gcc 4.5.1 / as 2.20.51.0.11 miscompiling drivers/char/i8k.c ?

[Andi Kleen]

After upgrading my Dell laptop, both OS+kernel the i8k interface was giving
nonsensical output. As it turned out it's not the kernel but compiler
upgrade which broke this.
Guys at Archlinux have found the underlying cause (but don't seem to have
https://bbs.archlinux.org/viewtopic.php?pid=780692#p780692
gcc seems to optimize the assembly statements away.
And indeed, applying this patch makes the i8k interface work again,
i.e. replacing the asm(..) construct by asm volatile(..)

The compiler really should not optimize the asm away, because
it has both input and output arguments which are later used.
"asm volatile" normally just means "don't move significantly"

I tested it with gcc version 4.5.0 20100604 [gcc-4_5-branch revision
160292] (SUSE Linux)
and the asm statement is there for both 32bit and 64bit
(with an allmodconfig, with both -O2 and -Os)

If gcc 4.5.1 broke that over 4.5.0 you should really file a bug report
for the compiler, it seems like a serious regression in 4.5.1

-Andi

[Andreas Schwab]

Post by Andi Kleen

After upgrading my Dell laptop, both OS+kernel the i8k interface was giving
nonsensical output. As it turned out it's not the kernel but compiler
upgrade which broke this.
Guys at Archlinux have found the underlying cause (but don't seem to have
https://bbs.archlinux.org/viewtopic.php?pid=780692#p780692
gcc seems to optimize the assembly statements away.
And indeed, applying this patch makes the i8k interface work again,
i.e. replacing the asm(..) construct by asm volatile(..)

The compiler really should not optimize the asm away, because
it has both input and output arguments which are later used.
"asm volatile" normally just means "don't move significantly"

The asm fails to mention that it modifies *regs.

Andreas.

[Andi Kleen]

Post by Andreas Schwab
The asm fails to mention that it modifies *regs.

It has a memory clobber, that should be enough, no?

Besides in any case it cannot be eliminated because it has
valid non dead inputs and outputs.

-Andi

[Richard Guenther]

Post by Andi Kleen

Post by Andreas Schwab
The asm fails to mention that it modifies *regs.

It has a memory clobber, that should be enough, no?

No. A memory clobber does not cover automatic storage.

Btw, I can't see a testcase anywhere so I just assume Andreas got
it right as usual.

Richard.

Post by Andi Kleen
Besides in any case it cannot be eliminated because it has
valid non dead inputs and outputs.
-Andi
--

[Andi Kleen]

Post by Richard Guenther

Post by Andi Kleen

Post by Andreas Schwab
The asm fails to mention that it modifies *regs.

It has a memory clobber, that should be enough, no?

No. A memory clobber does not cover automatic storage.

That's a separate problem.

Post by Richard Guenther
Btw, I can't see a testcase anywhere so I just assume Andreas got
it right as usual.

An asm with live inputs and outputs should never be optimized
way. If 4.5.1 started doing that it's seriously broken.

-Andi

[Richard Guenther]

Post by Andi Kleen

Post by Andi Kleen

Post by Andreas Schwab
The asm fails to mention that it modifies *regs.

It has a memory clobber, that should be enough, no?

No.  A memory clobber does not cover automatic storage.

That's a separate problem.

Btw, I can't see a testcase anywhere so I just assume Andreas got
it right as usual.

An asm with live inputs and outputs should never be optimized
way. If 4.5.1 started doing that it's seriously broken.

Please provide a testcase, such asms can be optimized if the
outputs are dead.

Richard.

Post by Andi Kleen
-Andi
--

[Paul Koning]

Post by Richard Guenther

Post by Andi Kleen

Post by Richard Guenther

Post by Andi Kleen

Post by Andreas Schwab
The asm fails to mention that it modifies *regs.

It has a memory clobber, that should be enough, no?

No. A memory clobber does not cover automatic storage.

That's a separate problem.

Post by Richard Guenther
Btw, I can't see a testcase anywhere so I just assume Andreas got
it right as usual.

An asm with live inputs and outputs should never be optimized
way. If 4.5.1 started doing that it's seriously broken.

Please provide a testcase, such asms can be optimized if the
outputs are dead.

I don't know about 4.5, but I noticed that with 4.6 (trunk), testcasese like gcc.c-torture/compile/20000804-1.c optimize away the asm and all the operand generation except for -O0.

paul

[Jakub Jelinek]

Post by Paul Koning
I don't know about 4.5, but I noticed that with 4.6 (trunk), testcasese
like gcc.c-torture/compile/20000804-1.c optimize away the asm and all the
operand generation except for -O0.

That's fine, the asm isn't volatile and the output is not used.

Jakub

[Michael Matz]

Hi,

Post by Andi Kleen

Post by Richard Guenther

Post by Andi Kleen

Post by Andreas Schwab
The asm fails to mention that it modifies *regs.

It has a memory clobber, that should be enough, no?

No. A memory clobber does not cover automatic storage.

That's a separate problem.

Post by Richard Guenther
Btw, I can't see a testcase anywhere so I just assume Andreas got
it right as usual.

An asm with live inputs and outputs should never be optimized
way. If 4.5.1 started doing that it's seriously broken.

You know the drill: testcase -> gcc.gnu.org/bugzilla/

(In particular up to now it's only speculation in some forum that the asm
really is optimized away, which I agree would be a bug, or if it isn't
merely that regs->eax isn't reloaded after the asm(), which would be
caused by the problem Andreas mentioned)


Ciao,
Michael.

[Dave Korn]

Post by Andi Kleen
An asm with live inputs and outputs should never be optimized
way. If 4.5.1 started doing that it's seriously broken.

I don't see that. Consider:

void foo (void)
{
int x, y, z;
x = 23;
y = x + 1;
z = y + 1;
}

So far, you'd agree the compiler may optimise the entire function away? So
why not this:

void foo (void)
{
int x, y, z;
x = 23;
asm ("do something" : "=r" (y) : "r" (x) );
z = y + 1;
}

?

cheers,
DaveK

[Michael Matz]

Hi,

Post by Dave Korn
void foo (void)
{
int x, y, z;
x = 23;
asm ("do something" : "=r" (y) : "r" (x) );
z = y + 1;
}

The case in i8k.c really is different. It does use the value by
influencing the return value and the callers use the returned value in
conditionals and the like. It really, really _is_ used :-) and if GCC
removes the asm (which up to now is only speculation) then it's a GCC bug.

