From 83de532563c6d922c6262dea757a22cb90d06101 Mon Sep 17 00:00:00 2001
From: grischka <grischka>
Date: Sat, 24 May 2025 23:53:37 +0200
Subject: [PATCH] revert "Save registers around attribute cleanup" (almost)

In fact, we don't need to save registers.  We need to
save the symbol if it is a SValue on vstack (the return
value in this case)

Replaces b6a16e3be4aff2d6b6e4870a1fde5e83707e9be4
---
 arm-gen.c     | 43 +++++-------------------------
 arm64-gen.c   | 42 +++++-------------------------
 c67-gen.c     | 15 +----------
 i386-gen.c    | 32 +++++------------------
 il-gen.c      | 13 +---------
 riscv64-gen.c | 57 ++++++----------------------------------
 tcc.h         |  4 +--
 tccgen.c      | 46 ++++++++++++++++++--------------
 x86_64-gen.c  | 72 +++++++++++----------------------------------------
 9 files changed, 73 insertions(+), 251 deletions(-)

diff --git a/arm-gen.c b/arm-gen.c
index 9aa6e7ae..ec4a2e25 100644
--- a/arm-gen.c
+++ b/arm-gen.c
@@ -837,32 +837,6 @@ static void gcall_or_jmp(int is_jmp)
   }
 }
 
-ST_FUNC void save_return_reg(CType *func_type)
-{
-    int freg = is_float(func_type->t & VT_BTYPE);
-    int ireg = !freg;
-
-    if ((func_type->t & VT_BTYPE) == VT_STRUCT)
-        ireg = freg = 1;
-    if (ireg)
-        o(0xe92d0003);  /* push {r0,r1} */
-    if (freg)
-        o(0xed2d0b04);  /* vpush {d0,d1} */
-}
-
-ST_FUNC void restore_return_reg(CType *func_type)
-{
-    int freg = is_float(func_type->t & VT_BTYPE);
-    int ireg = !freg;
-
-    if ((func_type->t & VT_BTYPE) == VT_STRUCT)
-        ireg = freg = 1;
-    if (freg)
-        o(0xecbd0b04); /* vpop {d0,d1} */
-    if (ireg)
-        o(0xe8bd0003); /* pop {r0,r1} */
-}
-
 #if defined(CONFIG_TCC_BCHECK)
 
 static void gen_bounds_call(int v)
@@ -886,13 +860,12 @@ static void gen_bounds_prolog(void)
     o(0xe1a00000);  /* call __bound_local_new */
 }
 
-static void gen_bounds_epilog(Sym *func_sym)
+static void gen_bounds_epilog(void)
 {
     addr_t saved_ind;
     addr_t *bounds_ptr;
     Sym *sym_data;
     int offset_modified = func_bound_offset != lbounds_section->data_offset;
-    CType *func_type = &func_sym->type.ref->type;
 
     if (!offset_modified && !func_bound_add_epilog)
         return;
@@ -918,16 +891,16 @@ static void gen_bounds_epilog(Sym *func_sym)
     }
 
     /* generate bound check local freeing */
-    if (func_type->t != VT_VOID)
-	save_return_reg(func_type);
+    o(0xe92d0003);  /* push {r0,r1} */
+    o(0xed2d0b04);  /* vpush {d0,d1} */
     o(0xe59f0000);  /* ldr r0, [pc] */
     o(0xea000000);  /* b $+4 */
     greloc(cur_text_section, sym_data, ind, R_ARM_REL32);
     o(-12);  /* lbounds_section->data_offset */
     o(0xe080000f);  /* add r0,r0,pc */
     gen_bounds_call(TOK___bound_local_delete);
-    if (func_type->t != VT_VOID)
-	restore_return_reg(func_type);
+    o(0xecbd0b04); /* vpop {d0,d1} */
+    o(0xe8bd0003); /* pop {r0,r1} */
 }
 #endif
 
@@ -1535,17 +1508,15 @@ from_stack:
 }
 
 /* generate function epilog */
-void gfunc_epilog(Sym *func_sym)
+void gfunc_epilog(void)
 {
   uint32_t x;
   int diff;
 
