Tests in abitest.c now work on Win32.

I expect that Linux-x86 is probably fine. All other architectures
except ARM are definitely broken since I haven't yet implemented
gfunc_sret for these, although replicating the current behaviour
should be straightforward.

7 files changed, 173 insertions, 66 deletions

diff --git a/Makefile b/Makefile
index fb94845..f487fbf 100644
--- a/Makefile
+++ b/Makefile
@@ -45,7 +45,13 @@ endif
 
 # make libtcc as static or dynamic library?
 ifdef DISABLE_STATIC
+ifndef CONFIG_WIN32
 LIBTCC=libtcc.so.1.0
+else
+LIBTCC=libtcc.dll
+LIBTCC_DLL=yes
+LIBTCC_EXTRA=libtcc.def libtcc.a
+endif
 LINK_LIBTCC=-Wl,-rpath,"$(libdir)"
 ifdef DISABLE_RPATH
 LINK_LIBTCC=
@@ -126,7 +132,7 @@ ifdef CONFIG_USE_LIBGCC
 LIBTCC1=
 endif
 
-TCCLIBS = $(LIBTCC1) $(LIBTCC)
+TCCLIBS = $(LIBTCC1) $(LIBTCC) $(LIBTCC_EXTRA)
 TCCDOCS = tcc.1 tcc-doc.html tcc-doc.info
 
 ifdef CONFIG_CROSS
@@ -185,14 +191,21 @@ endif
 $(LIBTCC_OBJ) tcc.o : %.o : %.c $(LIBTCC_INC)
 	$(CC) -o $@ -c $< $(NATIVE_DEFINES) $(CPPFLAGS) $(CFLAGS)
 
+ifndef LIBTCC_DLL
 libtcc.a: $(LIBTCC_OBJ)
 	$(AR) rcs $@ $^
+endif
 
 libtcc.so.1.0: $(LIBTCC_OBJ)
 	$(CC) -shared -Wl,-soname,$@ -o $@ $^ $(LDFLAGS)
 
 libtcc.so.1.0: CFLAGS+=-fPIC
 
+ifdef LIBTCC_DLL
+libtcc.dll libtcc.def libtcc.a: $(LIBTCC_OBJ)
+	$(CC) -shared $^ -o $@ $(LDFLAGS) -Wl,--output-def,libtcc.def,--out-implib,libtcc.a
+endif
+
 # windows utilities
 tiny_impdef$(EXESUF): win32/tools/tiny_impdef.c
 	$(CC) -o $@ $< $(CPPFLAGS) $(CFLAGS) $(LDFLAGS)
@@ -286,7 +299,8 @@ install: $(PROGS) $(TCCLIBS) $(TCCDOCS)
 	cp -r $(top_srcdir)/win32/examples/. "$(tccdir)/examples"
 	$(INSTALL) -m644 $(addprefix $(top_srcdir)/include/,$(TCC_INCLUDES)) "$(tccdir)/include"
 	$(INSTALL) -m644 tcc-doc.html $(top_srcdir)/win32/tcc-win32.txt "$(tccdir)/doc"
-	$(INSTALL) -m644 $(LIBTCC) $(top_srcdir)/libtcc.h "$(tccdir)/libtcc"
+	$(INSTALL) -m644 $(top_srcdir)/libtcc.h $(LIBTCC_EXTRA) "$(tccdir)/libtcc"
+	$(INSTALL) -m644 $(LIBTCC) $(tccdir)
 ifdef CONFIG_CROSS
 	mkdir -p "$(tccdir)/lib/32"
 	mkdir -p "$(tccdir)/lib/64"
diff --git a/arm-gen.c b/arm-gen.c
index 3e86d9e..03262aa 100644
--- a/arm-gen.c
+++ b/arm-gen.c
@@ -800,6 +800,19 @@ int assign_fpreg(struct avail_regs *avregs, int align, int size)
 }
 #endif
 
+/* Return 1 if this function returns via an sret pointer, 0 otherwise */
+ST_FUNC int gfunc_sret(CType *vt, CType *ret, int *align) {
+    size = type_size(vt, &align);
+    if (size > 4) {
+        return 1;
+    } else {
+        *align = 4;
+        ret->ref = NULL;
+        ret->t = VT_INT;
+    }
+    return 0;
+}
+
 /* Generate function call. The function address is pushed first, then
    all the parameters in call order. This functions pops all the
    parameters and the function address. */
diff --git a/i386-gen.c b/i386-gen.c
index a9e1977..67123d6 100644
--- a/i386-gen.c
+++ b/i386-gen.c
@@ -364,6 +364,28 @@ static void gcall_or_jmp(int is_jmp)
 static uint8_t fastcall_regs[3] = { TREG_EAX, TREG_EDX, TREG_ECX };
 static uint8_t fastcallw_regs[2] = { TREG_ECX, TREG_EDX };
 
