Disable big-endian simd in swap_nonoverlapping_bytes #43159
Conversation
This is a workaround for rust-lang#42778, which was git-bisected to rust-lang#40454's optimizations to `mem::swap`, later moved to `ptr` in rust-lang#42819. Natively compiled rustc couldn't even compile stage1 libcore on powerpc64 and s390x, but they work fine without this `repr(simd)`. Since powerpc64le works OK, it seems probably related to being big-endian. The underlying problem is not yet known, but this at least makes those architectures functional again in the meantime. cc @arielb1
|
Nice find! I'm not sure why Maybe I'll try to look at the LLVM IR differences - I don't think there should be any. |
|
I'll give this an r+ for now to get things working. I am not sure on that swap PR on non-x86 architectures anyway. @bors r+ |
|
📌 Commit 77bd4dc has been approved by |
|
BTW, the "broken" swap seemed to work on a ppc64 qemu (on which building the compiler failed). It would be really nice if we could find out the brokenness. |
arielb1
added
the
S-waiting-on-bors
|
Yeah, I'm not satisfied with this long-term -- just wanted to make sure to get something in place before the next beta branch. I've tried guessing a few minimal scenarios and not seen any difference in codegen, so we'll have to actually track down which parts are different in the full build. |
|
⌛ Testing commit 77bd4dc with merge b9cde66ebe43aa342de8979dbc3e800e6ba3b75c... |
|
💔 Test failed - status-travis |
|
@bors retry -- android spurious failure, hopefully fixed by clearing caches (I think this started too late, possibly) |
|
@bors rollup |
|
Restarted the tests - looks like we ran out of disk space. |
Disable big-endian simd in swap_nonoverlapping_bytes This is a workaround for rust-lang#42778, which was git-bisected to rust-lang#40454's optimizations to `mem::swap`, later moved to `ptr` in rust-lang#42819. Natively compiled rustc couldn't even compile stage1 libcore on powerpc64 and s390x, but they work fine without this `repr(simd)`. Since powerpc64le works OK, it seems probably related to being big-endian. The underlying problem is not yet known, but this at least makes those architectures functional again in the meantime. cc @arielb1
Revert "Disable big-endian simd in swap_nonoverlapping_bytes" This reverts commit 77bd4dc (#43159). Issue #42778 was formerly easy to reproduce on two big-endian targets, `powerpc64` and `s390x`, so we disabled SIMD on this function for all big-endian targets as a workaround. I have re-tested this code on `powerpc64` and `s390x`, each with the bundled LLVM 8 and with external LLVM 7 and LLVM 6, and the problems no longer appear. So it seems safe to remove this workaround, although I'm still a little uncomfortable that we never found a root-cause... Closes #42778. r? @arielb1
This is a workaround for #42778, which was git-bisected to #40454's
optimizations to
mem::swap, later moved toptrin #42819. Nativelycompiled rustc couldn't even compile stage1 libcore on powerpc64 and
s390x, but they work fine without this
repr(simd). Since powerpc64leworks OK, it seems probably related to being big-endian.
The underlying problem is not yet known, but this at least makes those
architectures functional again in the meantime.
cc @arielb1