On Tue, May 5, 2020 at 2:55 PM Jason A. Donenfeld <Jason@xxxxxxxxx> wrote:
>
> clang-10 has a broken optimization stage that doesn't enable the
> compiler to prove at compile time that certain memcpys are within
> bounds, and thus the outline memcpy is always called, resulting in
> horrific performance, and in some cases, excessive stack frame growth.
> Here's a simple reproducer:
>
> typedef unsigned long size_t;
> void *c(void *dest, const void *src, size_t n) __asm__("memcpy");
> extern inline __attribute__((gnu_inline)) void *memcpy(void *dest, const void *src, size_t n) { return c(dest, src, n); }
> void blah(char *a)
> {
> unsigned long long b[10], c[10];
> int i;
>
> memcpy(b, a, sizeof(b));
> for (i = 0; i < 10; ++i)
> c[i] = b[i] ^ b[9 - i];
> for (i = 0; i < 10; ++i)
> b[i] = c[i] ^ a[i];
> memcpy(a, b, sizeof(b));
> }
>
> Compile this with clang-9 and clang-10 and observe:
>
> zx2c4@thinkpad /tmp/curve25519-hacl64-stack-frame-size-test $ clang-10 -Wframe-larger-than=0 -O3 -c b.c -o c10.o
> b.c:5:6: warning: stack frame size of 104 bytes in function 'blah' [-Wframe-larger-than=]
> void blah(char *a)
> ^
> 1 warning generated.
> zx2c4@thinkpad /tmp/curve25519-hacl64-stack-frame-size-test $ clang-9 -Wframe-larger-than=0 -O3 -c b.c -o c9.o
>
> Looking at the disassembly of c10.o and c9.o, one can see that c9.o is
> properly optimized in the obvious way one would expect, while c10.o has
> blown up and includes extern calls to memcpy.
>
> This is present on the most trivial bits of code. Thus, for clang-10, we
> just set __NO_FORTIFY globally, so that this issue won't be incurred.
>
> Cc: Arnd Bergmann <arnd@xxxxxxxx>
> Cc: LKML <linux-kernel@xxxxxxxxxxxxxxx>
> Cc: clang-built-linux <clang-built-linux@xxxxxxxxxxxxxxxx>
> Cc: Kees Cook <keescook@xxxxxxxxxxxx>
> Cc: George Burgess <gbiv@xxxxxxxxxx>
> Cc: Nick Desaulniers <ndesaulniers@xxxxxxxxxx>
> Link: https://bugs.llvm.org/show_bug.cgi?id=45802
> Signed-off-by: Jason A. Donenfeld <Jason@xxxxxxxxx>
I'm going to request this not be merged until careful comment from
George and Kees. My hands are quite full at the moment with other
regressions. I suspect these threads may be relevant, though I
haven't had time to read through them myself.
https://github.com/ClangBuiltLinux/linux/issues/979
https://github.com/ClangBuiltLinux/linux/pull/980
> ---
> Makefile | 7 +++++++
> 1 file changed, 7 insertions(+)
>
> diff --git a/Makefile b/Makefile
> index 49b2709ff44e..f022f077591d 100644
> --- a/Makefile
> +++ b/Makefile
> @@ -768,6 +768,13 @@ KBUILD_CFLAGS += -Wno-gnu
> # source of a reference will be _MergedGlobals and not on of the whitelisted names.
> # See modpost pattern 2
> KBUILD_CFLAGS += -mno-global-merge
> +
> +# clang-10 has a broken optimization stage that causes memcpy to always be
> +# outline, resulting in excessive stack frame growth and poor performance.
> +ifeq ($(shell test $(CONFIG_CLANG_VERSION) -ge 100000 && test $(CONFIG_CLANG_VERSION) -lt 110000; echo $$?),0)
> +KBUILD_CFLAGS += -D__NO_FORTIFY
> +endif
> +
> else
>
> # These warnings generated too much noise in a regular build.
> --
> 2.26.2
>
--
Thanks,
~Nick Desaulniers
