diff --git a/src/lj_bcread.c b/src/lj_bcread.c
index fcc2aa1d..7bb16a60 100644
--- a/src/lj_bcread.c
+++ b/src/lj_bcread.c
@@ -59,7 +59,7 @@ static LJ_NOINLINE void bcread_fill(LexState *ls, MSize len, int need)
 	lua_assert(ls->pe == sbufP(&ls->sb));
 	if (ls->p != p) memmove(p, ls->p, n);
       } else {  /* Copy from buffer provided by reader. */
-	p = lj_buf_need(ls->L, &ls->sb, len);
+	p = lj_buf_need(&ls->sb, len);
 	memcpy(p, ls->p, n);
       }
       ls->p = p;
@@ -74,7 +74,7 @@ static LJ_NOINLINE void bcread_fill(LexState *ls, MSize len, int need)
     }
     if (n) {  /* Append to buffer. */
       n += (MSize)sz;
-      p = lj_buf_need(ls->L, &ls->sb, n < len ? len : n);
+      p = lj_buf_need(&ls->sb, n < len ? len : n);
       memcpy(sbufP(&ls->sb), buf, sz);
       setsbufP(&ls->sb, p + n);
       ls->p = p;
diff --git a/src/lj_bcwrite.c b/src/lj_bcwrite.c
index 71c86581..b3289a13 100644
--- a/src/lj_bcwrite.c
+++ b/src/lj_bcwrite.c
@@ -24,7 +24,6 @@
 /* Context for bytecode writer. */
 typedef struct BCWriteCtx {
   SBuf sb;			/* Output buffer. */
-  lua_State *L;			/* Lua state. */
   GCproto *pt;			/* Root prototype. */
   lua_Writer wfunc;		/* Writer callback. */
   void *wdata;			/* Writer callback data. */
@@ -37,11 +36,11 @@ typedef struct BCWriteCtx {
 /* Write a single constant key/value of a template table. */
 static void bcwrite_ktabk(BCWriteCtx *ctx, cTValue *o, int narrow)
 {
-  char *p = lj_buf_more(ctx->L, &ctx->sb, 1+10);
+  char *p = lj_buf_more(&ctx->sb, 1+10);
   if (tvisstr(o)) {
     const GCstr *str = strV(o);
     MSize len = str->len;
-    p = lj_buf_more(ctx->L, &ctx->sb, 5+len);
+    p = lj_buf_more(&ctx->sb, 5+len);
     p = lj_buf_wuleb128(p, BCDUMP_KTAB_STR+len);
     p = lj_buf_wmem(p, strdata(str), len);
   } else if (tvisint(o)) {
@@ -143,7 +142,7 @@ static void bcwrite_kgc(BCWriteCtx *ctx, GCproto *pt)
       need = 1+2*5;
     }
     /* Write constant type. */
-    p = lj_buf_more(ctx->L, &ctx->sb, need);
+    p = lj_buf_more(&ctx->sb, need);
     p = lj_buf_wuleb128(p, tp);
     /* Write constant data (if any). */
     if (tp >= BCDUMP_KGC_STR) {
@@ -171,7 +170,7 @@ static void bcwrite_knum(BCWriteCtx *ctx, GCproto *pt)
 {
   MSize i, sizekn = pt->sizekn;
   cTValue *o = mref(pt->k, TValue);
-  char *p = lj_buf_more(ctx->L, &ctx->sb, 10*sizekn);
+  char *p = lj_buf_more(&ctx->sb, 10*sizekn);
   for (i = 0; i < sizekn; i++, o++) {
     int32_t k;
     if (tvisint(o)) {
@@ -211,7 +210,7 @@ static char *bcwrite_bytecode(BCWriteCtx *ctx, char *p, GCproto *pt)
 #if LJ_HASJIT
   /* Unpatch modified bytecode containing ILOOP/JLOOP etc. */
   if ((pt->flags & PROTO_ILOOP) || pt->trace) {
-    jit_State *J = L2J(ctx->L);
+    jit_State *J = L2J(sbufL(&ctx->sb));
     MSize i;
     for (i = 0; i < nbc; i++, q += sizeof(BCIns)) {
       BCOp op = (BCOp)q[LJ_ENDIAN_SELECT(0, 3)];
@@ -249,7 +248,7 @@ static void bcwrite_proto(BCWriteCtx *ctx, GCproto *pt)
   }
 