The code outlines like so:

int i8k_smm (regs) {
int rc;
asm (... : "=r"(rc) ...);
if (rc != 0 || ...)
return -EINVAL;
return 0;
}

...
struct regs regs = {.eax = ...}
return i8k_smm(regs) ?: regs.eax;
...

My speculation is, that the asm is not removed but rather that regs.eax
isn't reloaded after the asm because the memory clobber doesn't clobber
automatic variables.


Ciao,
Michael.

[Andi Kleen]

Post by Michael Matz
My speculation is, that the asm is not removed but rather that regs.eax
isn't reloaded after the asm because the memory clobber doesn't clobber
automatic variables.

Yes that makes sense. I wasn't able to verify it so far though.

Maybe the original poster could try the obvious patch
instead of the volatile change.


i8k: tell gcc that regs gets clobbered

Signed-off-by: Andi Kleen <***@linux.intel.com>

diff --git a/drivers/char/i8k.c b/drivers/char/i8k.c
index 3bc0eef..f3bbf73 100644
--- a/drivers/char/i8k.c
+++ b/drivers/char/i8k.c
@@ -142,7 +142,7 @@ static int i8k_smm(struct smm_regs *regs)
"lahf\n\t"
"shrl $8,%%eax\n\t"
"andl $1,%%eax\n"
- :"=a"(rc)
+ :"=a"(rc), "=m" (*regs)
: "a"(regs)
: "%ebx", "%ecx", "%edx", "%esi", "%edi", "memory");
#else
@@ -167,7 +167,7 @@ static int i8k_smm(struct smm_regs *regs)
"movl %%edx,0(%%eax)\n\t"
"lahf\n\t"
"shrl $8,%%eax\n\t"
- "andl $1,%%eax\n":"=a"(rc)
+ "andl $1,%%eax\n":"=a"(rc), "=m" (*regs)
: "a"(regs)
: "%ebx", "%ecx", "%edx", "%esi", "%edi", "memory");
#endif

-Andi

[Andreas Schwab]

Post by Andi Kleen
@@ -142,7 +142,7 @@ static int i8k_smm(struct smm_regs *regs)
"lahf\n\t"
"shrl $8,%%eax\n\t"
"andl $1,%%eax\n"
- :"=a"(rc)
+ :"=a"(rc), "=m" (*regs)

I think this should be "+m".

Andreas.

[Jim]

Post by Andreas Schwab

Post by Andi Kleen
@@ -142,7 +142,7 @@ static int i8k_smm(struct smm_regs *regs)
"lahf\n\t"
"shrl $8,%%eax\n\t"
"andl $1,%%eax\n"
- :"=a"(rc)
+ :"=a"(rc), "=m" (*regs)

I think this should be "+m".
Andreas.

Just tested Andi's patch with Andreas' suggestion to make it +m,
i.e. like attached and can confirm it solves the issue.

Thanks guys,
Jim Bos

[Jim Bos]

More recent GCC caused the i8k driver to stop working, on Slackware
compiler was upgraded from gcc-4.4.4 to gcc-4.5.1 after which it
didn't work anymore, meaning the driver didn't load or gave total
nonsensical output.
As it turned out the asm(..) statement forgot to mention it modifies
the *regs variable.
Credits to Andi Kleen <***@firstfloor.org> and Andreas Schwab
<***@linux-m68k.org> for providing the fix.

Signed-off-by: Jim Bos <***@xs4all.nl>

[James Cloos]

Gcc 4.5.1 running on an amd64 box "cross"-compiling for a P3 i8k fails
to compile the module since commit 6b4e81db2552bad04100e7d5ddeed7e848f5=
3b48
with:

CC drivers/char/i8k.o
drivers/char/i8k.c: In function =E2=80=98i8k_smm=E2=80=99:
drivers/char/i8k.c:149:2: error: can't find a register in class =E2=80=98=
GENERAL_REGS=E2=80=99 while reloading =E2=80=98asm=E2=80=99
drivers/char/i8k.c:149:2: error: =E2=80=98asm=E2=80=99 operand has impo=
ssible constraints

-JimC
--=20
James Cloos <***@jhcloos.com> OpenPGP: 1024D/ED7DAEA6

[Linus Torvalds]

Gcc 4.5.1 running on an amd64 box "cross"-compiling for a P3 i8k fail=

s

to compile the module since commit 6b4e81db2552bad04100e7d5ddeed7e848=

f53b48

=A0CC =A0 =A0 =A0drivers/char/i8k.o
drivers/char/i8k.c:149:2: error: can't find a register in class =91GE=

NERAL_REGS=92 while reloading =91asm=92

drivers/char/i8k.c:149:2: error: =91asm=92 operand has impossible con=

straints

At this point, I think this falls clearly under "unresolvable gcc bug".

Quite frankly, I think gcc was buggy to begin with: since we had a
memory clobber, the "+m" (*regs) should not have mattered. The fact
that "*regs" may be some local variable doesn't make any difference
what-so-ever, since we took the address of the variable. So the memory
clobber _clearly_ can change that variable.

So when Richard Gunther says "a memory clobber doesn't cover automatic
storage", to me that very clearly spells "gcc is buggy as hell".
Because automatic storage with its address taken _very_ much gets
clobbered by things like memset etc. If the compiler doesn't
understand that, the compiler is just broken.

And now, if even the (superfluous) "+m" isn't working, it sounds like
we have no sane options left. Except to say that gcc-4.5.1 is totally
broken wrt asms.

Can we just get gcc to realize that when you pass the address of
automatic storage to an asm, that means that "memory" really does
clobber it? Because clearly that is the case.

Linus

[Jakub Jelinek]

Post by Linus Torvalds
So when Richard Gunther says "a memory clobber doesn't cover automatic
storage", to me that very clearly spells "gcc is buggy as hell".
Because automatic storage with its address taken _very_ much gets
clobbered by things like memset etc. If the compiler doesn't
understand that, the compiler is just broken.

