On 8/3/2026 10:24 AM, Matt Turner wrote:
On Mon, Jul 20, 2026 at 12:26 PM Matt Turner <[email protected]> wrote:
Alpha is one of the last targets still forcing TARGET_LRA_P to false, and
reload is slated for removal (PR113932). Add a transitional -mlra option,
following the hppa pattern (r16-5534), so that -mno-lra remains available as
an escape hatch until reload is deleted.
LRA is enabled by default only on BWX targets. On non-BWX targets a QImode or
HImode pseudo that has to be spilled reaches the target's reload_in<mode> /
reload_out<mode> secondary-reload path, which LRA does not yet drive correctly
(PR117185: LRA either loops generating reload insns or passes a pseudo where
get_unaligned_address expects a MEM). BWX has real byte/word memory
instructions and does not need that path, so it is unaffected. Gating the
default on TARGET_BWX enables LRA where it is known good and leaves non-BWX on
reload until PR117185 is resolved. -mlra / -mno-lra override the default in
either direction.
The default is resolved in alpha_option_override once TARGET_BWX is final, so
it tracks -mcpu (e.g. -mcpu=ev4 selects reload, -mcpu=ev56 selects LRA) rather
than only the configured default CPU.
Tested with an alpha-linux-gnu cross (both ev4/non-BWX and --with-cpu=ev56/BWX,
--enable-checking=yes,rtl): libgcc, libgomp and libstdc++ build cleanly under
LRA, and gcc.c-torture/execute under qemu-alpha at -O0 and -O2 gives identical
results with -mlra and -mno-lra on both configurations (3380 results each, no
differences). A native ev56 bootstrap with c,c++,m2 completes and, comparing a
full make check-gcc/check-g++ at the default (LRA) against --target_board
unix/-mno-lra (reload), LRA shows no correctness regressions: gcc unexpected
failures drop from 626 to 541 and g++ from 57 to 56, with the only LRA-only
differences being debug-info-quality (gcc.dg/guality) shifts in LRA's favor.
The prerequisite middle-end fix for the BWX m2 miscompile is PR117184.
gcc/
* config/alpha/alpha.opt (mlra): New option.
* config/alpha/alpha.cc (alpha_option_override): Default alpha_lra_p
to TARGET_BWX when not set explicitly.
(alpha_use_lra_p): New function.
(TARGET_LRA_P): Define to it instead of hook_bool_void_false.
---
Hi Vlad and Peter,
I could really use your review on this patch as well as your feedback
on the questions in
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=117185#c8
So at a high level, do we want to have LRA depend on subtarget
capabilities. I don't think we've ever done that, though I can
understand the motivation here. We don't have an active alpha
maintainer, so I'm happy to go when general consensus here.
I haven't gotten a clean bootstrap & regression on alpha in about a
month and my tester only runs in the default mode, so I don't think it's
really going to exercise this meaningfully or give us any sense of LRA
on alpha state going forward. Given it's a dead architecture one
possibility that we should at least consider is deprecating the non-BWX
variants at which point we could convert to LRA and having
bootstrap/regression testing in place.
For reference Matt, I have an alpha native root filesystem with enough
components to bootstrap GCC. My tester chroots into that RFS using user
mode QEMU emulation. So it looks native, but doesn't have as much
overhead as full system emulation. A bootstrap and regression test
takes 16-19 hours on my builders.
Jeff
