llvm-project

d2d620d4 - [MC][ARM] Don't set funcs to Thumb as a side effect of .hidden (#181156)

[MC][ARM] Don't set funcs to Thumb as a side effect of .hidden (#181156) When assembling a source file which switches between Arm and Thumb state using `.arm` and `.thumb`, if you defined a function in Arm state and mark it as hidden at dynamic link time using `.hidden`, but don't actually issue the `.hidden` directive until you have switched back to Thumb state, then the function would be accidentally marked as Thumb as a side effect of making it hidden. This happened in `ARMELFStreamer::emitSymbolAttribute`, and the comment suggests that it was semi-deliberate: it was intended to happen as a side effect of `.type foo,%function`, because the function label might have already been defined without a type, and shouldn't be marked as Thumb until it's known that it's a function. But I think it was an accident that the same behavior also applies to any other addition of a symbol attribute, such as `.hidden`: the call to `setIsThumbFunc` was conditioned on whether the symbol has function type after setting the attribute, not whether function type was the attribute _actually being set_. So if you set the symbol to function type and _then_ use `.hidden`, the condition would match again. Fixes #180358, in which this came up in real-world code: rustc's IR output included a top-level function in the form of inline asm, defined in Arm state, and marked it as hidden via the LLVM IR `declare`, so that LLVM didn't emit the `.hidden` directive until after it had switched back to Thumb state. The new test case includes a check for the _intended_ behavior of the code in `emitSymbolAttribute`, to confirm that that still works.