Support loongarch64 (#678)

* update config.{guess,sub}

* Support loongarch64

Co-Authored-By: Cheng Lulu <chenglulu@loongson.cn>
Co-Authored-By: Xi Ruoyao <xry111@mengyan1223.wang>
Co-Authored-By: Xu Hao
Co-Authored-By: Zhang Wenlong <zhangwenlong@loongson.cn>
Co-Authored-By: Pan Xuefeng <panxuefeng@loongson.cn>

Co-authored-by: panxuefeng <panxuefeng@loongson.cn>
Co-authored-by: Cheng Lulu <chenglulu@loongson.cn>
Co-authored-by: Xi Ruoyao <xry111@mengyan1223.wang>

3 files changed, 973 insertions, 0 deletions

diff --git a/src/loongarch64/ffi.c b/src/loongarch64/ffi.c
new file mode 100644
index 0000000..7a28892
--- /dev/null
+++ b/src/loongarch64/ffi.c
@@ -0,0 +1,595 @@
+/* -----------------------------------------------------------------------
+   ffi.c - Copyright (c) 2022 Xu Chenghua <xuchenghua@loongson.cn>
+                         2022 Cheng Lulu <chenglulu@loongson.cn>
+   Based on RISC-V port
+
+   LoongArch Foreign Function Interface
+
+   Permission is hereby granted, free of charge, to any person obtaining
+   a copy of this software and associated documentation files (the
+   ``Software''), to deal in the Software without restriction, including
+   without limitation the rights to use, copy, modify, merge, publish,
+   distribute, sublicense, and/or sell copies of the Software, and to
+   permit persons to whom the Software is furnished to do so, subject to
+   the following conditions:
+
+   The above copyright notice and this permission notice shall be included
+   in all copies or substantial portions of the Software.
+
+   THE SOFTWARE IS PROVIDED ``AS IS'', WITHOUT WARRANTY OF ANY KIND,
+   EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+   MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+   NONINFRINGEMENT.  IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT
+   HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY,
+   WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+   OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
+   DEALINGS IN THE SOFTWARE.
+   ----------------------------------------------------------------------- */
+
+#include <ffi.h>
+#include <ffi_common.h>
+
+#include <stdlib.h>
+#include <stdint.h>
+
+#if defined(__loongarch_soft_float)
+# define ABI_FRLEN 0
+#elif defined(__loongarch_single_float)
+# define ABI_FRLEN 32
+# define ABI_FLOAT float
+#elif defined(__loongarch_double_float)
+# define ABI_FRLEN 64
+# define ABI_FLOAT double
+#else
+#error unsupported LoongArch floating-point ABI
+#endif
+
+#define NARGREG 8
+#define STKALIGN 16
+#define MAXCOPYARG (2 * sizeof (double))
+
+/* call_context registers
+   - 8 floating point parameter/result registers.
+   - 8 integer parameter/result registers.
+   - 2 registers used by the assembly code to in-place construct its own
+     stack frame
+     - frame register
+     - return register
+*/
+typedef struct call_context
+{
+  ABI_FLOAT fa[8];
+  size_t a[10];
+} call_context;
+
+typedef struct call_builder
+{
+  call_context *aregs;
+  int used_integer;
+  int used_float;
+  size_t *used_stack;
+  size_t *stack;
+  size_t next_struct_area;
+} call_builder;
+
+/* Integer (not pointer) less than ABI GRLEN.  */
+/* FFI_TYPE_INT does not appear to be used.  */
+#if __SIZEOF_POINTER__ == 8
+# define IS_INT(type) ((type) >= FFI_TYPE_UINT8 && (type) <= FFI_TYPE_SINT64)
+#else
+# define IS_INT(type) ((type) >= FFI_TYPE_UINT8 && (type) <= FFI_TYPE_SINT32)
+#endif
+
+#if ABI_FRLEN
+typedef struct float_struct_info
+{
+  char as_elements;
+  char type1;
+  char offset2;
+  char type2;
+} float_struct_info;
+
+#if ABI_FRLEN >= 64
+# define IS_FLOAT(type) ((type) >= FFI_TYPE_FLOAT && (type) <= FFI_TYPE_DOUBLE)
+#else
+# define IS_FLOAT(type) ((type) == FFI_TYPE_FLOAT)
+#endif
+
+static ffi_type **
+flatten_struct (ffi_type *in, ffi_type **out, ffi_type **out_end)
+{
+  int i;
+
+  if (out == out_end)
+    return out;
+  if (in->type != FFI_TYPE_STRUCT)
+    *(out++) = in;
+  else
+    for (i = 0; in->elements[i]; i++)
