Merge pull request #34 from davidsch/armhf

Fix ARM hard-float support for large numbers of VFP arguments

ChangeLog

@@ -1,3 +1,14 @@
+2013-03-28  David Schneider  <david.schneider@bivab.de>
+
+	 * src/arm/ffi.c: Fix support for ARM hard-float calling convention.
+	 * src/arm/sysv.S: call different methods for SYSV and VFP ABIs.
+	 * testsuite/libffi.call/cls_many_mixed_args.c: testcase for a closure with
+	 mixed arguments, many doubles.
+	 * testsuite/libffi.call/many_double.c: testcase for calling a function using
+	 more than 8 doubles.
+	 * testcase/libffi.call/many.c: use absolute value to check result against an
+	 epsilon
+
 2013-03-17  Anthony Green  <green@moxielogic.com>
 
 	* README: Update for 3.0.13.

src/arm/ffi.c

@@ -37,6 +37,84 @@
 static int vfp_type_p (ffi_type *);
 static void layout_vfp_args (ffi_cif *);
 
+static char* ffi_align(ffi_type **p_arg, char *argp)
+{
+  /* Align if necessary */
+  register size_t alignment = (*p_arg)->alignment;
+  if (alignment < 4)
+  {
+    alignment = 4;
+  }
+#ifdef _WIN32_WCE
+  if (alignment > 4)
+  {
+    alignment = 4;
+  }
+#endif
+  if ((alignment - 1) & (unsigned) argp)
+  {
+    argp = (char *) ALIGN(argp, alignment);
+  }
+
+  if ((*p_arg)->type == FFI_TYPE_STRUCT)
+  {
+    argp = (char *) ALIGN(argp, 4);
+  }
+  return argp;
+}
+
+static size_t ffi_put_arg(ffi_type **arg_type, void **arg, char *stack)
+{
+	register char* argp = stack;
+	register ffi_type **p_arg = arg_type;
+	register void **p_argv = arg;
+	register size_t z = (*p_arg)->size;
+  if (z < sizeof(int))
+    {
+		z = sizeof(int);
+		switch ((*p_arg)->type)
+      {
+      case FFI_TYPE_SINT8:
+        *(signed int *) argp = (signed int)*(SINT8 *)(* p_argv);
+        break;
+        
+      case FFI_TYPE_UINT8:
+        *(unsigned int *) argp = (unsigned int)*(UINT8 *)(* p_argv);
+        break;
+        
+      case FFI_TYPE_SINT16:
+        *(signed int *) argp = (signed int)*(SINT16 *)(* p_argv);
+        break;
+        
+      case FFI_TYPE_UINT16:
+        *(unsigned int *) argp = (unsigned int)*(UINT16 *)(* p_argv);
+        break;
+        
+      case FFI_TYPE_STRUCT:
+        memcpy(argp, *p_argv, (*p_arg)->size);
+        break;
+
+      default:
+        FFI_ASSERT(0);
+      }
+    }
+  else if (z == sizeof(int))
+    {
+		if ((*p_arg)->type == FFI_TYPE_FLOAT)
+			*(float *) argp = *(float *)(* p_argv);
+		else
+			*(unsigned int *) argp = (unsigned int)*(UINT32 *)(* p_argv);
+    }
+	else if (z == sizeof(double) && (*p_arg)->type == FFI_TYPE_DOUBLE)
+		{
+			*(double *) argp = *(double *)(* p_argv);
+		}
+  else
+    {
+      memcpy(argp, *p_argv, z);
+    }
+  return z;
+}
 /* ffi_prep_args is called by the assembly routine once stack space
    has been allocated for the function's arguments
    
@@ -44,15 +122,15 @@ static void layout_vfp_args (ffi_cif *);
    value is cif->vfp_used (word bitset of VFP regs used for passing
    arguments). These are only used for the VFP hard-float ABI.
 */
-int ffi_prep_args(char *stack, extended_cif *ecif, float *vfp_space)
+int ffi_prep_args_SYSV(char *stack, extended_cif *ecif, float *vfp_space)
 {
-  register unsigned int i, vi = 0;
+  register unsigned int i;
   register void **p_argv;
   register char *argp;
   register ffi_type **p_arg;
-
   argp = stack;
   
