x86-64: chkstk, alloca

Makefile

@@ -11,6 +11,9 @@ LIBS_P=
 
 ifneq ($(GCC_MAJOR),2)
 CFLAGS+=-fno-strict-aliasing
+ifneq ($(GCC_MAJOR),3)
+CFLAGS+=-Wno-pointer-sign -Wno-sign-compare -D_FORTIFY_SOURCE=0
+endif
 endif
 
 ifeq ($(ARCH),i386)
@@ -19,15 +22,8 @@ ifeq ($(GCC_MAJOR),2)
 CFLAGS+=-m386 -malign-functions=0
 else
 CFLAGS+=-march=i386 -falign-functions=0
-ifneq ($(GCC_MAJOR),3)
-CFLAGS+=-Wno-pointer-sign -Wno-sign-compare -D_FORTIFY_SOURCE=0
 endif
 endif
-endif
-
-ifeq ($(ARCH),x86-64)
-CFLAGS+=-Wno-pointer-sign
-endif
 
 ifndef CONFIG_WIN32
 LIBS=-lm
@@ -36,26 +32,29 @@ LIBS+=-ldl
 endif
 endif
 
-ifdef CONFIG_WIN32
-NATIVE_TARGET=-DTCC_TARGET_PE
-LIBTCC1=libtcc1.a
-else
 ifeq ($(ARCH),i386)
 NATIVE_TARGET=-DTCC_TARGET_I386
 LIBTCC1=libtcc1.a
 BCHECK_O=bcheck.o
-else
+ALLOCA_O=alloca86.o alloca86-bt.o
-ifeq ($(ARCH),arm)
-NATIVE_TARGET=-DTCC_TARGET_ARM
-NATIVE_TARGET+=$(if $(wildcard /lib/ld-linux.so.3),-DTCC_ARM_EABI)
-NATIVE_TARGET+=$(if $(shell grep -l "^Features.* \(vfp\|iwmmxt\) " /proc/cpuinfo),-DTCC_ARM_VFP)
 else
 ifeq ($(ARCH),x86-64)
 NATIVE_TARGET=-DTCC_TARGET_X86_64
 LIBTCC1=libtcc1.a
+BCHECK_O=
+ALLOCA_O=alloca86_64.o
 endif
 endif
+
+ifeq ($(ARCH),arm)
+NATIVE_TARGET=-DTCC_TARGET_ARM
+NATIVE_TARGET+=$(if $(wildcard /lib/ld-linux.so.3),-DTCC_ARM_EABI)
+NATIVE_TARGET+=$(if $(shell grep -l "^Features.* \(vfp\|iwmmxt\) " /proc/cpuinfo),-DTCC_ARM_VFP)
 endif
+
+ifdef CONFIG_WIN32
+NATIVE_TARGET+=-DTCC_TARGET_PE
+BCHECK_O=
 endif
 
 ifneq ($(wildcard /lib/ld-uClibc.so.0),)
@@ -70,7 +69,6 @@ endif
 ifeq ($(TOP),.)
 
 PROGS=tcc$(EXESUF)
-
 I386_CROSS = i386-tcc$(EXESUF)
 WIN32_CROSS = i386-win32-tcc$(EXESUF)
 WIN64_CROSS = x86_64-win32-tcc$(EXESUF)
@@ -171,26 +169,19 @@ tiny_libmaker$(EXESUF): win32/tools/tiny_libmaker.c
 	$(CC) -o $@ $< $(CFLAGS)
 
 # TinyCC runtime libraries
-LIBTCC1_OBJS=libtcc1.o
+LIBTCC1_OBJS=libtcc1.o $(ALLOCA_O)
 LIBTCC1_CC=$(CC)
 VPATH+=lib
+
 ifdef CONFIG_WIN32
 # for windows, we must use TCC because we generate ELF objects
 LIBTCC1_OBJS+=crt1.o wincrt1.o dllcrt1.o dllmain.o chkstk.o
 LIBTCC1_CC=./tcc.exe -Bwin32 -Iinclude $(NATIVE_TARGET)
 VPATH+=win32/lib
 endif
-ifeq ($(ARCH),i386)
-LIBTCC1_OBJS+=alloca86.o alloca86-bt.o
-else
-ifeq ($(ARCH),x86-64)
-LIBTCC1_OBJS+=alloca86_64.o
-endif
-endif
 
 %.o: %.c
 	$(LIBTCC1_CC) -o $@ -c $< -O2 -Wall
-
 %.o: %.S
 	$(LIBTCC1_CC) -o $@ -c $<
 

lib/alloca86-bt.S

@@ -1,5 +1,5 @@
 /* ---------------------------------------------- */
-/* alloca86b.S */
+/* alloca86-bt.S */
 
 #include "../config.h"
 

lib/alloca86_64.S

@@ -7,13 +7,16 @@
 
 alloca:
     pop     %rdx
+#ifdef TCC_TARGET_PE
+    mov     %rcx,%rax
+#else
     mov     %rdi,%rax
+#endif
     add     $15,%rax
     and     $-16,%rax
     jz      p3
 