+      out = flatten_struct (in->elements[i], out, out_end);
+  return out;
+}
+
+/* Structs with at most two fields after flattening, one of which is of
+   floating point type, are passed in multiple registers if sufficient
+   registers are available.  */
+static float_struct_info
+struct_passed_as_elements (call_builder *cb, ffi_type *top)
+{
+  float_struct_info ret = {0, 0, 0, 0};
+  ffi_type *fields[3];
+  int num_floats, num_ints;
+  int num_fields = flatten_struct (top, fields, fields + 3) - fields;
+
+  if (num_fields == 1)
+    {
+      if (IS_FLOAT (fields[0]->type))
+	{
+	  ret.as_elements = 1;
+	  ret.type1 = fields[0]->type;
+	}
+    }
+  else if (num_fields == 2)
+    {
+      num_floats = IS_FLOAT (fields[0]->type) + IS_FLOAT (fields[1]->type);
+      num_ints = IS_INT (fields[0]->type) + IS_INT (fields[1]->type);
+      if (num_floats == 0 || num_floats + num_ints != 2)
+	return ret;
+      if (cb->used_float + num_floats > NARGREG
+	  || cb->used_integer + (2 - num_floats) > NARGREG)
+	return ret;
+      if (!IS_FLOAT (fields[0]->type) && !IS_FLOAT (fields[1]->type))
+	return ret;
+
+      ret.type1 = fields[0]->type;
+      ret.type2 = fields[1]->type;
+      ret.offset2 = FFI_ALIGN (fields[0]->size, fields[1]->alignment);
+      ret.as_elements = 1;
+    }
+  return ret;
+}
+#endif
+
+/* Allocates a single register, float register, or GRLEN-sized stack slot to a
+   datum.  */
+static void
+marshal_atom (call_builder *cb, int type, void *data)
+{
+  size_t value = 0;
+  switch (type)
+    {
+    case FFI_TYPE_UINT8:
+      value = *(uint8_t *) data;
+      break;
+    case FFI_TYPE_SINT8:
+      value = *(int8_t *) data;
+      break;
+    case FFI_TYPE_UINT16:
+      value = *(uint16_t *) data;
+      break;
+    case FFI_TYPE_SINT16:
+      value = *(int16_t *) data;
+      break;
+    /* 32-bit quantities are always sign-extended in the ABI.  */
+    case FFI_TYPE_UINT32:
+      value = *(int32_t *) data;
+      break;
+    case FFI_TYPE_SINT32:
+      value = *(int32_t *) data;
+      break;
+#if __SIZEOF_POINTER__ == 8
+    case FFI_TYPE_UINT64:
+      value = *(uint64_t *) data;
+      break;
+    case FFI_TYPE_SINT64:
+      value = *(int64_t *) data;
+      break;
+#endif
+    case FFI_TYPE_POINTER:
+      value = *(size_t *) data;
+      break;
+
+#if ABI_FRLEN >= 32
+    case FFI_TYPE_FLOAT:
+      *(float *)(cb->aregs->fa + cb->used_float++) = *(float *) data;
+      return;
+#endif
+#if ABI_FRLEN >= 64
+    case FFI_TYPE_DOUBLE:
+      (cb->aregs->fa[cb->used_float++]) = *(double *) data;
+      return;
+#endif
+    default:
+      FFI_ASSERT (0);
+      break;
+    }
+
+  if (cb->used_integer == NARGREG)
+    *cb->used_stack++ = value;
+  else
+    cb->aregs->a[cb->used_integer++] = value;
+}
+
+static void
+unmarshal_atom (call_builder *cb, int type, void *data)
+{
+  size_t value;
+  switch (type)
+    {
+#if ABI_FRLEN >= 32
+    case FFI_TYPE_FLOAT:
+      *(float *) data = *(float *)(cb->aregs->fa + cb->used_float++);
+      return;
+#endif
+#if ABI_FRLEN >= 64
+    case FFI_TYPE_DOUBLE:
+      *(double *) data = cb->aregs->fa[cb->used_float++];
+      return;
+#endif
+    }
+
+  if (cb->used_integer == NARGREG)
+    value = *cb->used_stack++;
+  else
+    value = cb->aregs->a[cb->used_integer++];
+
+  switch (type)
+    {
+    case FFI_TYPE_UINT8:
+    case FFI_TYPE_SINT8:
+    case FFI_TYPE_UINT16:
+    case FFI_TYPE_SINT16:
+    case FFI_TYPE_UINT32:
+    case FFI_TYPE_SINT32:
+#if __SIZEOF_POINTER__ == 8
+    case FFI_TYPE_UINT64:
+    case FFI_TYPE_SINT64:
+#endif
+    case FFI_TYPE_POINTER:
+      *(ffi_arg *)data = value;
+      break;
+    default:
+      FFI_ASSERT (0);
+      break;
+    }
+}
+
+/* Allocate and copy a structure that is passed by value on the stack and
+   return a pointer to it.  */
+static void *
+allocate_and_copy_struct_to_stack (call_builder *cb, void *data,
+				   ffi_type *type)
+{
+  size_t dest = cb->next_struct_area - type->size;
