Re: [Tarantool-patches] [PATCH luajit 04/19] MIPS64: Add soft-float support to JIT compiler backend.

[Maxim Kokryashkin via Tarantool-patches <tarantool-patches@dev.tarantool.org>]

Hi, Sergey!
LGTM, except for a few comments below.
On Wed, Aug 09, 2023 at 06:35:53PM +0300, Sergey Kaplun via Tarantool-patches wrote:
> From: Mike Pall <mike>
> 
> Contributed by Djordje Kovacevic and Stefan Pejic from RT-RK.com.
> Sponsored by Cisco Systems, Inc.
> 
> (cherry-picked from commit a057a07ab702e225e21848d4f918886c5b0ac06b)
> 
> The software floating point library is used on machines which do not
> have hardware support for floating point [1]. This patch enables
> support for such machines in JIT compiler backend for MIPS64.
Typo: s/in JIT/in the JIT/
> This includes:
> * `vm_tointg()` helper is added in <src/vm_mips64.dasm> to convert FP
>   number to integer with a check for the soft-float support (called from
>   JIT).
> * `sfmin/max()` helpers are added in <src/vm_mips64.dasm> for min/max
>   operations with a check for the soft-float support (called from JIT).
Typo: s/the soft-float/soft-float/
> * `LJ_SOFTFP32` macro is introduced to be used for 32-bit MIPS instead
>   `LJ_SOFTFP`.
> * All fp-depending paths are instrumented with `LJ_SOFTFP` or
Typo: s/fp-depending/fp-dependent/
>   `LJ_SOFTFP32` macro.
Typo: s/macro/macros/
> * The corresponding function calls in <src/lj_ircall.h> are marked as
>   `XA_FP32`, `XA2_FP32`, i.e. as required extra arguments on the stack
>   for soft-FP on 32-bit MIPS.

Shouldn't we also mention the `asm_tobit` function?
> 
> [1]: https://gcc.gnu.org/onlinedocs/gccint/Soft-float-library-routines.html
> 
> Sergey Kaplun:
> * added the description for the feature
> 
> Part of tarantool/tarantool#8825
> ---
>  src/lj_arch.h      |   4 +-
>  src/lj_asm.c       |   8 +-
>  src/lj_asm_mips.h  | 217 +++++++++++++++++++++++++++++++++++++--------
>  src/lj_crecord.c   |   4 +-
>  src/lj_emit_mips.h |   2 +
>  src/lj_ffrecord.c  |   2 +-
>  src/lj_ircall.h    |  43 ++++++---
>  src/lj_iropt.h     |   2 +-
>  src/lj_jit.h       |   4 +-
>  src/lj_obj.h       |   3 +
>  src/lj_opt_split.c |   2 +-
>  src/lj_snap.c      |  21 +++--
>  src/vm_mips64.dasc |  49 ++++++++++
>  13 files changed, 286 insertions(+), 75 deletions(-)
> 
> diff --git a/src/lj_arch.h b/src/lj_arch.h
> index 5276ae56..c39526ea 100644
> --- a/src/lj_arch.h
> +++ b/src/lj_arch.h
> @@ -349,9 +349,6 @@
>  #define LJ_ARCH_BITS		32
>  #define LJ_TARGET_MIPS32	1
>  #else
> -#if LJ_ABI_SOFTFP || !LJ_ARCH_HASFPU
> -#define LJ_ARCH_NOJIT		1	/* NYI */
> -#endif
>  #define LJ_ARCH_BITS		64
>  #define LJ_TARGET_MIPS64	1
>  #define LJ_TARGET_GC64		1
> @@ -528,6 +525,7 @@
>  #define LJ_ABI_SOFTFP		0
>  #endif
>  #define LJ_SOFTFP		(!LJ_ARCH_HASFPU)
> +#define LJ_SOFTFP32		(LJ_SOFTFP && LJ_32)
>  
>  #if LJ_ARCH_ENDIAN == LUAJIT_BE
>  #define LJ_LE			0
> diff --git a/src/lj_asm.c b/src/lj_asm.c
> index 0bfa44ed..15de7e33 100644
> --- a/src/lj_asm.c
> +++ b/src/lj_asm.c
> @@ -341,7 +341,7 @@ static Reg ra_rematk(ASMState *as, IRRef ref)
>    ra_modified(as, r);
>    ir->r = RID_INIT;  /* Do not keep any hint. */
>    RA_DBGX((as, "remat     $i $r", ir, r));
> -#if !LJ_SOFTFP
> +#if !LJ_SOFTFP32
>    if (ir->o == IR_KNUM) {
>      emit_loadk64(as, r, ir);
>    } else
> @@ -1356,7 +1356,7 @@ static void asm_call(ASMState *as, IRIns *ir)
>    asm_gencall(as, ci, args);
>  }
>  
> -#if !LJ_SOFTFP
> +#if !LJ_SOFTFP32
>  static void asm_fppow(ASMState *as, IRIns *ir, IRRef lref, IRRef rref)
>  {
>    const CCallInfo *ci = &lj_ir_callinfo[IRCALL_pow];
> @@ -1703,10 +1703,10 @@ static void asm_ir(ASMState *as, IRIns *ir)
>    case IR_MUL: asm_mul(as, ir); break;
>    case IR_MOD: asm_mod(as, ir); break;
>    case IR_NEG: asm_neg(as, ir); break;
> -#if LJ_SOFTFP
> +#if LJ_SOFTFP32
>    case IR_DIV: case IR_POW: case IR_ABS:
>    case IR_ATAN2: case IR_LDEXP: case IR_FPMATH: case IR_TOBIT:
> -    lua_assert(0);  /* Unused for LJ_SOFTFP. */
> +    lua_assert(0);  /* Unused for LJ_SOFTFP32. */
>      break;
>  #else
>    case IR_DIV: asm_div(as, ir); break;
> diff --git a/src/lj_asm_mips.h b/src/lj_asm_mips.h
> index 0e60fc07..a26a82cd 100644
> --- a/src/lj_asm_mips.h
> +++ b/src/lj_asm_mips.h
> @@ -290,7 +290,7 @@ static void asm_gencall(ASMState *as, const CCallInfo *ci, IRRef *args)
>  	  {
>  	    ra_leftov(as, gpr, ref);
>  	    gpr++;
> -#if LJ_64
> +#if LJ_64 && !LJ_SOFTFP
>  	    fpr++;
>  #endif
>  	  }
> @@ -301,7 +301,7 @@ static void asm_gencall(ASMState *as, const CCallInfo *ci, IRRef *args)
