[Bug target/126426] New: riscv rvv: Incorrect result with -march=rva23u64 -O2 in libgav1

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=126426

            Bug ID: 126426
           Summary: riscv rvv: Incorrect result with -march=rva23u64 -O2
                    in libgav1
           Product: gcc
           Version: 17.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: target
          Assignee: unassigned at gcc dot gnu.org
          Reporter: joel at jms dot id.au
                CC: bergner at gcc dot gnu.org, law at gcc dot gnu.org,
                    rdapp at gcc dot gnu.org
  Target Milestone: ---
            Target: riscv64-*-linux-gnu

Created attachment 65137
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=65137&action=edit
reduced test case

I was rebuilding some packages with Debian's GCC 16.1 and -march=rva23u64 -O2
and found a test suite failure in libgav1:

BitstreamUtils.GetMvProjection

  projection_mv.mv32:      0           spec_projection_mv.mv32: 1073676288 
(0x3FFF0000)
  projection_mv.mv32:      0           spec_projection_mv.mv32: 3221291008 
(0xC0010000)
  projection_mv.mv32:      4249419304  spec_projection_mv.mv32: 4294966824 
(0xFFFFFE28)

Attached is a reduce test case derived from libgav1 (manual hacking + cvise).

To reproduce build with vector and -O2 (the static is there to simplify running
under qemu-user). -march=rva32u64 and -march=rv64gcv both are sufficient.


 $ riscv64-unknown-linux-gnu-gcc -O2 -march=rv64gcv -static repro.c
 $ qemu-riscv64 -cpu rv64,v=true ./a.out ; echo $?
 BAD: fffd0000
 255


Some experimentation showed it goes away with -fno-tree-loop-vectorize, or just
-fno-ext-dce.


 $ riscv64-unknown-linux-gnu-gcc -O2 -march=rv64gcv -static -fno-ext-dce
repro.c
 $ qemu-riscv64 -cpu rv64,v=true ./a.out ; echo $?
 OK
 0


Initially discovered with Debian 16.1.0-2, reproduces with gcc 16 6afcc4f6d and
master f6b00aefc25a (from 2026-07-25).