+
+  dest = FFI_ALIGN_DOWN (dest, type->alignment);
+  cb->next_struct_area = dest;
+
+  return memcpy ((char *)cb->stack + dest, data, type->size);
+}
+
+/* Adds an argument to a call, or a not by reference return value.  */
+static void
+marshal (call_builder *cb, ffi_type *type, int var, void *data)
+{
+  size_t realign[2];
+
+#if ABI_FRLEN
+  if (!var && type->type == FFI_TYPE_STRUCT)
+    {
+      float_struct_info fsi = struct_passed_as_elements (cb, type);
+      if (fsi.as_elements)
+	{
+	  marshal_atom (cb, fsi.type1, data);
+	  if (fsi.offset2)
+	    marshal_atom (cb, fsi.type2, ((char *) data) + fsi.offset2);
+	  return;
+	}
+    }
+
+  if (!var && cb->used_float < NARGREG
+      && IS_FLOAT (type->type))
+    {
+      marshal_atom (cb, type->type, data);
+      return;
+    }
+
+  double promoted;
+  if (var && type->type == FFI_TYPE_FLOAT)
+    {
+      /* C standard requires promoting float -> double for variable arg.  */
+      promoted = *(float *) data;
+      type = &ffi_type_double;
+      data = &promoted;
+    }
+#endif
+
+  if (type->size > 2 * __SIZEOF_POINTER__)
+    /* Pass by reference.  */
+    {
+      allocate_and_copy_struct_to_stack (cb, data, type);
+      data = (char *)cb->stack + cb->next_struct_area;
+      marshal_atom (cb, FFI_TYPE_POINTER, &data);
+    }
+  else if (IS_INT (type->type) || type->type == FFI_TYPE_POINTER)
+    marshal_atom (cb, type->type, data);
+  else
+    {
+      /* Overlong integers, soft-float floats, and structs without special
+	 float handling are treated identically from this point on.  */
+
+      /* Variadics are aligned even in registers.  */
+      if (type->alignment > __SIZEOF_POINTER__)
+	{
+	  if (var)
+	    cb->used_integer = FFI_ALIGN (cb->used_integer, 2);
+	  cb->used_stack
+	    = (size_t *) FFI_ALIGN (cb->used_stack, 2 * __SIZEOF_POINTER__);
+	}
+
+      memcpy (realign, data, type->size);
+      if (type->size > 0)
+	marshal_atom (cb, FFI_TYPE_POINTER, realign);
+      if (type->size > __SIZEOF_POINTER__)
+	marshal_atom (cb, FFI_TYPE_POINTER, realign + 1);
+    }
+}
+
+/* For arguments passed by reference returns the pointer, otherwise the arg
+   is copied (up to MAXCOPYARG bytes).  */
+static void *
+unmarshal (call_builder *cb, ffi_type *type, int var, void *data)
+{
+  size_t realign[2];
+  void *pointer;
+
+#if ABI_FRLEN
+  if (!var && type->type == FFI_TYPE_STRUCT)
+    {
+      float_struct_info fsi = struct_passed_as_elements (cb, type);
+      if (fsi.as_elements)
+	{
+	  unmarshal_atom (cb, fsi.type1, data);
+	  if (fsi.offset2)
+	    unmarshal_atom (cb, fsi.type2, ((char *) data) + fsi.offset2);
+	  return data;
+	}
+    }
+
+  if (!var && cb->used_float < NARGREG
+      && IS_FLOAT (type->type))
+    {
+      unmarshal_atom (cb, type->type, data);
+      return data;
+    }
+
+  if (var && type->type == FFI_TYPE_FLOAT)
+    {
+      int m = cb->used_integer;
+      void *promoted
+	= m < NARGREG ? cb->aregs->a + m : cb->used_stack + m - NARGREG + 1;
+      *(float *) promoted = *(double *) promoted;
+    }
+#endif
+
+  if (type->size > 2 * __SIZEOF_POINTER__)
+    {
+      /* Pass by reference.  */
+      unmarshal_atom (cb, FFI_TYPE_POINTER, (char *) &pointer);
+      return pointer;
+    }
+  else if (IS_INT (type->type) || type->type == FFI_TYPE_POINTER)
+    {
+      unmarshal_atom (cb, type->type, data);
+      return data;
+    }
+  else
+    {
+      /* Overlong integers, soft-float floats, and structs without special
+	 float handling are treated identically from this point on.  */
+
+      /* Variadics are aligned even in registers.  */
+      if (type->alignment > __SIZEOF_POINTER__)
+	{
+	  if (var)
+	    cb->used_integer = FFI_ALIGN (cb->used_integer, 2);
+	  cb->used_stack
+	    = (size_t *) FFI_ALIGN (cb->used_stack, 2 * __SIZEOF_POINTER__);
+	}
+
+      if (type->size > 0)
+	unmarshal_atom (cb, FFI_TYPE_POINTER, realign);
+      if (type->size > __SIZEOF_POINTER__)