 #ifdef CONFIG_TCC_BCHECK
   if (tcc_state->do_bounds_check)
-    gen_bounds_epilog(func_sym);
+    gen_bounds_epilog();
 #endif
-  func_sym = NULL;
-
   /* Copy float return value to core register if base standard is used and
      float computation is made with VFP */
 #if defined(TCC_ARM_EABI) && defined(TCC_ARM_VFP)
diff --git a/arm64-gen.c b/arm64-gen.c
index a10a3d72..d15446d3 100644
--- a/arm64-gen.c
+++ b/arm64-gen.c
@@ -674,32 +674,6 @@ static void arm64_gen_bl_or_b(int b)
     }
 }
 
-ST_FUNC void save_return_reg(CType *func_type)
-{
-    int freg = is_float(func_type->t & VT_BTYPE);
-    int ireg = !freg;
-
-    if ((func_type->t & VT_BTYPE) == VT_STRUCT)
-        ireg = freg = 1;
-    if (ireg)
-        o(0xa9bf07e0); /* stp x0, x1, [sp, #-16]! */
-    if (freg)
-        o(0xadbf07e0); /* stp q0, q1, [sp, #-32]! */
-}   
-    
-ST_FUNC void restore_return_reg(CType *func_type)
-{   
-    int freg = is_float(func_type->t & VT_BTYPE);
-    int ireg = !freg;
-
-    if ((func_type->t & VT_BTYPE) == VT_STRUCT)
-        ireg = freg = 1;
-    if (freg)
-        o(0xacc107e0); /* ldp q0, q1, [sp], #32 */
-    if (ireg)
-        o(0xa8c107e0); /* ldp x0, x1, [sp], #16 */
-}
-
 #if defined(CONFIG_TCC_BCHECK)
 
 static void gen_bounds_call(int v)
@@ -722,13 +696,12 @@ static void gen_bounds_prolog(void)
     o(0xd503201f);  /* nop -> call __bound_local_new */
 }
 
-static void gen_bounds_epilog(Sym *func_sym)
+static void gen_bounds_epilog(void)
 {
     addr_t saved_ind;
     addr_t *bounds_ptr;
     Sym *sym_data;
     int offset_modified = func_bound_offset != lbounds_section->data_offset;
-    CType *func_type = &func_sym->type.ref->type;
 
     if (!offset_modified && !func_bound_add_epilog)
         return;
@@ -753,15 +726,15 @@ static void gen_bounds_epilog(Sym *func_sym)
     }
 
     /* generate bound check local freeing */
-    if (func_type->t != VT_VOID)
-        save_return_reg(func_type);
+    o(0xa9bf07e0); /* stp x0, x1, [sp, #-16]! */
+    o(0x3c9f0fe0); /* str q0, [sp, #-16]! */
     greloca(cur_text_section, sym_data, ind, R_AARCH64_ADR_GOT_PAGE, 0);
     o(0x90000000 | 0);            // adrp x0, #sym_data
     greloca(cur_text_section, sym_data, ind, R_AARCH64_LD64_GOT_LO12_NC, 0);
     o(0xf9400000 | 0 | (0 << 5)); // ld x0,[x0, #sym_data]
     gen_bounds_call(TOK___bound_local_delete);
-    if (func_type->t != VT_VOID)
-        restore_return_reg(func_type);
+    o(0x3cc107e0); /* ldr q0, [sp], #16 */
+    o(0xa8c107e0); /* ldp x0, x1, [sp], #16 */
 }
 #endif
 
@@ -1486,13 +1459,12 @@ ST_FUNC void gfunc_return(CType *func_type)
     vtop--;
 }
 
-ST_FUNC void gfunc_epilog(Sym *func_sym)
+ST_FUNC void gfunc_epilog(void)
 {
 #ifdef CONFIG_TCC_BCHECK
     if (tcc_state->do_bounds_check)
-        gen_bounds_epilog(func_sym);
+        gen_bounds_epilog();
 #endif
-    func_sym = NULL;
 
     if (loc) {
         // Insert instructions to subtract size of stack frame from SP.
diff --git a/c67-gen.c b/c67-gen.c
index bafedc01..9490a27f 100644
--- a/c67-gen.c
+++ b/c67-gen.c
@@ -2019,22 +2019,9 @@ void gfunc_prolog(Sym *func_sym)
     C67_PUSH(C67_B3);
 }
 
