fix mingw build and crashing bugs for Python Windows ARM64 (#496)

* fix mingw build and crashing bugs for Python Windows ARM64

* Fix issues found in PR review

9 files changed, 53 insertions, 15 deletions

diff --git a/.appveyor.yml b/.appveyor.yml
index 1c7fe9b..7a97c2a 100644
--- a/.appveyor.yml
+++ b/.appveyor.yml
@@ -13,6 +13,7 @@ platform:
   - x64
   - x86
   - arm
+  - arm64
 
 environment:
   global:
@@ -30,6 +31,12 @@ install:
           $env:HOST="i686-pc-cygwin"
           $env:MSVCC="/cygdrive/c/projects/libffi/msvcc.sh"
           $env:SRC_ARCHITECTURE="x86"
+        } ElseIf ($env:Platform -Match "arm64") {
+          $env:VCVARS_PLATFORM="x86_arm64"
+          $env:BUILD="i686-pc-cygwin"
+          $env:HOST="aarch64-w64-cygwin"
+          $env:MSVCC="/cygdrive/c/projects/libffi/msvcc.sh -marm64"
+          $env:SRC_ARCHITECTURE="aarch64"
         } ElseIf ($env:Platform -Match "arm") {
           $env:VCVARS_PLATFORM="x86_arm"
           $env:BUILD="i686-pc-cygwin"
diff --git a/Makefile.am b/Makefile.am
index b74b73b..07d928e 100644
--- a/Makefile.am
+++ b/Makefile.am
@@ -79,6 +79,7 @@ noinst_HEADERS = \
 
 EXTRA_libffi_la_SOURCES = \
 	src/aarch64/ffi.c src/aarch64/sysv.S				\
+	src/aarch64/win64_armasm.S						\
 	src/alpha/ffi.c src/alpha/osf.S					\
 	src/arc/ffi.c src/arc/arcompact.S				\
 	src/arm/ffi.c src/arm/sysv.S					\
diff --git a/configure.host b/configure.host
index 8181a61..6762eba 100644
--- a/configure.host
+++ b/configure.host
@@ -6,6 +6,11 @@
 # THIS TABLE IS SORTED.  KEEP IT THAT WAY.
 # Most of the time we can define all the variables all at once...
 case "${host}" in
+  aarch64*-*-cygwin* | aarch64*-*-mingw* | aarch64*-*-win* )
+	TARGET=ARM_WIN64; TARGETDIR=aarch64
+	MSVC=1
+	;;
+
   aarch64*-*-*)
 	TARGET=AARCH64; TARGETDIR=aarch64
 	SOURCES="ffi.c sysv.S"
@@ -250,6 +255,9 @@ case "${TARGET}" in
   ARM_WIN32)
 	SOURCES="ffi.c sysv_msvc_arm32.S"
 	;;
+  ARM_WIN64)
+	SOURCES="ffi.c win64_armasm.S"
+	;;
   MIPS)
 	SOURCES="ffi.c o32.S n32.S"
 	;;
diff --git a/msvc_build/aarch64/aarch64_include/ffi.h b/msvc_build/aarch64/aarch64_include/ffi.h
index 3adb5cb..02f26a2 100644
--- a/msvc_build/aarch64/aarch64_include/ffi.h
+++ b/msvc_build/aarch64/aarch64_include/ffi.h
@@ -197,7 +197,7 @@ FFI_EXTERN ffi_type ffi_type_float;
 FFI_EXTERN ffi_type ffi_type_double;
 FFI_EXTERN ffi_type ffi_type_pointer;
 
-#ifndef _M_ARM64 
+#ifndef _M_ARM64
 FFI_EXTERN ffi_type ffi_type_longdouble;
 #else
 #define ffi_type_longdouble ffi_type_double
@@ -227,7 +227,6 @@ typedef struct {
   ffi_type *rtype;
   unsigned bytes;
   unsigned flags;
-  unsigned isVariadic;
 #ifdef FFI_EXTRA_CIF_FIELDS
   FFI_EXTRA_CIF_FIELDS;
 #endif
diff --git a/msvcc.sh b/msvcc.sh
index 9c52f52..97facd6 100755
--- a/msvcc.sh
+++ b/msvcc.sh
@@ -85,6 +85,11 @@ do
       safeseh=
       shift 1
     ;;
+    -marm64)
+      ml='armasm64'
+      safeseh=
+      shift 1
+    ;;
     -clang-cl)
       cl="clang-cl"
       shift 1
@@ -299,6 +304,10 @@ if [ -n "$assembly" ]; then
       defines="$defines -D_M_ARM"
     fi
 
