x86: fix for gcc.c-torture/execute/va-arg-25.c execution at -Os

x86: fix for gcc.c-torture/execute/va-arg-25.c execution at -Os

[Jan Beulich]

[-- Attachment #1: Type: text/plain, Size: 384 bytes --]

(resending with patch only in attachment because it got rejected by
sourceware's spam checker)

However, the newly added test gcc.c-torture/execute/va-arg-26.c once
again
fails (on systems supporting SSE2) and will continue to do so until the
collision
between -Os and the required alignment of SSE/SSE2 vector types has
been
fixed.

boostrapped and tested on i686-pc-linux-gnu.

Jan

[-- Attachment #2: gcc-mainline-x86-vector-va-arg.patch --]
[-- Type: application/octet-stream, Size: 4749 bytes --]

This fixes gcc.c-torture/execute/va-arg-25.c execution at -Os. However, the
newly added test gcc.c-torture/execute/va-arg-26.c once again fails (on
systems supporting SSE2) and will continue to do so until the collision
between -Os and the required alignment of SSE/SSE2 vector types has been
fixed.

2004-07-08 Jan Beulich <jbeulich@novell.com>

	* config/i386/i386.c (init_cumulative_args): Set mmx/sse registers
	available for use only when TARGET_MMX/_SSE is set, and check for
	variable argument function if any register count is non-zero.
	(function_arg): Correctly suppress repeated warnings for passing SSE
	vectors are arguments without SSE enabled.
	(ix86_function_arg_boundary): Without SSE enabled, the use of what
	would be SSE register modes does no longer influence the alignment.

testsuite:
2004-07-08 Jan Beulich <jbeulich@novell.com>

	* gcc.c-torture/execute/va-arg-26.[cx]: New (x86-specific)
	test.

--- /home/jbeulich/src/gcc/mainline/2004-07-05.10.09/gcc/config/i386/i386.c	2004-07-02 15:20:42.000000000 +0200
+++ 2004-07-05.10.09/gcc/config/i386/i386.c	2004-07-07 15:43:51.817899296 +0200
@@ -1917,8 +1917,10 @@
     cum->nregs = ix86_function_regparm (fntype, fndecl);
   else
     cum->nregs = ix86_regparm;
-  cum->sse_nregs = SSE_REGPARM_MAX;
-  cum->mmx_nregs = MMX_REGPARM_MAX;
+  if (TARGET_SSE)
+    cum->sse_nregs = SSE_REGPARM_MAX;
+  if (TARGET_MMX)
+    cum->mmx_nregs = MMX_REGPARM_MAX;
   cum->warn_sse = true;
   cum->warn_mmx = true;
   cum->maybe_vaarg = false;
@@ -1937,9 +1939,9 @@
   /* Determine if this function has variable arguments.  This is
      indicated by the last argument being 'void_type_mode' if there
      are no variable arguments.  If there are variable arguments, then
-     we won't pass anything in registers */
+     we won't pass anything in registers in 32-bit mode. */
 
