On 1/14/2026 4:33 AM, Jakub Jelinek wrote:
Hi!
On Tue, Nov 04, 2025 at 12:59:03PM +0530, Kishan Parmar wrote:
PR rtl-optimization/93738
* simplify-rtx.cc (simplify_binary_operation_1): Canonicalize
SUBREG(LSHIFTRT) into LSHIFTRT(SUBREG) when valid.
This change regressed the following testcase on aarch64-linux.
From what I can see, the PR93738 change has been written with non-paradoxical
SUBREGs in mind but on this testcase on aarch64 we have a paradoxical SUBREG,
in particular simplify_binary_operation_1 is called with AND, SImode,
(subreg:SI (lshiftrt:HI (subreg:HI (reg/v:SI 108 [ x ]) 0)
(const_int 8 [0x8])) 0)
and op1 (const_int 32767 [0x7fff]) and simplifies that since the PR93738
optimization was added into
(and:SI (lshiftrt:SI (reg/v:SI 108 [ x ])
(const_int 8 [0x8]))
(const_int 32767 [0x7fff]))
This looks wrong to me.
Consider (reg/v:SI 108 [ x ]) 0) could have value 0x12345678U.
The original expression takes lowpart 16-bits from that, i.e. 0x5678U,
shifts that right logically by 8 bits, so 0x56U, makes a paradoxical SUBREG
from that, i.e. 0x????0056U and masks that with 0x7fff, i.e. result is 0x56U.
The new expression shifts 0x12345678U logically right by 8 bits, i.e. 0x123456U
and
masks it by 0x7fff, result 0x3456U.
Thus, I think we need to limit to non-paradoxical SUBREGs.
On the rlwimi-2.c testcase I see on powerpc64le-linux no differences in
emitted assembly without/with the patch.
Bootstrapped/regtested on x86_64-linux, i686-linux, aarch64-linux and
powerpc64le-linux, ok for trunk?
2026-01-14 Jakub Jelinek <[email protected]>
PR rtl-optimization/123544
* simplify-rtx.cc (simplify_context::simplify_binary_operation_1)
<case AND>: Don't canonicalize (subreg (lshiftrt (x cnt)) low) into
(lshiftrt (subreg x low) cnt) if the SUBREG is paradoxical.
* gcc.dg/pr123544.c: New test.
It's easy to forget that subreg_lowpart_p returns true for
paradoxicals. I'm pretty sure I've been bitten by that before.
OK for the trunk.
jeff
