On 9/2/26 11:51 AM, Wang Yaduo wrote:
Recovered scalar types are needed by the base cost hook, but they must not
trigger vector-specific statement cost adjustments.
gcc/ChangeLog:
* config/riscv/riscv-vector-costs.cc (costs::add_stmt_cost): Skip
statement cost adjustments for scalar costing.
---
When costing scalar IL, vectype is null and 32dc4bc recovers a type from
the statement LHS. The LHS can itself be vector-typed, however, so this
can produce a vectype that is wrong for the scalar cost being computed.
This happens for the 512-element constructor in combine-4.c. Its root
statement has a vector(512) long int LHS. Recovering that type while
computing SCALAR_COST lets adjust_stmt_cost apply M8 scaling, raising the
scalar cost from 1 to 8. An SLP alternative using 512-byte vectors, whose
cost is 6, then appears profitable and expands the constructor through a
4096-byte stack temporary. This removes one of the four expected
vslideup.vx instructions.
Require m_cost_type != SCALAR_COST before calling adjust_stmt_cost. The
recovered type is still passed to builtin_vectorization_cost, but cannot
trigger vector-specific adjustments during scalar costing.
Tested the full gcc.target/riscv testsuite on rv64gcv, including execution
tests; no new FAIL, XPASS, UNRESOLVED, or ERROR results versus 32dc4bc^.
Just to be clear, that is not sufficient for a patch to the generic part
of the compiler. For such changes the minimum is to do a bootstrap and
regression test on a platform like x86 or aarch64.
FTR, I had already thrown this into my tester several days ago. So it's
been through that bootstrap and regression test on x86_64, aarch64.
I'll take care of pushing it to the trunk. Thanks.
Jeff