   /* Start writing the prototype info to a buffer. */
-  p = lj_buf_need(ctx->L, &ctx->sb,
+  p = lj_buf_need(&ctx->sb,
 		  5+4+6*5+(pt->sizebc-1)*(MSize)sizeof(BCIns)+pt->sizeuv*2);
   p += 5;  /* Leave room for final size. */
 
@@ -282,7 +281,7 @@ static void bcwrite_proto(BCWriteCtx *ctx, GCproto *pt)
 
   /* Write debug info, if not stripped. */
   if (sizedbg) {
-    p = lj_buf_more(ctx->L, &ctx->sb, sizedbg);
+    p = lj_buf_more(&ctx->sb, sizedbg);
     p = lj_buf_wmem(p, proto_lineinfo(pt), sizedbg);
     setsbufP(&ctx->sb, p);
   }
@@ -294,7 +293,7 @@ static void bcwrite_proto(BCWriteCtx *ctx, GCproto *pt)
     char *q = sbufB(&ctx->sb) + (5 - nn);
     p = lj_buf_wuleb128(q, n);  /* Fill in final size. */
     lua_assert(p == sbufB(&ctx->sb) + 5);
-    ctx->status = ctx->wfunc(ctx->L, q, nn+n, ctx->wdata);
+    ctx->status = ctx->wfunc(sbufL(&ctx->sb), q, nn+n, ctx->wdata);
   }
 }
 
@@ -304,7 +303,7 @@ static void bcwrite_header(BCWriteCtx *ctx)
   GCstr *chunkname = proto_chunkname(ctx->pt);
   const char *name = strdata(chunkname);
   MSize len = chunkname->len;
-  char *p = lj_buf_need(ctx->L, &ctx->sb, 5+5+len);
+  char *p = lj_buf_need(&ctx->sb, 5+5+len);
   *p++ = BCDUMP_HEAD1;
   *p++ = BCDUMP_HEAD2;
   *p++ = BCDUMP_HEAD3;
@@ -316,7 +315,7 @@ static void bcwrite_header(BCWriteCtx *ctx)
     p = lj_buf_wuleb128(p, len);
     p = lj_buf_wmem(p, name, len);
   }
-  ctx->status = ctx->wfunc(ctx->L, sbufB(&ctx->sb),
+  ctx->status = ctx->wfunc(sbufL(&ctx->sb), sbufB(&ctx->sb),
 			   (MSize)(p - sbufB(&ctx->sb)), ctx->wdata);
 }
 
@@ -325,7 +324,7 @@ static void bcwrite_footer(BCWriteCtx *ctx)
 {
   if (ctx->status == 0) {
     uint8_t zero = 0;
-    ctx->status = ctx->wfunc(ctx->L, &zero, 1, ctx->wdata);
+    ctx->status = ctx->wfunc(sbufL(&ctx->sb), &zero, 1, ctx->wdata);
   }
 }
 
@@ -333,8 +332,8 @@ static void bcwrite_footer(BCWriteCtx *ctx)
 static TValue *cpwriter(lua_State *L, lua_CFunction dummy, void *ud)
 {
   BCWriteCtx *ctx = (BCWriteCtx *)ud;
-  UNUSED(dummy);
-  lj_buf_need(L, &ctx->sb, 1024);  /* Avoids resize for most prototypes. */
+  UNUSED(L); UNUSED(dummy);
+  lj_buf_need(&ctx->sb, 1024);  /* Avoids resize for most prototypes. */
   bcwrite_header(ctx);
   bcwrite_proto(ctx, ctx->pt);
   bcwrite_footer(ctx);
@@ -347,16 +346,15 @@ int lj_bcwrite(lua_State *L, GCproto *pt, lua_Writer writer, void *data,
 {
   BCWriteCtx ctx;
   int status;
-  ctx.L = L;
   ctx.pt = pt;
   ctx.wfunc = writer;
   ctx.wdata = data;
   ctx.strip = strip;
   ctx.status = 0;
-  lj_buf_init(&ctx.sb);
+  lj_buf_init(L, &ctx.sb);
   status = lj_vm_cpcall(L, NULL, &ctx, cpwriter);
   if (status == 0) status = ctx.status;
-  lj_buf_free(G(ctx.L), &ctx.sb);
+  lj_buf_free(G(sbufL(&ctx.sb)), &ctx.sb);
   return status;
 }
 
diff --git a/src/lj_buf.c b/src/lj_buf.c
index a05dc22e..c08d23c9 100644
--- a/src/lj_buf.c
+++ b/src/lj_buf.c
@@ -13,8 +13,9 @@
 #include "lj_err.h"
 #include "lj_buf.h"
 