I'll leave the discussion about meaning of "memory" clobber aside to
Richard,

Post by Linus Torvalds
And now, if even the (superfluous) "+m" isn't working, it sounds like
we have no sane options left. Except to say that gcc-4.5.1 is totally

just to say that of course there are sane options left.

:"=a"(rc), "+m" (*regs)
: "a"(regs)
: "%ebx", "%ecx", "%edx", "%esi", "%edi", "memory");

is simply too high register pressure for i386 if you force also
-fno-omit-frame-pointer, there is not a single register left.

Yes, reload should figure out it has address of regs already tied to %eax,
unfortunately starting with IRA it doesn't (I'll file a GCC bug about that;
so that leaves 4.4/4.5/4.6 currently not being able to compile it).

That said, changing the inline asm to just clobber one less register
would be completely sufficient to make it work well with all gccs out there,
just push/pop one of the register around the whole body. I doubt calling
out SMM BIOS is actually so performance critical that one push and one pop
would ruin it. Of course x86_64 version can stay as is, there are enough
registers left...

Jakub

[Andi Kleen]

Post by Jakub Jelinek
That said, changing the inline asm to just clobber one less register
would be completely sufficient to make it work well with all gccs out there,
just push/pop one of the register around the whole body. I doubt calling
out SMM BIOS is actually so performance critical that one push and one pop
would ruin it. Of course x86_64 version can stay as is, there are enough
registers left...

Yes traditionally clobbering all registers has been dangerous
and it clearly can be done inside the asm too.

Here's a untested patch to do some manual push/pops too. Could someone with
the hardware please test it? (running a 32bit kernel)

-Andi

---

i8k: Clobber less registers

gcc doesn't like inline assembler statements that clobber nearly
all registers. Save a few registers manually on i386 to avoid this
problem.

Fix suggested by Jakub Jelinek

Signed-off-by: Andi Kleen <***@linux.intel.com>

diff --git a/drivers/char/i8k.c b/drivers/char/i8k.c
index f0863be..a2da38b 100644
--- a/drivers/char/i8k.c
+++ b/drivers/char/i8k.c
@@ -146,7 +146,10 @@ static int i8k_smm(struct smm_regs *regs)
: "a"(regs)
: "%ebx", "%ecx", "%edx", "%esi", "%edi", "memory");
#else
- asm("pushl %%eax\n\t"
+ asm("pushl %%ebx\n\t"
+ "pushl %%ecx\n\t"
+ "pushl %%edx\n\t"
+ "pushl %%eax\n\t"
"movl 0(%%eax),%%edx\n\t"
"push %%edx\n\t"
"movl 4(%%eax),%%ebx\n\t"
@@ -167,10 +170,13 @@ static int i8k_smm(struct smm_regs *regs)
"movl %%edx,0(%%eax)\n\t"
"lahf\n\t"
"shrl $8,%%eax\n\t"
- "andl $1,%%eax\n"
+ "andl $1,%%eax\n\t"
+ "popl %%edx\n\t"
+ "popl %%ecx\n\t"
+ "popl %%ebx\n"
:"=a"(rc), "+m" (*regs)
: "a"(regs)
- : "%ebx", "%ecx", "%edx", "%esi", "%edi", "memory");
+ : "%esi", "%edi", "memory");
#endif
if (rc != 0 || (regs->eax & 0xffff) == 0xffff || regs->eax == eax)
return -EINVAL;

[Jakub Jelinek]

Post by Jakub Jelinek
Yes, reload should figure out it has address of regs already tied to %eax,
unfortunately starting with IRA it doesn't (I'll file a GCC bug about that;

http://gcc.gnu.org/PR46479

Jakub

[Jakub Jelinek]

Post by Jakub Jelinek

Post by Linus Torvalds
So when Richard Gunther says "a memory clobber doesn't cover automatic
storage", to me that very clearly spells "gcc is buggy as hell".
Because automatic storage with its address taken _very_ much gets
clobbered by things like memset etc. If the compiler doesn't
understand that, the compiler is just broken.

I'll leave the discussion about meaning of "memory" clobber aside to
Richard,

And for this the starting point should be what has been requested,
i.e. preprocessed source + gcc options + gcc version and some hints what
actually misbehaves (with the , "+m" (*regs) change reverted)
in gcc bugzilla. Only with that we can actually look at what has been
happening, see whether it is the tree optimizations or RTL and which one
makes a difference.
If I've missed a PR about this I apologize.

Jakub

[Andi Kleen]

Post by Jakub Jelinek
And for this the starting point should be what has been requested,
i.e. preprocessed source + gcc options + gcc version and some hints what
actually misbehaves (with the , "+m" (*regs) change reverted)
in gcc bugzilla. Only with that we can actually look at what has been
happening, see whether it is the tree optimizations or RTL and which one
makes a difference.
If I've missed a PR about this I apologize.

I tried to file one, but I can't reproduce it currently
(I don't have hardware, so have to rely on code reading and the 32bit
code looks correct to me even without the additional +m)

The preprocessed source is at
http://halobates.de/tmp/i8k.i

Options I used:

-D__KERNEL__ -Wall -Wundef -Wstrict-prototypes -Wno-trigraphs -fno-strict-aliasing -fno-common -Werror-implicit-function-declaration -Wno-format-security -fno-delete-null-pointer-checks -O2 -m32 -msoft-float -mregparm=3 -freg-struct-return -mpreferred-stack-boundary=2 -march=i686 -mtune=pentium3 -mtune=generic -maccumulate-outgoing-args -Wa,-mtune=generic32 -ffreestanding -DCONFIG_AS_CFI=1 -DCONFIG_AS_CFI_SIGNAL_FRAME=1 -DCONFIG_AS_CFI_SECTIONS=1 -pipe -Wno-sign-compare -fno-asynchronous-unwind-tables -mno-sse -mno-mmx -mno-sse2 -mno-3dnow -Wframe-larger-than=2048 -fno-stack-protector -fno-omit-frame-pointer -fno-optimize-sibling-calls -pg -Wdeclaration-after-statement -Wno-pointer-sign -fno-strict-overflow -fconserve-stack

-Andi

[Jakub Jelinek]

Post by Andi Kleen

Post by Jakub Jelinek
And for this the starting point should be what has been requested,
i.e. preprocessed source + gcc options + gcc version and some hints what
actually misbehaves (with the , "+m" (*regs) change reverted)
in gcc bugzilla. Only with that we can actually look at what has been
happening, see whether it is the tree optimizations or RTL and which one
makes a difference.
If I've missed a PR about this I apologize.

