IR parser accepts missing comma before align in global variable declaration and emits as attribute

[arsenm]

; RUN: llvm-as < %s | llvm-dis | FileCheck %s

; CHECK: @code_model_tiny_drop = global i32 0, code_model "tiny" #0
; CHECK: attributes #0 = { align=4 }

@missing_comma_after_code_model = global i32 0, code_model "tiny" align 4

There is a missing comma after the code_model marker, before align. This should probably be rejected. I'm not sure how this ends up getting interpreted, but when printed it looks like it was mishandled. The special align directive on the global declaration is missing, and replaced with an attribute group as if it were treated as an ordinary attribute on a function

[heiher]

I found this commit (f3d7904), and it intentionally allows global variable to have attributes, which doesn't seem to be a bug.

[arsenm]

I found this commit (f3d7904), and it intentionally allows global variable to have attributes, which doesn't seem to be a bug.

Global variables have string attributes, and no first class attributes apply to them. For globals, align is a first class feature and not an attribute. The verifier should fail if any non-string attribute applies

[heiher]

Refer to this test case:

https://github.com/llvm/llvm-project/blob/bcf0f8d8aa27910545762e2cc7733a11a00954d0/llvm/test/Bitcode/globalvariable-attributes.ll

@g4 = global i32 2, align 4 "key5" = "value5" #0

attributes #0 = { "string" = "value" nobuiltin norecurse }

For norecurse, Is the following incorrect?

@g4 = global i32 2 norecurse

[arsenm]

I don't believe any of the proper attributes apply, and most are intended only for functions or parameters / return values. norecurse does not make sense for a global.