-LJ_NOINLINE void lj_buf_grow(lua_State *L, SBuf *sb, char *en)
+LJ_NOINLINE void LJ_FASTCALL lj_buf_grow(SBuf *sb, char *en)
 {
+  lua_State *L = sbufL(sb);
   char *b = sbufB(sb);
   MSize sz = (MSize)(en - b);
   MSize osz = (MSize)(sbufE(sb) - b), nsz = osz;
@@ -29,12 +30,14 @@ LJ_NOINLINE void lj_buf_grow(lua_State *L, SBuf *sb, char *en)
   setmref(sb->e, b + nsz);
 }
 
-char *lj_buf_tmp(lua_State *L, MSize sz)
+char * LJ_FASTCALL lj_buf_tmp(lua_State *L, MSize sz)
 {
-  return lj_buf_need(L, &G(L)->tmpbuf, sz);
+  SBuf *sb = &G(L)->tmpbuf;
+  setmref(sb->L, L);
+  return lj_buf_need(sb, sz);
 }
 
-void lj_buf_shrink(lua_State *L, SBuf *sb)
+void LJ_FASTCALL lj_buf_shrink(lua_State *L, SBuf *sb)
 {
   char *b = sbufB(sb);
   MSize osz = (MSize)(sbufE(sb) - b);
@@ -54,14 +57,14 @@ char *lj_buf_wmem(char *p, const void *q, MSize len)
   return p;
 }
 
-void lj_buf_putmem(lua_State *L, SBuf *sb, const void *q, MSize len)
+void lj_buf_putmem(SBuf *sb, const void *q, MSize len)
 {
-  char *p = lj_buf_more(L, sb, len);
+  char *p = lj_buf_more(sb, len);
   p = lj_buf_wmem(p, q, len);
   setsbufP(sb, p);
 }
 
-uint32_t lj_buf_ruleb128(const char **pp)
+uint32_t LJ_FASTCALL lj_buf_ruleb128(const char **pp)
 {
   const uint8_t *p = (const uint8_t *)*pp;
   uint32_t v = *p++;
@@ -74,7 +77,7 @@ uint32_t lj_buf_ruleb128(const char **pp)
   return v;
 }
 
-char *lj_buf_wuleb128(char *p, uint32_t v)
+char * LJ_FASTCALL lj_buf_wuleb128(char *p, uint32_t v)
 {
   for (; v >= 0x80; v >>= 7)
     *p++ = (char)((v & 0x7f) | 0x80);
diff --git a/src/lj_buf.h b/src/lj_buf.h
index 5a7eae31..289eb01d 100644
--- a/src/lj_buf.h
+++ b/src/lj_buf.h
@@ -14,21 +14,24 @@
 #define sbufB(sb)	(mref((sb)->b, char))
 #define sbufP(sb)	(mref((sb)->p, char))
 #define sbufE(sb)	(mref((sb)->e, char))
+#define sbufL(sb)	(mref((sb)->L, lua_State))
 #define sbufsz(sb)	((MSize)(sbufE((sb)) - sbufB((sb))))
 #define sbuflen(sb)	((MSize)(sbufP((sb)) - sbufB((sb))))
 #define setsbufP(sb, q)	(setmref((sb)->p, (q)))
+#define setsbufL(sb, l)	(setmref((sb)->L, (l)))
 
-LJ_FUNC char *lj_buf_tmp(lua_State *L, MSize sz);
-LJ_FUNC void lj_buf_grow(lua_State *L, SBuf *sb, char *en);
-LJ_FUNC void lj_buf_shrink(lua_State *L, SBuf *sb);
+LJ_FUNC char * LJ_FASTCALL lj_buf_tmp(lua_State *L, MSize sz);
+LJ_FUNC void LJ_FASTCALL lj_buf_grow(SBuf *sb, char *en);
+LJ_FUNC void LJ_FASTCALL lj_buf_shrink(lua_State *L, SBuf *sb);
 
