@rnk, I agree that it is a performance optimization, and it actually is in the hot path. I can understand the motivation for the jump. I think that the trampoline might also be a problem for lldb as it (or used to?) statically analyzes to see if it is jumping to objc_msgSend (yes, it is ugly).

I don't mind the separation, but, I don't see what makes this target specific still. I really think that we should have this be out-side of the target specific content, but having this in LLVM instead of the readobj tool makes sense. At the very least, I think that we need to fix the big endian case.

SVN r230867.

The header doesn't have anything that is version specific, so I believe that this is safe for all the versions that we support.

Address rnk's comments.

Catch the constant folding in Sema (thanks for pointing out that was possible rnk!)

Emit inserts exactly a single instruction. It provides no guarantees of the state, so I don't think that we need to worry about LLVM doing something like inserting spills.

Address review comments from majnemer

Seems that I can't close the revision :-(.

Committed in SVN r221548 with the recommended change.

Address Jim's comments about reducing duplication and some error handling.

Seems that an alternative implementation has been committed.

Address a number of Jim's comments.

Given that the move will simplify the diff, I've addressed the inline comments and move into Utils from its own directory. I'll see about restructuring the descriptors to private from public. If clang ever needs to generate the descriptors, we can make them public then. Better to make as many of the types opaque as possible.

SVN r211716.

ping

Fix Intrin.h with minor oversights caught by Warren.

Address comments from rnk.

@rnk, Im a bit confused here. Why do I need the I constraint for __emit? Isnt I an immediate integer valid for a data operation? The __emit takes an unsigned long and emits that long constant as an instruction.

ping

SVN r211032 with a small fixup at SVN r211035.

Use __ARM_ARCH_ISA_THUMB introduced in SVN r210991.

It would be nice to have similar tests for the Itanium ABI, but I suspect that you are less interested in that.

Make getLinkerFileListName more similar to getDependencyFileName. Add an assertion. Tweak some variable names as per review comments.

As Jim mentioned at the beginning, that there is some overhead to this, and enabling this unconditionally seems a bit heavy handed. By improving the driver to be a bit more similar to gcc (adding support for -mlong-calls/-mno-long-calls), we can enable this as and when it is needed as the linker should indicate that the jump target is out of range.

This has been committed already, just closing off the revision.

I believe that should be correctly handled by the bundling approach (SVN r207608). It can always be resurrected if necessary.

I think everyone agrees that this is not a particularly pretty solution. Given that any case of the bundle being split up should be considered a bug, I think that there is no reason to hold up other changes on this. This addresses all but one case: the constant island pass. I've committed a change with the bundling approach with a bit more aggressive reorder prevention. The constant island handling will have to be addressed separately and we can discuss how to best handle that separately.

Committed as SVN r207246.

Committed as SVN r207243.

Committed with the suggested changes as SVN r207242.

Update to use new @llvm.arm.hint intrinsic as per the suggestion of Tim.

Hmm, seems that phabricator truncated your message Tim.