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.
--- gcc/simplify-rtx.cc.jj 2026-01-12 10:06:41.385871794 +0100
+++ gcc/simplify-rtx.cc 2026-01-13 18:12:57.320005833 +0100
@@ -4193,7 +4193,9 @@ simplify_context::simplify_binary_operat
Keeps shift and AND in the same mode, improving recognition.
Only applied when subreg is a lowpart, shift is valid,
and no precision is lost. */
- if (SUBREG_P (op0) && subreg_lowpart_p (op0)
+ if (SUBREG_P (op0)
+ && subreg_lowpart_p (op0)
+ && !paradoxical_subreg_p (op0)
&& GET_CODE (XEXP (op0, 0)) == LSHIFTRT
/* simplify_subreg asserts the object being accessed is not
VOIDmode or BLKmode. We may have a REG_EQUAL note which
--- gcc/testsuite/gcc.dg/pr123544.c.jj 2026-01-13 18:22:36.252113131 +0100
+++ gcc/testsuite/gcc.dg/pr123544.c 2026-01-13 18:22:22.438349152 +0100
@@ -0,0 +1,44 @@
+/* PR rtl-optimization/123544 */
+/* { dg-do run { target bitint } } */
+/* { dg-options "-O1 -w" } */
+
+typedef unsigned char V __attribute__((vector_size (4)));
+unsigned long long c;
+
+signed char
+foo ()
+{
+ return 0;
+}
+
+unsigned int
+bar (V x, short int y, long long z)
+{
+ _BitInt (15) a = 828;
+ a &= (unsigned) z < 5 ? 0 : (unsigned) x[1];
+ if (y)
+ return a;
+ for (unsigned short d; d; )
+ ;
+}
+
+int
+baz (unsigned char x)
+{
+ unsigned long long a[] = { };
+ unsigned long long b = bar ((V) { 1, 1, 1 }, 1, -1);
+ if (x)
+ {
+ c += b;
+ return 0;
+ }
+ a[bar ((V) {}, 0, foo ())];
+}
+
+int
+main ()
+{
+ baz (1);
+ if (c)
+ __builtin_abort ();
+}
Jakub