+	unmarshal_atom (cb, FFI_TYPE_POINTER, realign + 1);
+      memcpy (data, realign, type->size);
+      return data;
+    }
+}
+
+static int
+passed_by_ref (call_builder *cb, ffi_type *type, int var)
+{
+#if ABI_FRLEN
+  if (!var && type->type == FFI_TYPE_STRUCT)
+    {
+      float_struct_info fsi = struct_passed_as_elements (cb, type);
+      if (fsi.as_elements)
+	return 0;
+    }
+#endif
+
+  return type->size > 2 * __SIZEOF_POINTER__;
+}
+
+/* Perform machine dependent cif processing.  */
+ffi_status
+ffi_prep_cif_machdep (ffi_cif *cif)
+{
+  cif->loongarch_nfixedargs = cif->nargs;
+  return FFI_OK;
+}
+
+/* Perform machine dependent cif processing when we have a variadic
+   function.  */
+ffi_status
+ffi_prep_cif_machdep_var (ffi_cif *cif, unsigned int nfixedargs,
+			  unsigned int ntotalargs)
+{
+  cif->loongarch_nfixedargs = nfixedargs;
+  return FFI_OK;
+}
+
+/* Low level routine for calling functions.  */
+extern void ffi_call_asm (void *stack, struct call_context *regs,
+			  void (*fn) (void), void *closure) FFI_HIDDEN;
+
+static void
+ffi_call_int (ffi_cif *cif, void (*fn) (void), void *rvalue, void **avalue,
+	      void *closure)
+{
+  /* This is a conservative estimate, assuming a complex return value and
+     that all remaining arguments are long long / __int128 */
+  size_t arg_bytes = cif->bytes;
+  size_t rval_bytes = 0;
+  if (rvalue == NULL && cif->rtype->size > 2 * __SIZEOF_POINTER__)
+    rval_bytes = FFI_ALIGN (cif->rtype->size, STKALIGN);
+  size_t alloc_size = arg_bytes + rval_bytes + sizeof (call_context);
+
+  /* The assembly code will deallocate all stack data at lower addresses
+     than the argument region, so we need to allocate the frame and the
+     return value after the arguments in a single allocation.  */
+  size_t alloc_base;
+  /* Argument region must be 16-byte aligned in LP64 ABIs.  */
+  if (_Alignof(max_align_t) >= STKALIGN)
+    /* Since sizeof long double is normally 16, the compiler will
+       guarantee alloca alignment to at least that much.  */
+    alloc_base = (size_t) alloca (alloc_size);
+  else
+    alloc_base = FFI_ALIGN (alloca (alloc_size + STKALIGN - 1), STKALIGN);
+
+  if (rval_bytes)
+    rvalue = (void *) (alloc_base + arg_bytes);
+
+  call_builder cb;
+  cb.used_float = cb.used_integer = 0;
+  cb.aregs = (call_context *) (alloc_base + arg_bytes + rval_bytes);
+  cb.used_stack = (void *) alloc_base;
+  cb.stack = (void *) alloc_base;
+  cb.next_struct_area = arg_bytes;
+
+  int return_by_ref = passed_by_ref (&cb, cif->rtype, 0);
+  if (return_by_ref)
+    cb.aregs->a[cb.used_integer++] = (size_t)rvalue;
+
+  int i;
+  for (i = 0; i < cif->nargs; i++)
+    marshal (&cb, cif->arg_types[i], i >= cif->loongarch_nfixedargs,
+	     avalue[i]);
+
+  ffi_call_asm ((void *) alloc_base, cb.aregs, fn, closure);
+
+  cb.used_float = cb.used_integer = 0;
+  if (!return_by_ref && rvalue)
+    unmarshal (&cb, cif->rtype, 0, rvalue);
+}
+
+void
+ffi_call (ffi_cif *cif, void (*fn) (void), void *rvalue, void **avalue)
+{
+  ffi_call_int (cif, fn, rvalue, avalue, NULL);
+}
+
+void
+ffi_call_go (ffi_cif *cif, void (*fn) (void), void *rvalue, void **avalue,
+	     void *closure)
+{
+  ffi_call_int (cif, fn, rvalue, avalue, closure);
+}
+
+extern void ffi_closure_asm (void) FFI_HIDDEN;
+
+ffi_status
+ffi_prep_closure_loc (ffi_closure *closure, ffi_cif *cif,
+		      void (*fun) (ffi_cif *, void *, void **, void *),
+		      void *user_data, void *codeloc)
+{
+  uint32_t *tramp = (uint32_t *) &closure->tramp[0];
+  uint64_t fn = (uint64_t) (uintptr_t) ffi_closure_asm;
+
+  if (cif->abi <= FFI_FIRST_ABI || cif->abi >= FFI_LAST_ABI)
+    return FFI_BAD_ABI;
+
+  /* We will call ffi_closure_inner with codeloc, not closure, but as long
+     as the memory is readable it should work.  */
+  tramp[0] = 0x1800000c; /* pcaddi $t0, 0 (i.e. $t0 <- tramp) */
+  tramp[1] = 0x28c0418d; /* ld.d   $t1, $t0, 16 */
+  tramp[2] = 0x4c0001a0; /* jirl   $zero, $t1, 0 */
+  tramp[3] = 0x03400000; /* nop */