>  	  emit_spstore(as, ir, r, ofs);
>  	  ofs += irt_isnum(ir->t) ? 8 : 4;
>  #else
> -	  emit_spstore(as, ir, r, ofs + ((LJ_BE && (LJ_SOFTFP || r < RID_MAX_GPR) && !irt_is64(ir->t)) ? 4 : 0));
> +	  emit_spstore(as, ir, r, ofs + ((LJ_BE && !irt_isfp(ir->t) && !irt_is64(ir->t)) ? 4 : 0));
>  	  ofs += 8;
>  #endif
>  	}
> @@ -312,7 +312,7 @@ static void asm_gencall(ASMState *as, const CCallInfo *ci, IRRef *args)
>  #endif
>        if (gpr <= REGARG_LASTGPR) {
>  	gpr++;
> -#if LJ_64
> +#if LJ_64 && !LJ_SOFTFP
>  	fpr++;
>  #endif
>        } else {
> @@ -461,12 +461,36 @@ static void asm_tobit(ASMState *as, IRIns *ir)
>    emit_tg(as, MIPSI_MFC1, dest, tmp);
>    emit_fgh(as, MIPSI_ADD_D, tmp, left, right);
>  }
> +#elif LJ_64  /* && LJ_SOFTFP */
> +static void asm_tointg(ASMState *as, IRIns *ir, Reg r)
> +{
> +  /* The modified regs must match with the *.dasc implementation. */
> +  RegSet drop = RID2RSET(REGARG_FIRSTGPR)|RID2RSET(RID_RET)|RID2RSET(RID_RET+1)|
> +		RID2RSET(RID_R1)|RID2RSET(RID_R12);
> +  if (ra_hasreg(ir->r)) rset_clear(drop, ir->r);
> +  ra_evictset(as, drop);
> +  /* Return values are in RID_RET (converted value) and RID_RET+1 (status). */
> +  ra_destreg(as, ir, RID_RET);
> +  asm_guard(as, MIPSI_BNE, RID_RET+1, RID_ZERO);
> +  emit_call(as, (void *)lj_ir_callinfo[IRCALL_lj_vm_tointg].func, 0);
> +  if (r == RID_NONE)
> +    ra_leftov(as, REGARG_FIRSTGPR, ir->op1);
> +  else if (r != REGARG_FIRSTGPR)
> +    emit_move(as, REGARG_FIRSTGPR, r);
> +}
> +
> +static void asm_tobit(ASMState *as, IRIns *ir)
> +{
> +  Reg dest = ra_dest(as, ir, RSET_GPR);
> +  emit_dta(as, MIPSI_SLL, dest, dest, 0);
> +  asm_callid(as, ir, IRCALL_lj_vm_tobit);
> +}
>  #endif
>  
>  static void asm_conv(ASMState *as, IRIns *ir)
>  {
>    IRType st = (IRType)(ir->op2 & IRCONV_SRCMASK);
> -#if !LJ_SOFTFP
> +#if !LJ_SOFTFP32
>    int stfp = (st == IRT_NUM || st == IRT_FLOAT);
>  #endif
>  #if LJ_64
> @@ -477,12 +501,13 @@ static void asm_conv(ASMState *as, IRIns *ir)
>    lua_assert(!(irt_isint64(ir->t) ||
>  	       (st == IRT_I64 || st == IRT_U64))); /* Handled by SPLIT. */
>  #endif
> -#if LJ_32 && LJ_SOFTFP
> +#if LJ_SOFTFP32
>    /* FP conversions are handled by SPLIT. */
>    lua_assert(!irt_isfp(ir->t) && !(st == IRT_NUM || st == IRT_FLOAT));
>    /* Can't check for same types: SPLIT uses CONV int.int + BXOR for sfp NEG. */
>  #else
>    lua_assert(irt_type(ir->t) != st);
> +#if !LJ_SOFTFP
>    if (irt_isfp(ir->t)) {
>      Reg dest = ra_dest(as, ir, RSET_FPR);
>      if (stfp) {  /* FP to FP conversion. */
> @@ -608,6 +633,42 @@ static void asm_conv(ASMState *as, IRIns *ir)
>        }
>      }
>    } else
> +#else
> +  if (irt_isfp(ir->t)) {
> +#if LJ_64 && LJ_HASFFI
> +    if (stfp) {  /* FP to FP conversion. */
> +      asm_callid(as, ir, irt_isnum(ir->t) ? IRCALL_softfp_f2d :
> +					    IRCALL_softfp_d2f);
> +    } else {  /* Integer to FP conversion. */
> +      IRCallID cid = ((IRT_IS64 >> st) & 1) ?
> +	(irt_isnum(ir->t) ?
> +	 (st == IRT_I64 ? IRCALL_fp64_l2d : IRCALL_fp64_ul2d) :
> +	 (st == IRT_I64 ? IRCALL_fp64_l2f : IRCALL_fp64_ul2f)) :
> +	(irt_isnum(ir->t) ?
> +	 (st == IRT_INT ? IRCALL_softfp_i2d : IRCALL_softfp_ui2d) :
> +	 (st == IRT_INT ? IRCALL_softfp_i2f : IRCALL_softfp_ui2f));
> +      asm_callid(as, ir, cid);
> +    }
> +#else
> +    asm_callid(as, ir, IRCALL_softfp_i2d);
> +#endif
> +  } else if (stfp) {  /* FP to integer conversion. */
> +    if (irt_isguard(ir->t)) {
> +      /* Checked conversions are only supported from number to int. */
> +      lua_assert(irt_isint(ir->t) && st == IRT_NUM);
> +      asm_tointg(as, ir, RID_NONE);
> +    } else {
> +      IRCallID cid = irt_is64(ir->t) ?
> +	((st == IRT_NUM) ?
> +	 (irt_isi64(ir->t) ? IRCALL_fp64_d2l : IRCALL_fp64_d2ul) :
> +	 (irt_isi64(ir->t) ? IRCALL_fp64_f2l : IRCALL_fp64_f2ul)) :
> +	((st == IRT_NUM) ?
> +	 (irt_isint(ir->t) ? IRCALL_softfp_d2i : IRCALL_softfp_d2ui) :
> +	 (irt_isint(ir->t) ? IRCALL_softfp_f2i : IRCALL_softfp_f2ui));
> +      asm_callid(as, ir, cid);
> +    }
> +  } else
> +#endif
>  #endif
>    {
>      Reg dest = ra_dest(as, ir, RSET_GPR);
> @@ -665,7 +726,7 @@ static void asm_strto(ASMState *as, IRIns *ir)
>    const CCallInfo *ci = &lj_ir_callinfo[IRCALL_lj_strscan_num];
>    IRRef args[2];
>    int32_t ofs = 0;
> -#if LJ_SOFTFP
> +#if LJ_SOFTFP32
>    ra_evictset(as, RSET_SCRATCH);
>    if (ra_used(ir)) {
>      if (ra_hasspill(ir->s) && ra_hasspill((ir+1)->s) &&
> @@ -806,7 +867,7 @@ static void asm_href(ASMState *as, IRIns *ir, IROp merge)
>    MCLabel l_end, l_loop, l_next;
>  