+
   if ( ecif->cif->flags == FFI_TYPE_STRUCT ) {
     *(void **) argp = ecif->rvalue;
     argp += 4;
@@ -62,81 +140,89 @@ int ffi_prep_args(char *stack, extended_cif *ecif, float *vfp_space)
 
   for (i = ecif->cif->nargs, p_arg = ecif->cif->arg_types;
        (i != 0);
-       i--, p_arg++)
+       i--, p_arg++, p_argv++)
     {
-      size_t z;
+    argp = ffi_align(p_arg, argp);
-      size_t alignment;
+    argp += ffi_put_arg(p_arg, p_argv, argp);
-
-      /* Allocated in VFP registers. */
-      if (ecif->cif->abi == FFI_VFP
-	  && vi < ecif->cif->vfp_nargs && vfp_type_p (*p_arg))
-	{
-	  float* vfp_slot = vfp_space + ecif->cif->vfp_args[vi++];
-	  if ((*p_arg)->type == FFI_TYPE_FLOAT)
-	    *((float*)vfp_slot) = *((float*)*p_argv);
-	  else if ((*p_arg)->type == FFI_TYPE_DOUBLE)
-	    *((double*)vfp_slot) = *((double*)*p_argv);
-	  else
-	    memcpy(vfp_slot, *p_argv, (*p_arg)->size);
-	  p_argv++;
-	  continue;
-	}
-
-      /* Align if necessary */
-      alignment = (*p_arg)->alignment;
-#ifdef _WIN32_WCE
-      if (alignment > 4)
-	alignment = 4;
-#endif
-      if ((alignment - 1) & (unsigned) argp) {
-	argp = (char *) ALIGN(argp, alignment);
-      }
-
-      if ((*p_arg)->type == FFI_TYPE_STRUCT)
-	argp = (char *) ALIGN(argp, 4);
-
-	  z = (*p_arg)->size;
-	  if (z < sizeof(int))
-	    {
-	      z = sizeof(int);
-	      switch ((*p_arg)->type)
-		{
-		case FFI_TYPE_SINT8:
-		  *(signed int *) argp = (signed int)*(SINT8 *)(* p_argv);
-		  break;
-		  
-		case FFI_TYPE_UINT8:
-		  *(unsigned int *) argp = (unsigned int)*(UINT8 *)(* p_argv);
-		  break;
-		  
-		case FFI_TYPE_SINT16:
-		  *(signed int *) argp = (signed int)*(SINT16 *)(* p_argv);
-		  break;
-		  
-		case FFI_TYPE_UINT16:
-		  *(unsigned int *) argp = (unsigned int)*(UINT16 *)(* p_argv);
-		  break;
-		  
-		case FFI_TYPE_STRUCT:
-		  memcpy(argp, *p_argv, (*p_arg)->size);
-		  break;
-
-		default:
-		  FFI_ASSERT(0);
-		}
-	    }
-	  else if (z == sizeof(int))
-	    {
-	      *(unsigned int *) argp = (unsigned int)*(UINT32 *)(* p_argv);
-	    }
-	  else
-	    {
-	      memcpy(argp, *p_argv, z);
-	    }
-	  p_argv++;
-	  argp += z;
     }
 
+  return 0;
+}
+
+int ffi_prep_args_VFP(char *stack, extended_cif *ecif, float *vfp_space)
+{
+  // make sure we are using FFI_VFP
+  FFI_ASSERT(ecif->cif->abi == FFI_VFP);
+
+  register unsigned int i, vi = 0;
+  register void **p_argv;
+  register char *argp, *regp, *eo_regp;
+  register ffi_type **p_arg;
+  char stack_used = 0;
+  char done_with_regs = 0;
+  char is_vfp_type;
+
+  /* the first 4 words on the stack are used for values passed in core
+   * registers. */
+  regp = stack;
+  eo_regp = argp = regp + 16;
+  
+
+  /* if the function returns an FFI_TYPE_STRUCT in memory, that address is
+   * passed in r0 to the function */
+  if ( ecif->cif->flags == FFI_TYPE_STRUCT ) {
+    *(void **) regp = ecif->rvalue;
+    regp += 4;
+  }
+
+  p_argv = ecif->avalue;
+
+  for (i = ecif->cif->nargs, p_arg = ecif->cif->arg_types;
+       (i != 0);
+       i--, p_arg++, p_argv++)
+    {
+      is_vfp_type = vfp_type_p (*p_arg);
+
+      /* Allocated in VFP registers. */
+      if(vi < ecif->cif->vfp_nargs && is_vfp_type)
+        {
+          char *vfp_slot = (char *)(vfp_space + ecif->cif->vfp_args[vi++]);
+          ffi_put_arg(p_arg, p_argv, vfp_slot);
+          continue;
+        }
+      /* Try allocating in core registers. */
+      else if (!done_with_regs && !is_vfp_type)
+        {
+          char *tregp = ffi_align(p_arg, regp);
+          size_t size = (*p_arg)->size; 
+          size = (size < 4)? 4 : size; // pad
+          /* Check if there is space left in the aligned register area to place
+           * the argument */
+          if(tregp + size <= eo_regp)
+            {
+              regp = tregp + ffi_put_arg(p_arg, p_argv, tregp);
+              done_with_regs = (regp == argp);
+              // ensure we did not write into the stack area
+              FFI_ASSERT(regp <= argp);
+              continue;
+            }
+          /* In case there are no arguments in the stack area yet, 
+          the argument is passed in the remaining core registers and on the
+          stack. */
+          else if (!stack_used) 
+            {
+              stack_used = 1;
+              done_with_regs = 1;
+              argp = tregp + ffi_put_arg(p_arg, p_argv, tregp);
+              FFI_ASSERT(eo_regp < argp);
+              continue;
+            }
+        }
+      /* Base case, arguments are passed on the stack */
+      stack_used = 1;
+      argp = ffi_align(p_arg, argp);
+      argp += ffi_put_arg(p_arg, p_argv, argp);
+    }
   /* Indicate the VFP registers used. */
   return ecif->cif->vfp_used;
 }
@@ -271,6 +357,9 @@ void ffi_call(ffi_cif *cif, void (*fn)(void), void *rvalue, void **avalue)
 static void ffi_prep_incoming_args_SYSV (char *stack, void **ret,
 					 void** args, ffi_cif* cif, float *vfp_stack);
 