I tried to file one, but I can't reproduce it currently
(I don't have hardware, so have to rely on code reading and the 32bit
code looks correct to me even without the additional +m)
The preprocessed source is at
http://halobates.de/tmp/i8k.i
-D__KERNEL__ -Wall -Wundef -Wstrict-prototypes -Wno-trigraphs -fno-strict-aliasing -fno-common -Werror-implicit-function-declaration -Wno-format-security -fno-delete-null-pointer-checks -O2 -m32 -msoft-float -mregparm=3 -freg-struct-return -mpreferred-stack-boundary=2 -march=i686 -mtune=pentium3 -mtune=generic -maccumulate-outgoing-args -Wa,-mtune=generic32 -ffreestanding -DCONFIG_AS_CFI=1 -DCONFIG_AS_CFI_SIGNAL_FRAME=1 -DCONFIG_AS_CFI_SECTIONS=1 -pipe -Wno-sign-compare -fno-asynchronous-unwind-tables -mno-sse -mno-mmx -mno-sse2 -mno-3dnow -Wframe-larger-than=2048 -fno-stack-protector -fno-omit-frame-pointer -fno-optimize-sibling-calls -pg -Wdeclaration-after-statement -Wno-pointer-sign -fno-strict-overflow -fconserve-stack

Indeed, with this and 4.5.2 20101111 (prerelease) from SVN as well as
gcc-4.5.1-5.fc14:

...
movl %eax, -16(%ebp) # regs, %sfp
movl (%eax), %eax # regs_2(D)->eax,
movl %eax, -20(%ebp) #, %sfp
movl -16(%ebp), %eax # %sfp,
#APP
# 149 "/home/lsrc/git/linux-work2/drivers/char/i8k.c" 1
...
#NO_APP
testl %eax, %eax #
movl $-22, %edx #, D.18378
movl %eax, -24(%ebp) #, %sfp
je .L7 #,
.L2:
movl -12(%ebp), %ebx #,
movl %edx, %eax # D.18378,
movl -8(%ebp), %esi #,
movl -4(%ebp), %edi #,
movl %ebp, %esp #,
popl %ebp #
ret
.p2align 4,,7
.p2align 3
.L7:
movl -16(%ebp), %eax # %sfp,
movl (%eax), %ecx # regs_2(D)->eax, D.18371
cmpw $-1, %cx #, D.18371
je .L2 #,
cmpl %ecx, -20(%ebp) # D.18371, %sfp
cmovne -24(%ebp), %edx # %sfp,, D.18378
jmp .L2 #
.size i8k_smm, .-i8k_smm

I don't see any problems on the assembly level. i8k_smm is
not inlined in this case and checks all 3 conditions.

Guess we need somebody who actually reported the problem, state what
gcc was actually used and post preprocessed source, gcc options
from his case.

Jakub

[Andi Kleen]

Post by Jakub Jelinek
Guess we need somebody who actually reported the problem, state what
gcc was actually used and post preprocessed source, gcc options
from his case.

Jim Bos,
Can you please supply that?

Please use

rm drivers/char/i8k.o
make V=1 drivers/char/i8k.o
make drivers/char/i8k.i

and supply the .i file and the output of the first make line

Thanks,
-Andi

[Jim Bos]

Post by Andi Kleen

Post by Jakub Jelinek
Guess we need somebody who actually reported the problem, state what
gcc was actually used and post preprocessed source, gcc options
from his case.

Jim Bos,
Can you please supply that?
Please use
rm drivers/char/i8k.o
make V=1 drivers/char/i8k.o
make drivers/char/i8k.i
and supply the .i file and the output of the first make line
Thanks,
-Andi

Andi,

See attached, note this is the vanilla 2.6.36 i8k.c (without any patch).
And to be 100% sure, if I build this (make drivers/char/i8k.ko) it won't
work.

[ The i8k.i is rather big, even gzipped 80k, not sure if it'll bounce ]

_
Jim

[Jakub Jelinek]

Post by Jim Bos
See attached, note this is the vanilla 2.6.36 i8k.c (without any patch).
And to be 100% sure, if I build this (make drivers/char/i8k.ko) it won't
work.
[ The i8k.i is rather big, even gzipped 80k, not sure if it'll bounce ]

Please also say which exact gcc you are using.

Note, I've compiled it with current 4.5 branch and made the function
always_inline and still didn't see any issues in the *.optimized dump,
regs.eax after the inline asm has always been compared to the constant
that has been stored into regs.eax before the inline asm.

Jakub

[Jim Bos]

Post by Jakub Jelinek

Post by Jim Bos
See attached, note this is the vanilla 2.6.36 i8k.c (without any patch).
And to be 100% sure, if I build this (make drivers/char/i8k.ko) it won't
work.
[ The i8k.i is rather big, even gzipped 80k, not sure if it'll bounce ]

Please also say which exact gcc you are using.
Note, I've compiled it with current 4.5 branch and made the function
always_inline and still didn't see any issues in the *.optimized dump,
regs.eax after the inline asm has always been compared to the constant
that has been stored into regs.eax before the inline asm.
Jakub
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

# gcc -v
Reading specs from /usr/lib/gcc/i486-slackware-linux/4.5.1/specs
COLLECT_GCC=gcc
COLLECT_LTO_WRAPPER=/usr/libexec/gcc/i486-slackware-linux/4.5.1/lto-wrapper
Target: i486-slackware-linux
Configured with: ../gcc-4.5.1/configure --prefix=/usr --libdir=/usr/lib
--mandir=/usr/man --infodir=/usr/info --enable-shared --enable-bootstrap
--enable-languages=ada,c,c++,fortran,java,objc --enable-threads=posix
--enable-checking=release --with-system-zlib
--with-python-dir=/lib/python2.6/site-packages
--disable-libunwind-exceptions --enable-__cxa_atexit --enable-libssp
--with-gnu-ld --verbose --with-arch=i486 --target=i486-slackware-linux
--build=i486-slackware-linux --host=i486-slackware-linux
Thread model: posix
gcc version 4.5.1 (GCC)

I'm re-reading this thread where I found the asm-> asm volatine suggestion:
https://bbs.archlinux.org/viewtopic.php?pid=752099#p752099
but nobody there reported their gcc version (but apparently first
people started complaining May 1st).

