https://gcc.gnu.org/bugzilla/show_bug.cgi?id=117955
Jeffrey A. Law <law at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Last reconfirmed| |2025-02-04
Summary|[15 Regression] GCC |GCC generate illegal riscv
|generate illegal riscv |instruction `vsetvli
|instruction `vsetvli |zero,zero,e64,mf4,ta,ma`
|zero,zero,e64,mf4,ta,ma` |with -O2 and -O3
|with -O2 and -O3 |
Ever confirmed|0 |1
Status|UNCONFIRMED |NEW
--- Comment #5 from Jeffrey A. Law <law at gcc dot gnu.org> ---
Bisection to the commit that "fixed" this bug points to:
commit b82a5810e7bcc82b933e16f9067879b9d22b85c7
Author: Robin Dapp <rd...@ventanamicro.com>
Date: Thu Nov 21 14:49:53 2024 +0100
RISC-V: Ensure vtype for full-register moves [PR117544].
As discussed in PR117544 the VTYPE register is not preserved across
function calls. Even though vmv1r-like instructions operate
independently of the actual vtype they still require a valid vtype. As
we cannot guarantee that the vtype is valid we must make sure to emit a
vsetvl between a function call and a vmv1r.v.
This patch makes the necessary changes by splitting the full-reg-move
insns into patterns that use the vtype register and adding vmov to the
types of instructions requiring a vset.
PR target/117544
gcc/ChangeLog:
* config/riscv/vector.md (*mov<mode>_whole): Split.
(*mov<mode>_fract): Ditto.
(*mov<mode>): Ditto.
(*mov<mode>_vls): Ditto.
(*mov<mode>_reg_whole_vtype): New pattern with vtype use.
(*mov<mode>_fract_vtype): Ditto.
(*mov<mode>_vtype): Ditto.
(*mov<mode>_vls_vtype): Ditto.
gcc/testsuite/ChangeLog:
* gcc.target/riscv/rvv/base/abi-call-args-4.c: Expect vsetvl.
* gcc.target/riscv/rvv/base/pr117544.c: New test.
But I strongly suspect that just made whatever the underlying issue is go
latent. Conceptually that just makes us do vlen/vtype analysis on additional
instructions which could certainly perturb vsetvl generation.
I'm removing the regression marker since it's not triggering on the trunk.
However, I'm leaving the bug open as I have no confidence it's actually fixed.