 LJ_FUNC char *lj_buf_wmem(char *p, const void *q, MSize len);
-LJ_FUNC void lj_buf_putmem(lua_State *L, SBuf *sb, const void *q, MSize len);
-LJ_FUNC uint32_t lj_buf_ruleb128(const char **pp);
-LJ_FUNC char *lj_buf_wuleb128(char *p, uint32_t v);
+LJ_FUNC void lj_buf_putmem(SBuf *sb, const void *q, MSize len);
+LJ_FUNC uint32_t LJ_FASTCALL lj_buf_ruleb128(const char **pp);
+LJ_FUNC char * LJ_FASTCALL lj_buf_wuleb128(char *p, uint32_t v);
 
-static LJ_AINLINE void lj_buf_init(SBuf *sb)
+static LJ_AINLINE void lj_buf_init(lua_State *L, SBuf *sb)
 {
+  setsbufL(sb, L);
   setmref(sb->p, NULL); setmref(sb->e, NULL); setmref(sb->b, NULL);
 }
 
@@ -47,25 +50,25 @@ static LJ_AINLINE GCstr *lj_buf_str(lua_State *L, SBuf *sb)
   return lj_str_new(L, sbufB(sb), sbuflen(sb));
 }
 
-static LJ_AINLINE char *lj_buf_need(lua_State *L, SBuf *sb, MSize sz)
+static LJ_AINLINE char *lj_buf_need(SBuf *sb, MSize sz)
 {
   char *en = sbufB(sb) + sz;
   if (LJ_UNLIKELY(en > sbufE(sb)))
-    lj_buf_grow(L, sb, en);
+    lj_buf_grow(sb, en);
   return sbufB(sb);
 }
 
-static LJ_AINLINE char *lj_buf_more(lua_State *L, SBuf *sb, MSize sz)
+static LJ_AINLINE char *lj_buf_more(SBuf *sb, MSize sz)
 {
   char *en = sbufP(sb) + sz;
   if (LJ_UNLIKELY(en > sbufE(sb)))
-    lj_buf_grow(L, sb, en);
+    lj_buf_grow(sb, en);
   return sbufP(sb);
 }
 
-static LJ_AINLINE void lj_buf_putb(lua_State *L, SBuf *sb, int c)
+static LJ_AINLINE void lj_buf_putb(SBuf *sb, int c)
 {
-  char *p = lj_buf_more(L, sb, 1);
+  char *p = lj_buf_more(sb, 1);
   *p++ = (char)c;
   setsbufP(sb, p);
 }
diff --git a/src/lj_cparse.c b/src/lj_cparse.c
index b8e31820..b8c95bd3 100644
--- a/src/lj_cparse.c
+++ b/src/lj_cparse.c
@@ -89,7 +89,7 @@ static LJ_AINLINE CPChar cp_get(CPState *cp)
 /* Save character in buffer. */
 static LJ_AINLINE void cp_save(CPState *cp, CPChar c)
 {
-  lj_buf_putb(cp->L, &cp->sb, c);
+  lj_buf_putb(&cp->sb, c);
 }
 
 /* Skip line break. Handles "\n", "\r", "\r\n" or "\n\r". */
@@ -367,7 +367,7 @@ static void cp_init(CPState *cp)
   cp->depth = 0;
   cp->curpack = 0;
   cp->packstack[0] = 255;
-  lj_buf_init(&cp->sb);
+  lj_buf_init(cp->L, &cp->sb);
   lua_assert(cp->p != NULL);
   cp_get(cp);  /* Read-ahead first char. */
   cp->tok = 0;
diff --git a/src/lj_lex.c b/src/lj_lex.c
index 5a8f1003..c988a6c1 100644
--- a/src/lj_lex.c
+++ b/src/lj_lex.c
@@ -61,7 +61,7 @@ static LJ_AINLINE LexChar lex_next(LexState *ls)
 /* Save character. */
 static LJ_AINLINE void lex_save(LexState *ls, LexChar c)
 {
-  lj_buf_putb(ls->L, &ls->sb, c);
+  lj_buf_putb(&ls->sb, c);
 }
 