_
Jim

[Jakub Jelinek]

Post by Jim Bos
# gcc -v
Reading specs from /usr/lib/gcc/i486-slackware-linux/4.5.1/specs
COLLECT_GCC=gcc
COLLECT_LTO_WRAPPER=/usr/libexec/gcc/i486-slackware-linux/4.5.1/lto-wrapper
Target: i486-slackware-linux
Configured with: ../gcc-4.5.1/configure --prefix=/usr --libdir=/usr/lib
--mandir=/usr/man --infodir=/usr/info --enable-shared --enable-bootstrap
--enable-languages=ada,c,c++,fortran,java,objc --enable-threads=posix
--enable-checking=release --with-system-zlib
--with-python-dir=/lib/python2.6/site-packages
--disable-libunwind-exceptions --enable-__cxa_atexit --enable-libssp
--with-gnu-ld --verbose --with-arch=i486 --target=i486-slackware-linux
--build=i486-slackware-linux --host=i486-slackware-linux
Thread model: posix
gcc version 4.5.1 (GCC)

Does it have any patches applied? The gcc options look the same as what
I've been already trying earlier.
Thus, can you run gcc with those options on i8k.i and add -fverbose-asm
to make it easier to read and post i8k.s you get?

Jakub

[Jim Bos]

Post by Jakub Jelinek

Post by Jim Bos
# gcc -v
Reading specs from /usr/lib/gcc/i486-slackware-linux/4.5.1/specs
COLLECT_GCC=gcc
COLLECT_LTO_WRAPPER=/usr/libexec/gcc/i486-slackware-linux/4.5.1/lto-wrapper
Target: i486-slackware-linux
Configured with: ../gcc-4.5.1/configure --prefix=/usr --libdir=/usr/lib
--mandir=/usr/man --infodir=/usr/info --enable-shared --enable-bootstrap
--enable-languages=ada,c,c++,fortran,java,objc --enable-threads=posix
--enable-checking=release --with-system-zlib
--with-python-dir=/lib/python2.6/site-packages
--disable-libunwind-exceptions --enable-__cxa_atexit --enable-libssp
--with-gnu-ld --verbose --with-arch=i486 --target=i486-slackware-linux
--build=i486-slackware-linux --host=i486-slackware-linux
Thread model: posix
gcc version 4.5.1 (GCC)

Does it have any patches applied? The gcc options look the same as what
I've been already trying earlier.
Thus, can you run gcc with those options on i8k.i and add -fverbose-asm
to make it easier to read and post i8k.s you get?
Jakub

Slackware is typically not patching much (and I'm just using the
pre-compiled binary). Here is the link to how it's built:
http://slackware.osuosl.org/slackware-current/source/d/gcc/
there doesn't appear to be anything relevant changed.

I already posted the .s files, plain 2.6.36 and the one with working
patch, I =think= that's already using -fverbose-asm, at least that shows
in the output.

_
Jim

[Linus Torvalds]

I don't see any problems on the assembly level.  i8k_smm is
not inlined in this case and checks all 3 conditions.

If it really is related to gcc not understanding that "*regs" has
changed due to the memory being an automatic variable, and passing in
"regs" itself as a pointer to that automatic variable together with
the "memory" clobber not being sufficient, than I think it's the lack
of inlining that will automatically hide the bug.

(Side note: and I think this does show how much of a gcc bug it is not
to consider "memory" together with passing in a pointer to an asm to
always be a clobber).

Because if it isn't inlined, then "regs" will be seen a a real pointer
to some external memory (the caller) rather than being optimized to
just be the auto structure on the stack. Because *mem is auto only
within the context of the caller.

Which actually points to a possible simpler:
- remove the "+m" since it adds too much register pressure
- mark the i8k_smm() as "noinline" instead.

Quite frankly, I'd hate to add even more crud to that inline asm (to
save/restore the registers manually). It's already not the prettiest
thing around.

So does the attached patch work for everybody?

Linus

[Jim Bos]

Post by Linus Torvalds

Post by Jakub Jelinek
I don't see any problems on the assembly level. i8k_smm is
not inlined in this case and checks all 3 conditions.

If it really is related to gcc not understanding that "*regs" has
changed due to the memory being an automatic variable, and passing in
"regs" itself as a pointer to that automatic variable together with
the "memory" clobber not being sufficient, than I think it's the lack
of inlining that will automatically hide the bug.
(Side note: and I think this does show how much of a gcc bug it is not
to consider "memory" together with passing in a pointer to an asm to
always be a clobber).
Because if it isn't inlined, then "regs" will be seen a a real pointer
to some external memory (the caller) rather than being optimized to
just be the auto structure on the stack. Because *mem is auto only
within the context of the caller.
- remove the "+m" since it adds too much register pressure
- mark the i8k_smm() as "noinline" instead.
Quite frankly, I'd hate to add even more crud to that inline asm (to
save/restore the registers manually). It's already not the prettiest
thing around.
So does the attached patch work for everybody?
Linus

Hmm, that doesn't work.

[ Not sure if you read to whole thread but initial workaround was to
change the asm(..) to asm volatile(..) which did work. ]

Jim.

[Linus Torvalds]

Post by Jim Bos
Hmm, that doesn't work.
[ Not sure if you read to whole thread but initial workaround was to
change the asm(..) to asm volatile(..) which did work. ]

Since I have a different gcc than yours (and I'm not going to compile
my own), have you posted your broken .s file anywhere? In fact, with
the noinline (and the removal of the "+m" thing - iow just the patch
you tried), what does just the "i8k_smm" function assembly look like
for you after you've done a "make drivers/char/i8k.s"?

If the asm just doesn't exist AT ALL, that's just odd. Because every
single call-site of i8k_smm() clearly looks at the return value. So
the volatile really shouldn't make any difference from that
standpoint. Odd.

