tcg/optimize: Fix s_mask for rotate by constant

[PATCH 0/3] tcg/optimize: Fix s_mask for rotate by constant

Posted by Dominic P 3 weeks, 2 days ago

fold_shift rotates s_mask for rotl/rotr by a constant.  For I32 the
INT32_MIN OR in fold_masks_zosa_int can turn the result into a bogus
sign mask, and the next sign extension is dropped (ARM sxtb; ror #1;
sxtb, i386 movsbl; roll $31; movsbl).  Patch 1 handles rotates like
5fe51606146 handled the logical right shift, patch 2 adds tests,
patch 3 builds the test-i386-opt-shr test that the meson conversion
dropped.

Found by differential testing of GCC output under qemu-arm.

Tested: make check-tcg for arm, aarch64, i386 and x86_64 linux-user
on 9c88811b81, 312/312 pass.  The two new tests fail without patch 1.

AI disclosure: these patches were prepared with an AI assistant
(Claude, Anthropic); each commit carries an Assisted-by: trailer.  I
know docs/devel/code-provenance.rst declines such contributions and
that checkpatch.pl rejects the trailer.  I am sending them openly and
leave acceptance to the maintainers.  If they cannot be taken, please
treat this as a bug report; the reproducers are in patch 1.

Dominic P (3):
  tcg/optimize: Do not derive s_mask from rotated s_mask
  tests/tcg: Add tests for sign extension after rotate
  tests/tcg/i386: Build test-i386-opt-shr

 tcg/optimize.c                     |  7 +++++--
 tests/tcg/arm/meson.build          |  1 +
 tests/tcg/arm/test-arm-opt-ror.c   | 21 +++++++++++++++++++++
 tests/tcg/i386/meson.build         |  2 ++
 tests/tcg/i386/test-i386-opt-rol.c | 21 +++++++++++++++++++++
 5 files changed, 50 insertions(+), 2 deletions(-)
 create mode 100644 tests/tcg/arm/test-arm-opt-ror.c
 create mode 100644 tests/tcg/i386/test-i386-opt-rol.c

-- 
2.55.0

Re: [PATCH 0/3] tcg/optimize: Fix s_mask for rotate by constant

Posted by Richard Henderson 1 week, 6 days ago

On 9/30/26 20:51, Dominic P wrote:
> fold_shift rotates s_mask for rotl/rotr by a constant.  For I32 the
> INT32_MIN OR in fold_masks_zosa_int can turn the result into a bogus
> sign mask, and the next sign extension is dropped (ARM sxtb; ror #1;
> sxtb, i386 movsbl; roll $31; movsbl).  Patch 1 handles rotates like
> 5fe51606146 handled the logical right shift, patch 2 adds tests,
> patch 3 builds the test-i386-opt-shr test that the meson conversion
> dropped.
> 
> Found by differential testing of GCC output under qemu-arm.
> 
> Tested: make check-tcg for arm, aarch64, i386 and x86_64 linux-user
> on 9c88811b81, 312/312 pass.  The two new tests fail without patch 1.
> 
> AI disclosure: these patches were prepared with an AI assistant
> (Claude, Anthropic); each commit carries an Assisted-by: trailer.  I
> know docs/devel/code-provenance.rst declines such contributions and
> that checkpatch.pl rejects the trailer.  I am sending them openly and
> leave acceptance to the maintainers.  If they cannot be taken, please
> treat this as a bug report; the reproducers are in patch 1.
> 
> Dominic P (3):
>    tcg/optimize: Do not derive s_mask from rotated s_mask
>    tests/tcg: Add tests for sign extension after rotate
>    tests/tcg/i386: Build test-i386-opt-shr
Thanks.  I've taken the first two patches into tcg-next, since the third 
has already been merged.  I tweaked the if condition into a switch.


r~