-ST_FUNC void save_return_reg(CType *func_type)
-{
-    func_type = NULL;
-    // TODO
-}
-
-ST_FUNC void restore_return_reg(CType *func_type)
-{
-    func_type = NULL;
-    // TODO
-}
-
 /* generate function epilog */
-void gfunc_epilog(Sym *func_sym)
+void gfunc_epilog(void)
 {
-    func_sym = NULL;
     {
 	int local = (-loc + 7) & -8;	// stack must stay aligned to 8 bytes for LDDW instr
 	C67_POP(C67_B3);
diff --git a/i386-gen.c b/i386-gen.c
index 5e9a9e5c..5814a398 100644
--- a/i386-gen.c
+++ b/i386-gen.c
@@ -104,7 +104,7 @@ static addr_t func_bound_offset;
 static unsigned long func_bound_ind;
 ST_DATA int func_bound_add_epilog;
 static void gen_bounds_prolog(void);
-static void gen_bounds_epilog(Sym *func_sym);
+static void gen_bounds_epilog(void);
 #endif
 
 /* XXX: make it faster ? */
@@ -595,15 +595,14 @@ ST_FUNC void gfunc_prolog(Sym *func_sym)
 }
 
 /* generate function epilog */
-ST_FUNC void gfunc_epilog(Sym *func_sym)
+ST_FUNC void gfunc_epilog(void)
 {
     addr_t v, saved_ind;
 
 #ifdef CONFIG_TCC_BCHECK
     if (tcc_state->do_bounds_check)
-        gen_bounds_epilog(func_sym);
+        gen_bounds_epilog();
 #endif
-    func_sym = NULL;
 
     /* align local size to word & save local variables */
     v = (-loc + 3) & -4;
@@ -1041,22 +1040,6 @@ ST_FUNC void ggoto(void)
     vtop--;
 }
 
-ST_FUNC void save_return_reg(CType *func_type)
-{
-    int ireg = !is_float(func_type->t & VT_BTYPE);
-
-    if (ireg)
-        o(0x5250); /* push %rax; %push %rdx */
-}
-
-ST_FUNC void restore_return_reg(CType *func_type)
-{
-    int ireg = !is_float(func_type->t & VT_BTYPE);
-
-    if (ireg)
-        o(0x585a); /* pop %rdx; pop %rax */
-}
-
 /* bound check support functions */
 #ifdef CONFIG_TCC_BCHECK
 
@@ -1070,13 +1053,12 @@ static void gen_bounds_prolog(void)
     oad(0xb8, 0); /* call to function */
 }
 
-static void gen_bounds_epilog(Sym *func_sym)
+static void gen_bounds_epilog(void)
 {
     addr_t saved_ind;
     addr_t *bounds_ptr;
     Sym *sym_data;
     int offset_modified = func_bound_offset != lbounds_section->data_offset;
-    CType *func_type = &func_sym->type.ref->type;
 
     if (!offset_modified && !func_bound_add_epilog)
         return;
@@ -1099,13 +1081,11 @@ static void gen_bounds_epilog(Sym *func_sym)
     }
 
     /* generate bound check local freeing */
-    if (func_type->t != VT_VOID)
-        save_return_reg(func_type);
+    o(0x5250); /* save returned value, if any */
     greloc(cur_text_section, sym_data, ind + 1, R_386_32);
     oad(0xb8, 0); /* mov %eax, xxx */
     gen_static_call(TOK___bound_local_delete);
-    if (func_type->t != VT_VOID)
-        restore_return_reg(func_type);
+    o(0x585a); /* restore returned value, if any */
 }
 #endif
 
diff --git a/il-gen.c b/il-gen.c
index a00d6786..bb670ccb 100644
--- a/il-gen.c
+++ b/il-gen.c
@@ -457,20 +457,9 @@ void gfunc_prolog(int t)
     }
 }
 