>    rset_clear(allow, tab);
> -#if LJ_32 && LJ_SOFTFP
> +#if LJ_SOFTFP32
>    if (!isk) {
>      key = ra_alloc1(as, refkey, allow);
>      rset_clear(allow, key);
> @@ -826,7 +887,7 @@ static void asm_href(ASMState *as, IRIns *ir, IROp merge)
>      }
>    }
>  #else
> -  if (irt_isnum(kt)) {
> +  if (!LJ_SOFTFP && irt_isnum(kt)) {
>      key = ra_alloc1(as, refkey, RSET_FPR);
>      tmpnum = ra_scratch(as, rset_exclude(RSET_FPR, key));
>    } else if (!irt_ispri(kt)) {
> @@ -882,6 +943,9 @@ static void asm_href(ASMState *as, IRIns *ir, IROp merge)
>      emit_dta(as, MIPSI_DSRA32, tmp1, tmp1, 15);
>      emit_tg(as, MIPSI_DMTC1, tmp1, tmpnum);
>      emit_tsi(as, MIPSI_LD, tmp1, dest, (int32_t)offsetof(Node, key.u64));
> +  } else if (LJ_SOFTFP && irt_isnum(kt)) {
> +    emit_branch(as, MIPSI_BEQ, tmp1, key, l_end);
> +    emit_tsi(as, MIPSI_LD, tmp1, dest, (int32_t)offsetof(Node, key.u64));
>    } else if (irt_isaddr(kt)) {
>      Reg refk = tmp2;
>      if (isk) {
> @@ -960,7 +1024,7 @@ static void asm_href(ASMState *as, IRIns *ir, IROp merge)
>        emit_dta(as, MIPSI_ROTR, dest, tmp1, (-HASH_ROT1)&31);
>        if (irt_isnum(kt)) {
>  	emit_dst(as, MIPSI_ADDU, tmp1, tmp1, tmp1);
> -	emit_dta(as, MIPSI_DSRA32, tmp1, tmp1, 0);
> +	emit_dta(as, MIPSI_DSRA32, tmp1, LJ_SOFTFP ? key : tmp1, 0);
>  	emit_dta(as, MIPSI_SLL, tmp2, LJ_SOFTFP ? key : tmp1, 0);
>  #if !LJ_SOFTFP
>  	emit_tg(as, MIPSI_DMFC1, tmp1, key);
> @@ -1123,7 +1187,7 @@ static MIPSIns asm_fxloadins(IRIns *ir)
>    case IRT_U8: return MIPSI_LBU;
>    case IRT_I16: return MIPSI_LH;
>    case IRT_U16: return MIPSI_LHU;
> -  case IRT_NUM: lua_assert(!LJ_SOFTFP); return MIPSI_LDC1;
> +  case IRT_NUM: lua_assert(!LJ_SOFTFP32); if (!LJ_SOFTFP) return MIPSI_LDC1;
>    case IRT_FLOAT: if (!LJ_SOFTFP) return MIPSI_LWC1;
>    default: return (LJ_64 && irt_is64(ir->t)) ? MIPSI_LD : MIPSI_LW;
>    }
> @@ -1134,7 +1198,7 @@ static MIPSIns asm_fxstoreins(IRIns *ir)
>    switch (irt_type(ir->t)) {
>    case IRT_I8: case IRT_U8: return MIPSI_SB;
>    case IRT_I16: case IRT_U16: return MIPSI_SH;
> -  case IRT_NUM: lua_assert(!LJ_SOFTFP); return MIPSI_SDC1;
> +  case IRT_NUM: lua_assert(!LJ_SOFTFP32); if (!LJ_SOFTFP) return MIPSI_SDC1;
>    case IRT_FLOAT: if (!LJ_SOFTFP) return MIPSI_SWC1;
>    default: return (LJ_64 && irt_is64(ir->t)) ? MIPSI_SD : MIPSI_SW;
>    }
> @@ -1199,7 +1263,7 @@ static void asm_xstore_(ASMState *as, IRIns *ir, int32_t ofs)
>  
>  static void asm_ahuvload(ASMState *as, IRIns *ir)
>  {
> -  int hiop = (LJ_32 && LJ_SOFTFP && (ir+1)->o == IR_HIOP);
> +  int hiop = (LJ_SOFTFP32 && (ir+1)->o == IR_HIOP);
>    Reg dest = RID_NONE, type = RID_TMP, idx;
>    RegSet allow = RSET_GPR;
>    int32_t ofs = 0;
> @@ -1212,7 +1276,7 @@ static void asm_ahuvload(ASMState *as, IRIns *ir)
>      }
>    }
>    if (ra_used(ir)) {
> -    lua_assert((LJ_SOFTFP ? 0 : irt_isnum(ir->t)) ||
> +    lua_assert((LJ_SOFTFP32 ? 0 : irt_isnum(ir->t)) ||
>  	       irt_isint(ir->t) || irt_isaddr(ir->t));
>      dest = ra_dest(as, ir, (!LJ_SOFTFP && irt_isnum(t)) ? RSET_FPR : allow);
>      rset_clear(allow, dest);
> @@ -1261,10 +1325,10 @@ static void asm_ahustore(ASMState *as, IRIns *ir)
>    int32_t ofs = 0;
>    if (ir->r == RID_SINK)
>      return;
> -  if (!LJ_SOFTFP && irt_isnum(ir->t)) {
> -    src = ra_alloc1(as, ir->op2, RSET_FPR);
> +  if (!LJ_SOFTFP32 && irt_isnum(ir->t)) {
> +    src = ra_alloc1(as, ir->op2, LJ_SOFTFP ? RSET_GPR : RSET_FPR);
>      idx = asm_fuseahuref(as, ir->op1, &ofs, allow);
> -    emit_hsi(as, MIPSI_SDC1, src, idx, ofs);
> +    emit_hsi(as, LJ_SOFTFP ? MIPSI_SD : MIPSI_SDC1, src, idx, ofs);
>    } else {
>  #if LJ_32
>      if (!irt_ispri(ir->t)) {
> @@ -1312,7 +1376,7 @@ static void asm_sload(ASMState *as, IRIns *ir)
>    IRType1 t = ir->t;
>  #if LJ_32
>    int32_t ofs = 8*((int32_t)ir->op1-1) + ((ir->op2 & IRSLOAD_FRAME) ? 4 : 0);
> -  int hiop = (LJ_32 && LJ_SOFTFP && (ir+1)->o == IR_HIOP);
> +  int hiop = (LJ_SOFTFP32 && (ir+1)->o == IR_HIOP);
>    if (hiop)
>      t.irt = IRT_NUM;
>  #else
> @@ -1320,7 +1384,7 @@ static void asm_sload(ASMState *as, IRIns *ir)
>  #endif
>    lua_assert(!(ir->op2 & IRSLOAD_PARENT));  /* Handled by asm_head_side(). */
>    lua_assert(irt_isguard(ir->t) || !(ir->op2 & IRSLOAD_TYPECHECK));
> -#if LJ_32 && LJ_SOFTFP
> +#if LJ_SOFTFP32
>    lua_assert(!(ir->op2 & IRSLOAD_CONVERT));  /* Handled by LJ_SOFTFP SPLIT. */
>    if (hiop && ra_used(ir+1)) {
>      type = ra_dest(as, ir+1, allow);
> @@ -1328,29 +1392,44 @@ static void asm_sload(ASMState *as, IRIns *ir)
>    }
>  #else
