Re: target/6526: [SH4] sdivsi3_i4 can clobber xd0/xd2

[Marcus Comstedt <marcus@mc.pp.se>]

The following reply was made to PR target/6526; it has been noted by GNATS.

From: Marcus Comstedt <marcus@mc.pp.se>
To: Dara Hazeghi <dhazeghi@yahoo.com>
Cc: gcc-gnats@gcc.gnu.org
Subject: Re: target/6526: [SH4] sdivsi3_i4 can clobber xd0/xd2
Date: 10 May 2003 17:53:25 +0200

 Dara Hazeghi <dhazeghi@yahoo.com> writes:
 
 > Hello,
 > 
 > could the submitter please verify if this problem still occurs in a
 > more current version of gcc (ie 3.2.3) ? If so, could the submitter
 > please send a self-contained testcase demonstrating this problem?
 > Thanks,
 
 I don't have a 3.2.3 compiled, but I can verify that it still occurs
 with the 20030421 snapshot of 3.3.
 
 Here is a testcase:
 
 --8<-- testcase.c --8<--
 void write_xf0(float f)
 {
   __asm__ __volatile__("frchg; fmov.s @%0,fr0; frchg" : : "r" (&f));
 }
 
 float read_xf0()
 {
   float f;
   __asm__ __volatile__("frchg; fmov.s fr0,@%0; frchg" : : "r" (&f));
   return f;
 }
 
 int do_stuff(int a, int b)
 {
   int z;
 
   write_xf0(7.0);
 
   z = a % b;
 
   return z+(int)read_xf0();
 }
 
 int testfunc()
 {
   /* Set FPSCR.FR to 1, selecting register bank 1 */
   __set_fpscr(0x200000);
 
   return do_stuff(42, 17);
 }
 --8<--
 
 Compile with `sh-elf-gcc -m4-single-only -c testcase.c', use
 `-m4-single-only' when linking as well.
 
 With the current libgcc implementation, testfunc() will return 10, not
 15 as expected.
 
 Also, you can change the declaration of do_stuff() to make both args
 unsigned int, to see that udivsi3_i4 has the same problem.
 
 
 Btw, here is a patch to gcc/config/sh/lib1funcs.asm which fixes the
 problem for both functions concerned:
 
 ---8<---
 --- lib1funcs.asm.orig	Sat May 10 17:48:43 2003
 +++ lib1funcs.asm	Sat May 10 17:50:28 2003
 @@ -893,18 +893,21 @@
  #endif
  	.global	GLOBAL(sdivsi3_i4)
  GLOBAL(sdivsi3_i4):
 -	sts.l fpscr,@-r15
 -	mov #8,r2
 -	swap.w r2,r2
 -	lds r2,fpscr
 +	mov.l r3,@-r15
 +	sts fpscr,r2
 +	mov #8,r3
 +	swap.w r3,r3
 +	or r2,r3
 +	lds r3,fpscr
  	lds r4,fpul
  	float fpul,dr0
  	lds r5,fpul
  	float fpul,dr2
  	fdiv dr2,dr0
  	ftrc dr0,fpul
 +	lds r2,fpscr
  	rts
 -	lds.l @r15+,fpscr
 +	mov.l @r15+,r3
  
  #endif /* ! __SH5__ || __SH5__ == 32 */
  #endif /* ! __SH4__ */
 @@ -1243,11 +1246,20 @@
  	mov #1,r1
  	cmp/hi r1,r5
  	bf trivial
 -	sts.l fpscr,@-r15
 -	mova L1,r0
 -	lds.l @r0+,fpscr
 +	sts fpscr,r0
  	rotr r1
 +	mov.l r0,@-r15
  	xor r1,r4
 +	xor r1,r5
 +#ifndef FMOVD_WORKS
 +	mov #0x8,r1
 +#else
 +	mov #0x18,r1
 +#endif
 +	swap.w r1,r1
 +	or r0,r1
 +	mova L1,r0
 +	lds r1,fpscr	
  	lds r4,fpul
  #ifdef FMOVD_WORKS
  	fmov.d @r0+,dr4
 @@ -1261,7 +1273,6 @@
  #endif
  #endif
  	float fpul,dr0
 -	xor r1,r5
  	lds r5,fpul
  	float fpul,dr2
  	fadd dr4,dr0
 @@ -1271,20 +1282,16 @@
  	rts
  	lds.l @r15+,fpscr
  
 -#ifdef FMOVD_WORKS
 -	.align 3	! make double below 8 byte aligned.
 -#endif
  trivial:
  	rts
  	lds r4,fpul
  
 -	.align 2
 -L1:
 -#ifndef FMOVD_WORKS
 -	.long 0x80000
 +#ifdef FMOVD_WORKS
 +	.align 3	! make double below 8 byte aligned.
  #else
 -	.long 0x180000
 +	.align 2
  #endif
 +L1:
  	.double 2147483648
  
  #endif /* ! __SH4__ */
 ---8<---
 
 Hope this helps.
 
 
   // Marcus