-void save_return_reg(CType *func_type)
-{
-    func_type = NULL;
-}
-
-void restore_return_reg(CType *func_type)
-{
-    func_type = NULL;
-}
-
 /* generate function epilog */
-void gfunc_epilog(Sym *func_sym)
+void gfunc_epilog(void)
 {
-    func_sym = NULL;
     out_op(IL_OP_RET);
     fprintf(il_outfile, "}\n\n");
 }
diff --git a/riscv64-gen.c b/riscv64-gen.c
index c38c996f..99c94dc3 100644
--- a/riscv64-gen.c
+++ b/riscv64-gen.c
@@ -436,46 +436,6 @@ static void gcall_or_jmp(int docall)
     }
 }
 
-ST_FUNC void save_return_reg(CType *func_type)
-{
-    int freg = is_float(func_type->t & VT_BTYPE) &&
-	       (func_type->t & VT_BTYPE) != VT_LDOUBLE;
-    int ireg = !freg;
-
-    if ((func_type->t & VT_BTYPE) == VT_STRUCT)
-        ireg = freg = 1;
-    if (ireg && freg) {
-        o(0xe02a1101); /* addi sp,sp,-32  sd   a0,0(sp)   */
-        o(0xa82ae42e); /* sd   a1,8(sp)   fsd  fa0,16(sp) */
-    }
-    else if (ireg) {
-        o(0xe02a1141); /* addi sp,sp,-16  sd   a0,0(sp)   */
-        o(0x0001e42e); /* sd   a1,8(sp)   nop */
-    }
-    else if (freg)
-        o(0xa02a1141); /* addi sp,sp,-16  fsd  fa0,0(sp)   */
-}
-
-ST_FUNC void restore_return_reg(CType *func_type)
-{
-    int freg = is_float(func_type->t & VT_BTYPE) &&
-	       (func_type->t & VT_BTYPE) != VT_LDOUBLE;
-    int ireg = !freg;
-
-    if ((func_type->t & VT_BTYPE) == VT_STRUCT)
-        ireg = freg = 1;
-    if (ireg && freg) {
-        o(0x65a26502); /* ld   a0,0(sp)   ld   a1,8(sp)   */
-        o(0x61052542); /* fld  fa0,16(sp) addi sp,sp,32   */
-    }
-    else if (ireg) {
-        o(0x65a26502); /* ld   a0,0(sp)   ld   a1,8(sp)   */
-        o(0x00010141); /* addi sp,sp,16   nop */
-    }
-    else if (freg)
-        o(0x01412502); /* fld  fa0,0(sp)  addi sp,sp,16   */
-}
-
 #if defined(CONFIG_TCC_BCHECK)
 
 static void gen_bounds_call(int v)
@@ -499,14 +459,14 @@ static void gen_bounds_prolog(void)
     o(0x00000013);
 }
 
-static void gen_bounds_epilog(Sym *func_sym)
+static void gen_bounds_epilog(void)
 {
     addr_t saved_ind;
     addr_t *bounds_ptr;
     Sym *sym_data;
     Sym label = {0};
+
     int offset_modified = func_bound_offset != lbounds_section->data_offset;
-    CType *func_type = &func_sym->type.ref->type;
 
     if (!offset_modified && !func_bound_add_epilog)
         return;
@@ -534,16 +494,16 @@ static void gen_bounds_epilog(Sym *func_sym)
     }
 
     /* generate bound check local freeing */
-    if (func_type->t != VT_VOID)
-        save_return_reg(func_type);
+    o(0xe02a1101); /* addi sp,sp,-32  sd   a0,0(sp)   */
+    o(0xa82ae42e); /* sd   a1,8(sp)   fsd  fa0,16(sp) */
     put_extern_sym(&label, cur_text_section, ind, 0);
     greloca(cur_text_section, sym_data, ind, R_RISCV_GOT_HI20, 0);
     o(0x17 | (10 << 7));    // auipc a0, 0 %pcrel_hi(sym)+addend
     greloca(cur_text_section, &label, ind, R_RISCV_PCREL_LO12_I, 0);
     EI(0x03, 3, 10, 10, 0); // ld a0, 0(a0)
     gen_bounds_call(TOK___bound_local_delete);
-    if (func_type->t != VT_VOID)
-        restore_return_reg(func_type);
+    o(0x65a26502); /* ld   a0,0(sp)   ld   a1,8(sp)   */
+    o(0x61052542); /* fld  fa0,16(sp) addi sp,sp,32   */
 }
 #endif
 