>    if ((ir->op2 & IRSLOAD_CONVERT) && irt_isguard(t) && irt_isint(t)) {
> -    dest = ra_scratch(as, RSET_FPR);
> +    dest = ra_scratch(as, LJ_SOFTFP ? allow : RSET_FPR);
>      asm_tointg(as, ir, dest);
>      t.irt = IRT_NUM;  /* Continue with a regular number type check. */
>    } else
>  #endif
>    if (ra_used(ir)) {
> -    lua_assert((LJ_SOFTFP ? 0 : irt_isnum(ir->t)) ||
> +    lua_assert((LJ_SOFTFP32 ? 0 : irt_isnum(ir->t)) ||
>  	       irt_isint(ir->t) || irt_isaddr(ir->t));
>      dest = ra_dest(as, ir, (!LJ_SOFTFP && irt_isnum(t)) ? RSET_FPR : allow);
>      rset_clear(allow, dest);
>      base = ra_alloc1(as, REF_BASE, allow);
>      rset_clear(allow, base);
> -    if (!LJ_SOFTFP && (ir->op2 & IRSLOAD_CONVERT)) {
> +    if (!LJ_SOFTFP32 && (ir->op2 & IRSLOAD_CONVERT)) {
>        if (irt_isint(t)) {
> -	Reg tmp = ra_scratch(as, RSET_FPR);
> +	Reg tmp = ra_scratch(as, LJ_SOFTFP ? RSET_GPR : RSET_FPR);
> +#if LJ_SOFTFP
> +	ra_evictset(as, rset_exclude(RSET_SCRATCH, dest));
> +	ra_destreg(as, ir, RID_RET);
> +	emit_call(as, (void *)lj_ir_callinfo[IRCALL_softfp_d2i].func, 0);
> +	if (tmp != REGARG_FIRSTGPR)
> +	  emit_move(as, REGARG_FIRSTGPR, tmp);
> +#else
>  	emit_tg(as, MIPSI_MFC1, dest, tmp);
>  	emit_fg(as, MIPSI_TRUNC_W_D, tmp, tmp);
> +#endif
>  	dest = tmp;
>  	t.irt = IRT_NUM;  /* Check for original type. */
>        } else {
>  	Reg tmp = ra_scratch(as, RSET_GPR);
> +#if LJ_SOFTFP
> +	ra_evictset(as, rset_exclude(RSET_SCRATCH, dest));
> +	ra_destreg(as, ir, RID_RET);
> +	emit_call(as, (void *)lj_ir_callinfo[IRCALL_softfp_i2d].func, 0);
> +	emit_dta(as, MIPSI_SLL, REGARG_FIRSTGPR, tmp, 0);
> +#else
>  	emit_fg(as, MIPSI_CVT_D_W, dest, dest);
>  	emit_tg(as, MIPSI_MTC1, tmp, dest);
> +#endif
>  	dest = tmp;
>  	t.irt = IRT_INT;  /* Check for original type. */
>        }
> @@ -1399,7 +1478,7 @@ dotypecheck:
>        if (irt_isnum(t)) {
>  	asm_guard(as, MIPSI_BEQ, RID_TMP, RID_ZERO);
>  	emit_tsi(as, MIPSI_SLTIU, RID_TMP, RID_TMP, (int32_t)LJ_TISNUM);
> -	if (ra_hasreg(dest))
> +	if (!LJ_SOFTFP && ra_hasreg(dest))
>  	  emit_hsi(as, MIPSI_LDC1, dest, base, ofs);
>        } else {
>  	asm_guard(as, MIPSI_BNE, RID_TMP,
> @@ -1409,7 +1488,7 @@ dotypecheck:
>      }
>      emit_tsi(as, MIPSI_LD, type, base, ofs);
>    } else if (ra_hasreg(dest)) {
> -    if (irt_isnum(t))
> +    if (!LJ_SOFTFP && irt_isnum(t))
>        emit_hsi(as, MIPSI_LDC1, dest, base, ofs);
>      else
>        emit_tsi(as, irt_isint(t) ? MIPSI_LW : MIPSI_LD, dest, base,
> @@ -1554,26 +1633,40 @@ static void asm_fpunary(ASMState *as, IRIns *ir, MIPSIns mi)
>    Reg left = ra_hintalloc(as, ir->op1, dest, RSET_FPR);
>    emit_fg(as, mi, dest, left);
>  }
> +#endif
>  
> +#if !LJ_SOFTFP32
>  static void asm_fpmath(ASMState *as, IRIns *ir)
>  {
>    if (ir->op2 == IRFPM_EXP2 && asm_fpjoin_pow(as, ir))
>      return;
> +#if !LJ_SOFTFP
>    if (ir->op2 <= IRFPM_TRUNC)
>      asm_callround(as, ir, IRCALL_lj_vm_floor + ir->op2);
>    else if (ir->op2 == IRFPM_SQRT)
>      asm_fpunary(as, ir, MIPSI_SQRT_D);
>    else
> +#endif
>      asm_callid(as, ir, IRCALL_lj_vm_floor + ir->op2);
>  }
>  #endif
>  
> +#if !LJ_SOFTFP
> +#define asm_fpadd(as, ir)	asm_fparith(as, ir, MIPSI_ADD_D)
> +#define asm_fpsub(as, ir)	asm_fparith(as, ir, MIPSI_SUB_D)
> +#define asm_fpmul(as, ir)	asm_fparith(as, ir, MIPSI_MUL_D)
> +#elif LJ_64  /* && LJ_SOFTFP */
> +#define asm_fpadd(as, ir)	asm_callid(as, ir, IRCALL_softfp_add)
> +#define asm_fpsub(as, ir)	asm_callid(as, ir, IRCALL_softfp_sub)
> +#define asm_fpmul(as, ir)	asm_callid(as, ir, IRCALL_softfp_mul)
> +#endif
> +
>  static void asm_add(ASMState *as, IRIns *ir)
>  {
>    IRType1 t = ir->t;
> -#if !LJ_SOFTFP
> +#if !LJ_SOFTFP32
>    if (irt_isnum(t)) {
> -    asm_fparith(as, ir, MIPSI_ADD_D);
> +    asm_fpadd(as, ir);
>    } else
>  #endif
>    {
> @@ -1595,9 +1688,9 @@ static void asm_add(ASMState *as, IRIns *ir)
>  
>  static void asm_sub(ASMState *as, IRIns *ir)
>  {
> -#if !LJ_SOFTFP
> +#if !LJ_SOFTFP32
>    if (irt_isnum(ir->t)) {
> -    asm_fparith(as, ir, MIPSI_SUB_D);
> +    asm_fpsub(as, ir);
>    } else
>  #endif
>    {
> @@ -1611,9 +1704,9 @@ static void asm_sub(ASMState *as, IRIns *ir)
>  
>  static void asm_mul(ASMState *as, IRIns *ir)
>  {
> -#if !LJ_SOFTFP
> +#if !LJ_SOFTFP32
>    if (irt_isnum(ir->t)) {
> -    asm_fparith(as, ir, MIPSI_MUL_D);
> +    asm_fpmul(as, ir);
>    } else
>  #endif
>    {
> @@ -1640,7 +1733,7 @@ static void asm_mod(ASMState *as, IRIns *ir)
>      asm_callid(as, ir, IRCALL_lj_vm_modi);
>  }
>  
> -#if !LJ_SOFTFP
> +#if !LJ_SOFTFP32