+  tramp[4] = fn;
+  tramp[5] = fn >> 32;
+
+  closure->cif = cif;
+  closure->fun = fun;
+  closure->user_data = user_data;
+
+  __builtin___clear_cache (codeloc, codeloc + FFI_TRAMPOLINE_SIZE);
+  return FFI_OK;
+}
+
+extern void ffi_go_closure_asm (void) FFI_HIDDEN;
+
+ffi_status
+ffi_prep_go_closure (ffi_go_closure *closure, ffi_cif *cif,
+		     void (*fun) (ffi_cif *, void *, void **, void *))
+{
+  if (cif->abi <= FFI_FIRST_ABI || cif->abi >= FFI_LAST_ABI)
+    return FFI_BAD_ABI;
+
+  closure->tramp = (void *) ffi_go_closure_asm;
+  closure->cif = cif;
+  closure->fun = fun;
+  return FFI_OK;
+}
+
+/* Called by the assembly code with aregs pointing to saved argument registers
+   and stack pointing to the stacked arguments.  Return values passed in
+   registers will be reloaded from aregs.  */
+void FFI_HIDDEN
+ffi_closure_inner (ffi_cif *cif,
+		   void (*fun) (ffi_cif *, void *, void **, void *),
+		   void *user_data, size_t *stack, call_context *aregs)
+{
+  void **avalue = alloca (cif->nargs * sizeof (void *));
+  /* Storage for arguments which will be copied by unmarshal().  We could
+     theoretically avoid the copies in many cases and use at most 128 bytes
+     of memory, but allocating disjoint storage for each argument is
+     simpler.  */
+  char *astorage = alloca (cif->nargs * MAXCOPYARG);
+  void *rvalue;
+  call_builder cb;
+  int return_by_ref;
+  int i;
+
+  cb.aregs = aregs;
+  cb.used_integer = cb.used_float = 0;
+  cb.used_stack = stack;
+
+  return_by_ref = passed_by_ref (&cb, cif->rtype, 0);
+  if (return_by_ref)
+    unmarshal (&cb, &ffi_type_pointer, 0, &rvalue);
+  else
+    rvalue = alloca (cif->rtype->size);
+
+  for (i = 0; i < cif->nargs; i++)
+    avalue[i]
+      = unmarshal (&cb, cif->arg_types[i], i >= cif->loongarch_nfixedargs,
+		   astorage + i * MAXCOPYARG);
+
+  fun (cif, rvalue, avalue, user_data);
+
+  if (!return_by_ref && cif->rtype->type != FFI_TYPE_VOID)
+    {
+      cb.used_integer = cb.used_float = 0;
+      marshal (&cb, cif->rtype, 0, rvalue);
+    }
+}
diff --git a/src/loongarch64/ffitarget.h b/src/loongarch64/ffitarget.h
new file mode 100644
index 0000000..5a4698a
--- /dev/null
+++ b/src/loongarch64/ffitarget.h
@@ -0,0 +1,82 @@
+/* -----------------------------------------------------------------*-C-*-
+   ffitarget.h - Copyright (c) 2022 Xu Chenghua <xuchenghua@loongson.cn>
+                               2022 Cheng Lulu <chenglulu@loongson.cn>
+
+   Target configuration macros for LoongArch.
+
+   Permission is hereby granted, free of charge, to any person obtaining
+   a copy of this software and associated documentation files (the
+   ``Software''), to deal in the Software without restriction, including
+   without limitation the rights to use, copy, modify, merge, publish,
+   distribute, sublicense, and/or sell copies of the Software, and to
+   permit persons to whom the Software is furnished to do so, subject to
+   the following conditions:
+
+   The above copyright notice and this permission notice shall be included
+   in all copies or substantial portions of the Software.
+
+   THE SOFTWARE IS PROVIDED ``AS IS'', WITHOUT WARRANTY OF ANY KIND,
+   EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+   MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+   NONINFRINGEMENT.  IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT
+   HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY,
+   WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+   OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
+   DEALINGS IN THE SOFTWARE.
+
+   ----------------------------------------------------------------------- */
+
+#ifndef LIBFFI_TARGET_H
+#define LIBFFI_TARGET_H
+
+#ifndef LIBFFI_H
+#error \
+  "Please do not include ffitarget.h directly into your source.  Use ffi.h instead."
+#endif
+
+#ifndef __loongarch__
+#error \
+  "libffi was configured for a LoongArch target but this does not appear to be a LoongArch compiler."