@@ -933,15 +893,14 @@ ST_FUNC void arch_transfer_ret_regs(int aftercall)
     vtop--;
 }
 
-ST_FUNC void gfunc_epilog(Sym *func_sym)
+ST_FUNC void gfunc_epilog(void)
 {
     int v, saved_ind, d, large_ofs_ind;
 
 #ifdef CONFIG_TCC_BCHECK
     if (tcc_state->do_bounds_check)
-        gen_bounds_epilog(func_sym);
+        gen_bounds_epilog();
 #endif
-    func_sym = NULL;
 
     loc = (loc - num_va_regs * 8);
     d = v = (-loc + 15) & -16;
diff --git a/tcc.h b/tcc.h
index debc96d7..452a723a 100644
--- a/tcc.h
+++ b/tcc.h
@@ -1607,9 +1607,7 @@ ST_FUNC void store(int r, SValue *v);
 ST_FUNC int gfunc_sret(CType *vt, int variadic, CType *ret, int *align, int *regsize);
 ST_FUNC void gfunc_call(int nb_args);
 ST_FUNC void gfunc_prolog(Sym *func_sym);
-ST_FUNC void gfunc_epilog(Sym *func_sym);
-ST_FUNC void save_return_reg(CType *func_type);
-ST_FUNC void restore_return_reg(CType *func_type);
+ST_FUNC void gfunc_epilog(void);
 ST_FUNC void gen_fill_nops(int);
 ST_FUNC int gjmp(int t);
 ST_FUNC void gjmp_addr(int a);
diff --git a/tccgen.c b/tccgen.c
index a1b1b80c..5829037c 100644
--- a/tccgen.c
+++ b/tccgen.c
@@ -6839,18 +6839,28 @@ static void end_switch(void)
 /* ------------------------------------------------------------------------- */
 /* __attribute__((cleanup(fn))) */
 
-static void try_call_scope_cleanup(Sym *stop, CType *func_type)
+/* save SValue of symbol to local stack */
+static void save_cleanup_sym(Sym *s)
+{
+    SValue *sv = vtop;
+    while (sv >= vstack) {
+        if (sv->sym == s) {
+            int align, size = type_size(&sv->type, &align);
+            loc = (loc - size) & -align;
+            vset(&sv->type, VT_LOCAL | VT_LVAL, loc);
+            vpushv(sv), *sv = vtop[-1], vstore(), --vtop;
+        }
+        --sv;
+    }
+}
+
+static void try_call_scope_cleanup(Sym *stop)
 {
     Sym *cls = cur_scope->cl.s;
-
-    if (cls == stop)
-	func_type = NULL;
-    if (func_type && func_type->t != VT_VOID)
-	save_return_reg(func_type);
     for (; cls != stop; cls = cls->next) {
 	Sym *fs = cls->cleanup_func;
 	Sym *vs = cls->prev_tok;
-
+        save_cleanup_sym(vs);
 	vpushsym(&fs->type, fs);
 	vset(&vs->type, vs->r, vs->c);
 	vtop->sym = vs;
@@ -6858,8 +6868,6 @@ static void try_call_scope_cleanup(Sym *stop, CType *func_type)
 	gaddrof();
 	gfunc_call(1);
     }
-    if (func_type && func_type->t != VT_VOID)
-	restore_return_reg(func_type);
 }
 
 static void try_call_cleanup_goto(Sym *cleanupstate)
@@ -6879,7 +6887,7 @@ static void try_call_cleanup_goto(Sym *cleanupstate)
     for (; cc != oc; cc = cc->next, oc = oc->next, --ccd)
       ;
 