>  static void asm_pow(ASMState *as, IRIns *ir)
>  {
>  #if LJ_64 && LJ_HASFFI
> @@ -1660,7 +1753,11 @@ static void asm_div(ASMState *as, IRIns *ir)
>  					  IRCALL_lj_carith_divu64);
>    else
>  #endif
> +#if !LJ_SOFTFP
>      asm_fparith(as, ir, MIPSI_DIV_D);
> +#else
> +  asm_callid(as, ir, IRCALL_softfp_div);
> +#endif
>  }
>  #endif
>  
> @@ -1670,6 +1767,13 @@ static void asm_neg(ASMState *as, IRIns *ir)
>    if (irt_isnum(ir->t)) {
>      asm_fpunary(as, ir, MIPSI_NEG_D);
>    } else
> +#elif LJ_64  /* && LJ_SOFTFP */
> +  if (irt_isnum(ir->t)) {
> +    Reg dest = ra_dest(as, ir, RSET_GPR);
> +    Reg left = ra_hintalloc(as, ir->op1, dest, RSET_GPR);
> +    emit_dst(as, MIPSI_XOR, dest, left,
> +	    ra_allock(as, 0x8000000000000000ll, rset_exclude(RSET_GPR, dest)));
> +  } else
>  #endif
>    {
>      Reg dest = ra_dest(as, ir, RSET_GPR);
> @@ -1679,7 +1783,17 @@ static void asm_neg(ASMState *as, IRIns *ir)
>    }
>  }
>  
> +#if !LJ_SOFTFP
>  #define asm_abs(as, ir)		asm_fpunary(as, ir, MIPSI_ABS_D)
> +#elif LJ_64   /* && LJ_SOFTFP */
> +static void asm_abs(ASMState *as, IRIns *ir)
> +{
> +  Reg dest = ra_dest(as, ir, RSET_GPR);
> +  Reg left = ra_alloc1(as, ir->op1, RSET_GPR);
> +  emit_tsml(as, MIPSI_DEXTM, dest, left, 30, 0);
> +}
> +#endif
> +
>  #define asm_atan2(as, ir)	asm_callid(as, ir, IRCALL_atan2)
>  #define asm_ldexp(as, ir)	asm_callid(as, ir, IRCALL_ldexp)
>  
> @@ -1924,15 +2038,21 @@ static void asm_bror(ASMState *as, IRIns *ir)
>    }
>  }
>  
> -#if LJ_32 && LJ_SOFTFP
> +#if LJ_SOFTFP
>  static void asm_sfpmin_max(ASMState *as, IRIns *ir)
>  {
>    CCallInfo ci = lj_ir_callinfo[(IROp)ir->o == IR_MIN ? IRCALL_lj_vm_sfmin : IRCALL_lj_vm_sfmax];
> +#if LJ_64
> +  IRRef args[2];
> +  args[0] = ir->op1;
> +  args[1] = ir->op2;
> +#else
>    IRRef args[4];
>    args[0^LJ_BE] = ir->op1;
>    args[1^LJ_BE] = (ir+1)->op1;
>    args[2^LJ_BE] = ir->op2;
>    args[3^LJ_BE] = (ir+1)->op2;
> +#endif
>    asm_setupresult(as, ir, &ci);
>    emit_call(as, (void *)ci.func, 0);
>    ci.func = NULL;
> @@ -1942,7 +2062,10 @@ static void asm_sfpmin_max(ASMState *as, IRIns *ir)
>  
>  static void asm_min_max(ASMState *as, IRIns *ir, int ismax)
>  {
> -  if (!LJ_SOFTFP && irt_isnum(ir->t)) {
> +  if (!LJ_SOFTFP32 && irt_isnum(ir->t)) {
> +#if LJ_SOFTFP
> +    asm_sfpmin_max(as, ir);
> +#else
>      Reg dest = ra_dest(as, ir, RSET_FPR);
>      Reg right, left = ra_alloc2(as, ir, RSET_FPR);
>      right = (left >> 8); left &= 255;
> @@ -1953,6 +2076,7 @@ static void asm_min_max(ASMState *as, IRIns *ir, int ismax)
>        if (dest != right) emit_fg(as, MIPSI_MOV_D, dest, right);
>      }
>      emit_fgh(as, MIPSI_C_OLT_D, 0, ismax ? left : right, ismax ? right : left);
> +#endif
>    } else {
>      Reg dest = ra_dest(as, ir, RSET_GPR);
>      Reg right, left = ra_alloc2(as, ir, RSET_GPR);
> @@ -1973,18 +2097,24 @@ static void asm_min_max(ASMState *as, IRIns *ir, int ismax)
>  
>  /* -- Comparisons --------------------------------------------------------- */
>  
> -#if LJ_32 && LJ_SOFTFP
> +#if LJ_SOFTFP
>  /* SFP comparisons. */
>  static void asm_sfpcomp(ASMState *as, IRIns *ir)
>  {
>    const CCallInfo *ci = &lj_ir_callinfo[IRCALL_softfp_cmp];
>    RegSet drop = RSET_SCRATCH;
>    Reg r;
> +#if LJ_64
> +  IRRef args[2];
> +  args[0] = ir->op1;
> +  args[1] = ir->op2;
> +#else
>    IRRef args[4];
>    args[LJ_LE ? 0 : 1] = ir->op1; args[LJ_LE ? 1 : 0] = (ir+1)->op1;
>    args[LJ_LE ? 2 : 3] = ir->op2; args[LJ_LE ? 3 : 2] = (ir+1)->op2;
> +#endif
>  
> -  for (r = REGARG_FIRSTGPR; r <= REGARG_FIRSTGPR+3; r++) {
> +  for (r = REGARG_FIRSTGPR; r <= REGARG_FIRSTGPR+(LJ_64?1:3); r++) {
>      if (!rset_test(as->freeset, r) &&
>  	regcost_ref(as->cost[r]) == args[r-REGARG_FIRSTGPR])
>        rset_clear(drop, r);
> @@ -2038,11 +2168,15 @@ static void asm_comp(ASMState *as, IRIns *ir)
>  {
>    /* ORDER IR: LT GE LE GT  ULT UGE ULE UGT. */
>    IROp op = ir->o;
> -  if (!LJ_SOFTFP && irt_isnum(ir->t)) {
> +  if (!LJ_SOFTFP32 && irt_isnum(ir->t)) {
> +#if LJ_SOFTFP
> +    asm_sfpcomp(as, ir);
> +#else
>      Reg right, left = ra_alloc2(as, ir, RSET_FPR);
>      right = (left >> 8); left &= 255;
>      asm_guard(as, (op&1) ? MIPSI_BC1T : MIPSI_BC1F, 0, 0);
>      emit_fgh(as, MIPSI_C_OLT_D + ((op&3) ^ ((op>>2)&1)), 0, left, right);
> +#endif
>    } else {
>      Reg right, left = ra_alloc1(as, ir->op1, RSET_GPR);
>      if (op == IR_ABC) op = IR_UGT;
> @@ -2074,9 +2208,13 @@ static void asm_equal(ASMState *as, IRIns *ir)
>    Reg right, left = ra_alloc2(as, ir, (!LJ_SOFTFP && irt_isnum(ir->t)) ?
