[Bug middle-end/126701] [16/17 Regression] ICE at -O2 on aarch64 during RTL pass: expand

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=126701

--- Comment #5 from ktkachov at gcc dot gnu.org ---
(In reply to Jakub Jelinek from comment #4)
> To answer myself, I can reproduce the ICE this testcase on 16 branch with a
> x86_64-linux to aarch64-linux cross (just removed stdint.h include and
> replaced uint64_t with unsigned long and uint8_t with unsigned char):
> ./cc1 -quiet -nostdinc -O2 pr126701.c -march=armv9-a+sve -fchecking=1
> during RTL pass: expand
> pr126701.c: In function ‘f30’:
> pr126701.c:8:6: internal compiler error: in verify_partition_mem_exprs, at
> cfgexpand.cc:2484
>     8 | void f30(unsigned long a0, _Bool a1, v16u64 a2)
>       |      ^~~
> 0x1bcfc0d internal_error(char const*, ...)
> 	../../gcc/diagnostic-global-context.cc:787
> 0x4429d9 fancy_abort(char const*, int, char const*)
> 	../../gcc/diagnostics/context.cc:1813
> 0x5d28bb verify_partition_mem_exprs
> 	../../gcc/cfgexpand.cc:2484
> 0x5e0c14 execute
> 	../../gcc/cfgexpand.cc:7199
> and I can reproduce it also if I revert on the branch the r16-9426 and
> r16-9425 commits (but keep the r16-9427).
> So, the sharing problem was there already before (or the verification commit
> has issues).  Thus I'd say the right fix for 16.2 is to revert r16-9427.

Ok, I'm fine with reverting that commit on the branch, but it's intended to
catch real silent wrong-code that pair-fusion introduces if the checking code
does not catch it first.