Linus

[Jim Bos]

Post by Linus Torvalds

Post by Jim Bos
Hmm, that doesn't work.
[ Not sure if you read to whole thread but initial workaround was to
change the asm(..) to asm volatile(..) which did work. ]

Since I have a different gcc than yours (and I'm not going to compile
my own), have you posted your broken .s file anywhere? In fact, with
the noinline (and the removal of the "+m" thing - iow just the patch
you tried), what does just the "i8k_smm" function assembly look like
for you after you've done a "make drivers/char/i8k.s"?
If the asm just doesn't exist AT ALL, that's just odd. Because every
single call-site of i8k_smm() clearly looks at the return value. So
the volatile really shouldn't make any difference from that
standpoint. Odd.
Linus

Attached version with plain 2.6.36 source and version with the committed
patch, i.e with the '"+m" (*regs)'


_
Jim

[Jim Bos]

Post by Jim Bos

Post by Linus Torvalds

Post by Jim Bos
Hmm, that doesn't work.
[ Not sure if you read to whole thread but initial workaround was to
change the asm(..) to asm volatile(..) which did work. ]

Since I have a different gcc than yours (and I'm not going to compile
my own), have you posted your broken .s file anywhere? In fact, with
the noinline (and the removal of the "+m" thing - iow just the patch
you tried), what does just the "i8k_smm" function assembly look like
for you after you've done a "make drivers/char/i8k.s"?
If the asm just doesn't exist AT ALL, that's just odd. Because every
single call-site of i8k_smm() clearly looks at the return value. So
the volatile really shouldn't make any difference from that
standpoint. Odd.
Linus

Attached version with plain 2.6.36 source and version with the committed
patch, i.e with the '"+m" (*regs)'
_
Jim

And I just tried with your noninline patch which results in exactly the
same .s file as with plain 2.6.36 source, i.e. the noninline patch is
not doing anything here.

_
Jim

[Linus Torvalds]

Post by Jim Bos
Attached version with plain 2.6.36 source and version with the committed
patch, i.e with the '"+m" (*regs)'

Looks 100% identical in i8k_smm() itself, and I'm not seeing anything
bad. The asm has certainly not been optimized away as implied in the
archlinux thread.

There are differences, but they are with code generation *elsewhere*.

To me it is starting to look like the real problem is that gcc has
decided that the "i8k_smm()" function is "__attribute__((const))".

Which is clearly totally bogus. If a function has an inline asm that
has a memory clobber, it is clearly *not* 'const'. But that does
explain the bug, and does explain why "+m" makes a difference and why
"noinline" does not.

So what I _think_ happens is that

- gcc logic for the automatic 'const' attribute for functions is
broken, so it marks that function 'const'.

- since the rule for a const function is that it only _looks_ at its
attributes and has no side effects, now the callers will decide that
'i8k_smm()' cannot change the passed-in structure, so they'll happily
optimize away all the accesses to it.

Linus

[Jakub Jelinek]

Post by Jim Bos

Post by Linus Torvalds

Post by Jim Bos
Hmm, that doesn't work.
[ Not sure if you read to whole thread but initial workaround was to
change the asm(..) to asm volatile(..) which did work. ]

Since I have a different gcc than yours (and I'm not going to compile
my own), have you posted your broken .s file anywhere? In fact, with
the noinline (and the removal of the "+m" thing - iow just the patch
you tried), what does just the "i8k_smm" function assembly look like
for you after you've done a "make drivers/char/i8k.s"?
If the asm just doesn't exist AT ALL, that's just odd. Because every
single call-site of i8k_smm() clearly looks at the return value. So
the volatile really shouldn't make any difference from that
standpoint. Odd.
Linus

Attached version with plain 2.6.36 source and version with the committed
patch, i.e with the '"+m" (*regs)'

Thanks, this actually helped to see the problem.
The problem is not inside of i8k_smm, which is not inlined, but in the
callers.
ipa-pure-const.c pass thinks i8k_smm is a pure function, thus
regs = {};
regs.eax = 166;
x = i8k_smm (&regs);
if (!x) x = regs.eax;
in the callers is optimized into
regs = {}
regs.eax = 166;
x = i8k_smm (&regs);
if (!x) x = 166;
Now, not sure why this happens, as there is
case GIMPLE_ASM:
for (i = 0; i < gimple_asm_nclobbers (stmt); i++)
{
tree op = gimple_asm_clobber_op (stmt, i);
if (simple_cst_equal(TREE_VALUE (op), memory_identifier_string) == 1)
{
if (dump_file)
fprintf (dump_file, " memory asm clobber is not const/pure");
/* Abandon all hope, ye who enter here. */
local->pure_const_state = IPA_NEITHER;
}
}
Debugging...

Jakub

[Jakub Jelinek]

Post by Jakub Jelinek
Now, not sure why this happens, as there is
for (i = 0; i < gimple_asm_nclobbers (stmt); i++)
{
tree op = gimple_asm_clobber_op (stmt, i);
if (simple_cst_equal(TREE_VALUE (op), memory_identifier_string) == 1)
{
if (dump_file)
fprintf (dump_file, " memory asm clobber is not const/pure");
/* Abandon all hope, ye who enter here. */
local->pure_const_state = IPA_NEITHER;
}
}
Debugging...

Ah, the problem is that memory_identifier_string is only initialized in
ipa-reference.c's initialization, so it can be (and is in this case) NULL in
ipa-pure-const.c.

Two possible fixes (the latter is apparently what is used in
tree-ssa-operands.c, so is probably sufficient). Guess ipa-reference.c
should be changed to do the same and just drop memory_identifier_string.

Jakub

[Linus Torvalds]

Post by Jakub Jelinek
Ah, the problem is that memory_identifier_string is only initialized in
ipa-reference.c's initialization, so it can be (and is in this case) NULL in
ipa-pure-const.c.

Ok. And I guess you can verify that all versions of gcc do this
correctly for "asm volatile"?

Because since we'll have to work around this problem in the kernel, I
suspect the simplest solution is to remove the "+m" that causes
register pressure problems, and then use "asm volatile" to work around
the const-function bug.

And add a large comment about why "asm volatile" is probably always a
good idea when you have a memory clobber and don't have any other
visible memory modifications.

I do wonder if this explains some of the problems we had with the
bitop asms too.

Hmm?

Linus

[Jakub Jelinek]

Post by Linus Torvalds

Post by Jakub Jelinek
Ah, the problem is that memory_identifier_string is only initialized in
ipa-reference.c's initialization, so it can be (and is in this case) NULL in
ipa-pure-const.c.