-  if (cum->nregs || !TARGET_MMX || !TARGET_SSE)
+  if (cum->nregs || cum->mmx_nregs || cum->sse_nregs)
     {
       for (param = (fntype) ? TYPE_ARG_TYPES (fntype) : 0;
 	   param != 0; param = next_param)
@@ -2671,7 +2673,7 @@
       case V2DFmode:
 	if (!type || !AGGREGATE_TYPE_P (type))
 	  {
-	    if (!TARGET_SSE && !warnedmmx && cum->warn_sse)
+	    if (!TARGET_SSE && !warnedsse && cum->warn_sse)
 	      {
 		warnedsse = true;
 		warning ("SSE vector argument without SSE enabled "
@@ -2819,7 +2821,9 @@
 	 The handling here differs from field_alignment.  ICC aligns MMX
 	 arguments to 4 byte boundaries, while structure fields are aligned
 	 to 8 byte boundaries.  */
-      if (!type)
+      if (!TARGET_SSE)
+	align = PARM_BOUNDARY;
+      else if (!type)
 	{
 	  if (!SSE_REG_MODE_P (mode))
 	    align = PARM_BOUNDARY;
--- /home/jbeulich/src/gcc/mainline/2004-07-05.10.09/gcc/testsuite/gcc.c-torture/execute/va-arg-26.c	1970-01-01 01:00:00.000000000 +0100
+++ 2004-07-05.10.09/gcc/testsuite/gcc.c-torture/execute/va-arg-26.c	2004-06-15 09:37:03.000000000 +0200
@@ -0,0 +1,62 @@
+/* Varargs and vectors!  */
+
+#include <stdarg.h>
+
+#define vector __attribute__((vector_size(16)))
+
+const vector unsigned int v1 = {10,11,12,13};
+const vector unsigned int v2 = {20,21,22,23};
+
+void foo(int a, ...)
+{
+  va_list args;
+  vector unsigned int v;
+
+  va_start (args, a);
+  v = va_arg (args, vector unsigned int);
+  if (a != 1 || memcmp (&v, &v1, sizeof (v)) != 0)
+    abort ();
+  a = va_arg (args, int);
+  if (a != 2)
+    abort ();
+  v = va_arg (args, vector unsigned int);
+  if (memcmp (&v, &v2, sizeof (v)) != 0)
+    abort ();
+  va_end (args);
+}
+
+static void bail_if_no_sse (void)
+{
+  int fl1, fl2;
+
+  /* See if we can use cpuid.  */
+  __asm__ ("pushfl; pushfl; popl %0; movl %0,%1; xorl %2,%0;"
+	   "pushl %0; popfl; pushfl; popl %0; popfl"
+	   : "=&r" (fl1), "=&r" (fl2)
+	   : "i" (0x00200000));
+  if (((fl1 ^ fl2) & 0x00200000) == 0)
+    exit (0);
+
+  /* See if cpuid gives capabilities.  */
+  __asm__ ("cpuid" : "=a" (fl1) : "0" (0) : "ebx", "ecx", "edx");
+  if (fl1 == 0)
+    exit (0);
+
+  /* See if capabilities include SSE (25th bit; 26 for SSE2).  */
+  __asm__ ("cpuid" : "=a" (fl1), "=d" (fl2) : "0" (1) : "ebx", "ecx");
+#ifdef __OPTIMIZE_SIZE__
+  if ((fl2 & (1 << 25)) == 0)
+#else
+  if ((fl2 & (1 << 26)) == 0)
+#endif
+    exit (0);
+}
+
+int main(void)
+{
+  bail_if_no_sse ();
+
+  foo (1, (vector unsigned int){10,11,12,13}, 2,
+       (vector unsigned int){20,21,22,23});
+  return 0;
+}
--- /home/jbeulich/src/gcc/mainline/2004-07-05.10.09/gcc/testsuite/gcc.c-torture/execute/va-arg-26.x	1970-01-01 01:00:00.000000000 +0100
+++ 2004-07-05.10.09/gcc/testsuite/gcc.c-torture/execute/va-arg-26.x	2004-06-09 16:33:56.000000000 +0200
@@ -0,0 +1,5 @@
+if { [istarget "i?86-*-*"] } {
+    set additional_flags "-msse2"
+    return 0
+}
+return 1

Re: x86: fix for gcc.c-torture/execute/va-arg-25.c execution at -Os

[Richard Henderson]

On Thu, Jul 08, 2004 at 03:40:48PM +0200, Jan Beulich wrote:
> 	* config/i386/i386.c (init_cumulative_args): Set mmx/sse registers
> 	available for use only when TARGET_MMX/_SSE is set, and check for
> 	variable argument function if any register count is non-zero.
> 	(function_arg): Correctly suppress repeated warnings for passing SSE
> 	vectors are arguments without SSE enabled.
> 	(ix86_function_arg_boundary): Without SSE enabled, the use of what
> 	would be SSE register modes does no longer influence the alignment.

Applied.

> 	* gcc.c-torture/execute/va-arg-26.[cx]: New (x86-specific)
> 	test.

Not applied.  If this is going to fail basically forever,
we need to XFAIL it, or just open a PR for it.


r~