+#endif
+
+#ifndef LIBFFI_ASM
+
+typedef unsigned long ffi_arg;
+typedef signed long ffi_sarg;
+
+typedef enum ffi_abi
+{
+  FFI_FIRST_ABI = 0,
+  FFI_LP64S,
+  FFI_LP64F,
+  FFI_LP64D,
+  FFI_LAST_ABI,
+
+#if defined(__loongarch64)
+#if defined(__loongarch_soft_float)
+  FFI_DEFAULT_ABI = FFI_LP64S
+#elif defined(__loongarch_single_float)
+  FFI_DEFAULT_ABI = FFI_LP64F
+#elif defined(__loongarch_double_float)
+  FFI_DEFAULT_ABI = FFI_LP64D
+#else
+#error unsupported LoongArch floating-point ABI
+#endif
+#else
+#error unsupported LoongArch base architecture
+#endif
+} ffi_abi;
+
+#endif /* LIBFFI_ASM */
+
+/* ---- Definitions for closures ----------------------------------------- */
+
+#define FFI_CLOSURES 1
+#define FFI_GO_CLOSURES 1
+#define FFI_TRAMPOLINE_SIZE 24
+#define FFI_NATIVE_RAW_API 0
+#define FFI_EXTRA_CIF_FIELDS \
+  unsigned loongarch_nfixedargs; \
+  unsigned loongarch_unused;
+#define FFI_TARGET_SPECIFIC_VARIADIC
+#endif
diff --git a/src/loongarch64/sysv.S b/src/loongarch64/sysv.S
new file mode 100644
index 0000000..9e0da11
--- /dev/null
+++ b/src/loongarch64/sysv.S
@@ -0,0 +1,296 @@
+/* -----------------------------------------------------------------------
+  sysv.S  - Copyright (c) 2022 Xu Chenghua <xuchenghua@loongson.cn>
+                          2022 Cheng Lulu <chenglulu@loongson.cn>
+
+   LoongArch Foreign Function Interface
+
+   Permission is hereby granted, free of charge, to any person obtaining
+   a copy of this software and associated documentation files (the
+   ``Software''), to deal in the Software without restriction, including
+   without limitation the rights to use, copy, modify, merge, publish,
+   distribute, sublicense, and/or sell copies of the Software, and to
+   permit persons to whom the Software is furnished to do so, subject to
+   the following conditions:
+
+   The above copyright notice and this permission notice shall be included
+   in all copies or substantial portions of the Software.
+
+   THE SOFTWARE IS PROVIDED ``AS IS'', WITHOUT WARRANTY OF ANY KIND,
+   EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+   MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+   NONINFRINGEMENT.  IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT
+   HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY,
+   WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+   OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
+   DEALINGS IN THE SOFTWARE.
+   ----------------------------------------------------------------------- */
+
+#define LIBFFI_ASM
+#include <fficonfig.h>
+#include <ffi.h>
+
+/* Define aliases so that we can handle all ABIs uniformly.  */
+
+#if __SIZEOF_POINTER__ == 8
+# define PTRS 8
+# define LARG ld.d
+# define SARG st.d
+#else
+# define PTRS 4
+# define LARG ld.w
+# define SARG st.w
+#endif
+
+#if defined(__loongarch_single_float)
+# define FLTS 4
+# define FLD fld.w
+# define FST fst.w
+#elif defined(__loongarch_double_float)
+# define FLTS 8
+# define FLARG fld.d
+# define FSARG fst.d
+#elif defined(__loongarch_soft_float)
+# define FLTS 0
+#else
+#error unsupported LoongArch floating-point ABI
+#endif
+
+    .text
+    .globl  ffi_call_asm
+    .type   ffi_call_asm, @function
+    .hidden ffi_call_asm
+/* struct call_context
+   {
+     ABI_FLOAT fa[8];
+     size_t a[10];
+   }
+
+   - 8 floating point parameter/result registers (fa[0] - fa[7])
+   - 8 integer parameter/result registers (a[0] - a[7])
+   - 2 registers used by the assembly code to in-place construct its own stack
+     frame.
+    - frame pointer (a[8])
+    - return address (a[9])
+
+   void ffi_call_asm (size_t *stackargs, struct call_context *regargs,
+		      void (*fn)(void), void *closure); */
+
+#define FRAME_LEN (8 * FLTS + 10 * PTRS)
+
+ffi_call_asm:
+	.cfi_startproc
+
+	/* We are NOT going to set up an ordinary stack frame.  In order to pass
+	   the stacked args to the called function, we adjust our stack pointer
+	   to a0, which is in the _caller's_ alloca area.  We establish our own
+	   stack frame at the end of the call_context.
+
+	   Anything below the arguments will be freed at this point, although
+	   we preserve the call_context so that it can be read back in the
+	   caller.  */
+
+	.cfi_def_cfa	5, FRAME_LEN # Interim CFA based on a1.
