[Qemu-devel] [PATCH 2/3] kvm-userspace: kvmppc: fix hostlonbits detection when cross compiling v2

[Qemu-devel] [PATCH 2/3] kvm-userspace: kvmppc: fix hostlonbits detection when cross compiling v2

[ehrhardt]

From: Christian Ehrhardt <ehrhardt@linux.vnet.ibm.com>

*update*
further debugging according to some requests revealed that ARCH_CFLAGS does
not contain all CFLAGS that might be needed, especially those supplied via
extra-cflags. Therefore people supplying things via extra-cflags instead of an
environment variable might have had issues.

A recent kvm merge with qemu brought code for 64bit power that broke cross
compilation. The issue is caused by configure trying to execute target
architecture binaries where configure is executed.

I tried to change that detection so that it works with&without cross
compilation with only a small change and especially without an addtional
configure command line switch. Including the bits/wordsize.h header a platform
usually can check its wordsize and by doing that configure can check the
hostlongbits without executing the binary. Instead it now stops after
preprocessing stage which resolved the __WORDSIZE constant and retrieves
that value.

I don't like my new check style, but it is at least less broken than before.
Another approach that was suggested was that qemu might end up needing
something like asm-offsets in the kernel to manage architecture sizes etc.
Comments and other approaches welcome.
 
Signed-off-by: Christian Ehrhardt <ehrhardt@linux.vnet.ibm.com>
---

[diffstat]
 configure |   13 +++++++------
 1 file changed, 7 insertions(+), 6 deletions(-)

[diff]

diff --git a/qemu/configure b/qemu/configure
--- a/qemu/configure
+++ b/qemu/configure
@@ -685,14 +685,15 @@
 # ppc specific hostlongbits selection
 if test "$cpu" = "powerpc" ; then
     cat > $TMPC <<EOF
-int main(void){return sizeof(long);}
+#include <bits/wordsize.h>
+__WORDSIZE
 EOF
 
-    if $cc $ARCH_CFLAGS -o $TMPE $TMPC 2> /dev/null; then
-        $TMPE
-        case $? in
-            4) hostlongbits="32";;
-            8) hostlongbits="64";;
+    if $cc $ARCH_CFLAGS $CFLAGS -E -o $TMPE.E $TMPC 2> /dev/null; then
+        wordsize=`tail -n 1 ${TMPE}.E`
+        case $wordsize in
+            32) hostlongbits="32";;
+            64) hostlongbits="64";;
             *) echo "Couldn't determine bits per long value"; exit 1;;
         esac
     else

[Qemu-devel] Re: [PATCH 2/3] kvm-userspace: kvmppc: fix hostlonbits detection when cross compiling v2

[Hollis Blanchard]

On Tue, 2008-09-30 at 11:10 +0200, ehrhardt@linux.vnet.ibm.com wrote:
> diff --git a/qemu/configure b/qemu/configure
> --- a/qemu/configure
> +++ b/qemu/configure
> @@ -685,14 +685,15 @@
>  # ppc specific hostlongbits selection
>  if test "$cpu" = "powerpc" ; then
>      cat > $TMPC <<EOF
> -int main(void){return sizeof(long);}
> +#include <bits/wordsize.h>
> +__WORDSIZE
>  EOF
> 
> -    if $cc $ARCH_CFLAGS -o $TMPE $TMPC 2> /dev/null; then
> -        $TMPE
> -        case $? in
> -            4) hostlongbits="32";;
> -            8) hostlongbits="64";;
> +    if $cc $ARCH_CFLAGS $CFLAGS -E -o $TMPE.E $TMPC 2> /dev/null; then
> +        wordsize=`tail -n 1 ${TMPE}.E`
> +        case $wordsize in
> +            32) hostlongbits="32";;
> +            64) hostlongbits="64";;
>              *) echo "Couldn't determine bits per long value"; exit 1;;
>          esac
>      else

Yeah, trying to execute target-specific code is simply incompatible with
cross-compiling, so clearly this needs to be changed.

<bits/wordsize.h> might be too Linux-specific, not sure. Also, I worry
about the preprocessor being allowed to insert all the whitespace it
wants, so I'm not sure if you can assume that __WORDSIZE will be on the
last output line.