>  				       RSET_FPR : RSET_GPR);
>    right = (left >> 8); left &= 255;
> -  if (!LJ_SOFTFP && irt_isnum(ir->t)) {
> +  if (!LJ_SOFTFP32 && irt_isnum(ir->t)) {
> +#if LJ_SOFTFP
> +    asm_sfpcomp(as, ir);
> +#else
>      asm_guard(as, (ir->o & 1) ? MIPSI_BC1T : MIPSI_BC1F, 0, 0);
>      emit_fgh(as, MIPSI_C_EQ_D, 0, left, right);
> +#endif
>    } else {
>      asm_guard(as, (ir->o & 1) ? MIPSI_BEQ : MIPSI_BNE, left, right);
>    }
> @@ -2269,7 +2407,7 @@ static void asm_stack_restore(ASMState *as, SnapShot *snap)
>      if ((sn & SNAP_NORESTORE))
>        continue;
>      if (irt_isnum(ir->t)) {
> -#if LJ_SOFTFP
> +#if LJ_SOFTFP32
>        Reg tmp;
>        RegSet allow = rset_exclude(RSET_GPR, RID_BASE);
>        lua_assert(irref_isk(ref));  /* LJ_SOFTFP: must be a number constant. */
> @@ -2278,6 +2416,9 @@ static void asm_stack_restore(ASMState *as, SnapShot *snap)
>        if (rset_test(as->freeset, tmp+1)) allow = RID2RSET(tmp+1);
>        tmp = ra_allock(as, (int32_t)ir_knum(ir)->u32.hi, allow);
>        emit_tsi(as, MIPSI_SW, tmp, RID_BASE, ofs+(LJ_BE?0:4));
> +#elif LJ_SOFTFP  /* && LJ_64 */
> +      Reg src = ra_alloc1(as, ref, rset_exclude(RSET_GPR, RID_BASE));
> +      emit_tsi(as, MIPSI_SD, src, RID_BASE, ofs);
>  #else
>        Reg src = ra_alloc1(as, ref, RSET_FPR);
>        emit_hsi(as, MIPSI_SDC1, src, RID_BASE, ofs);
> diff --git a/src/lj_crecord.c b/src/lj_crecord.c
> index ffe995f4..804cdbf4 100644
> --- a/src/lj_crecord.c
> +++ b/src/lj_crecord.c
> @@ -212,7 +212,7 @@ static void crec_copy_emit(jit_State *J, CRecMemList *ml, MSize mlp,
>      ml[i].trval = emitir(IRT(IR_XLOAD, ml[i].tp), trsptr, 0);
>      ml[i].trofs = trofs;
>      i++;
> -    rwin += (LJ_SOFTFP && ml[i].tp == IRT_NUM) ? 2 : 1;
> +    rwin += (LJ_SOFTFP32 && ml[i].tp == IRT_NUM) ? 2 : 1;
>      if (rwin >= CREC_COPY_REGWIN || i >= mlp) {  /* Flush buffered stores. */
>        rwin = 0;
>        for ( ; j < i; j++) {
> @@ -1152,7 +1152,7 @@ static TRef crec_call_args(jit_State *J, RecordFFData *rd,
>  	else
>  	  tr = emitconv(tr, IRT_INT, d->size==1 ? IRT_I8 : IRT_I16,IRCONV_SEXT);
>        }
> -    } else if (LJ_SOFTFP && ctype_isfp(d->info) && d->size > 4) {
> +    } else if (LJ_SOFTFP32 && ctype_isfp(d->info) && d->size > 4) {
>        lj_needsplit(J);
>      }
>  #if LJ_TARGET_X86
> diff --git a/src/lj_emit_mips.h b/src/lj_emit_mips.h
> index 8a9ee24d..bb6593ae 100644
> --- a/src/lj_emit_mips.h
> +++ b/src/lj_emit_mips.h
> @@ -12,6 +12,8 @@ static intptr_t get_k64val(IRIns *ir)
>      return (intptr_t)ir_kgc(ir);
>    } else if (ir->o == IR_KPTR || ir->o == IR_KKPTR) {
>      return (intptr_t)ir_kptr(ir);
> +  } else if (LJ_SOFTFP && ir->o == IR_KNUM) {
> +    return (intptr_t)ir_knum(ir)->u64;
>    } else {
>      lua_assert(ir->o == IR_KINT || ir->o == IR_KNULL);
>      return ir->i;  /* Sign-extended. */
> diff --git a/src/lj_ffrecord.c b/src/lj_ffrecord.c
> index 8af9da1d..0746ec64 100644
> --- a/src/lj_ffrecord.c
> +++ b/src/lj_ffrecord.c
> @@ -986,7 +986,7 @@ static void LJ_FASTCALL recff_string_format(jit_State *J, RecordFFData *rd)
>      handle_num:
>        tra = lj_ir_tonum(J, tra);
>        tr = lj_ir_call(J, id, tr, trsf, tra);
> -      if (LJ_SOFTFP) lj_needsplit(J);
> +      if (LJ_SOFTFP32) lj_needsplit(J);
>        break;
>      case STRFMT_STR:
>        if (!tref_isstr(tra)) {
> diff --git a/src/lj_ircall.h b/src/lj_ircall.h
> index aa06b273..c1ac29d1 100644
> --- a/src/lj_ircall.h
> +++ b/src/lj_ircall.h
> @@ -52,7 +52,7 @@ typedef struct CCallInfo {
>  #define CCI_XARGS(ci)		(((ci)->flags >> CCI_XARGS_SHIFT) & 3)
>  #define CCI_XA			(1u << CCI_XARGS_SHIFT)
>  
> -#if LJ_SOFTFP || (LJ_32 && LJ_HASFFI)
> +#if LJ_SOFTFP32 || (LJ_32 && LJ_HASFFI)
>  #define CCI_XNARGS(ci)		(CCI_NARGS((ci)) + CCI_XARGS((ci)))
>  #else
>  #define CCI_XNARGS(ci)		CCI_NARGS((ci))
> @@ -79,13 +79,19 @@ typedef struct CCallInfo {
>  #define IRCALLCOND_SOFTFP_FFI(x)	NULL
>  #endif
>  
> -#if LJ_SOFTFP && LJ_TARGET_MIPS32
> +#if LJ_SOFTFP && LJ_TARGET_MIPS
>  #define IRCALLCOND_SOFTFP_MIPS(x)	x
>  #else
>  #define IRCALLCOND_SOFTFP_MIPS(x)	NULL
>  #endif
>  
> -#define LJ_NEED_FP64	(LJ_TARGET_ARM || LJ_TARGET_PPC || LJ_TARGET_MIPS32)
> +#if LJ_SOFTFP && LJ_TARGET_MIPS64
> +#define IRCALLCOND_SOFTFP_MIPS64(x)	x
> +#else
> +#define IRCALLCOND_SOFTFP_MIPS64(x)	NULL
> +#endif
> +
> +#define LJ_NEED_FP64	(LJ_TARGET_ARM || LJ_TARGET_PPC || LJ_TARGET_MIPS)
>  
>  #if LJ_HASFFI && (LJ_SOFTFP || LJ_NEED_FP64)
>  #define IRCALLCOND_FP64_FFI(x)		x
> @@ -113,6 +119,14 @@ typedef struct CCallInfo {
>  #define XA2_FP		0
>  #endif
>  
> +#if LJ_SOFTFP32
> +#define XA_FP32		CCI_XA
> +#define XA2_FP32	(CCI_XA+CCI_XA)
> +#else
> +#define XA_FP32		0
> +#define XA2_FP32	0
> +#endif
> +
>  #if LJ_32
>  #define XA_64		CCI_XA
>  #define XA2_64		(CCI_XA+CCI_XA)