+	SARG	$fp, $a1, FRAME_LEN - 2*PTRS
+	.cfi_offset	22, -2*PTRS
+	SARG	$ra, $a1, FRAME_LEN - 1*PTRS
+	.cfi_offset	1, -1*PTRS
+
+	addi.d	$fp, $a1, FRAME_LEN
+	move	$sp, $a0
+	.cfi_def_cfa	22, 0 # Our frame is fully set up.
+
+	# Load arguments.
+	move	$t1, $a2
+	move	$t2, $a3
+
+#if FLTS
+	FLARG	$fa0, $fp, -FRAME_LEN+0*FLTS
+	FLARG	$fa1, $fp, -FRAME_LEN+1*FLTS
+	FLARG	$fa2, $fp, -FRAME_LEN+2*FLTS
+	FLARG	$fa3, $fp, -FRAME_LEN+3*FLTS
+	FLARG	$fa4, $fp, -FRAME_LEN+4*FLTS
+	FLARG	$fa5, $fp, -FRAME_LEN+5*FLTS
+	FLARG	$fa6, $fp, -FRAME_LEN+6*FLTS
+	FLARG	$fa7, $fp, -FRAME_LEN+7*FLTS
+#endif
+
+	LARG	$a0, $fp, -FRAME_LEN+8*FLTS+0*PTRS
+	LARG	$a1, $fp, -FRAME_LEN+8*FLTS+1*PTRS
+	LARG	$a2, $fp, -FRAME_LEN+8*FLTS+2*PTRS
+	LARG	$a3, $fp, -FRAME_LEN+8*FLTS+3*PTRS
+	LARG	$a4, $fp, -FRAME_LEN+8*FLTS+4*PTRS
+	LARG	$a5, $fp, -FRAME_LEN+8*FLTS+5*PTRS
+	LARG	$a6, $fp, -FRAME_LEN+8*FLTS+6*PTRS
+	LARG	$a7, $fp, -FRAME_LEN+8*FLTS+7*PTRS
+
+	/* Call */
+	jirl	$ra, $t1, 0
+
+#if FLTS
+	/* Save return values - only a0/a1 (fa0/fa1) are used.  */
+	FSARG	$fa0, $fp, -FRAME_LEN+0*FLTS
+	FSARG	$fa1, $fp, -FRAME_LEN+1*FLTS
+#endif
+
+	SARG	$a0, $fp, -FRAME_LEN+8*FLTS+0*PTRS
+	SARG	$a1, $fp, -FRAME_LEN+8*FLTS+1*PTRS
+
+	/* Restore and return.  */
+	addi.d	$sp, $fp, -FRAME_LEN
+	.cfi_def_cfa	3, FRAME_LEN
+	LARG	$ra, $fp, -1*PTRS
+	.cfi_restore	1
+	LARG	$fp, $fp, -2*PTRS
+	.cfi_restore	22
+	jr	$ra
+	.cfi_endproc
+	.size	ffi_call_asm, .-ffi_call_asm
+
+
+/* ffi_closure_asm. Expects address of the passed-in ffi_closure in t1.
+   void ffi_closure_inner (ffi_cif *cif,
+			   void (*fun)(ffi_cif *, void *, void **, void *),
+			   void *user_data,
+			   size_t *stackargs, struct call_context *regargs) */
+
+	.globl	ffi_closure_asm
+	.hidden	ffi_closure_asm
+	.type	ffi_closure_asm, @function
+
+ffi_closure_asm:
+	.cfi_startproc
+	addi.d	$sp, $sp, -FRAME_LEN
+	.cfi_def_cfa_offset FRAME_LEN
+
+	/* Make a frame.  */
+	SARG	$fp, $sp, FRAME_LEN - 2*PTRS
+	.cfi_offset	22, -2*PTRS
+	SARG	$ra, $sp, FRAME_LEN - 1*PTRS
+	.cfi_offset	1, -1*PTRS
+	addi.d	$fp, $sp, FRAME_LEN
+
+	/* Save arguments.  */
+#if FLTS
+	FSARG	$fa0, $sp, 0*FLTS
+	FSARG	$fa1, $sp, 1*FLTS
+	FSARG	$fa2, $sp, 2*FLTS
+	FSARG	$fa3, $sp, 3*FLTS
+	FSARG	$fa4, $sp, 4*FLTS
+	FSARG	$fa5, $sp, 5*FLTS
+	FSARG	$fa6, $sp, 6*FLTS
+	FSARG	$fa7, $sp, 7*FLTS
+#endif
+
+	SARG	$a0, $sp, 8*FLTS+0*PTRS
+	SARG	$a1, $sp, 8*FLTS+1*PTRS
+	SARG	$a2, $sp, 8*FLTS+2*PTRS
+	SARG	$a3, $sp, 8*FLTS+3*PTRS
+	SARG	$a4, $sp, 8*FLTS+4*PTRS
+	SARG	$a5, $sp, 8*FLTS+5*PTRS