+    if [ $ml = "armasm64" ]; then
+      defines="$defines -D_M_ARM64"
+    fi
+
     if test -n "$verbose"; then
       echo "$cl -nologo -EP $includes $defines $src > $ppsrc"
     fi
@@ -307,6 +316,8 @@ if [ -n "$assembly" ]; then
     output="$(echo $output | sed 's%/F[dpa][^ ]*%%g')"
     if [ $ml = "armasm" ]; then
       args="-nologo -g -oldit $armasm_output $ppsrc -errorReport:prompt"
+    elif [ $ml = "armasm64" ]; then
+      args="-nologo -g $armasm_output $ppsrc -errorReport:prompt"
     else
       args="-nologo $safeseh $single $output $ppsrc"
     fi
diff --git a/src/aarch64/ffi.c b/src/aarch64/ffi.c
index 4225f72..1ebf43c 100644
--- a/src/aarch64/ffi.c
+++ b/src/aarch64/ffi.c
@@ -666,7 +666,7 @@ ffi_call_int (ffi_cif *cif, void (*fn)(void), void *orig_rvalue,
 	      {
 		int elems = 4 - (h & 3);
 #ifdef _M_ARM64 /* for handling armasm calling convention */
-                if (cif->isVariadic)
+                if (cif->is_variadic)
                   {
                     if (state.ngrn + elems <= N_X_ARG_REG)
                       {
@@ -808,7 +808,13 @@ ffi_prep_closure_loc (ffi_closure *closure,
   ffi_clear_cache(tramp, tramp + FFI_TRAMPOLINE_SIZE);
 
   /* Also flush the cache for code mapping.  */
+#ifdef _M_ARM64
+  // Not using dlmalloc.c for Windows ARM64 builds
+  // so calling ffi_data_to_code_pointer() isn't necessary
+  unsigned char *tramp_code = tramp;
+  #else
   unsigned char *tramp_code = ffi_data_to_code_pointer (tramp);
+  #endif
   ffi_clear_cache (tramp_code, tramp_code + FFI_TRAMPOLINE_SIZE);
 #endif
 
@@ -909,7 +915,7 @@ ffi_closure_SYSV_inner (ffi_cif *cif,
 	    {
 	      n = 4 - (h & 3);
 #ifdef _M_ARM64  /* for handling armasm calling convention */
-              if (cif->isVariadic)
+              if (cif->is_variadic)
                 {
                   if (state.ngrn + n <= N_X_ARG_REG)
                     {
@@ -948,7 +954,7 @@ ffi_closure_SYSV_inner (ffi_cif *cif,
                       avalue[i] = allocate_to_stack(&state, stack,
                                                    ty->alignment, s);
                     }
-#ifdef _M_ARM64  /* for handling armasm calling convention */    
+#ifdef _M_ARM64  /* for handling armasm calling convention */
                 }
 #endif  /* for handling armasm calling convention */
             }
diff --git a/src/aarch64/ffitarget.h b/src/aarch64/ffitarget.h
index 32c4f08..ecb6d2d 100644
--- a/src/aarch64/ffitarget.h
+++ b/src/aarch64/ffitarget.h
@@ -69,6 +69,10 @@ typedef enum ffi_abi
 #define FFI_TRAMPOLINE_CLOSURE_OFFSET FFI_TRAMPOLINE_SIZE
 #endif
 
+#ifdef _M_ARM64
+#define FFI_EXTRA_CIF_FIELDS unsigned is_variadic
+#endif
+
 /* ---- Internal ---- */
 
 #if defined (__APPLE__)
diff --git a/src/aarch64/win64_armasm.S b/src/aarch64/win64_armasm.S
index 906ceff..8ebc444 100644
--- a/src/aarch64/win64_armasm.S
+++ b/src/aarch64/win64_armasm.S
@@ -57,12 +57,14 @@ SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.  */
    x5 closure

 */

 

-	NESTED_ENTRY ffi_call_SYSV

-	/* Use a stack frame allocated by our caller. */

-	PROLOG_NOP	stp	x29, x30, [x1, #32]

-    /* For unwind information, Windows has to store fp and lr  */

+	NESTED_ENTRY ffi_call_SYSV_fake

+

+	/* For unwind information, Windows has to store fp and lr  */

 	PROLOG_SAVE_REG_PAIR	x29, x30, #-32!

-	

+

+	ALTERNATE_ENTRY ffi_call_SYSV

+	/* Use a stack frame allocated by our caller. */

+	stp	x29, x30, [x1]

 	mov	x29, x1

 	mov	sp, x0

 

@@ -97,8 +99,8 @@ ffi_call_SYSV_L1
 

 	/* Partially deconstruct the stack frame. */

 	mov     sp, x29 

-	ldp     x29, x30, [x29, #32]

-	

+	ldp     x29, x30, [x29]

+

 	/* Save the return value as directed.  */

 	adr	x5, ffi_call_SYSV_return

 	and	w4, w4, #AARCH64_RET_MASK

@@ -177,7 +179,7 @@ ffi_call_SYSV_return
 	nop

 	

 	

-	NESTED_END ffi_call_SYSV	

+	NESTED_END ffi_call_SYSV_fake

 	

 

 /* ffi_closure_SYSV

diff --git a/src/prep_cif.c b/src/prep_cif.c
index 406e4b6..06c6544 100644
--- a/src/prep_cif.c
+++ b/src/prep_cif.c
@@ -129,8 +129,8 @@ ffi_status FFI_HIDDEN ffi_prep_cif_core(ffi_cif *cif, ffi_abi abi,
   cif->rtype = rtype;
 
   cif->flags = 0;
- #ifdef _M_ARM64
-  cif->isVariadic = isvariadic;
+#ifdef _M_ARM64
+  cif->is_variadic = isvariadic;
 #endif
 #if HAVE_LONG_DOUBLE_VARIANT
   ffi_prep_types (abi);