> @@ -185,20 +199,21 @@ typedef struct CCallInfo {
>    _(ANY,	pow,			2,   N, NUM, XA2_FP) \
>    _(ANY,	atan2,			2,   N, NUM, XA2_FP) \
>    _(ANY,	ldexp,			2,   N, NUM, XA_FP) \
> -  _(SOFTFP,	lj_vm_tobit,		2,   N, INT, 0) \
> -  _(SOFTFP,	softfp_add,		4,   N, NUM, 0) \
> -  _(SOFTFP,	softfp_sub,		4,   N, NUM, 0) \
> -  _(SOFTFP,	softfp_mul,		4,   N, NUM, 0) \
> -  _(SOFTFP,	softfp_div,		4,   N, NUM, 0) \
> -  _(SOFTFP,	softfp_cmp,		4,   N, NIL, 0) \
> +  _(SOFTFP,	lj_vm_tobit,		1,   N, INT, XA_FP32) \
> +  _(SOFTFP,	softfp_add,		2,   N, NUM, XA2_FP32) \
> +  _(SOFTFP,	softfp_sub,		2,   N, NUM, XA2_FP32) \
> +  _(SOFTFP,	softfp_mul,		2,   N, NUM, XA2_FP32) \
> +  _(SOFTFP,	softfp_div,		2,   N, NUM, XA2_FP32) \
> +  _(SOFTFP,	softfp_cmp,		2,   N, NIL, XA2_FP32) \
>    _(SOFTFP,	softfp_i2d,		1,   N, NUM, 0) \
> -  _(SOFTFP,	softfp_d2i,		2,   N, INT, 0) \
> -  _(SOFTFP_MIPS, lj_vm_sfmin,		4,   N, NUM, 0) \
> -  _(SOFTFP_MIPS, lj_vm_sfmax,		4,   N, NUM, 0) \
> +  _(SOFTFP,	softfp_d2i,		1,   N, INT, XA_FP32) \
> +  _(SOFTFP_MIPS, lj_vm_sfmin,		2,   N, NUM, XA2_FP32) \
> +  _(SOFTFP_MIPS, lj_vm_sfmax,		2,   N, NUM, XA2_FP32) \
> +  _(SOFTFP_MIPS64, lj_vm_tointg,	1,   N, INT, 0) \
>    _(SOFTFP_FFI,	softfp_ui2d,		1,   N, NUM, 0) \
>    _(SOFTFP_FFI,	softfp_f2d,		1,   N, NUM, 0) \
> -  _(SOFTFP_FFI,	softfp_d2ui,		2,   N, INT, 0) \
> -  _(SOFTFP_FFI,	softfp_d2f,		2,   N, FLOAT, 0) \
> +  _(SOFTFP_FFI,	softfp_d2ui,		1,   N, INT, XA_FP32) \
> +  _(SOFTFP_FFI,	softfp_d2f,		1,   N, FLOAT, XA_FP32) \
>    _(SOFTFP_FFI,	softfp_i2f,		1,   N, FLOAT, 0) \
>    _(SOFTFP_FFI,	softfp_ui2f,		1,   N, FLOAT, 0) \
>    _(SOFTFP_FFI,	softfp_f2i,		1,   N, INT, 0) \
> diff --git a/src/lj_iropt.h b/src/lj_iropt.h
> index 73aef0ef..a59ba3f4 100644
> --- a/src/lj_iropt.h
> +++ b/src/lj_iropt.h
> @@ -150,7 +150,7 @@ LJ_FUNC IRType lj_opt_narrow_forl(jit_State *J, cTValue *forbase);
>  /* Optimization passes. */
>  LJ_FUNC void lj_opt_dce(jit_State *J);
>  LJ_FUNC int lj_opt_loop(jit_State *J);
> -#if LJ_SOFTFP || (LJ_32 && LJ_HASFFI)
> +#if LJ_SOFTFP32 || (LJ_32 && LJ_HASFFI)
>  LJ_FUNC void lj_opt_split(jit_State *J);
>  #else
>  #define lj_opt_split(J)		UNUSED(J)
> diff --git a/src/lj_jit.h b/src/lj_jit.h
> index cc8efd20..c06829ab 100644
> --- a/src/lj_jit.h
> +++ b/src/lj_jit.h
> @@ -375,7 +375,7 @@ enum {
>    ((TValue *)(((intptr_t)&J->ksimd[2*(n)] + 15) & ~(intptr_t)15))
>  
>  /* Set/reset flag to activate the SPLIT pass for the current trace. */
> -#if LJ_SOFTFP || (LJ_32 && LJ_HASFFI)
> +#if LJ_SOFTFP32 || (LJ_32 && LJ_HASFFI)
>  #define lj_needsplit(J)		(J->needsplit = 1)
>  #define lj_resetsplit(J)	(J->needsplit = 0)
>  #else
> @@ -438,7 +438,7 @@ typedef struct jit_State {
>    MSize sizesnapmap;	/* Size of temp. snapshot map buffer. */
>  
>    PostProc postproc;	/* Required post-processing after execution. */
> -#if LJ_SOFTFP || (LJ_32 && LJ_HASFFI)
> +#if LJ_SOFTFP32 || (LJ_32 && LJ_HASFFI)
>    uint8_t needsplit;	/* Need SPLIT pass. */
>  #endif
>    uint8_t retryrec;	/* Retry recording. */
> diff --git a/src/lj_obj.h b/src/lj_obj.h
> index 45507e0d..bf95e1eb 100644
> --- a/src/lj_obj.h
> +++ b/src/lj_obj.h
> @@ -984,6 +984,9 @@ static LJ_AINLINE void copyTV(lua_State *L, TValue *o1, const TValue *o2)
>  
>  #if LJ_SOFTFP
>  LJ_ASMF int32_t lj_vm_tobit(double x);
> +#if LJ_TARGET_MIPS64
> +LJ_ASMF int32_t lj_vm_tointg(double x);
> +#endif
>  #endif
>  
>  static LJ_AINLINE int32_t lj_num2bit(lua_Number n)
> diff --git a/src/lj_opt_split.c b/src/lj_opt_split.c
> index c0788106..2fc36b8d 100644
> --- a/src/lj_opt_split.c
> +++ b/src/lj_opt_split.c
> @@ -8,7 +8,7 @@
>  
>  #include "lj_obj.h"
>  
> -#if LJ_HASJIT && (LJ_SOFTFP || (LJ_32 && LJ_HASFFI))
> +#if LJ_HASJIT && (LJ_SOFTFP32 || (LJ_32 && LJ_HASFFI))
>  
>  #include "lj_err.h"
>  #include "lj_buf.h"
> diff --git a/src/lj_snap.c b/src/lj_snap.c
> index a063c316..9146cddc 100644
> --- a/src/lj_snap.c
> +++ b/src/lj_snap.c
> @@ -93,7 +93,7 @@ static MSize snapshot_slots(jit_State *J, SnapEntry *map, BCReg nslots)
>  	    (ir->op2 & (IRSLOAD_READONLY|IRSLOAD_PARENT)) != IRSLOAD_PARENT)
>  	  sn |= SNAP_NORESTORE;
>        }
> -      if (LJ_SOFTFP && irt_isnum(ir->t))
> +      if (LJ_SOFTFP32 && irt_isnum(ir->t))
>  	sn |= SNAP_SOFTFPNUM;
>        map[n++] = sn;
>      }
> @@ -379,7 +379,7 @@ IRIns *lj_snap_regspmap(GCtrace *T, SnapNo snapno, IRIns *ir)
>  	  break;
>  	}
>        }
> -    } else if (LJ_SOFTFP && ir->o == IR_HIOP) {
> +    } else if (LJ_SOFTFP32 && ir->o == IR_HIOP) {
>        ref++;
>      } else if (ir->o == IR_PVAL) {
>        ref = ir->op1 + REF_BIAS;
> @@ -491,7 +491,7 @@ void lj_snap_replay(jit_State *J, GCtrace *T)
>      } else {
>        IRType t = irt_type(ir->t);
>        uint32_t mode = IRSLOAD_INHERIT|IRSLOAD_PARENT;
> -      if (LJ_SOFTFP && (sn & SNAP_SOFTFPNUM)) t = IRT_NUM;