-    try_call_scope_cleanup(cc, NULL);
+    try_call_scope_cleanup(cc);
 }
 
 /* call 'func' for each __attribute__((cleanup(func))) */
@@ -6893,7 +6901,7 @@ static void block_cleanup(struct scope *o)
             if (!jmp)
                 jmp = gjmp(0);
             gsym(pcl->jnext);
-            try_call_scope_cleanup(o->cl.s, NULL);
+            try_call_scope_cleanup(o->cl.s);
             pcl->jnext = gjmp(0);
             if (!o->cl.n)
                 goto remove_pending;
@@ -6906,7 +6914,7 @@ static void block_cleanup(struct scope *o)
         }
     }
     gsym(jmp);
-    try_call_scope_cleanup(o->cl.s, NULL);
+    try_call_scope_cleanup(o->cl.s);
 }
 
 /* ------------------------------------------------------------------------- */
@@ -6970,11 +6978,11 @@ static void prev_scope(struct scope *o, int is_expr)
 }
 
 /* leave a scope via break/continue(/goto) */
-static void leave_scope(struct scope *o, CType *func_type)
+static void leave_scope(struct scope *o)
 {
     if (!o)
         return;
-    try_call_scope_cleanup(o->cl.s, func_type);
+    try_call_scope_cleanup(o->cl.s);
     vla_leave(o);
 }
 
@@ -7116,9 +7124,9 @@ again:
             tcc_warning("'return' with no value");
             b = 0;
         }
+        leave_scope(root_scope);
         if (b)
             gfunc_return(&func_vt);
-        leave_scope(root_scope, &func_vt);
         skip(';');
         /* jump unless last stmt in top-level block */
         if (tok != '}' || local_scope != 1)
@@ -7132,9 +7140,9 @@ again:
         if (!cur_scope->bsym)
             tcc_error("cannot break");
         if (cur_switch && cur_scope->bsym == cur_switch->bsym)
-            leave_scope(cur_switch->scope, NULL);
+            leave_scope(cur_switch->scope);
         else
-            leave_scope(loop_scope, NULL);
+            leave_scope(loop_scope);
         *cur_scope->bsym = gjmp(*cur_scope->bsym);
         skip(';');
 
@@ -7142,7 +7150,7 @@ again:
         /* compute jump */
         if (!cur_scope->csym)
             tcc_error("cannot continue");
-        leave_scope(loop_scope, NULL);
+        leave_scope(loop_scope);
         *cur_scope->csym = gjmp(*cur_scope->csym);
         skip(';');
 
@@ -8406,7 +8414,7 @@ static void gen_function(Sym *sym)
     /* reset local stack */
     pop_local_syms(NULL, 0);
     tcc_debug_prolog_epilog(tcc_state, 1);
-    gfunc_epilog(sym);
+    gfunc_epilog();
 
     /* end of function */
     tcc_debug_funcend(tcc_state, ind - func_ind);
diff --git a/x86_64-gen.c b/x86_64-gen.c
index d4c4c650..3a430585 100644
--- a/x86_64-gen.c
+++ b/x86_64-gen.c
@@ -655,50 +655,6 @@ static void gcall_or_jmp(int is_jmp)
     }
 }
 
-ST_FUNC void save_return_reg(CType *func_type)
-{
-    int freg = is_float(func_type->t & VT_BTYPE);
-    int ireg = !freg;
-
-    if ((func_type->t & VT_BTYPE) == VT_STRUCT)
-        ireg = freg = 1;
-    if (ireg)
-        o(0x5250); /* push %rax; %push %rdx */
-    if (freg) {
-	if ((func_type->t & VT_BTYPE) == VT_LDOUBLE ||
-	    (func_type->t & VT_BTYPE) == VT_STRUCT) {
-            o(0x10ec8348); /* sub $16,%rsp */
-	    o(0x243cdb);   /* fstpt (%rsp) */
-	}
-        o(0x20ec8348); /* sub $32,%rsp */
-        o(0x290f);     /* movaps %xmm0,0x10(%rsp) */
-        o(0x102444);
-        o(0x240c290f); /* movaps %xmm1,(%rsp) */
-    }
-}
-
-ST_FUNC void restore_return_reg(CType *func_type)
-{
-    int freg = is_float(func_type->t & VT_BTYPE);
-    int ireg = !freg;
-
-    if ((func_type->t & VT_BTYPE) == VT_STRUCT)
-        ireg = freg = 1;
-    if (freg) {
-        o(0x280f);     /* movaps 0x10(%rsp),%xmm0 */
-        o(0x102444);
-        o(0x240c280f); /* movaps (%rsp),%xmm1 */
-        o(0x20c48348); /* add $32,%rsp */
-	if ((func_type->t & VT_BTYPE) == VT_LDOUBLE ||
-	    (func_type->t & VT_BTYPE) == VT_STRUCT) {
-	    o(0x242cdb);   /* fldt (%rsp) */
-            o(0x10c48348); /* add $16,%rsp */
-	}
-    }
-    if (ireg)
-        o(0x585a); /* pop %rdx; pop %rax */
-}
-
 #if defined(CONFIG_TCC_BCHECK)
 
