[PULL,1/1] target/openrisc: Set EPCR to next PC on FPE exceptions

The architecture specification calls for the EPCR to be set to "Address
of next not executed instruction" when there is a floating point
exception (FPE).  This was not being done, so fix it by using the same
pattern as syscall.  Also, we move this logic down to be done for
instructions not in the delay slot as called for by the architecture
manual.

Without this patch FPU exceptions will loop, as the exception handling
will always return back to the failed floating point instruction.

This was not noticed in earlier testing because:

 1. The compiler usually generates code which clobbers the input operand
    such as:

      lf.div.s r19,r17,r19

 2. The target will store the operation output before to the register
    before handling the exception.  So an operation such as:

      float a = 100.0f;
      float b = 0.0f;
      float c = a / b;    /* lf.div.s r19,r17,r19 */

    Will first execute:

      100 / 0    -> Store inf to c (r19)
                 -> triggering divide by zero exception
                 -> handle and return

    Then it will execute:

      100 / inf  -> Store 0 to c  (no exception)

To confirm the looping behavior and the fix I used the following:

    float fpu_div(float a, float b) {
	float c;
	asm volatile("lf.div.s %0, %1, %2"
		      : "+r" (c)
		      : "r" (a), "r" (b));
	return c;
    }

Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Signed-off-by: Stafford Horne <shorne@gmail.com>
---
 target/openrisc/interrupt.c | 7 ++++---
 1 file changed, 4 insertions(+), 3 deletions(-)

09.08.2023 23:34, Stafford Horne пишет:
> The architecture specification calls for the EPCR to be set to "Address
> of next not executed instruction" when there is a floating point
> exception (FPE).  This was not being done, so fix it by using the same
> pattern as syscall.  Also, we move this logic down to be done for
> instructions not in the delay slot as called for by the architecture
> manual.
> 
> Without this patch FPU exceptions will loop, as the exception handling
> will always return back to the failed floating point instruction.
> 
> This was not noticed in earlier testing because:
> 
>   1. The compiler usually generates code which clobbers the input operand
>      such as:
> 
>        lf.div.s r19,r17,r19
> 
>   2. The target will store the operation output before to the register
>      before handling the exception.  So an operation such as:
> 
>        float a = 100.0f;
>        float b = 0.0f;
>        float c = a / b;    /* lf.div.s r19,r17,r19 */
> 
>      Will first execute:
> 
>        100 / 0    -> Store inf to c (r19)
>                   -> triggering divide by zero exception
>                   -> handle and return
> 
>      Then it will execute:
> 
>        100 / inf  -> Store 0 to c  (no exception)
> 
> To confirm the looping behavior and the fix I used the following:
> 
>      float fpu_div(float a, float b) {
> 	float c;
> 	asm volatile("lf.div.s %0, %1, %2"
> 		      : "+r" (c)
> 		      : "r" (a), "r" (b));
> 	return c;
>      }

Is it a -stable material?  It applies cleanly to 8.0 and 7.2.
Or maybe it is not needed on older versions, not being noticed before?

/mjt

On Thu, Aug 10, 2023 at 09:35:18AM +0300, Michael Tokarev wrote:
> 09.08.2023 23:34, Stafford Horne пишет:
> > The architecture specification calls for the EPCR to be set to "Address
> > of next not executed instruction" when there is a floating point
> > exception (FPE).  This was not being done, so fix it by using the same
> > pattern as syscall.  Also, we move this logic down to be done for
> > instructions not in the delay slot as called for by the architecture
> > manual.
> > 
> > Without this patch FPU exceptions will loop, as the exception handling
> > will always return back to the failed floating point instruction.
> > 
> > This was not noticed in earlier testing because:
> > 
> >   1. The compiler usually generates code which clobbers the input operand
> >      such as:
> > 
> >        lf.div.s r19,r17,r19
> > 
> >   2. The target will store the operation output before to the register
> >      before handling the exception.  So an operation such as:
> > 
> >        float a = 100.0f;
> >        float b = 0.0f;
> >        float c = a / b;    /* lf.div.s r19,r17,r19 */
> > 
> >      Will first execute:
> > 
> >        100 / 0    -> Store inf to c (r19)
> >                   -> triggering divide by zero exception
> >                   -> handle and return
> > 
> >      Then it will execute:
> > 
> >        100 / inf  -> Store 0 to c  (no exception)
> > 
> > To confirm the looping behavior and the fix I used the following:
> > 
> >      float fpu_div(float a, float b) {
> > 	float c;
> > 	asm volatile("lf.div.s %0, %1, %2"
> > 		      : "+r" (c)
> > 		      : "r" (a), "r" (b));
> > 	return c;
> >      }
> 
> Is it a -stable material?  It applies cleanly to 8.0 and 7.2.
> Or maybe it is not needed on older versions, not being noticed before?

I would say no, it will work on 8.0 an 7.2 but this code path is not very useful
withouth the other 8.1 Floating Point Exception handling updates.

-Stafford

10.08.2023 22:50, Stafford Horne wrote:
> On Thu, Aug 10, 2023 at 09:35:18AM +0300, Michael Tokarev wrote:
..
>> Is it a -stable material?  It applies cleanly to 8.0 and 7.2.
>> Or maybe it is not needed on older versions, not being noticed before?
> 
> I would say no, it will work on 8.0 an 7.2 but this code path is not very useful
> withouth the other 8.1 Floating Point Exception handling updates.

Thank you for letting me know. This makes good sense, and shares my expectations
too.  This particular situation is rather interesting, that's why I asked.

/mjt