| Message ID | CAFULd4aPtMsjPz6f5QWv-RXCWVHf1MJpCA8HYXDB3ySo9DjG5g@mail.gmail.com |
|---|---|
| State | New |
| Headers | show |
On Fri, Feb 8, 2013 at 6:28 AM, Uros Bizjak <ubizjak@gmail.com> wrote: > > Attached patch fixes several go testcase failures on CentOS 5.9 with > ancient (GNU ld version 2.17.50.0.6-20.el5_8.3 20061020) binutils. > These do not have CFI personality directive, so -fsplit-stack should > not be used there. > > The compiler detects this situation through > TARGET_SUPPORTS_SPLIT_STACK target hook, but the go driver > nevertheless emits -fsplit-stack option, depending solely on the > definition of TARGET_CAN_SPLIT_STACK. Forcing -fsplit-stack from the > driver triggers the error from the above target hook. > > To fix this situation, attached patch declares TARGET_CAN_SPLIT_STACK > only in case HAVE_GAS_CFI_PERSONALITY is set (please note that it > should be *set*, not only *defined*). > > 2013-02-08 Uros Bizjak <ubizjak@gmail.com> > > * config/i386/gnu-user.h (TARGET_CAN_SPLIT_STACK): Define only > when HAVE_GAS_CFI_PERSONALITY_DIRECTIVE is set. > * config/i386/gnu-user64.h (TARGET_CAN_SPLIT_STACK): Ditto. > > Patch was bootstrapped and regression tested on x86_64-pc-linux-gnu > {,-m32} on Fedora17 (GNU ld version 2.22.52.0.1-10.fc17 20120131) and > CentOS 5.9 (GNU ld version 2.17.50.0.6-20.el5_8.3 20061020), where it > fixes many go testsuite failures. > > Ian, although this is purely x86 patch, can you please approve it from > go perspective (go is the only user of TARGET_CAN_SPLIT_STACK)? This looks fine to me. You could perhaps add a comment pointing to the corresponding test in libgcc/config.host. Thanks. Ian
Index: config/i386/gnu-user64.h =================================================================== --- config/i386/gnu-user64.h (revision 195875) +++ config/i386/gnu-user64.h (working copy) @@ -85,8 +85,10 @@ #define TARGET_THREAD_SSP_OFFSET \ (TARGET_64BIT ? (TARGET_X32 ? 0x18 : 0x28) : 0x14) +#if HAVE_GAS_CFI_PERSONALITY_DIRECTIVE +#define TARGET_CAN_SPLIT_STACK +#endif /* We steal the last transactional memory word. */ -#define TARGET_CAN_SPLIT_STACK #define TARGET_THREAD_SPLIT_STACK_OFFSET \ (TARGET_64BIT ? (TARGET_X32 ? 0x40 : 0x70) : 0x30) #endif Index: config/i386/gnu-user.h =================================================================== --- config/i386/gnu-user.h (revision 195875) +++ config/i386/gnu-user.h (working copy) @@ -154,7 +154,9 @@ /* i386 glibc provides __stack_chk_guard in %gs:0x14. */ #define TARGET_THREAD_SSP_OFFSET 0x14 +#if HAVE_GAS_CFI_PERSONALITY_DIRECTIVE +#define TARGET_CAN_SPLIT_STACK +#endif /* We steal the last transactional memory word. */ -#define TARGET_CAN_SPLIT_STACK #define TARGET_THREAD_SPLIT_STACK_OFFSET 0x30 #endif
Hello! Attached patch fixes several go testcase failures on CentOS 5.9 with ancient (GNU ld version 2.17.50.0.6-20.el5_8.3 20061020) binutils. These do not have CFI personality directive, so -fsplit-stack should not be used there. The compiler detects this situation through TARGET_SUPPORTS_SPLIT_STACK target hook, but the go driver nevertheless emits -fsplit-stack option, depending solely on the definition of TARGET_CAN_SPLIT_STACK. Forcing -fsplit-stack from the driver triggers the error from the above target hook. To fix this situation, attached patch declares TARGET_CAN_SPLIT_STACK only in case HAVE_GAS_CFI_PERSONALITY is set (please note that it should be *set*, not only *defined*). 2013-02-08 Uros Bizjak <ubizjak@gmail.com> * config/i386/gnu-user.h (TARGET_CAN_SPLIT_STACK): Define only when HAVE_GAS_CFI_PERSONALITY_DIRECTIVE is set. * config/i386/gnu-user64.h (TARGET_CAN_SPLIT_STACK): Ditto. Patch was bootstrapped and regression tested on x86_64-pc-linux-gnu {,-m32} on Fedora17 (GNU ld version 2.22.52.0.1-10.fc17 20120131) and CentOS 5.9 (GNU ld version 2.17.50.0.6-20.el5_8.3 20061020), where it fixes many go testsuite failures. Ian, although this is purely x86 patch, can you please approve it from go perspective (go is the only user of TARGET_CAN_SPLIT_STACK)? Uros.