Poor i128 shift lowering

[alion02]

https://godbolt.org/z/dYxxKEaqG

g and h lowering should be the same (alive2). Reproduces with all shift variants and on both x86-64 and aarch64.

[alion02]

For context, in case it changes anything, in my real code the range information comes from a sub nuw nsw i32 6, %0. Notably, LLVM just deletes my attempted and 63 workaround.

[llvmbot]

@llvm/issue-subscribers-backend-aarch64

Author: None (alion02)

https://godbolt.org/z/dYxxKEaqG

g and h lowering should be the same (alive2). Reproduces with all shift variants and on both x86-64 and aarch64.

[llvmbot]

@llvm/issue-subscribers-backend-x86

Author: None (alion02)

https://godbolt.org/z/dYxxKEaqG

g and h lowering should be the same (alive2). Reproduces with all shift variants and on both x86-64 and aarch64.

[dtcxzyw]

@alion02 With #156710, the range information comes from sub nuw nsw i32 6, %0 will be preserved during optimization. But @h is still suboptimal as we don't preserve llvm.assume in the backend.