 #ifdef TCC_TARGET_PE
-    /* XXX: not tested */
 p1:
     cmp     $4096,%rax
     jle     p2
@@ -26,6 +29,10 @@ p2:
 
     sub     %rax,%rsp
     mov     %rsp,%rax
+#ifdef TCC_TARGET_PE
+    add     $32,%rax
+#endif
+
 p3:
     push    %rdx
     ret

tccpe.c

@@ -350,12 +350,7 @@ ST_DATA struct pe_header pe_header = {
     0x0002, /*WORD    Subsystem; */
     0x0000, /*WORD    DllCharacteristics; */
     0x00100000, /*DWORD   SizeOfStackReserve; */
-#ifdef TCC_TARGET_X86_64
-    // need to have a __chkstk eventually
-    0x00008000, /*DWORD   SizeOfStackCommit; */
-#else
     0x00001000, /*DWORD   SizeOfStackCommit; */
-#endif
     0x00100000, /*DWORD   SizeOfHeapReserve; */
     0x00001000, /*DWORD   SizeOfHeapCommit; */
     0x00000000, /*DWORD   LoaderFlags; */

tests/Makefile

@@ -81,8 +81,8 @@ test4: test.ref
 	@if diff -u test.ref test4.out ; then echo "BCheck Auto Test OK"; fi
 
 # memory and bound check auto test
-BOUNDS_OK  = 1 4 8 10
+BOUNDS_OK  = 1 4 8 10 14
-BOUNDS_FAIL= 2 5 7 9 11 12 13
+BOUNDS_FAIL= 2 5 7 9 11 12 13 15
 
 btest: boundtest.c
 	@echo ------------ $@ ------------

tests/boundtest.c

@@ -169,6 +169,23 @@ int test13(void)
     return strlen(tab);
 }
 
+int test14(void)
+{
+    char *p = alloca(TAB_SIZE);
+    memset(p, 'a', TAB_SIZE);
+    p[TAB_SIZE-1] = 0;
+    return strlen(p);
+}
+
+/* error */
+int test15(void)
+{
+    char *p = alloca(TAB_SIZE-1);
+    memset(p, 'a', TAB_SIZE);
+    p[TAB_SIZE-1] = 0;
+    return strlen(p);
+}
+
 int (*table_test[])(void) = {
     test1,
     test1,
@@ -184,6 +201,8 @@ int (*table_test[])(void) = {
     test11,
     test12,
     test13,
+    test14,
+    test15,
 };
 
 int main(int argc, char **argv)

win32/lib/chkstk.S

@@ -1,14 +1,19 @@
-// =================================================
+/* ---------------------------------------------- */
-// chkstk.s
+/* chkstk86.s */
+
+#include "../../config.h"
+
+/* ---------------------------------------------- */
+#ifndef TCC_TARGET_X86_64
+/* ---------------------------------------------- */
 
-.text
 .globl __chkstk
 
 __chkstk:
-	xchg    (%esp), %ebp   // store ebp, get ret.addr
+    xchg    (%esp),%ebp     /* store ebp, get ret.addr */
-	push    %ebp           // push ret.addr
+    push    %ebp            /* push ret.addr */
-	lea     4(%esp), %ebp  // setup frame ptr
+    lea     4(%esp),%ebp    /* setup frame ptr */
-	push    %ecx           // save ecx
+    push    %ecx            /* save ecx */
     mov     %ebp,%ecx
 P0:
     sub     $4096,%ecx
@@ -16,14 +21,41 @@ P0:
     sub     $4096,%eax
     cmp     $4096,%eax
     jge     P0
-
     sub     %eax,%ecx
-	mov     %esp,%eax
     test    %eax,(%ecx)
+
+    mov     %esp,%eax
     mov     %ecx,%esp
+    mov     (%eax),%ecx     /* restore ecx */
+    jmp     *4(%eax)
 
-	mov     (%eax),%ecx     // restore ecx
+/* ---------------------------------------------- */
-	mov     4(%eax),%eax
+#else
-	push    %eax
+/* ---------------------------------------------- */
-	ret
 
+.globl __chkstk
+
+__chkstk:
+    xchg    (%rsp),%rbp     /* store ebp, get ret.addr */
+    push    %rbp            /* push ret.addr */
+    lea     8(%rsp),%rbp    /* setup frame ptr */
+    push    %rcx            /* save ecx */
+    mov     %rbp,%rcx
+    movslq  %eax,%rax
+P0:
+    sub     $4096,%rcx
+    test    %rax,(%rcx)
+    sub     $4096,%rax
+    cmp     $4096,%rax
+    jge     P0
+    sub     %rax,%rcx
+    test    %rax,(%rcx)
+
+    mov     %rsp,%rax
+    mov     %rcx,%rsp
+    mov     (%rax),%rcx     /* restore ecx */
+    jmp     *8(%rax)
+
+/* ---------------------------------------------- */
+#endif
+/* ---------------------------------------------- */

x86_64-gen.c

@@ -919,12 +919,13 @@ void gfunc_epilog(void)
     v = (-loc + 15) & -16;
     saved_ind = ind;
     ind = func_sub_sp_offset - FUNC_PROLOG_SIZE;
-#if 0 // def TCC_TARGET_PE - don't have __chkstk yet, because assembler does not work
+#ifdef TCC_TARGET_PE
     if (v >= 4096) {
         Sym *sym = external_global_sym(TOK___chkstk, &func_old_type, 0);
         oad(0xb8, v); /* mov stacksize, %eax */
         oad(0xe8, -4); /* call __chkstk, (does the stackframe too) */
         greloc(cur_text_section, sym, ind-4, R_X86_64_PC32);
+	o(0x90); /* fill for FUNC_PROLOG_SIZE = 11 bytes */
     } else
 #endif
     {