> +      if (LJ_SOFTFP32 && (sn & SNAP_SOFTFPNUM)) t = IRT_NUM;
>        if (ir->o == IR_SLOAD) mode |= (ir->op2 & IRSLOAD_READONLY);
>        tr = emitir_raw(IRT(IR_SLOAD, t), s, mode);
>      }
> @@ -525,7 +525,7 @@ void lj_snap_replay(jit_State *J, GCtrace *T)
>  	    if (irs->r == RID_SINK && snap_sunk_store(T, ir, irs)) {
>  	      if (snap_pref(J, T, map, nent, seen, irs->op2) == 0)
>  		snap_pref(J, T, map, nent, seen, T->ir[irs->op2].op1);
> -	      else if ((LJ_SOFTFP || (LJ_32 && LJ_HASFFI)) &&
> +	      else if ((LJ_SOFTFP32 || (LJ_32 && LJ_HASFFI)) &&
>  		       irs+1 < irlast && (irs+1)->o == IR_HIOP)
>  		snap_pref(J, T, map, nent, seen, (irs+1)->op2);
>  	    }
> @@ -584,10 +584,10 @@ void lj_snap_replay(jit_State *J, GCtrace *T)
>  		lua_assert(irc->o == IR_CONV && irc->op2 == IRCONV_NUM_INT);
>  		val = snap_pref(J, T, map, nent, seen, irc->op1);
>  		val = emitir(IRTN(IR_CONV), val, IRCONV_NUM_INT);
> -	      } else if ((LJ_SOFTFP || (LJ_32 && LJ_HASFFI)) &&
> +	      } else if ((LJ_SOFTFP32 || (LJ_32 && LJ_HASFFI)) &&
>  			 irs+1 < irlast && (irs+1)->o == IR_HIOP) {
>  		IRType t = IRT_I64;
> -		if (LJ_SOFTFP && irt_type((irs+1)->t) == IRT_SOFTFP)
> +		if (LJ_SOFTFP32 && irt_type((irs+1)->t) == IRT_SOFTFP)
>  		  t = IRT_NUM;
>  		lj_needsplit(J);
>  		if (irref_isk(irs->op2) && irref_isk((irs+1)->op2)) {
> @@ -645,7 +645,7 @@ static void snap_restoreval(jit_State *J, GCtrace *T, ExitState *ex,
>      int32_t *sps = &ex->spill[regsp_spill(rs)];
>      if (irt_isinteger(t)) {
>        setintV(o, *sps);
> -#if !LJ_SOFTFP
> +#if !LJ_SOFTFP32
>      } else if (irt_isnum(t)) {
>        o->u64 = *(uint64_t *)sps;
>  #endif
> @@ -670,6 +670,9 @@ static void snap_restoreval(jit_State *J, GCtrace *T, ExitState *ex,
>  #if !LJ_SOFTFP
>      } else if (irt_isnum(t)) {
>        setnumV(o, ex->fpr[r-RID_MIN_FPR]);
> +#elif LJ_64  /* && LJ_SOFTFP */
> +    } else if (irt_isnum(t)) {
> +      o->u64 = ex->gpr[r-RID_MIN_GPR];
>  #endif
>  #if LJ_64 && !LJ_GC64
>      } else if (irt_is64(t)) {
> @@ -823,7 +826,7 @@ static void snap_unsink(jit_State *J, GCtrace *T, ExitState *ex,
>  	  val = lj_tab_set(J->L, t, &tmp);
>  	  /* NOBARRIER: The table is new (marked white). */
>  	  snap_restoreval(J, T, ex, snapno, rfilt, irs->op2, val);
> -	  if (LJ_SOFTFP && irs+1 < T->ir + T->nins && (irs+1)->o == IR_HIOP) {
> +	  if (LJ_SOFTFP32 && irs+1 < T->ir + T->nins && (irs+1)->o == IR_HIOP) {
>  	    snap_restoreval(J, T, ex, snapno, rfilt, (irs+1)->op2, &tmp);
>  	    val->u32.hi = tmp.u32.lo;
>  	  }
> @@ -884,7 +887,7 @@ const BCIns *lj_snap_restore(jit_State *J, void *exptr)
>  	continue;
>        }
>        snap_restoreval(J, T, ex, snapno, rfilt, ref, o);
> -      if (LJ_SOFTFP && (sn & SNAP_SOFTFPNUM) && tvisint(o)) {
> +      if (LJ_SOFTFP32 && (sn & SNAP_SOFTFPNUM) && tvisint(o)) {
>  	TValue tmp;
>  	snap_restoreval(J, T, ex, snapno, rfilt, ref+1, &tmp);
>  	o->u32.hi = tmp.u32.lo;
> diff --git a/src/vm_mips64.dasc b/src/vm_mips64.dasc
> index 04be38f0..9839b5ac 100644
> --- a/src/vm_mips64.dasc
> +++ b/src/vm_mips64.dasc
> @@ -1984,6 +1984,38 @@ static void build_subroutines(BuildCtx *ctx)
>    |1:
>    |  jr ra
>    |.  move CRET1, r0
> +  |
> +  |// FP number to int conversion with a check for soft-float.
> +  |// Modifies CARG1, CRET1, CRET2, TMP0, AT.
> +  |->vm_tointg:
> +  |.if JIT
> +  |  dsll CRET2, CARG1, 1
> +  |  beqz CRET2, >2
> +  |.  li TMP0, 1076
> +  |  dsrl AT, CRET2, 53
> +  |  dsubu TMP0, TMP0, AT
> +  |  sltiu AT, TMP0, 54
> +  |  beqz AT, >1
> +  |.  dextm CRET2, CRET2, 0, 20
> +  |  dinsu CRET2, AT, 21, 21
> +  |  slt AT, CARG1, r0
> +  |  dsrlv CRET1, CRET2, TMP0
> +  |  dsubu CARG1, r0, CRET1
> +  |  movn CRET1, CARG1, AT
> +  |  li CARG1, 64
> +  |  subu TMP0, CARG1, TMP0
> +  |  dsllv CRET2, CRET2, TMP0	// Integer check.
> +  |  sextw AT, CRET1
> +  |  xor AT, CRET1, AT		// Range check.
> +  |  jr ra
> +  |.  movz CRET2, AT, CRET2
> +  |1:
> +  |  jr ra
> +  |.  li CRET2, 1
> +  |2:
> +  |  jr ra
> +  |.  move CRET1, r0
> +  |.endif
>    |.endif
>    |
>    |.macro .ffunc_bit, name
> @@ -2669,6 +2701,23 @@ static void build_subroutines(BuildCtx *ctx)
>    |.  li CRET1, 0
>    |.endif
>    |
> +  |.macro sfmin_max, name, intins
> +  |->vm_sf .. name:
> +  |.if JIT and not FPU
> +  |  move TMP2, ra
> +  |  bal ->vm_sfcmpolt
> +  |.  nop
> +  |  move ra, TMP2
> +  |  move TMP0, CRET1
> +  |  move CRET1, CARG1
> +  |  jr ra
> +  |.  intins CRET1, CARG2, TMP0
> +  |.endif
> +  |.endmacro
> +  |
> +  |  sfmin_max min, movz
> +  |  sfmin_max max, movn
> +  |
>    |//-----------------------------------------------------------------------
>    |//-- Miscellaneous functions --------------------------------------------
>    |//-----------------------------------------------------------------------
> -- 
> 2.41.0
>