+static void ffi_prep_incoming_args_VFP (char *stack, void **ret,
+					 void** args, ffi_cif* cif, float *vfp_stack);
+
 void ffi_closure_SYSV (ffi_closure *);
 
 void ffi_closure_VFP (ffi_closure *);
@@ -278,7 +367,7 @@ void ffi_closure_VFP (ffi_closure *);
 /* This function is jumped to by the trampoline */
 
 unsigned int
-ffi_closure_SYSV_inner (closure, respp, args, vfp_args)
+ffi_closure_inner (closure, respp, args, vfp_args)
      ffi_closure *closure;
      void **respp;
      void *args;
@@ -296,8 +385,10 @@ ffi_closure_SYSV_inner (closure, respp, args, vfp_args)
    * value on the stack; and if the function returns
    * a structure, it will re-set RESP to point to the
    * structure return address.  */
-
+  if (cif->abi == FFI_VFP)
-  ffi_prep_incoming_args_SYSV(args, respp, arg_area, cif, vfp_args);
+    ffi_prep_incoming_args_VFP(args, respp, arg_area, cif, vfp_args);
+  else
+    ffi_prep_incoming_args_SYSV(args, respp, arg_area, cif, vfp_args);
 
   (closure->fun) (cif, *respp, arg_area, closure->user_data);
 
@@ -312,7 +403,7 @@ ffi_prep_incoming_args_SYSV(char *stack, void **rvalue,
 			    float *vfp_stack)
 /*@=exportheader@*/
 {
-  register unsigned int i, vi = 0;
+  register unsigned int i;
   register void **p_argv;
   register char *argp;
   register ffi_type **p_arg;
@@ -329,27 +420,8 @@ ffi_prep_incoming_args_SYSV(char *stack, void **rvalue,
   for (i = cif->nargs, p_arg = cif->arg_types; (i != 0); i--, p_arg++)
     {
       size_t z;
-      size_t alignment;
 
-      if (cif->abi == FFI_VFP
+      argp = ffi_align(p_arg, argp);
-	  && vi < cif->vfp_nargs && vfp_type_p (*p_arg))
-	{
-	  *p_argv++ = (void*)(vfp_stack + cif->vfp_args[vi++]);
-	  continue;
-	}
-
-      alignment = (*p_arg)->alignment;
-      if (alignment < 4)
-	alignment = 4;
-#ifdef _WIN32_WCE
-      else
-	if (alignment > 4)
-	  alignment = 4;
-#endif
-      /* Align if necessary */
-      if ((alignment - 1) & (unsigned) argp) {
-	argp = (char *) ALIGN(argp, alignment);
-      }
 
       z = (*p_arg)->size;
 
@@ -364,6 +436,90 @@ ffi_prep_incoming_args_SYSV(char *stack, void **rvalue,
   return;
 }
 