+/* Return 1 if this function returns via an sret pointer, 0 otherwise */
+ST_FUNC int gfunc_sret(CType *vt, CType *ret, int *ret_align) {
+    *ret_align = 1; // Never have to re-align return values for x86
+#ifdef TCC_TARGET_PE
+    int size, align;
+    size = type_size(vt, &align);
+    if (size > 8) {
+        return 1;
+    } else if (size > 4) {
+        ret->ref = NULL;
+        ret->t = VT_LLONG;
+        return 0;
+    } else {
+        ret->ref = NULL;
+        ret->t = VT_INT;
+        return 0;
+    }
+#else
+    return 1;
+#endif
+}
+
 /* Generate function call. The function address is pushed first, then
    all the parameters in call order. This functions pops all the
    parameters and the function address. */
@@ -444,10 +466,6 @@ ST_FUNC void gfunc_call(int nb_args)
     }
     gcall_or_jmp(0);
 
-#ifdef TCC_TARGET_PE
-    if ((func_sym->type.t & VT_BTYPE) == VT_STRUCT)
-        args_size -= 4;
-#endif
     if (args_size && func_call != FUNC_STDCALL)
         gadd_sp(args_size);
     vtop--;
@@ -491,7 +509,12 @@ ST_FUNC void gfunc_prolog(CType *func_type)
     /* if the function returns a structure, then add an
        implicit pointer parameter */
     func_vt = sym->type;
+#ifdef TCC_TARGET_PE
+    size = type_size(&func_vt,&align);
+    if (((func_vt.t & VT_BTYPE) == VT_STRUCT) && (size > 8)) {
+#else
     if ((func_vt.t & VT_BTYPE) == VT_STRUCT) {
+#endif
         /* XXX: fastcall case ? */
         func_vc = addr;
         addr += 4;
@@ -526,10 +549,6 @@ ST_FUNC void gfunc_prolog(CType *func_type)
     /* pascal type call ? */
     if (func_call == FUNC_STDCALL)
         func_ret_sub = addr - 8;
-#ifdef TCC_TARGET_PE
-    else if (func_vc)
-        func_ret_sub = 4;
-#endif
 
 #ifdef CONFIG_TCC_BCHECK
     /* leave some room for bound checking code */
diff --git a/tcc.h b/tcc.h
index ef645b1..f243ed0 100644
--- a/tcc.h
+++ b/tcc.h
@@ -1267,6 +1267,7 @@ ST_FUNC void gsym_addr(int t, int a);
 ST_FUNC void gsym(int t);
 ST_FUNC void load(int r, SValue *sv);
 ST_FUNC void store(int r, SValue *v);
+ST_FUNC int gfunc_sret(CType *vt, CType *ret, int *align);
 ST_FUNC void gfunc_call(int nb_args);
 ST_FUNC void gfunc_prolog(CType *func_type);
 ST_FUNC void gfunc_epilog(void);
diff --git a/tccgen.c b/tccgen.c
index 0a6250c..febab11 100644
--- a/tccgen.c
+++ b/tccgen.c
@@ -3850,7 +3850,7 @@ ST_FUNC void unary(void)
         } else if (tok == '(') {
             SValue ret;
             Sym *sa;
-            int nb_args;
+            int nb_args, sret;
 
