Add IR_TOI64.

4 files changed, 52 insertions, 2 deletions

diff --git a/src/lj_asm.c b/src/lj_asm.c
index 81529a56..e05de41e 100644
--- a/src/lj_asm.c
+++ b/src/lj_asm.c
@@ -1595,6 +1595,24 @@ static void asm_tobit(ASMState *as, IRIns *ir)
   ra_left(as, tmp, ir->op1);
 }
 
+static void asm_toi64(ASMState *as, IRIns *ir)
+{
+  Reg dest = ra_dest(as, ir, RSET_GPR);
+  IRRef lref = ir->op1;
+  lua_assert(LJ_64);  /* NYI: 32 bit register pairs. */
+  if (ir->op2 == IRTOINT_TRUNCI64) {
+    Reg left = asm_fuseload(as, lref, RSET_FPR);
+    emit_mrm(as, XO_CVTTSD2SI, dest|REX_64, left);
+  } else if (ir->op2 == IRTOINT_ZEXT64) {
+    /* Nothing to do. This assumes 32 bit regs are already zero-extended. */
+    ra_left(as, dest, lref);  /* But may need to move regs. */
+  } else {
+    Reg left = asm_fuseload(as, lref, RSET_GPR);
+    emit_mrm(as, XO_MOVSXd, dest|REX_64, left);
+    lua_assert(ir->op2 == IRTOINT_SEXT64);
+  }
+}
+
 static void asm_strto(ASMState *as, IRIns *ir)
 {
   /* Force a spill slot for the destination register (if any). */
@@ -3531,6 +3549,7 @@ static void asm_ir(ASMState *as, IRIns *ir)
       asm_toint(as, ir); break;
     break;
   case IR_TOBIT: asm_tobit(as, ir); break;
+  case IR_TOI64: asm_toi64(as, ir); break;
   case IR_TOSTR: asm_tostr(as, ir); break;
   case IR_STRTO: asm_strto(as, ir); break;
 
diff --git a/src/lj_ir.h b/src/lj_ir.h
index 72a4d6b4..72c37343 100644
--- a/src/lj_ir.h
+++ b/src/lj_ir.h
@@ -119,6 +119,7 @@
   _(TONUM,      N , ref, ___) \
   _(TOINT,      N , ref, lit) \
   _(TOBIT,      N , ref, ref) \
+  _(TOI64,      N , ref, lit) \
   _(TOSTR,      N , ref, ___) \
   _(STRTO,      N , ref, ___) \
   \
@@ -204,11 +205,14 @@ IRFLDEF(FLENUM)
 #define IRXLOAD_READONLY        1       /* Load from read-only data. */
 #define IRXLOAD_UNALIGNED       2       /* Unaligned load. */
 
-/* TOINT mode, stored in op2. Ordered by strength of the checks. */
+/* TOINT/TOI64 mode, stored in op2. Ordered by strength of the checks. */
 #define IRTOINT_CHECK           0       /* Number checked for integerness. */
 #define IRTOINT_INDEX           1       /* Checked + special backprop rules. */
 #define IRTOINT_ANY             2       /* Any FP number is ok. */
-#define IRTOINT_TOBIT           3       /* Cache only: TOBIT conversion. */
+#define IRTOINT_ZEXT64          3       /* Convert uint32_t to int64_t. */
+#define IRTOINT_SEXT64          4       /* Convert int32_t to int64_t. */
+#define IRTOINT_TRUNCI64        5       /* Truncate number to int64_t. */
+#define IRTOINT_TOBIT           6       /* Cache only: TOBIT conversion. */
 
 /* C call info for CALL* instructions. */
 typedef struct CCallInfo {
diff --git a/src/lj_opt_fold.c b/src/lj_opt_fold.c
index feddba87..cf5de82f 100644
--- a/src/lj_opt_fold.c
+++ b/src/lj_opt_fold.c
@@ -270,6 +270,23 @@ LJFOLDF(kfold_toint)
   return INTFOLD(k);
 }
 
+LJFOLD(TOI64 KINT any)
+LJFOLDF(kfold_toi64_kint)
+{
+  lua_assert(fins->op2 == IRTOINT_ZEXT64 || fins->op2 == IRTOINT_SEXT64);
+  if (fins->op2 == IRTOINT_ZEXT64)
+    return lj_ir_kint64(J, (int64_t)(uint32_t)fleft->i);
+  else
+    return lj_ir_kint64(J, (int64_t)(int32_t)fleft->i);
+}
+
+LJFOLD(TOI64 KNUM any)
+LJFOLDF(kfold_toi64_knum)
+{
+  lua_assert(fins->op2 == IRTOINT_TRUNCI64);
+  return lj_ir_kint64(J, (int64_t)knumleft);
+}
+
 LJFOLD(TOSTR KNUM)
 LJFOLDF(kfold_tostr_knum)
 {
@@ -471,6 +488,15 @@ LJFOLDF(shortcut_leftleft_across_phi)
   return fleft->op1;  /* f(g(x)) ==> x */
 }
 
+LJFOLD(TOI64 TONUM any)
+LJFOLDF(shortcut_leftleft_toint64)
+{
+  /* Fold even across PHI to avoid expensive int->num->int64 conversions. */
+  fins->op1 = fleft->op1;   /* (int64_t)(double)(int)x ==> (int64_t)x */
+  fins->op2 = IRTOINT_SEXT64;
+  return RETRYFOLD;
+}
+
 /* -- FP algebraic simplifications ---------------------------------------- */
 
 /* FP arithmetic is tricky -- there's not much to simplify.
diff --git a/src/lj_target_x86.h b/src/lj_target_x86.h
index d3956a09..bf58d67c 100644
--- a/src/lj_target_x86.h
+++ b/src/lj_target_x86.h
@@ -230,6 +230,7 @@ typedef enum {
   XO_MOVZXw =   XO_0f(b7),
   XO_MOVSXb =   XO_0f(be),
   XO_MOVSXw =   XO_0f(bf),
+  XO_MOVSXd =   XO_(63),
   XO_BSWAP =    XO_0f(c8),
 
   XO_MOVSD =    XO_f20f(10),