+/*@-exportheader@*/
+static void 
+ffi_prep_incoming_args_VFP(char *stack, void **rvalue,
+			    void **avalue, ffi_cif *cif,
+			    /* Used only under VFP hard-float ABI. */
+			    float *vfp_stack)
+/*@=exportheader@*/
+{
+  register unsigned int i, vi = 0;
+  register void **p_argv;
+  register char *argp, *regp, *eo_regp;
+  register ffi_type **p_arg;
+  char done_with_regs = 0;
+  char stack_used = 0;
+  char is_vfp_type;
+
+  FFI_ASSERT(cif->abi == FFI_VFP);
+  regp = stack;
+  eo_regp = argp = regp + 16;
+
+  if ( cif->flags == FFI_TYPE_STRUCT ) {
+    *rvalue = *(void **) regp;
+    regp += 4;
+  }
+
+  p_argv = avalue;
+
+  for (i = cif->nargs, p_arg = cif->arg_types; (i != 0); i--, p_arg++)
+    {
+    size_t z;
+    is_vfp_type = vfp_type_p (*p_arg); 
+
+    if(vi < cif->vfp_nargs && is_vfp_type)
+      {
+        *p_argv++ = (void*)(vfp_stack + cif->vfp_args[vi++]);
+        continue;
+      }
+    else if (!done_with_regs && !is_vfp_type)
+      {
+        char* tregp = ffi_align(p_arg, regp);
+
+        z = (*p_arg)->size; 
+        z = (z < 4)? 4 : z; // pad
+        
+        /* if the arguments either fits into the registers or uses registers
+         * and stack, while we haven't read other things from the stack */
+        if(tregp + z <= eo_regp || !stack_used) 
+          {
+          /* because we're little endian, this is what it turns into. */
+          *p_argv = (void*) tregp;
+
+          p_argv++;
+          regp = tregp + z;
+          /* if regp points above the end of the register area */
+          if(regp >= eo_regp)
+            {
+              /* sanity check that we haven't read from the stack area before
+               * reaching this point */
+              FFI_ASSERT(argp <= regp);
+              FFI_ASSERT(argp == stack + 16);
+              argp = regp;
+              done_with_regs = 1;
+              stack_used = 1;
+            }
+            continue;
+          }
+      }
+    stack_used = 1;
+
+    argp = ffi_align(p_arg, argp);
+
+    z = (*p_arg)->size;
+
+    /* because we're little endian, this is what it turns into.   */
+
+    *p_argv = (void*) argp;
+
+    p_argv++;
+    argp += z;
+    }
+  
+  return;
+}
+
 /* How to make a trampoline.  */
 
 extern unsigned int ffi_arm_trampoline[3];

src/arm/sysv.S

@@ -187,7 +187,7 @@ ARM_FUNC_START ffi_call_SYSV
 	@     r1 already set
 
 	@ Call ffi_prep_args(stack, &ecif)
-	bl	CNAME(ffi_prep_args)
+	bl	CNAME(ffi_prep_args_SYSV)
 
 	@ move first 4 parameters in registers
 	ldmia	sp, {r0-r3}
@@ -260,7 +260,7 @@ LSYM(Lepilogue):
 
 /*
 	unsigned int FFI_HIDDEN
-	ffi_closure_SYSV_inner (closure, respp, args)
+	ffi_closure_inner (closure, respp, args)
 	     ffi_closure *closure;
 	     void **respp;
   	     void *args;
@@ -276,7 +276,7 @@ ARM_FUNC_START ffi_closure_SYSV
 	sub	sp, sp, #16
 	str	sp, [sp, #8]
 	add	r1, sp, #8
-	bl	CNAME(ffi_closure_SYSV_inner)
+	bl	CNAME(ffi_closure_inner)
 	cmp	r0, #FFI_TYPE_INT
 	beq	.Lretint
 
@@ -364,7 +364,7 @@ ARM_FUNC_START ffi_call_VFP
 	sub	r2, fp, #64   @ VFP scratch space
 
 	@ Call ffi_prep_args(stack, &ecif, vfp_space)
-	bl	CNAME(ffi_prep_args)
+	bl	CNAME(ffi_prep_args_VFP)
 
 	@ Load VFP register args if needed
 	cmp	r0, #0
@@ -446,7 +446,7 @@ ARM_FUNC_START ffi_closure_VFP
 	sub	sp, sp, #72
 	str	sp, [sp, #64]
 	add	r1, sp, #64
-	bl	CNAME(ffi_closure_SYSV_inner)
+	bl	CNAME(ffi_closure_inner)
 
