[v2,6/6] sim/common: Fix issue with wrong byte order on BE targets
Commit Message
* Stafford Horne <shorne@gmail.com> [2019-04-10 06:39:25 +0900]:
> Currently only the OpenRISC sim uses this JOINSIDF() function to compose a
> double float from 2 registers. The old code doesn't seem to work as the
> work order gets swapped when running on a x86_64 host. This change
> fixes that, but I am not sure if its the best thing to do.
>
> On mips they do similar reg pair floating point operations composing
> doubles from 2 32-bit registers in sim/mips/cp1.c value_fpr().
>
> sim/common/ChangeLog:
>
> * cgen-ops.h (JOINSIDF): Fix big endian check.
> ---
> sim/common/cgen-ops.h | 8 ++------
> 1 file changed, 2 insertions(+), 6 deletions(-)
>
> diff --git a/sim/common/cgen-ops.h b/sim/common/cgen-ops.h
> index 841552066f..d718394723 100644
> --- a/sim/common/cgen-ops.h
> +++ b/sim/common/cgen-ops.h
> @@ -431,12 +431,8 @@ JOINSIDI (SI x0, SI x1)
> SEMOPS_INLINE DF
> JOINSIDF (SI x0, SI x1)
> {
> - union { SI in[2]; DF out; } x;
> - if (CURRENT_TARGET_BYTE_ORDER == BFD_ENDIAN_BIG)
I think this is the problem with the existing code, we're using memory
on the HOST to perform packing / unpacking, and so its the byte
ordering of the HOST that we care about here, not the target.
If I change the above line to instead be:
if (HOST_BYTE_ORDER == BFD_ENDIAN_BIG)
then everything works fine.
On inspection I believe all of the uses of CURRENT_TARGET_BYTE_ORDER
in this file should be similarly replaced.
I've attached a patch for this change. If you agree I'll go ahead and
push it.
Thanks,
Andrew
--
[PATCH] sim: Use host not target byte order for merging and splitting values
When using writes to memory through a struct to merge and extract
multi-word value, it is the endianness of the host, not the target
that affects which order the component words need to be written into
the structure.
Of the 5 functions adjusted here 4 of them are unused. The 5th,
JOINSIDF will soon be used by the or1k target.
For or1k, simulated on x86-64, this change fixes this function so that
the correct bytes are now returned.
sim/common/ChangeLog:
* cgen-ops.h (SUBWORDXFSI): Compare HOST_BYTE_ORDER not
CURRENT_TARGET_BYTE_ORDER.
(SUBWORDTFSI): Likewise.
(JOINSIDF): Likewise.
(JOINSIXF): Likewise.
(JOINSITF): Likewise.
---
sim/common/ChangeLog | 9 +++++++++
sim/common/cgen-ops.h | 10 +++++-----
2 files changed, 14 insertions(+), 5 deletions(-)
Comments
Hi Andrew,
On Thu, Apr 11, 2019 at 11:27:49PM +0100, Andrew Burgess wrote:
> * Stafford Horne <shorne@gmail.com> [2019-04-10 06:39:25 +0900]:
>
> > Currently only the OpenRISC sim uses this JOINSIDF() function to compose a
> > double float from 2 registers. The old code doesn't seem to work as the
> > work order gets swapped when running on a x86_64 host. This change
> > fixes that, but I am not sure if its the best thing to do.
> >
> > On mips they do similar reg pair floating point operations composing
> > doubles from 2 32-bit registers in sim/mips/cp1.c value_fpr().
> >
> > sim/common/ChangeLog:
> >
> > * cgen-ops.h (JOINSIDF): Fix big endian check.
> > ---
> > sim/common/cgen-ops.h | 8 ++------
> > 1 file changed, 2 insertions(+), 6 deletions(-)
> >
> > diff --git a/sim/common/cgen-ops.h b/sim/common/cgen-ops.h
> > index 841552066f..d718394723 100644
> > --- a/sim/common/cgen-ops.h
> > +++ b/sim/common/cgen-ops.h
> > @@ -431,12 +431,8 @@ JOINSIDI (SI x0, SI x1)
> > SEMOPS_INLINE DF
> > JOINSIDF (SI x0, SI x1)
> > {
> > - union { SI in[2]; DF out; } x;
> > - if (CURRENT_TARGET_BYTE_ORDER == BFD_ENDIAN_BIG)
>
> I think this is the problem with the existing code, we're using memory
> on the HOST to perform packing / unpacking, and so its the byte
> ordering of the HOST that we care about here, not the target.
>
> If I change the above line to instead be:
>
> if (HOST_BYTE_ORDER == BFD_ENDIAN_BIG)
>
> then everything works fine.
>
> On inspection I believe all of the uses of CURRENT_TARGET_BYTE_ORDER
> in this file should be similarly replaced.
>
> I've attached a patch for this change. If you agree I'll go ahead and
> push it.
I tested this with the or1k testsuite and it does seem to work fine, thank you.
Please go ahead.
