callr:
                callr(rn(node->u.w), no_flag);
-- 
2.39.0