             /* function call  */
             if ((vtop->type.t & VT_BTYPE) != VT_FUNC) {
@@ -3874,18 +3874,26 @@ ST_FUNC void unary(void)
             ret.r2 = VT_CONST;
             /* compute first implicit argument if a structure is returned */
             if ((s->type.t & VT_BTYPE) == VT_STRUCT) {
-                /* get some space for the returned structure */
-                size = type_size(&s->type, &align);
-                loc = (loc - size) & -align;
-                ret.type = s->type;
-                ret.r = VT_LOCAL | VT_LVAL;
-                /* pass it as 'int' to avoid structure arg passing
-                   problems */
-                vseti(VT_LOCAL, loc);
-                ret.c = vtop->c;
-                nb_args++;
+                int ret_align;
+                sret = gfunc_sret(&s->type, &ret.type, &ret_align);
+                if (sret) {
+                    /* get some space for the returned structure */
+                    size = type_size(&s->type, &align);
+                    loc = (loc - size) & -align;
+                    ret.type = s->type;
+                    ret.r = VT_LOCAL | VT_LVAL;
+                    /* pass it as 'int' to avoid structure arg passing
+                       problems */
+                    vseti(VT_LOCAL, loc);
+                    ret.c = vtop->c;
+                    nb_args++;
+                }
             } else {
-                ret.type = s->type; 
+                sret = 0;
+                ret.type = s->type;
+            }
+
+            if (!sret) {
                 /* return in register */
                 if (is_float(ret.type.t)) {
                     ret.r = reg_fret(ret.type.t);
@@ -3919,6 +3927,17 @@ ST_FUNC void unary(void)
             /* return value */
             vsetc(&ret.type, ret.r, &ret.c);
             vtop->r2 = ret.r2;
+            /* handle packed struct return */
+            if (((s->type.t & VT_BTYPE) == VT_STRUCT) && !sret) {
+                size = type_size(&s->type, &align);
+                loc = (loc - size) & -align;
+                int addr = loc;
+                vset(&ret.type, VT_LOCAL | VT_LVAL, addr);
+                vswap();
+                vstore();
+                vtop--;
+                vset(&s->type, VT_LOCAL | VT_LVAL, addr);
+            }
         } else {
             break;
         }
@@ -4466,40 +4485,38 @@ static void block(int *bsym, int *csym, int *case_sym, int *def_sym,
             gexpr();
             gen_assign_cast(&func_vt);
             if ((func_vt.t & VT_BTYPE) == VT_STRUCT) {
-                CType type;
-                /* if returning structure, must copy it to implicit
-                   first pointer arg location */
-#ifdef TCC_ARM_EABI
-                int align, size;
-                size = type_size(&func_vt,&align);
-                if(size <= 4)
-                {
-                    if((vtop->r != (VT_LOCAL | VT_LVAL) || (vtop->c.i & 3))
-                       && (align & 3))
-                    {
-                        int addr;
-                        loc = (loc - size) & -4;
+                CType type, ret_type;
+                int ret_align;
+                if (gfunc_sret(&func_vt, &ret_type, &ret_align)) {
+                    /* if returning structure, must copy it to implicit
+                       first pointer arg location */
+                    type = func_vt;
+                    mk_pointer(&type);
+                    vset(&type, VT_LOCAL | VT_LVAL, func_vc);
+                    indir();
+                    vswap();
+                    /* copy structure value to pointer */
+                    vstore();
+                } else {
+                    /* returning structure packed into registers */
+                    int size, addr, align;
+                    size = type_size(&func_vt,&align);
+                    if ((vtop->r != (VT_LOCAL | VT_LVAL) || (vtop->c.i & (ret_align-1)))
+                        && (align & (ret_align-1))) {
+                        loc = (loc - size) & -align;
                         addr = loc;
                         type = func_vt;
                         vset(&type, VT_LOCAL | VT_LVAL, addr);
                         vswap();
                         vstore();
-                        vset(&int_type, VT_LOCAL | VT_LVAL, addr);
+                        vset(&ret_type, VT_LOCAL | VT_LVAL, addr);
                     }
-                    vtop->type = int_type;
-                    gv(RC_IRET);
-                } else {
-#endif
-                type = func_vt;
-                mk_pointer(&type);
-                vset(&type, VT_LOCAL | VT_LVAL, func_vc);
-                indir();
-                vswap();
-                /* copy structure value to pointer */
-                vstore();
-#ifdef TCC_ARM_EABI
+                    vtop->type = ret_type;
+                    if (is_float(ret_type.t))
+                        gv(rc_fret(ret_type.t));
+                    else
+                        gv(RC_IRET);
                 }
-#endif
             } else if (is_float(func_vt.t)) {
                 gv(rc_fret(func_vt.t));
             } else {
diff --git a/tests/Makefile b/tests/Makefile
index f47ede4..ae5d47d 100644
--- a/tests/Makefile
+++ b/tests/Makefile
@@ -13,7 +13,7 @@ TESTS = \
  hello-run \
  libtest \
  test3 \
- abitest-exe \
+ abitest \
  moretests
 
