This is due to a limitation in CodeGen. Currently, for aggregate return values,
it only supports aggregates that fit in the return value registers designated
by the ABI.

It's likely that you're working with an ABI that defines only 2 registers for
returning integer values, while the Vector type in your example requires 3,
at least according to the way aggregate return values are currently lowered.

The general fix is to add support in CodeGen for rewriting the code to return
the value in memory.

For this particular testcase on targets like x86-64, the C ABI may have
the return value packed in registers, so teaching CodeGen to follow the
ABI more closely for aggregate return values would also allow this test to
work, on those targets.

Alternatively, you can work around this in your front-end by using the
sret mechanism for returning Vector values.

Dan, do we have a bugzilla for "codegen doesn't support arbitrary aggregate result types" yet? If not, please repurpose this one, otherwise please close this as a dupe, thanks!

Re-title this bugzilla entry.

*** Bug #2691 has been marked as a duplicate of this bug. ***

*** Bug #2584 has been marked as a duplicate of this bug. ***

*** Bug #1937 has been marked as a duplicate of this bug. ***

*** Bug #2044 has been marked as a duplicate of this bug. ***

*** Bug llvm/llvm-bugzilla-archive#3318 has been marked as a duplicate of this bug. ***

The 2.4 release notes claim "full support" for first-class aggregates, so perhaps they should be amended. It's frustrating to learn about limitations this way.

It might also be helpful to mention this limitation in the section on function types in the language reference manual, which gives an example of a two-word aggregate return value.

I added a note about this to LangRef.html.
http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20090112/072053.html

*** Bug llvm/llvm-bugzilla-archive#3388 has been marked as a duplicate of this bug. ***

*** Bug llvm/llvm-bugzilla-archive#3397 has been marked as a duplicate of this bug. ***

*** Bug llvm/llvm-bugzilla-archive#3736 has been marked as a duplicate of this bug. ***

*** Bug llvm/llvm-bugzilla-archive#4110 has been marked as a duplicate of this bug. ***

*** Bug llvm/llvm-bugzilla-archive#4773 has been marked as a duplicate of this bug. ***

*** Bug llvm/llvm-bugzilla-archive#5070 has been marked as a duplicate of this bug. ***

*** Bug llvm/llvm-bugzilla-archive#6439 has been marked as a duplicate of this bug. ***

Kenneth Uildriks implemented target-independent infrastructure, and x86/x86-64 support for this.

*** Bug llvm/llvm-bugzilla-archive#6547 has been marked as a duplicate of this bug. ***

mentioned in issue llvm/llvm-bugzilla-archive#3318

mentioned in issue llvm/llvm-bugzilla-archive#3388

mentioned in issue llvm/llvm-bugzilla-archive#3397

mentioned in issue llvm/llvm-bugzilla-archive#3736

mentioned in issue llvm/llvm-bugzilla-archive#4110

mentioned in issue llvm/llvm-bugzilla-archive#4773

mentioned in issue llvm/llvm-bugzilla-archive#5070

mentioned in issue llvm/llvm-bugzilla-archive#6439

mentioned in issue llvm/llvm-bugzilla-archive#6547

This already works by implicitly converting types that don't fit in registers to sret