[Mono-dev] Arm9 NS9215 floating point troubles

[Geoff Norton]

Is your system armeb or armel?

It could be a endian bug in our softfloat impl somewhere.

-g

On 2010-06-17, at 7:54 PM, Trevor Ackerman wrote:

> I have more to report.
> 
> I wrote a quick native C program to print out the bytes of a float and double variable that were assigned the literal value 1.0f. 
> 
> Then in the mono 2.6.4 routine mono_method_to_ir in source code file mono/mini/method_to_ir.c I dumped out the bytes of ip (instruction pointer) used for the double value in the case for CEE_LDC_R8.
> 
> I discovered that the bytes in the double value used on mono had the high and low 32 bits swapped compared to those produced by the native C program. I hacked the routine decompose_soft_float to swap the high and low words and now I have no troubles and the basic-float regression test passes 100%.
> 
> Although this happens to work, I have a hard time believing that this is the correct solution to my problem. I feel that others are probably using ARM9 without floating point issues and that I am probably missing something in how I built mono for my platform. If anyone can shed some light on what I did wrong with building mono that'd be great. Of course if this is the correct action to take please let me know that too and how I may contribute the change back to the trunk (assuming that the trunk doesn't work which I haven't had time to test yet).
> 
> In the meantime here's my hack to decompose_soft_float in method-to-ir.c
> 
> 
>  5073                 case OP_R8CONST: {
>  5074             unsigned char *ucp = (unsigned char *) ins->inst_p0;
>  5075             unsigned char rawval[8];
>  5076                     printf("decompose_soft_float OP_R8CONST\n");
>  5077             rawval[0] = ucp[4];
>  5078             rawval[1] = ucp[5];
>  5079             rawval[2] = ucp[6];
>  5080             rawval[3] = ucp[7];
>  5081             rawval[4] = ucp[0];
>  5082             rawval[5] = ucp[1];
>  5083             rawval[6] = ucp[2];
>  5084             rawval[7] = ucp[3];
>  5085                     DVal d;
>  5086 //                  d.vald = *(double*)ins->inst_p0;
>  5087                     d.vald = *(double*)rawval;
>  5088                     MONO_EMIT_NEW_I8CONST (cfg, ins->dreg, d.vall);
>  5089                     break;
>  5090                 }
> 
> --- On Thu, 6/17/10, Trevor Ackerman <t_ackerman at yahoo.com> wrote:
> 
> From: Trevor Ackerman <t_ackerman at yahoo.com>
> Subject: Re: [Mono-dev] Arm9 NS9215 floating point troubles
> To: mono-devel-list at lists.ximian.com
> Date: Thursday, June 17, 2010, 11:31 AM
> 
> Good suggestion but that did not change the results.
> 
> --- On Thu, 6/17/10, Robert Jordan <robertj at gmx.net> wrote:
> 
> From: Robert Jordan <robertj at gmx.net>
> Subject: Re: [Mono-dev] Arm9 NS9215 floating point troubles
> To: mono-devel-list at lists.ximian.com
> Date: Thursday, June 17, 2010, 10:53 AM
> 
> On 17.06.2010 18:07, Trevor Ackerman wrote:
> > I have been able to cross-compile Mono 2.6.4 for the NS9215 (no fpu
> > afaik) and I'm having trouble with floats when executing code.
> 
> ...
> 
> > My CFLAGS and CPPFLAGS environment variables are
> >   both
> > -DARM_FPU_NONE=1 -DMONO_ARCH_SOFT_FLOAT=1
> 
> -DNO_UNALIGNED_ACCESS is probably needed as well.
> 
> Robert
> 
> _______________________________________________
> Mono-devel-list mailing list
> Mono-devel-list at lists.ximian.com
> http://lists.ximian.com/mailman/listinfo/mono-devel-list
> 
> 
> -----Inline Attachment Follows-----
> 
> _______________________________________________
> Mono-devel-list mailing list
> Mono-devel-list at lists.ximian.com
> http://lists.ximian.com/mailman/listinfo/mono-devel-list
> 
> _______________________________________________
> Mono-devel-list mailing list
> Mono-devel-list at lists.ximian.com
> http://lists.ximian.com/mailman/listinfo/mono-devel-list

-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://lists.ximian.com/pipermail/mono-devel-list/attachments/20100617/1d55fd7c/attachment.html