clang-15's ability to elide loops completely became more aggressive when
it can deduce how a variable is being updated in a loop. Counting down
one variable by an increment of another can be replaced by a modulo
operation.

For 64b variables on 32b ARM EABI targets, this can result in the
compiler generating calls to __aeabi_uldivmod, which it does for a do
while loop in float64_rem().

For the kernel, we'd generally prefer that developers not open code 64b
division via binary / operators and instead use the more explicit
helpers from div64.h. On arm-linux-gnuabi targets, failure to do so can
result in linkage failures due to undefined references to
__aeabi_uldivmod().

While developers can avoid open coding divisions on 64b variables, the
compiler doesn't know that the Linux kernel has a partial implementation
of a compiler runtime (--rtlib) to enforce this convention.

It's also undecidable for the compiler whether the code in question
would be faster to execute the loop vs elide it and do the 64b division.

While I actively avoid using the internal -mllvm command line flags, I
think we get better code than using barrier() here, which will force
reloads+spills in the loop for all toolchains.

Link: https://github.com/ClangBuiltLinux/linux/issues/1666
Reported-by: Nathan Chancellor <nathan@kernel.org>
Signed-off-by: Nick Desaulniers <ndesaulniers@google.com>
---
 arch/arm/nwfpe/Makefile | 6 ++++++
 1 file changed, 6 insertions(+)

diff --git a/arch/arm/nwfpe/Makefile b/arch/arm/nwfpe/Makefile
index 303400fa2cdf..2aec85ab1e8b 100644
--- a/arch/arm/nwfpe/Makefile
+++ b/arch/arm/nwfpe/Makefile
@@ -11,3 +11,9 @@ nwfpe-y				+= fpa11.o fpa11_cpdo.o fpa11_cpdt.o \
 				   entry.o
 
 nwfpe-$(CONFIG_FPE_NWFPE_XP)	+= extended_cpdo.o
+
+# Try really hard to avoid generating calls to __aeabi_uldivmod() from
+# float64_rem() due to loop elision.
+ifdef CONFIG_CC_IS_CLANG
+CFLAGS_softfloat.o	+= -mllvm -replexitval=never
+endif
-- 
2.38.0.rc2.412.g84df46c1b4-goog

On Mon, Oct 10, 2022 at 03:53:42PM -0700, Nick Desaulniers wrote:
> clang-15's ability to elide loops completely became more aggressive when
> it can deduce how a variable is being updated in a loop. Counting down
> one variable by an increment of another can be replaced by a modulo
> operation.
> 
> For 64b variables on 32b ARM EABI targets, this can result in the
> compiler generating calls to __aeabi_uldivmod, which it does for a do
> while loop in float64_rem().
> 
> For the kernel, we'd generally prefer that developers not open code 64b
> division via binary / operators and instead use the more explicit
> helpers from div64.h. On arm-linux-gnuabi targets, failure to do so can
> result in linkage failures due to undefined references to
> __aeabi_uldivmod().
> 
> While developers can avoid open coding divisions on 64b variables, the
> compiler doesn't know that the Linux kernel has a partial implementation
> of a compiler runtime (--rtlib) to enforce this convention.
> 
> It's also undecidable for the compiler whether the code in question
> would be faster to execute the loop vs elide it and do the 64b division.
> 
> While I actively avoid using the internal -mllvm command line flags, I
> think we get better code than using barrier() here, which will force
> reloads+spills in the loop for all toolchains.
> 
> Link: https://github.com/ClangBuiltLinux/linux/issues/1666
> Reported-by: Nathan Chancellor <nathan@kernel.org>
> Signed-off-by: Nick Desaulniers <ndesaulniers@google.com>

I built ARCH=arm allmodconfig + CONFIG_WERROR=n without this patch and
saw the link failure then applied it and the error went away. Thanks for
all the investigation done into fixing this! I think you put this in the
patch tracker already but just for posterity:

Tested-by: Nathan Chancellor <nathan@kernel.org>

> ---
>  arch/arm/nwfpe/Makefile | 6 ++++++
>  1 file changed, 6 insertions(+)
> 
> diff --git a/arch/arm/nwfpe/Makefile b/arch/arm/nwfpe/Makefile
> index 303400fa2cdf..2aec85ab1e8b 100644
> --- a/arch/arm/nwfpe/Makefile
> +++ b/arch/arm/nwfpe/Makefile
> @@ -11,3 +11,9 @@ nwfpe-y				+= fpa11.o fpa11_cpdo.o fpa11_cpdt.o \
>  				   entry.o
>  
>  nwfpe-$(CONFIG_FPE_NWFPE_XP)	+= extended_cpdo.o
> +
> +# Try really hard to avoid generating calls to __aeabi_uldivmod() from
> +# float64_rem() due to loop elision.
> +ifdef CONFIG_CC_IS_CLANG
> +CFLAGS_softfloat.o	+= -mllvm -replexitval=never
> +endif
> -- 
> 2.38.0.rc2.412.g84df46c1b4-goog
>

On Tue, Oct 11, 2022, at 12:53 AM, Nick Desaulniers wrote:
> clang-15's ability to elide loops completely became more aggressive when
> it can deduce how a variable is being updated in a loop. Counting down
> one variable by an increment of another can be replaced by a modulo
> operation.
>
> For 64b variables on 32b ARM EABI targets, this can result in the
> compiler generating calls to __aeabi_uldivmod, which it does for a do
> while loop in float64_rem().
>
> For the kernel, we'd generally prefer that developers not open code 64b
> division via binary / operators and instead use the more explicit
> helpers from div64.h. On arm-linux-gnuabi targets, failure to do so can
> result in linkage failures due to undefined references to
> __aeabi_uldivmod().
>
> While developers can avoid open coding divisions on 64b variables, the
> compiler doesn't know that the Linux kernel has a partial implementation
> of a compiler runtime (--rtlib) to enforce this convention.
>
> It's also undecidable for the compiler whether the code in question
> would be faster to execute the loop vs elide it and do the 64b division.
>
> While I actively avoid using the internal -mllvm command line flags, I
> think we get better code than using barrier() here, which will force
> reloads+spills in the loop for all toolchains.
>
> Link: https://github.com/ClangBuiltLinux/linux/issues/1666
> Reported-by: Nathan Chancellor <nathan@kernel.org>
> Signed-off-by: Nick Desaulniers <ndesaulniers@google.com>

Works for me,

Reviewed-by: Arnd Bergmann <arnd@arndb.de>

I would have been fine with disallowing NWFPE for clang, or with
adding a barrier in the loop as well, i.e. any approach that
doesn't cause invalid behavior or a maintenance burden, given that
there is probably nobody that actually needs nwfpe on a clang built
kernel.

    Arnd