 /* Save previous character and get next character. */
diff --git a/src/lj_load.c b/src/lj_load.c
index 2572e79b..b5cbb3ba 100644
--- a/src/lj_load.c
+++ b/src/lj_load.c
@@ -55,7 +55,7 @@ LUA_API int lua_loadx(lua_State *L, lua_Reader reader, void *data,
   ls.rdata = data;
   ls.chunkarg = chunkname ? chunkname : "?";
   ls.mode = mode;
-  lj_buf_init(&ls.sb);
+  lj_buf_init(L, &ls.sb);
   status = lj_vm_cpcall(L, NULL, &ls, cpparser);
   lj_lex_cleanup(L, &ls);
   lj_gc_check(L);
diff --git a/src/lj_obj.h b/src/lj_obj.h
index 4d7b9262..33237119 100644
--- a/src/lj_obj.h
+++ b/src/lj_obj.h
@@ -124,6 +124,7 @@ typedef struct SBuf {
   MRef p;		/* String buffer pointer. */
   MRef e;		/* String buffer end pointer. */
   MRef b;		/* String buffer base. */
+  MRef L;		/* lua_State, used for buffer resizing. */
 } SBuf;
 
 /* -- Tags and values ----------------------------------------------------- */
@@ -516,7 +517,7 @@ typedef struct global_State {
   lua_Alloc allocf;	/* Memory allocator. */
   void *allocd;		/* Memory allocator data. */
   GCState gc;		/* Garbage collector. */
-  SBuf tmpbuf;		/* Temporary buffer for string concatenation. */
+  SBuf tmpbuf;		/* Temporary string buffer. */
   Node nilnode;		/* Fallback 1-element hash part (nil key and value). */
   GCstr strempty;	/* Empty string. */
   uint8_t stremptyz;	/* Zero terminator of empty string. */
diff --git a/src/lj_parse.c b/src/lj_parse.c
index 39a01e6f..c1ef2593 100644
--- a/src/lj_parse.c
+++ b/src/lj_parse.c
@@ -1441,7 +1441,7 @@ static size_t fs_prep_var(LexState *ls, FuncState *fs, size_t *ofsvar)
   for (i = 0, n = fs->nuv; i < n; i++) {
     GCstr *s = strref(vs[fs->uvmap[i]].name);
     MSize len = s->len+1;
-    char *p = lj_buf_more(ls->L, &ls->sb, len);
+    char *p = lj_buf_more(&ls->sb, len);
     p = lj_buf_wmem(p, strdata(s), len);
     setsbufP(&ls->sb, p);
   }
@@ -1454,11 +1454,11 @@ static size_t fs_prep_var(LexState *ls, FuncState *fs, size_t *ofsvar)
       BCPos startpc;
       char *p;
       if ((uintptr_t)s < VARNAME__MAX) {
-	p = lj_buf_more(ls->L, &ls->sb, 1 + 2*5);
+	p = lj_buf_more(&ls->sb, 1 + 2*5);
 	*p++ = (char)(uintptr_t)s;
       } else {
 	MSize len = s->len+1;
-	p = lj_buf_more(ls->L, &ls->sb, len + 2*5);
+	p = lj_buf_more(&ls->sb, len + 2*5);
 	p = lj_buf_wmem(p, strdata(s), len);
       }
       startpc = vs->startpc;
@@ -1468,7 +1468,7 @@ static size_t fs_prep_var(LexState *ls, FuncState *fs, size_t *ofsvar)
       lastpc = startpc;
     }
   }
-  lj_buf_putb(ls->L, &ls->sb, '\0');  /* Terminator for varinfo. */
+  lj_buf_putb(&ls->sb, '\0');  /* Terminator for varinfo. */
   return sbuflen(&ls->sb);
 }
 
