Description of problem:
Starting with Rust 1.43, we found that programs configured to build with full LTO were crashing rustc on s390x with a segmentation fault. It also happens with upstream Rust binaries. This is locally reproducible with a stage2 rustc, but stage1 (bootstrapped from 1.42) is fine.

Version-Release number of selected component (if applicable):
rust-1.43.0-1.fc31
llvm-9.0.1-5.fc31

How reproducible:
100%

Steps to Reproduce:
1. Rebuild an LTO rust package, like rust-afterburn.

Actual results:
SIGSEGV

Expected results:
Successful compilation.

Additional info:
The code which processes FatLTO modules was seeing a 0x1 data pointer for the name String. Ulrich Weigand (IBM) found a miscompilation in "Drop for Bomb" in src/librustc_codegen_ssa/back/write.rs, where it copied the wrong range into the newly Boxed value.

I bisected the problem in Rust to commit e82ec2315e5a, but the effect of that only simplified the LLVM IR. This did lead SROA to break up those memcpys more than they were before. Since F32 with LLVM 10 is working, I compared "opt" from 9 and 10 output and found a difference in shufflevector arguments. Bisecting LLVM led me to commit a9d6b0e5, "[InstCombine] Fix big-endian miscompile of (bitcast (zext/trunc (bitcast)))".

https://github.com/rust-lang/rust/commit/e82ec2315e5adb1c291c3702cd2ac1f46ecd0fcf
https://github.com/llvm/llvm-project/commit/a9d6b0e5444741d08ff1df7cf71d1559e7fefc1f

I have confirmed that backporting that change to LLVM 9 fixes the build for rust-afterburn, and the asm of that rustc code looks correct to me.