Instead you could use this:
	$cc -E -dM - < /dev/null | grep __powerpc64__

-- 
Hollis Blanchard
IBM Linux Technology Center

Re: [Qemu-devel] [PATCH 2/3] kvm-userspace: kvmppc: fix hostlonbits detection when cross compiling v2

[malc]

On Tue, 30 Sep 2008, ehrhardt@linux.vnet.ibm.com wrote:

> From: Christian Ehrhardt <ehrhardt@linux.vnet.ibm.com>
>
> *update*
> further debugging according to some requests revealed that ARCH_CFLAGS does
> not contain all CFLAGS that might be needed, especially those supplied via
> extra-cflags. Therefore people supplying things via extra-cflags instead of an
> environment variable might have had issues.

This part i don't get, there are few more checks before/after 
hostlongbits where no CFLAGS are added to the $cc argument list. What
makes hostlongbits selection "special"? Do people specify -m32/-m64 via
--extra-cflags?

>
> A recent kvm merge with qemu brought code for 64bit power that broke cross
> compilation. The issue is caused by configure trying to execute target
> architecture binaries where configure is executed.

Yes, i never thought about cross-compilation, my bad.

> I tried to change that detection so that it works with&without cross
> compilation with only a small change and especially without an addtional
> configure command line switch. Including the bits/wordsize.h header a platform
> usually can check its wordsize and by doing that configure can check the
> hostlongbits without executing the binary. Instead it now stops after
> preprocessing stage which resolved the __WORDSIZE constant and retrieves
> that value.
>
> I don't like my new check style, but it is at least less broken than before.
> Another approach that was suggested was that qemu might end up needing
> something like asm-offsets in the kernel to manage architecture sizes etc.
> Comments and other approaches welcome.
>

I think Hollis Blanchard's method is sound,

Thank you for bringing this up.

-- 
mailto:av1474@comtv.ru

Re: [Qemu-devel] [PATCH 2/3] kvm-userspace: kvmppc: fix hostlonbits detection when cross compiling v2

[malc]

On Tue, 30 Sep 2008, ehrhardt@linux.vnet.ibm.com wrote:

> From: Christian Ehrhardt <ehrhardt@linux.vnet.ibm.com>

The commit message of r5364, the revision which basically exists
because of your posts, was somehow butchered and as such doesn't
include any references, i'm very sorry about that.

The above line was there in the commit message editing buffer, but
then i somehow screwed it all up, probably when trying to reformat
stuff suffling text around.

Once again, i'm sorry.

[..snip..]

-- 
mailto:av1474@comtv.ru

Re: [Qemu-devel] [PATCH 2/3] kvm-userspace: kvmppc: fix hostlonbits detection when cross compiling v2

[Christian Ehrhardt]

malc wrote:
> On Tue, 30 Sep 2008, ehrhardt@linux.vnet.ibm.com wrote:
>
>> From: Christian Ehrhardt <ehrhardt@linux.vnet.ibm.com>
>>
>> *update*
>> further debugging according to some requests revealed that 
>> ARCH_CFLAGS does
>> not contain all CFLAGS that might be needed, especially those 
>> supplied via
>> extra-cflags. Therefore people supplying things via extra-cflags 
>> instead of an
>> environment variable might have had issues.
>
> This part i don't get, there are few more checks before/after 
> hostlongbits where no CFLAGS are added to the $cc argument list. What
> makes hostlongbits selection "special"? Do people specify -m32/-m64 via
> --extra-cflags?
>
it was there to ensure availability of the needed include paths to reach 
wordsize.h.
But Hollis approach is much simpler, better and more reliable so never 
mind :-)

>>
>> A recent kvm merge with qemu brought code for 64bit power that broke 
>> cross
>> compilation. The issue is caused by configure trying to execute target
>> architecture binaries where configure is executed.
>
> Yes, i never thought about cross-compilation, my bad.
np, now it's fixed - thanks for quickly applying it.
>
>> I tried to change that detection so that it works with&without cross
>> compilation with only a small change and especially without an addtional
>> configure command line switch. Including the bits/wordsize.h header a 
>> platform
>> usually can check its wordsize and by doing that configure can check the
>> hostlongbits without executing the binary. Instead it now stops after
>> preprocessing stage which resolved the __WORDSIZE constant and retrieves
>> that value.
>>
>> I don't like my new check style, but it is at least less broken than 
>> before.
>> Another approach that was suggested was that qemu might end up needing
>> something like asm-offsets in the kernel to manage architecture sizes 
>> etc.
>> Comments and other approaches welcome.
>>
>
> I think Hollis Blanchard's method is sound,
>
> Thank you for bringing this up.
>


-- 

Grüsse / regards, 
Christian Ehrhardt
IBM Linux Technology Center, Open Virtualization