 static void gen_bounds_call(int v)
@@ -725,13 +681,12 @@ static void gen_bounds_prolog(void)
     oad(0xb8, 0); /* call to function */
 }
 
-static void gen_bounds_epilog(Sym *func_sym)
+static void gen_bounds_epilog(void)
 {
     addr_t saved_ind;
     addr_t *bounds_ptr;
     Sym *sym_data;
     int offset_modified = func_bound_offset != lbounds_section->data_offset;
-    CType *func_type = &func_sym->type.ref->type;
 
     if (!offset_modified && !func_bound_add_epilog)
         return;
@@ -754,14 +709,20 @@ static void gen_bounds_epilog(Sym *func_sym)
     }
 
     /* generate bound check local freeing */
-    if (func_type->t != VT_VOID)
-	save_return_reg(func_type);
+    o(0x5250); /* save returned value, if any */
+    o(0x20ec8348); /* sub $32,%rsp */
+    o(0x290f);     /* movaps %xmm0,0x10(%rsp) */
+    o(0x102444);
+    o(0x240c290f); /* movaps %xmm1,(%rsp) */
     greloca(cur_text_section, sym_data, ind + 3, R_X86_64_PC32, -4);
     o(0x0d8d48 + ((TREG_FASTCALL_1 == TREG_RDI) * 0x300000)); /* lea xxx(%rip), %rcx/rdi */
     gen_le32 (0);
     gen_bounds_call(TOK___bound_local_delete);
-    if (func_type->t != VT_VOID)
-	restore_return_reg(func_type);
+    o(0x280f);     /* movaps 0x10(%rsp),%xmm0 */
+    o(0x102444);
+    o(0x240c280f); /* movaps (%rsp),%xmm1 */
+    o(0x20c48348); /* add $32,%rsp */
+    o(0x585a); /* restore returned value, if any */
 }
 #endif
 
@@ -1044,7 +1005,7 @@ void gfunc_prolog(Sym *func_sym)
 }
 
 /* generate function epilog */
-void gfunc_epilog(Sym *func_sym)
+void gfunc_epilog(void)
 {
     int v, start;
 
@@ -1054,9 +1015,8 @@ void gfunc_epilog(Sym *func_sym)
 
 #ifdef CONFIG_TCC_BCHECK
     if (tcc_state->do_bounds_check)
-        gen_bounds_epilog(func_sym);
+        gen_bounds_epilog();
 #endif
-    func_sym = NULL;
 
     o(0xc9); /* leave */
     if (func_ret_sub == 0) {
@@ -1640,16 +1600,14 @@ void gfunc_prolog(Sym *func_sym)
 }
 
 /* generate function epilog */
-void gfunc_epilog(Sym *func_sym)
+void gfunc_epilog(void)
 {
     int v, saved_ind;
 
 #ifdef CONFIG_TCC_BCHECK
     if (tcc_state->do_bounds_check)
-        gen_bounds_epilog(func_sym);
+        gen_bounds_epilog();
 #endif
-    func_sym = NULL;
-
     o(0xc9); /* leave */
     if (func_ret_sub == 0) {
         o(0xc3); /* ret */