 	cmp	r0, #FFI_TYPE_INT
 	beq	.Lretint_vfp

testsuite/libffi.call/cls_many_mixed_args.c

@@ -0,0 +1,68 @@
+/* Area:	closure_call
+   Purpose:	Check closures called with many args of mixed types
+   Limitations:	none.
+   PR:		none.
+   Originator:	<david.schneider@picle.org> */
+
+/* { dg-do run } */
+#include "ffitest.h"
+#include <float.h>
+#include <math.h>
+
+#define NARGS 12
+
+static void cls_ret_double_fn(ffi_cif* cif __UNUSED__, void* resp, void** args,
+			      void* userdata __UNUSED__)
+{
+  int i;
+  double r;
+  double t;
+  for(i = 0; i < NARGS; i++)
+    {
+    if(i == 8) 
+      {
+      t = *(long int *)args[i];
+      CHECK(t == i+1);
+      }
+    else
+      {
+      t = *(double *)args[i];
+      CHECK(fabs(t - ((i+1) * 0.1)) < FLT_EPSILON);
+      }
+    r += t;
+    }
+  *(double *)resp = r;
+}
+typedef double (*cls_ret_double)(double, double, double, double, double, double, double, double, long int, double, double, double);
+
+int main (void)
+{
+  ffi_cif cif;
+  void *code;
+  ffi_closure *pcl = ffi_closure_alloc(sizeof(ffi_closure), &code);
+  ffi_type * cl_arg_types[NARGS];
+  double res;
+  int i;
+  double expected = 15.9;
+
+  for(i = 0; i < NARGS; i++)
+    {
+    if(i == 8) 
+      cl_arg_types[i] = &ffi_type_slong;
+    else
+      cl_arg_types[i] = &ffi_type_double;
+    }
+
+  /* Initialize the cif */
+  CHECK(ffi_prep_cif(&cif, FFI_DEFAULT_ABI, NARGS,
+		     &ffi_type_double, cl_arg_types) == FFI_OK);
+
+  CHECK(ffi_prep_closure_loc(pcl, &cif, cls_ret_double_fn, NULL, code) == FFI_OK);
+
+  res = (((cls_ret_double)code))(0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 9,
+                                                                1.0, 1.1, 1.2);
+  if (abs(res - expected) < FLT_EPSILON)
+    exit(0);
+  else
+    abort();
+}

testsuite/libffi.call/many.c

@@ -7,7 +7,9 @@
 /* { dg-do run } */
 #include "ffitest.h"
 
+#include <stdlib.h>
 #include <float.h>
+#include <math.h>
 
 static float many(float f1,
 		  float f2,
@@ -62,7 +64,7 @@ int main (void)
 	       fa[8], fa[9],
 	       fa[10],fa[11],fa[12]);
 
-    if (f - ff < FLT_EPSILON)
+    if (fabs(f - ff) < FLT_EPSILON)
       exit(0);
     else
       abort();

testsuite/libffi.call/many_double.c

@@ -0,0 +1,70 @@
+/* Area:	ffi_call
+   Purpose:	Check return value double, with many arguments
+   Limitations:	none.
+   PR:		none.
+   Originator:	From the original ffitest.c  */
+
+/* { dg-do run } */
+#include "ffitest.h"
+
+#include <stdlib.h>
+#include <float.h>
+#include <math.h>
+
+static double many(double f1,
+		  double f2,
+		  double f3,
+		  double f4,
+		  double f5,
+		  double f6,
+		  double f7,
+		  double f8,
+		  double f9,
+		  double f10,
+		  double f11,
+		  double f12,
+		  double f13)
+{
+#if 0
+  printf("%f %f %f %f %f %f %f %f %f %f %f %f %f\n",
+	 (double) f1, (double) f2, (double) f3, (double) f4, (double) f5, 
+	 (double) f6, (double) f7, (double) f8, (double) f9, (double) f10,
+	 (double) f11, (double) f12, (double) f13);
+#endif
+
+  return ((f1/f2+f3/f4+f5/f6+f7/f8+f9/f10+f11/f12) * f13);
+}
+
+int main (void)
+{
+  ffi_cif cif;
+  ffi_type *args[13];
+  void *values[13];
+  double fa[13];
+  double f, ff;
+  int i;
+
+  for (i = 0; i < 13; i++)
+    {
+      args[i] = &ffi_type_double;
+      values[i] = &fa[i];
+      fa[i] = (double) i;
+    }
+
+    /* Initialize the cif */
+    CHECK(ffi_prep_cif(&cif, FFI_DEFAULT_ABI, 13, 
+		       &ffi_type_double, args) == FFI_OK);
+
+    ffi_call(&cif, FFI_FN(many), &f, values);
+
+    ff =  many(fa[0], fa[1],
+	       fa[2], fa[3],
+	       fa[4], fa[5],
+	       fa[6], fa[7],
+	       fa[8], fa[9],
+	       fa[10],fa[11],fa[12]);
+    if (fabs(f - ff) < FLT_EPSILON)
+      exit(0);
+    else
+      abort();
+}