[ARM] Default behavior is different about mtp with gcc

[Zhenhang1213]

If using ARMv7-a and later architectures， gcc use cp15 and llvm uses soft, Can I modify this default behavior? I'm sure that helps with performance

https://github.com/gcc-mirror/gcc/blob/e9bd9d42dddd7a9b1c080426b0b16d3704673cbf/gcc/config/arm/arm.cc#L3968-L3974

demo:https://godbolt.org/z/f48hdenj8

[llvmbot]

@llvm/issue-subscribers-clang-driver

Author: Austin (Zhenhang1213)

If using ARMv7-a and later architectures， gcc use cp15 and llvm uses soft, Can I modify this default behavior? I'm sure that helps with performance

https://github.com/gcc-mirror/gcc/blob/e9bd9d42dddd7a9b1c080426b0b16d3704673cbf/gcc/config/arm/arm.cc#L3968-L3974

demo:https://godbolt.org/z/6cP16naW8

[llvmbot]

@llvm/issue-subscribers-backend-arm

Author: Austin (Zhenhang1213)

If using ARMv7-a and later architectures， gcc use cp15 and llvm uses soft, Can I modify this default behavior? I'm sure that helps with performance

https://github.com/gcc-mirror/gcc/blob/e9bd9d42dddd7a9b1c080426b0b16d3704673cbf/gcc/config/arm/arm.cc#L3968-L3974

demo:https://godbolt.org/z/6cP16naW8

[dongjianqiang2]

Yes，I think we should follow the gcc's behavior.

[Zhenhang1213]

@davemgreen Do you have time to look at it? I look forward to your reply.

[davemgreen]

@smithp35 Any thoughts/opinions on changing this?

[smithp35]

I think this is a duplicate of https://bugs.llvm.org/show_bug.cgi?id=38394 probably a historic issue in there.

I think the only way we could change this would be to implement gcc's -mtp=auto and make that the default.

It would be a breaking change for anyone building a platform on Armv7-A/Armv7-R clang only, with their own implementation of __aeabi_read_tp that doesn't use cp15 to store the thread pointer. Obviously they could add -mtp=soft to their build options for future versions.

I can't find any known v7-A open-source platform that doesn't use cp15 in their implementation of __aeabi_read_tp so I suspect that the number of people suffering a breakage would be small to non-existent.

In summary.

• Someone would need to make an implementation of -mtp=auto, with that becoming the new default.
• The patch would need to update the release notes as this could be a breaking change.

[Zhenhang1213]

I think this is a duplicate of https://bugs.llvm.org/show_bug.cgi?id=38394 probably a historic issue in there.

I think the only way we could change this would be to implement gcc's -mtp=auto and make that the default.

It would be a breaking change for anyone building a platform on Armv7-A/Armv7-R clang only, with their own implementation of __aeabi_read_tp that doesn't use cp15 to store the thread pointer. Obviously they could add -mtp=soft to their build options for future versions.

I can't find any known v7-A open-source platform that doesn't use cp15 in their implementation of __aeabi_read_tp so I suspect that the number of people suffering a breakage would be small to non-existent.

In summary.

• Someone would need to make an implementation of -mtp=auto, with that becoming the new default.
• The patch would need to update the release notes as this could be a breaking change.

thx, do all ARMv7 chips support CP15? I seem to have not found the relevant manual, which relates to code implementation？