diff options
| author | Kyrylo Tkachov <ktkachov@nvidia.com> | 2025-03-24 01:53:06 -0700 |
|---|---|---|
| committer | Kyrylo Tkachov <ktkachov@nvidia.com> | 2025-03-31 11:15:28 +0200 |
| commit | 70391e3958db791edea4e877636592de47a785e7 (patch) | |
| tree | dcb330af493935438861899fb4a30ec7406e1339 /gcc/expr.cc | |
| parent | 5f2078ca11843e3fe506c2fb72b9260a3cb9853e (diff) | |
| download | gcc-70391e3958db791edea4e877636592de47a785e7.zip gcc-70391e3958db791edea4e877636592de47a785e7.tar.gz gcc-70391e3958db791edea4e877636592de47a785e7.tar.bz2 | |
PR middle-end/119442: expr.cc: Fix vec_duplicate into vector boolean modes
In this testcase GCC tries to expand a VNx4BI vector:
vector(4) <signed-boolean:4> _40;
_39 = (<signed-boolean:4>) _24;
_40 = {_39, _39, _39, _39};
This ends up in a scalarised sequence of bitfield insert operations.
This is despite the fact that AArch64 provides a vec_duplicate pattern
specifically for vec_duplicate into VNx4BI.
The store_constructor code is overly conservative when trying vec_duplicate
as it sees a requested VNx4BImode and an element mode of QImode, which I guess
is the storage mode of BImode objects.
The vec_duplicate expander in aarch64-sve.md explicitly allows QImode element
modes so it should be safe to use it. This patch extends that mode check
to allow such expanders.
The testcase is heavily auto-reduced from a real application but in itself is
nonsensical, but it does demonstrate the current problematic codegen.
This the testcase goes from:
pfalse p15.b
str p15, [sp, #6, mul vl]
mov w0, 0
ldr w2, [sp, 12]
bfi w2, w0, 0, 4
uxtw x2, w2
bfi w2, w0, 4, 4
uxtw x2, w2
bfi w2, w0, 8, 4
uxtw x2, w2
bfi w2, w0, 12, 4
str w2, [sp, 12]
ldr p15, [sp, #6, mul vl]
into:
whilelo p15.s, wzr, wzr
The whilelo could be optimised away into a pfalse of course, but the important
part is that the bfis are gones.
Bootstrapped and tested on aarch64-none-linux-gnu.
Signed-off-by: Kyrylo Tkachov <ktkachov@nvidia.com>
gcc/
PR middle-end/119442
* expr.cc (store_constructor): Also allow element modes explicitly
accepted by target vec_duplicate pattern.
gcc/testsuite/
PR middle-end/119442
* gcc.target/aarch64/vls_sve_vec_dup_1.c: New test.
Diffstat (limited to 'gcc/expr.cc')
| -rw-r--r-- | gcc/expr.cc | 11 |
1 files changed, 8 insertions, 3 deletions
diff --git a/gcc/expr.cc b/gcc/expr.cc index 9f4382d..2147eed 100644 --- a/gcc/expr.cc +++ b/gcc/expr.cc @@ -7920,11 +7920,16 @@ store_constructor (tree exp, rtx target, int cleared, poly_int64 size, gcc_assert (eltmode != BLKmode); /* Try using vec_duplicate_optab for uniform vectors. */ + icode = optab_handler (vec_duplicate_optab, mode); if (!TREE_SIDE_EFFECTS (exp) && VECTOR_MODE_P (mode) - && eltmode == GET_MODE_INNER (mode) - && ((icode = optab_handler (vec_duplicate_optab, mode)) - != CODE_FOR_nothing) + && icode != CODE_FOR_nothing + /* If the vec_duplicate target pattern does not specify an element + mode check that eltmode is the normal inner mode of the + requested vector mode. But if the target allows eltmode + explicitly go ahead and use it. */ + && (eltmode == GET_MODE_INNER (mode) + || insn_data[icode].operand[1].mode == eltmode) && (elt = uniform_vector_p (exp)) && !VECTOR_TYPE_P (TREE_TYPE (elt))) { |