+	SARG	$a6, $sp, 8*FLTS+6*PTRS
+	SARG	$a7, $sp, 8*FLTS+7*PTRS
+
+	/* Enter C */
+	LARG	$a0, $t0, FFI_TRAMPOLINE_SIZE+0*PTRS
+	LARG	$a1, $t0, FFI_TRAMPOLINE_SIZE+1*PTRS
+	LARG	$a2, $t0, FFI_TRAMPOLINE_SIZE+2*PTRS
+	addi.d	$a3, $sp, FRAME_LEN
+	move	$a4, $sp
+
+	bl	ffi_closure_inner
+
+	/* Return values.  */
+#if FLTS
+	FLARG	$fa0, $sp, 0*FLTS
+	FLARG	$fa1, $sp, 1*FLTS
+#endif
+
+	LARG	$a0, $sp, 8*FLTS+0*PTRS
+	LARG	$a1, $sp, 8*FLTS+1*PTRS
+
+	/* Restore and return.  */
+	LARG	$ra, $sp, FRAME_LEN-1*PTRS
+	.cfi_restore	1
+	LARG	$fp, $sp, FRAME_LEN-2*PTRS
+	.cfi_restore	22
+	addi.d	$sp, $sp, FRAME_LEN
+	.cfi_def_cfa_offset 0
+	jr	$ra
+	.cfi_endproc
+	.size	ffi_closure_asm, .-ffi_closure_asm
+
+/* ffi_go_closure_asm.  Expects address of the passed-in ffi_go_closure in t2.
+   void ffi_closure_inner (ffi_cif *cif,
+			   void (*fun)(ffi_cif *, void *, void **, void *),
+			   void *user_data,
+			   size_t *stackargs, struct call_context *regargs) */
+
+	.globl	ffi_go_closure_asm
+	.hidden	ffi_go_closure_asm
+	.type	ffi_go_closure_asm, @function
+
+ffi_go_closure_asm:
+	.cfi_startproc
+	addi.d	$sp, $sp, -FRAME_LEN
+	.cfi_def_cfa_offset FRAME_LEN
+
+	/* Make a frame.  */
+	SARG	$fp, $sp, FRAME_LEN - 2*PTRS
+	.cfi_offset	22, -2*PTRS
+	SARG	$ra, $sp, FRAME_LEN - 1*PTRS
+	.cfi_offset	1, -1*PTRS
+	addi.d	$fp, $sp, FRAME_LEN
+
+	/* Save arguments.  */
+#if FLTS
+	FSARG	$fa0, $sp, 0*FLTS
+	FSARG	$fa1, $sp, 1*FLTS
+	FSARG	$fa2, $sp, 2*FLTS
+	FSARG	$fa3, $sp, 3*FLTS
+	FSARG	$fa4, $sp, 4*FLTS
+	FSARG	$fa5, $sp, 5*FLTS
+	FSARG	$fa6, $sp, 6*FLTS
+	FSARG	$fa7, $sp, 7*FLTS
+#endif
+
+	SARG	$a0, $sp, 8*FLTS+0*PTRS
+	SARG	$a1, $sp, 8*FLTS+1*PTRS
+	SARG	$a2, $sp, 8*FLTS+2*PTRS
+	SARG	$a3, $sp, 8*FLTS+3*PTRS
+	SARG	$a4, $sp, 8*FLTS+4*PTRS
+	SARG	$a5, $sp, 8*FLTS+5*PTRS
+	SARG	$a6, $sp, 8*FLTS+6*PTRS
+	SARG	$a7, $sp, 8*FLTS+7*PTRS
+
+	/* Enter C */
+	LARG	$a0, $t2, 1*PTRS
+	LARG	$a1, $t2, 2*PTRS
+	move	$a2, $t2
+	addi.d	$a3, $sp, FRAME_LEN
+	move	$a4, $sp
+
+	bl	ffi_closure_inner
+
+	/* Return values.  */
+#if FLTS
+	FLARG	$fa0, $sp, 0*FLTS
+	FLARG	$fa1, $sp, 1*FLTS
+#endif
+
+	LARG	$a0, $sp, 8*FLTS+0*PTRS
+	LARG	$a1, $sp, 8*FLTS+1*PTRS
+
+	/* Restore and return.  */
+	LARG	$ra, $sp, FRAME_LEN-1*PTRS
+	.cfi_restore	1
+	LARG	$fp, $sp, FRAME_LEN-2*PTRS
+	.cfi_restore	22
+	addi.d	$sp, $sp, FRAME_LEN
+	.cfi_def_cfa_offset 0
+	jr	$ra
+	.cfi_endproc
+	.size	ffi_go_closure_asm, .-ffi_go_closure_asm
+
+#if defined __ELF__ && defined __linux__
+	.section .note.GNU-stack,"",%progbits
+#endif