[PATCH] match: move a mask across a following right shift
<[email protected]> Wed, 5 Aug 2026 13:58:57 +0200
| Newsgroups | gmane.comp.gcc.patches |
|---|---|
| Message-ID | <[email protected]> |
From: Kyrylo Tkachov <[email protected]> This is a respin of https://gcc.gnu.org/pipermail/gcc-patches/2026-August/726305.html with a subject change as this is more general now after Andrea's feedback. For an unsigned type, (X & (A << Y)) >> Y is equal to (X >> Y) & A. The same identity holds for a signed type when X is known nonnegative or when the possible high bits of A cannot affect the sign-extension bits. An all-ones A is a special case in which the mask disappears completely. unsigned f (unsigned x, int y) { return (x & (~0u << y)) >> y; } aarch64 -O2: before after mov w2, -1 lsr w0, w0, w1 lsl w2, w2, w1 and w0, w2, w0 lsr w0, w0, w1 Bootstrapped and tested on aarch64-none-linux-gnu. Ok for trunk? Thanks, Kyrill gcc/ChangeLog: * match.pd ((X & (A << Y)) >> Y): New simplification. gcc/testsuite/ChangeLog: * gcc.dg/tree-ssa/shift-mask-1.c: New test. Signed-off-by: Kyrylo Tkachov <[email protected]> --- gcc/match.pd | 17 ++++++++++++++++ gcc/testsuite/gcc.dg/tree-ssa/shift-mask-1.c | 21 ++++++++++++++++++++ 2 files changed, 38 insertions(+) create mode 100644 gcc/testsuite/gcc.dg/tree-ssa/shift-mask-1.c diff --git a/gcc/match.pd b/gcc/match.pd index 44ee449b61f..2e16ed0f7dc 100644 --- a/gcc/match.pd +++ b/gcc/match.pd @@ -5234,6 +5234,23 @@ DEFINE_INT_AND_FLOAT_ROUND_FN (RINT) - TYPE_PRECISION (TREE_TYPE (@2))))) (bit_and (convert @0) (lshift { build_minus_one_cst (type); } @1)))) +/* Move a mask across a right shift when this does not change the sign bits. + An all-ones mask only clears bits that the shift discards. */ +(simplify + (rshift (bit_and:cs @0 (lshift @2 @1)) @1) + (if (integer_all_onesp (@2)) + (rshift @0 @1) + (if (ANY_INTEGRAL_TYPE_P (type) + && (TYPE_UNSIGNED (type) + || tree_expr_nonnegative_p (@0) + || (INTEGRAL_TYPE_P (type) + && tree_fits_uhwi_p (@1) + && tree_to_uhwi (@1) < TYPE_PRECISION (type) + && (tree_nonzero_bits (@2) + & wi::mask (TYPE_PRECISION (type) + - tree_to_uhwi (@1) - 1, + true, TYPE_PRECISION (type))) == 0))) + (bit_and (rshift @0 @1) @2)))) /* (X op Y) >> C -> X >> C when every set bit of Y lies below bit C. Neither an inclusive nor an exclusive or can carry into the bits the shift keeps, so Y contributes nothing to the result. */ diff --git a/gcc/testsuite/gcc.dg/tree-ssa/shift-mask-1.c b/gcc/testsuite/gcc.dg/tree-ssa/shift-mask-1.c new file mode 100644 index 00000000000..d98d9ed870b --- /dev/null +++ b/gcc/testsuite/gcc.dg/tree-ssa/shift-mask-1.c @@ -0,0 +1,21 @@ +/* { dg-do compile } */ +/* { dg-options "-O2 -fdump-tree-optimized" } */ + +/* Masking off the bits that the following right shift discards is a no-op. */ + +unsigned int f1 (unsigned int x, int y) { return (x & (~0u << y)) >> y; } +int f2 (int x, int y) { return (x & (~0 << y)) >> y; } +unsigned long f3 (unsigned long x, int y) { return (x & (~0ul << y)) >> y; } +unsigned int f4 (unsigned int x, int y) { return ((~0u << y) & x) >> y; } + +unsigned int g1 (unsigned int x, int y, unsigned int a) +{ return (x & (a << y)) >> y; } +int g2 (unsigned short x, int y, int a) { return (x & (a << y)) >> y; } +int g3 (int x, unsigned short a) { return (x & (a << 3)) >> 3; } + +/* A variable signed value and mask can change the sign-extension bits. */ +int keep (int x, int y, int a) { return (x & (a << y)) >> y; } + +/* { dg-final { scan-tree-dump-times " & " 4 "optimized" } } */ +/* { dg-final { scan-tree-dump-times " << " 1 "optimized" } } */ +/* { dg-final { scan-tree-dump-times " >> " 8 "optimized" } } */ -- 2.50.1 (Apple Git-155)