https://gcc.gnu.org/bugzilla/show_bug.cgi?id=91927
Bug ID: 91927
Summary: -mstrict-align doesn't prevent unaligned accesses at
-O2 and -O3 on AARCH64 targets
Product: gcc
Version: 8.3.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: gr.audio at gmail dot com
Target Milestone: ---
Target: aarch64-elf
Hello,
I'm writing code for an AARCH64 target, and ran into the following issue.
First off, at the stage my code runs, the MMU is not enabled, so the CPU
doesn't allow unaligned accesses. I thus enabled the "-mstrict-align" option.
The GCC (8.3.0) binaries I'm using are from ARM:
https://developer.arm.com/tools-and-software/open-source-software/developer-tools/gnu-toolchain/gnu-a/downloads
but I built GCC 9.2.0 for aarch64-elf and got the same issue.
I get an exception (unaligned access) when my code runs, and investigated. Here
is how it happens.
I'm accessing successively two successive uint32_t members of a packed struct:
eg:
---------------------------------------------------------------
typedef struct __attribute__((__packed__))
{
...
uint32_t nWidth;
uint32_t nHeight;
...
} MyStruct_t;
...
MyStruct_t sThisStruct;
...
// My code accesses them just like so:
... = sThisStruct.nWidth;
... = sThisStruct.nHeight;
---------------------------------------------------------------
At -O2 and -O3, despite the "-mstrict-align" option, GCC emits code that
creates an unaligned access:
ldr d0, [sp, 52]
(SP + 52 is 4-byte aligned obviously, but not 8-byte aligned, and thus, this
64-bit access yields an exception.)
GCC optimizes the two 32-bit accesses in one 64-bit access, but it shouldn't in
this case, with the "-mstrict-align" option.
At -O1 and below, GCC correctly emits:
ldr w0, [sp, 52]
...
ldr w0, [sp, 56]
instead, which are correctly aligned accesses.
This issue looks like a cousin to the following bug:
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71727
It looks like the "-mstrict-align" option is not correctly enforced at certain
optimization levels, for different optimizations (but with the same overall
problem of potentially unaligned accesses).
---------------------------------------------------------------
GCC version:
Target: aarch64-elf
Configured with:
/tmp/dgboter/bbs/dsggnu-vm-1-x86_64--mingw32-i686/buildbot/mingw32-i686--aarch64-elf/build/src/gcc/configure
--target=aarch64-elf
--prefix=/tmp/dgboter/bbs/dsggnu-vm-1-x86_64--mingw32-i686/buildbot/mingw32-i686--aarch64-elf/build/bui
ld-mingw-aarch64-elf/install//
--with-gmp=/tmp/dgboter/bbs/dsggnu-vm-1-x86_64--mingw32-i686/buildbot/mingw32-i686--aarch64-el
f/build/build-mingw-aarch64-elf/host-tools
--with-mpfr=/tmp/dgboter/bbs/dsggnu-vm-1-x86_64--mingw32-i686/buildbot/mingw32-i68
6--aarch64-elf/build/build-mingw-aarch64-elf/host-tools
--with-mpc=/tmp/dgboter/bbs/dsggnu-vm-1-x86_64--mingw32-i686/buildbot
/mingw32-i686--aarch64-elf/build/build-mingw-aarch64-elf/host-tools
--with-isl=/tmp/dgboter/bbs/dsggnu-vm-1-x86_64--mingw32-i
686/buildbot/mingw32-i686--aarch64-elf/build/build-mingw-aarch64-elf/host-tools
--disable-shared --disable-nls --disable-thre
ads --disable-tls --enable-checking=release --enable-languages=c,c++,fortran
--with-newlib --with-libiconv-prefix=/tmp/dgbote
r/bbs/dsggnu-vm-1-x86_64--mingw32-i686/buildbot/mingw32-i686--aarch64-elf/build/build-mingw-aarch64-elf/host-tools
--host=i68
6-w64-mingw32 --with-pkgversion='GNU Toolchain for the A-profile Architecture
8.3-2019.03 (arm-rel-8.36)' --with-bugurl=https
://bugs.linaro.org/
Thread model: single
gcc version 8.3.0 (GNU Toolchain for the A-profile Architecture 8.3-2019.03
(arm-rel-8.36))
---------------------------------------------------------------
Also happens with aarch-64-elf-gcc 9.2.0 (custom built).