I thought I had tried this fix as well, I guess not.
-Stafford
* Stafford Horne <shorne@gmail.com> [2019-04-13 05:21:45 +0900]:
> Hi Andrew,
>
> On Thu, Apr 11, 2019 at 11:27:49PM +0100, Andrew Burgess wrote:
> > * Stafford Horne <shorne@gmail.com> [2019-04-10 06:39:25 +0900]:
> >
> > > Currently only the OpenRISC sim uses this JOINSIDF() function to compose a
> > > double float from 2 registers. The old code doesn't seem to work as the
> > > work order gets swapped when running on a x86_64 host. This change
> > > fixes that, but I am not sure if its the best thing to do.
> > >
> > > On mips they do similar reg pair floating point operations composing
> > > doubles from 2 32-bit registers in sim/mips/cp1.c value_fpr().
> > >
> > > sim/common/ChangeLog:
> > >
> > > * cgen-ops.h (JOINSIDF): Fix big endian check.
> > > ---
> > > sim/common/cgen-ops.h | 8 ++------
> > > 1 file changed, 2 insertions(+), 6 deletions(-)
> > >
> > > diff --git a/sim/common/cgen-ops.h b/sim/common/cgen-ops.h
> > > index 841552066f..d718394723 100644
> > > --- a/sim/common/cgen-ops.h
> > > +++ b/sim/common/cgen-ops.h
> > > @@ -431,12 +431,8 @@ JOINSIDI (SI x0, SI x1)
> > > SEMOPS_INLINE DF
> > > JOINSIDF (SI x0, SI x1)
> > > {
> > > - union { SI in[2]; DF out; } x;
> > > - if (CURRENT_TARGET_BYTE_ORDER == BFD_ENDIAN_BIG)
> >
> > I think this is the problem with the existing code, we're using memory
> > on the HOST to perform packing / unpacking, and so its the byte
> > ordering of the HOST that we care about here, not the target.
> >
> > If I change the above line to instead be:
> >
> > if (HOST_BYTE_ORDER == BFD_ENDIAN_BIG)
> >
> > then everything works fine.
> >
> > On inspection I believe all of the uses of CURRENT_TARGET_BYTE_ORDER
> > in this file should be similarly replaced.
> >
> > I've attached a patch for this change. If you agree I'll go ahead and
> > push it.
>
> I tested this with the or1k testsuite and it does seem to work fine, thank you.
> Please go ahead.
>
> I thought I had tried this fix as well, I guess not.
Thanks for testing this.
I've pushed this patch now.
Andrew
@@ -404,7 +404,7 @@ SUBWORDXFSI (XF in, int word)
/* Note: typedef struct { SI parts[3]; } XF; */
union { XF in; SI out[3]; } x;
x.in = in;
- if (CURRENT_TARGET_BYTE_ORDER == BFD_ENDIAN_BIG)
+ if (HOST_BYTE_ORDER == BFD_ENDIAN_BIG)
return x.out[word];
else
return x.out[2 - word];
@@ -416,7 +416,7 @@ SUBWORDTFSI (TF in, int word)
/* Note: typedef struct { SI parts[4]; } TF; */
union { TF in; SI out[4]; } x;
x.in = in;
- if (CURRENT_TARGET_BYTE_ORDER == BFD_ENDIAN_BIG)
+ if (HOST_BYTE_ORDER == BFD_ENDIAN_BIG)
return x.out[word];
else
return x.out[3 - word];
@@ -432,7 +432,7 @@ SEMOPS_INLINE DF
JOINSIDF (SI x0, SI x1)
{
union { SI in[2]; DF out; } x;
- if (CURRENT_TARGET_BYTE_ORDER == BFD_ENDIAN_BIG)
+ if (HOST_BYTE_ORDER == BFD_ENDIAN_BIG)
x.in[0] = x0, x.in[1] = x1;
else
x.in[1] = x0, x.in[0] = x1;
@@ -443,7 +443,7 @@ SEMOPS_INLINE XF
JOINSIXF (SI x0, SI x1, SI x2)
{
union { SI in[3]; XF out; } x;
- if (CURRENT_TARGET_BYTE_ORDER == BFD_ENDIAN_BIG)
+ if (HOST_BYTE_ORDER == BFD_ENDIAN_BIG)
x.in[0] = x0, x.in[1] = x1, x.in[2] = x2;
else
x.in[2] = x0, x.in[1] = x1, x.in[0] = x2;
@@ -454,7 +454,7 @@ SEMOPS_INLINE TF
JOINSITF (SI x0, SI x1, SI x2, SI x3)
{
union { SI in[4]; TF out; } x;
- if (CURRENT_TARGET_BYTE_ORDER == BFD_ENDIAN_BIG)
+ if (HOST_BYTE_ORDER == BFD_ENDIAN_BIG)
x.in[0] = x0, x.in[1] = x1, x.in[2] = x2, x.in[3] = x3;
else
x.in[3] = x0, x.in[2] = x1, x.in[1] = x2, x.in[0] = x3;