 # test4 -- problem with -static
@@ -180,12 +180,16 @@ asmtest: asmtest.ref
 
 # Check that code generated by libtcc is binary compatible with
 # that generated by CC
-abitest$(EXESUF): abitest.c $(top_builddir)/$(LIBTCC)
-	$(CC) -o $@ $^ $(CPPFLAGS) $(CFLAGS) $(NATIVE_DEFINES) $(LIBS) $(LINK_LIBTCC) $(LDFLAGS) -I$(top_srcdir) -g -O0
+abitest-cc$(EXESUF): abitest.c $(top_builddir)/$(LIBTCC)
+	$(CC) -o $@ $^ $(CPPFLAGS) $(CFLAGS) $(NATIVE_DEFINES) $(LIBS) $(LINK_LIBTCC) $(LDFLAGS) -I$(top_srcdir)
+
+abitest-tcc$(EXESUF): abitest.c $(top_builddir)/$(LIBTCC)
+	$(TCC) -o $@ $^ $(CPPFLAGS) $(CFLAGS) $(NATIVE_DEFINES) $(LIBS) $(LINK_LIBTCC) $(LDFLAGS) -I$(top_srcdir)
 
-abitest-exe: abitest$(EXESUF)
+abitest: abitest-cc$(EXESUF) abitest-tcc$(EXESUF)
 	@echo ------------ $@ ------------
-	abitest$(EXESUF) lib_path=..
+	abitest-cc$(EXESUF) lib_path=..
+	abitest-tcc$(EXESUF) lib_path=..
 
 # targets for development
 %.bin: %.c tcc
diff --git a/tests/abitest.c b/tests/abitest.c
index 51a30ba..90f4cb8 100644
--- a/tests/abitest.c
+++ b/tests/abitest.c
@@ -37,24 +37,62 @@ static int run_callback(const char *src, callback_type callback) {
   return result;
 }
 
-typedef struct test1_type_s {int x, y;} test1_type;
-typedef test1_type (*test1_function_type) ();
+/*
+ * reg_pack_test: return a small struct which should be packed into
+ * registers (Win32) during return.
+ */
+typedef struct reg_pack_test_type_s {int x, y;} reg_pack_test_type;
+typedef reg_pack_test_type (*reg_pack_test_function_type) (reg_pack_test_type);
+
+static int reg_pack_test_callback(void *ptr) {
+  reg_pack_test_function_type f = (reg_pack_test_function_type)ptr;
+  reg_pack_test_type a = {10, 35};
+  reg_pack_test_type r;
+  r = f(a);
+  return ((r.x == a.x*5) && (r.y == a.y*3)) ? 0 : -1;
+}
+
+static int reg_pack_test(void) {
+  const char *src =
+  "typedef struct reg_pack_test_type_s {int x, y;} reg_pack_test_type;"
+  "reg_pack_test_type f(reg_pack_test_type a) {\n"
+  "  reg_pack_test_type r = {a.x*5, a.y*3};\n"
+  "  return r;\n"
+  "}\n";
+  
+  return run_callback(src, reg_pack_test_callback);
+}
+
+/*
+ * sret_test: Create a struct large enough to be returned via sret
+ * (hidden pointer as first function argument)
+ */
+typedef struct sret_test_type_s {
+  long long a;
+  long long b;
+} sret_test_type;
+
+typedef sret_test_type (*sret_test_function_type) (sret_test_type);
 
-static int test1_callback(void *ptr) {
-  test1_type r;
-  r  = ((test1_function_type)ptr)();
-  return ((r.x == 10) && (r.y == 35)) ? 0 : -1;
+static int sret_test_callback(void *ptr) {
+  sret_test_function_type f = (sret_test_function_type)(ptr);
+  sret_test_type x = {5436LL, 658277698LL};
+  sret_test_type r = f(x);
+  return ((r.a==x.a*35)&&(r.b==x.b*19)) ? 0 : -1;
 }
 
-static int test1() {
+static int sret_test(void) {
   const char *src =
-  "typedef struct test1_type_s {int x, y;} test1_type;"
-  "test1_type f() {\n"
-  "  test1_type r = {10, 35};\n"
+  "typedef struct sret_test_type_s {\n"
+  "  long long a;\n"
+  "  long long b;\n"
+  "} sret_test_type;\n"
+  "sret_test_type f(sret_test_type x) {\n"
+  "  sret_test_type r = {x.a*35, x.b*19};\n"
   "  return r;\n"
   "}\n";
   
-  return run_callback(src, test1_callback);
+  return run_callback(src, sret_test_callback);
 }
 
 #define RUN_TEST(t) \
@@ -75,6 +113,7 @@ int main(int argc, char **argv) {
   if (argc == 2 && !memcmp(argv[1], "lib_path=",9))
     tccdir = argv[1] + 9;
   
-  RUN_TEST(test1);
+  RUN_TEST(reg_pack_test);
+  RUN_TEST(sret_test);
   return retval;
 }