diff --git a/src/lj_state.c b/src/lj_state.c
index 6c3e97b1..604ff886 100644
--- a/src/lj_state.c
+++ b/src/lj_state.c
@@ -204,7 +204,7 @@ LUA_API lua_State *lua_newstate(lua_Alloc f, void *ud)
   setnilV(&g->nilnode.val);
   setnilV(&g->nilnode.key);
   setmref(g->nilnode.freetop, &g->nilnode);
-  lj_buf_init(&g->tmpbuf);
+  lj_buf_init(NULL, &g->tmpbuf);
   g->gc.state = GCSpause;
   setgcref(g->gc.root, obj2gco(L));
   setmref(g->gc.sweep, &g->gc.root);
diff --git a/src/lj_str.c b/src/lj_str.c
index 84bab864..67353154 100644
--- a/src/lj_str.c
+++ b/src/lj_str.c
@@ -222,27 +222,28 @@ GCstr * LJ_FASTCALL lj_str_fromnumber(lua_State *L, cTValue *o)
 const char *lj_str_pushvf(lua_State *L, const char *fmt, va_list argp)
 {
   SBuf *sb = &G(L)->tmpbuf;
-  lj_buf_need(L, sb, (MSize)strlen(fmt));
+  setsbufL(sb, L);
+  lj_buf_need(sb, (MSize)strlen(fmt));
   lj_buf_reset(sb);
   for (;;) {
     const char *e = strchr(fmt, '%');
     if (e == NULL) break;
-    lj_buf_putmem(L, sb, fmt, (MSize)(e-fmt));
+    lj_buf_putmem(sb, fmt, (MSize)(e-fmt));
     /* This function only handles %s, %c, %d, %f and %p formats. */
     switch (e[1]) {
     case 's': {
       const char *s = va_arg(argp, char *);
       if (s == NULL) s = "(null)";
-      lj_buf_putmem(L, sb, s, (MSize)strlen(s));
+      lj_buf_putmem(sb, s, (MSize)strlen(s));
       break;
       }
     case 'c':
-      lj_buf_putb(L, sb, va_arg(argp, int));
+      lj_buf_putb(sb, va_arg(argp, int));
       break;
     case 'd': {
       char buf[LJ_STR_INTBUF];
       char *p = lj_str_bufint(buf, va_arg(argp, int32_t));
-      lj_buf_putmem(L, sb, p, (MSize)(buf+LJ_STR_INTBUF-p));
+      lj_buf_putmem(sb, p, (MSize)(buf+LJ_STR_INTBUF-p));
       break;
       }
     case 'f': {
@@ -251,7 +252,7 @@ const char *lj_str_pushvf(lua_State *L, const char *fmt, va_list argp)
       MSize len;
       tv.n = (lua_Number)(va_arg(argp, LUAI_UACNUMBER));
       len = (MSize)lj_str_bufnum(buf, &tv);
-      lj_buf_putmem(L, sb, buf, len);
+      lj_buf_putmem(sb, buf, len);
       break;
       }
     case 'p': {
@@ -260,7 +261,7 @@ const char *lj_str_pushvf(lua_State *L, const char *fmt, va_list argp)
       ptrdiff_t p = (ptrdiff_t)(va_arg(argp, void *));
       ptrdiff_t i, lasti = 2+FMTP_CHARS;
       if (p == 0) {
-	lj_buf_putmem(L, sb, "NULL", 4);
+	lj_buf_putmem(sb, "NULL", 4);
 	break;
       }
 #if LJ_64
@@ -271,20 +272,20 @@ const char *lj_str_pushvf(lua_State *L, const char *fmt, va_list argp)
       buf[1] = 'x';
       for (i = lasti-1; i >= 2; i--, p >>= 4)
 	buf[i] = "0123456789abcdef"[(p & 15)];
-      lj_buf_putmem(L, sb, buf, (MSize)lasti);
+      lj_buf_putmem(sb, buf, (MSize)lasti);
       break;
       }
     case '%':
-      lj_buf_putb(L, sb, '%');
+      lj_buf_putb(sb, '%');
       break;
     default:
-      lj_buf_putb(L, sb, '%');
-      lj_buf_putb(L, sb, e[1]);
+      lj_buf_putb(sb, '%');
+      lj_buf_putb(sb, e[1]);
       break;
     }
     fmt = e+2;
   }
-  lj_buf_putmem(L, sb, fmt, (MSize)strlen(fmt));
+  lj_buf_putmem(sb, fmt, (MSize)strlen(fmt));
   setstrV(L, L->top, lj_buf_str(L, sb));
   incr_top(L);
   return strVdata(L->top - 1);