Ok. And I guess you can verify that all versions of gcc do this
correctly for "asm volatile"?

Yes, reading 4.1/4.2/4.3/4.4/4.5/4.6 code ipa-pure-const.c handles
asm volatile correctly, in each case the function is no longer assumed to be
pure or const in the discovery (of course, user can still say the
function is const or pure). 4.0 and earlier didn't have ipa-pure-const.c.

Using the simplified

extern void abort (void);

__attribute__((noinline)) int
foo (int *p)
{
int r;
asm ("movl $6, (%1)\n\txorl %0, %0" : "=r" (r) : "r" (p) : "memory");
return r;
}

int
main (void)
{
int p = 8;
if ((foo (&p) ? : p) != 6)
abort ();
return 0;
}

testcase shows that in 4.1/4.2/4.3/4.4 this is miscompiled only when using
-fno-ipa-reference, in 4.5 it is miscompiled always when optimizing
unless -fno-ipa-pure-const (as 4.5 added local-pure-const pass which is run
before ipa-reference) and in 4.6 this has been fixed by Honza when
doing ipa cleanups.

Post by Linus Torvalds
Because since we'll have to work around this problem in the kernel, I
suspect the simplest solution is to remove the "+m" that causes
register pressure problems, and then use "asm volatile" to work around
the const-function bug.

Yes.

Jakub

[Jim Bos]

Post by Jakub Jelinek

Post by Linus Torvalds

Post by Jakub Jelinek
Ah, the problem is that memory_identifier_string is only initialized in
ipa-reference.c's initialization, so it can be (and is in this case) NULL in
ipa-pure-const.c.

Ok. And I guess you can verify that all versions of gcc do this
correctly for "asm volatile"?

Yes, reading 4.1/4.2/4.3/4.4/4.5/4.6 code ipa-pure-const.c handles
asm volatile correctly, in each case the function is no longer assumed to be
pure or const in the discovery (of course, user can still say the
function is const or pure). 4.0 and earlier didn't have ipa-pure-const.c.
Using the simplified
extern void abort (void);
__attribute__((noinline)) int
foo (int *p)
{
int r;
asm ("movl $6, (%1)\n\txorl %0, %0" : "=r" (r) : "r" (p) : "memory");
return r;
}
int
main (void)
{
int p = 8;
if ((foo (&p) ? : p) != 6)
abort ();
return 0;
}
testcase shows that in 4.1/4.2/4.3/4.4 this is miscompiled only when using
-fno-ipa-reference, in 4.5 it is miscompiled always when optimizing
unless -fno-ipa-pure-const (as 4.5 added local-pure-const pass which is run
before ipa-reference) and in 4.6 this has been fixed by Honza when
doing ipa cleanups.

Post by Linus Torvalds
Because since we'll have to work around this problem in the kernel, I
suspect the simplest solution is to remove the "+m" that causes
register pressure problems, and then use "asm volatile" to work around
the const-function bug.

Yes.
Jakub

Linus,

In case you didn't already fixed this, here's the follow-up patch.
---

The fix to work around the gcc miscompiling i8k.c to add "+m (*regs)"
caused register pressure problems. Changing the 'asm' statement to
'asm volatile' instead should prevent that and works around the gcc
bug as well.

Signed-off-by: Jim Bos <***@xs4all.nl>

[Andi Kleen]

Post by Jakub Jelinek
testcase shows that in 4.1/4.2/4.3/4.4 this is miscompiled only when using
-fno-ipa-reference, in 4.5 it is miscompiled always when optimizing
unless -fno-ipa-pure-const (as 4.5 added local-pure-const pass which is run
before ipa-reference) and in 4.6 this has been fixed by Honza when
doing ipa cleanups.

Maybe it would be better to simply change the kernel Makefiles to pass
-fno-ipa-pure-const instead of adding volatiles everywhere.

-Andi

[Jakub Jelinek]

Post by Andi Kleen

Post by Jakub Jelinek
testcase shows that in 4.1/4.2/4.3/4.4 this is miscompiled only when using
-fno-ipa-reference, in 4.5 it is miscompiled always when optimizing
unless -fno-ipa-pure-const (as 4.5 added local-pure-const pass which is run
before ipa-reference) and in 4.6 this has been fixed by Honza when
doing ipa cleanups.

Maybe it would be better to simply change the kernel Makefiles to pass
-fno-ipa-pure-const instead of adding volatiles everywhere.

If you do this, please do it for 4.5.[012] only. If you disable all gcc
passes that ever had any bugs in it, you'd need to disable most of them if
not all.

Jakub

[Richard Henderson]

- if (simple_cst_equal(TREE_VALUE (op), memory_identifier_string) == 1)
+ if (strcmp (TREE_STRING_POINTER (TREE_VALUE (link)), "memory") == 0)

I prefer this solution. I think memory_identifier_string is over-engineering.
Patch to remove it entirely is pre-approved.


r~

[Jakub Jelinek]

Post by Richard Henderson

- if (simple_cst_equal(TREE_VALUE (op), memory_identifier_string) == 1)
+ if (strcmp (TREE_STRING_POINTER (TREE_VALUE (link)), "memory") == 0)

I prefer this solution. I think memory_identifier_string is over-engineering.
Patch to remove it entirely is pre-approved.

Honza even committed this to the trunk in May, it is just release branches
that are broken (and only in 4.5 it matters a lot because it happens with
the default flags).

Jakub

[Richard Guenther]

Post by Jakub Jelinek

Post by Linus Torvalds
So when Richard Gunther says "a memory clobber doesn't cover automatic
storage", to me that very clearly spells "gcc is buggy as hell".
Because automatic storage with its address taken _very_ much gets
clobbered by things like memset etc. If the compiler doesn't
understand that, the compiler is just broken.

I'll leave the discussion about meaning of "memory" clobber aside to
Richard,

Of course GCC handles memset just fine. Note that I was refering
to non-address taken automatic storage for "memory" (even though
when double-checking the current implementation GCC even thinks
that all address-taken memory is clobbered by asms as soon as
they have at least one memory operand or a "memory" clobber).

