https://gcc.gnu.org/g:6ef27c311e317485760a6ffc1759cde241e396de
commit r17-5990-g6ef27c311e317485760a6ffc1759cde241e396de Author: Robin Dapp <[email protected]> Date: Wed Oct 28 21:21:37 2026 +0100 RISC-V: Adjust VLS subreg handling. Since the regmode_natural_size changes we'd spill subregs like (subreg:V4SI (reg:V8SI) 0) even though a lowpart subreg is safe in a VLA setting. This is caused by a mismatch in can_change_mode_class and regmode_natural_size. can_change_mode_class doesn't look at the subreg byte but would prohibit all such subregs because we cannot statically tell whether the inner reg is a register pair or a larger register. A better way is to always consider VLS vectors unsplittable but don't always prohibit them in can_change_mode_class. That way, lowpart subregs are still allowed but non-lowpart subregs aren't as they won't be generated in the first place. gcc/ChangeLog: * config/riscv/riscv.cc (riscv_can_change_mode_class): Don't prohibit ambiguous VLS subregs. (riscv_regmode_natural_size): Make VLS regs unsplittable when the vector size is variable. gcc/testsuite/ChangeLog: * gcc.target/riscv/rvv/autovec/vec-concat-1.c: Expect slides for all functions. * gcc.target/riscv/rvv/autovec/vls/extract-3.c: New test. Diff: --- gcc/config/riscv/riscv.cc | 17 ++-------------- .../gcc.target/riscv/rvv/autovec/vec-concat-1.c | 6 +----- .../gcc.target/riscv/rvv/autovec/vls/extract-3.c | 23 ++++++++++++++++++++++ 3 files changed, 26 insertions(+), 20 deletions(-) diff --git a/gcc/config/riscv/riscv.cc b/gcc/config/riscv/riscv.cc index a2a6f3c223c5..9f1f4df3ea64 100644 --- a/gcc/config/riscv/riscv.cc +++ b/gcc/config/riscv/riscv.cc @@ -13378,18 +13378,6 @@ riscv_can_change_mode_class (machine_mode from, machine_mode to, && !ordered_p (GET_MODE_PRECISION (from), GET_MODE_PRECISION (to))) return false; - /* Subregs of modes larger than one vector are ambiguous. - A V4DImode with rv64gcv_zvl128b could, for example, span two registers/one - register group of two at VLEN = 128 or one register at VLEN >= 256 and - we cannot, statically, determine which part of it to extract. - Therefore prevent that. */ - if (reg_classes_intersect_p (V_REGS, rclass) - && VECTOR_MODE_P (from) - && GET_MODE_NUNITS (from).is_constant () - && maybe_ne (GET_MODE_PRECISION (from), GET_MODE_PRECISION (to)) - && !ordered_p (BITS_PER_RISCV_VECTOR, GET_MODE_PRECISION (from))) - return false; - return !reg_classes_intersect_p (FP_REGS, rclass); } @@ -13816,10 +13804,9 @@ riscv_regmode_natural_size (machine_mode mode) if (known_lt (size, BYTES_PER_RISCV_VECTOR)) return size; } - if (!size.is_constant ()) + if (ordered_p (size, BYTES_PER_RISCV_VECTOR)) return BYTES_PER_RISCV_VECTOR; - else - return TARGET_MIN_VLEN / BITS_PER_UNIT; + return size; } return UNITS_PER_WORD; } diff --git a/gcc/testsuite/gcc.target/riscv/rvv/autovec/vec-concat-1.c b/gcc/testsuite/gcc.target/riscv/rvv/autovec/vec-concat-1.c index 0fdd169bcef5..8a320e1ffdc2 100644 --- a/gcc/testsuite/gcc.target/riscv/rvv/autovec/vec-concat-1.c +++ b/gcc/testsuite/gcc.target/riscv/rvv/autovec/vec-concat-1.c @@ -60,9 +60,5 @@ TEST_MODES (DEF_CONCAT) -/* LMUL <= 1: 11 funcs x 3 slideups (using vec_concat, from spills + 1 slideup - to no spills + 3 slideups). - LMUL > 1: 24 funcs x 1 slideup (no vec_concat, using fallback instead, - nothing changed from trunk). */ -/* { dg-final { scan-assembler-times "vslideup" 57 } } */ +/* { dg-final { scan-assembler-times "vslideup" 105 } } */ /* { dg-final { scan-assembler-not "vslide1down" } } */ diff --git a/gcc/testsuite/gcc.target/riscv/rvv/autovec/vls/extract-3.c b/gcc/testsuite/gcc.target/riscv/rvv/autovec/vls/extract-3.c new file mode 100644 index 000000000000..b6c8c913a90b --- /dev/null +++ b/gcc/testsuite/gcc.target/riscv/rvv/autovec/vls/extract-3.c @@ -0,0 +1,23 @@ +/* { dg-do compile } */ +/* { dg-options "-march=rv64gcv -mabi=lp64d -O2" } */ + +typedef int v8si __attribute__ ((vector_size (32))); +typedef int v4si __attribute__ ((vector_size (16))); + +void +lo (v8si *in, v4si *out) +{ + v8si a = *in + *in; + *out = __builtin_shufflevector (a, a, 0, 1, 2, 3); +} + +void +hi (v8si *in, v4si *out) +{ + v8si a = *in + *in; + *out = __builtin_shufflevector (a, a, 4, 5, 6, 7); +} + +/* { dg-final { scan-assembler-not {v[ls]e[0-9]+\.v\s+v[0-9]+,\s*-?[0-9]*\(sp\)} } } */ +/* { dg-final { scan-assembler-not {v[ls][0-9]r\.v} } } */ +/* { dg-final { scan-assembler-times {vslidedown\.vi} 1 } } */