It's just that in future we might want to improve this and I think
not covering non-address taken automatic storage for "memory"
is sensible. And I see that you don't see address-taken automatic
storage as a sensible choice to exclude from "memory", and I
have noted that.

Btw, I still haven't seen an testcase for the actual problem we are
talking about.

Richard.

[Jeff Law]

Post by Richard Guenther

Post by Andi Kleen

Post by Andreas Schwab
The asm fails to mention that it modifies *regs.

It has a memory clobber, that should be enough, no?

No. A memory clobber does not cover automatic storage.

A memory clobber should clobber anything in memory, including autos in
memory; if it doesn't, then that seems like a major problem. I'd like
to see the rationale behind not clobbering autos in memory.

Jeff

[Linus Torvalds]

Post by Jeff Law
A memory clobber should clobber anything in memory, including autos in
memory; if it doesn't, then that seems like a major problem.  I'd like to
see the rationale behind not clobbering autos in memory.

Yes. It turns out that the "asm optimized away" was entirely wrong (we
never saw that, it was just a report on another mailing list).

Looking at the asm posted, it seems to me that gcc actually compiles
the asm itself 100% correctly, and the "memory" clobber is working
fine inside that function. So the code generated for i8k_smm() itself
is all good.

But _while_ generating the good code, gcc doesn't seem to realize that
it writes to anything, so it decides to mark the function
"__attribute__((const))", which is obviously wrong (a memory clobber
definitely implies that it's not const). And as a result, the callers
will be mis-optimized, because they do things like

static int i8k_get_bios_version(void)
{
struct smm_regs regs = { .eax = I8K_SMM_BIOS_VERSION, };

return i8k_smm(&regs) ? : regs.eax;
}

and since gcc has (incorrectly) decided that "i8k_smm()" is a const
function, it thinks that "regs.eax" hasn't changed, so it doesn't
bother to reload it: it "knows" that it is still I8K_SMM_BIOS_VERSION
that it initialized it with. So it will basically have rewritten that
final return statement as

return i8k_smm(&regs) ? : I8K_SMM_BIOS_VERSION;

which obviously doesn't really work.

This also explains why adding "volatile" worked. The "asm volatile"
triggered "this is not a const function".

Similarly, the "+m" works, because it also makes clear that the asm is
writing to memory, and isn't a const function.

Now, the "memory" clobber should clearly also have done that, but I'd
be willing to bet that some version of gcc (possibly extra slackware
patches) had forgotten the trivial logic to say "a memory clobber also
makes the user function non-const".

Linus

[Richard Guenther]

Post by Andi Kleen

Post by Andreas Schwab
The asm fails to mention that it modifies *regs.

It has a memory clobber, that should be enough, no?

No. =A0A memory clobber does not cover automatic storage.

A memory clobber should clobber anything in memory, including autos i=

n

memory; if it doesn't, then that seems like a major problem. =A0I'd l=

ike to

see the rationale behind not clobbering autos in memory.

Non-address taken automatic storage. (note that we don't excercise thi=
s
in optimization yet)

It's difficult to model thins kind of non-aliased memory with this kind
of aliasing mechanism (apart from taking all asms as clobbering
everything as we currently do).

Richard.

Jeff

[Jeff Law]

Post by Jeff Law

Post by Richard Guenther

Post by Andi Kleen

Post by Andreas Schwab
The asm fails to mention that it modifies *regs.

It has a memory clobber, that should be enough, no?

No. A memory clobber does not cover automatic storage.

A memory clobber should clobber anything in memory, including autos in
memory; if it doesn't, then that seems like a major problem. I'd like to
see the rationale behind not clobbering autos in memory.

Non-address taken automatic storage. (note that we don't excercise this
in optimization yet)

If the address of the auto isn't taken, then why is the object in memory
to begin with (with the obvious exception for aggregates).

Jeff

[Richard Guenther]

Post by Andi Kleen

Post by Andreas Schwab
The asm fails to mention that it modifies *regs.

It has a memory clobber, that should be enough, no?

No. =A0A memory clobber does not cover automatic storage.

A memory clobber should clobber anything in memory, including autos=

in

memory; if it doesn't, then that seems like a major problem. =A0I'd=

like to

see the rationale behind not clobbering autos in memory.

Non-address taken automatic storage. =A0(note that we don't excercis=

e this

in optimization yet)

If the address of the auto isn't taken, then why is the object in mem=

ory to

begin with (with the obvious exception for aggregates).

Exactly sort of my point. If people pass the address of &x to an asm
and modify &x + 8 expecting the "adjacent" stack location to be changed
I want to tell them that's not a supported way to get to another stack
variable (even if they clobber "memory"). Or consider the C-decl guy
who wants to access adjacent parameters by address arithmetic on
the address of the first param ...

Richard.

Jeff

[Jeff Law]

If the address of the auto isn't taken, then why is the object in memory to
begin with (with the obvious exception for aggregates).

Exactly sort of my point. If people pass the address of&x to an asm
and modify&x + 8 expecting the "adjacent" stack location to be changed
I want to tell them that's not a supported way to get to another stack
variable (even if they clobber "memory"). Or consider the C-decl guy
who wants to access adjacent parameters by address arithmetic on
the address of the first param ...

Well, in that case, I think we can easily say that the programmer has
gone off the deep end and has entered the realm of undefined behavior.

Presumably we rooted out all relevant instances of the latter over the
last 20 years... It was fairly common in the past, but I doubt anyone
worth caring about is still writing code assuming they can take the
address of parameter A, offset it and get parameters B, C, D, etc.

jeff

[Jeff Law]

Post by Andreas Schwab

Post by Andi Kleen

After upgrading my Dell laptop, both OS+kernel the i8k interface was giving
nonsensical output. As it turned out it's not the kernel but compiler
upgrade which broke this.
Guys at Archlinux have found the underlying cause (but don't seem to have
https://bbs.archlinux.org/viewtopic.php?pid=780692#p780692
gcc seems to optimize the assembly statements away.
And indeed, applying this patch makes the i8k interface work again,
i.e. replacing the asm(..) construct by asm volatile(..)

The compiler really should not optimize the asm away, because
it has both input and output arguments which are later used.
"asm volatile" normally just means "don't move significantly"

The asm fails to mention that it modifies *regs.

But there's a memory clobber, that should be sufficient to indicate
*regs is modified.

jeff