On Thu, Feb 5, 2026 at 9:51 PM Robin Dapp <[email protected]> wrote:
>
> > It seems to be the fix is to check ops[] properly? But maybe I
> > mis-understand the failure mode.
>
> Yes, that's sufficient. The attached was actually my first try but
> then I started re-using code.
>
> Bootstrapped and regtested as before, aarch64 still running.
OK.
Thanks,
Richard.
> --
> Regards
> Robin
>
> [PATCH v2] math-opts: Only build FMA when use is in addends. [PR123940]
>
> When looking for an FMA opportunity we can find a multiplication use in an
> else operand:
>
> vect_pretmp_50.23_121 = MEM <vector(2) charD.2> [(charD.2 *)&dD.2916 + 14B];
> vect__28.25_123 = vect_pretmp_50.23_121 * { 2, 2 };
> vect_patt_99.26_124 = .COND_ADD ({ -1, -1 }, vect_pretmp_50.23_121, { 14,
> 15 }, vect__28.25_123);
>
> and build it:
> vect_pretmp_50.23_121 = MEM <vector(2) charD.2> [(charD.2 *)&dD.2916 + 14B];
> vect_patt_99.26_124 = .FMA (vect_pretmp_50.23_121, { 2, 2 },
> vect_pretmp_50.23_121);
>
> This patch checks if the use is in one of the addends.
>
> PR tree-optimization/123940
>
> gcc/ChangeLog:
>
> * tree-ssa-math-opts.cc (convert_mult_to_fma): Check
> multiplication result is an addend.
>
> gcc/testsuite/ChangeLog:
>
> * gcc.target/riscv/rvv/autovec/pr123940.c: New test.
> ---
> .../gcc.target/riscv/rvv/autovec/pr123940.c | 25 +++++++++++++++++++
> gcc/tree-ssa-math-opts.cc | 5 ++++
> 2 files changed, 30 insertions(+)
> create mode 100644 gcc/testsuite/gcc.target/riscv/rvv/autovec/pr123940.c
>
> diff --git a/gcc/testsuite/gcc.target/riscv/rvv/autovec/pr123940.c
> b/gcc/testsuite/gcc.target/riscv/rvv/autovec/pr123940.c
> new file mode 100644
> index 00000000000..329e598f068
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/riscv/rvv/autovec/pr123940.c
> @@ -0,0 +1,25 @@
> +/* { dg-do run } */
> +/* { dg-options "-O3 -march=rv64gcv -mabi=lp64d -std=gnu99
> -fdump-tree-widening_mul" } */
> +
> +long a;
> +long long b;
> +_Bool c[16];
> +char d[16];
> +char e = 0;
> +int f = 1;
> +
> +int main ()
> +{
> + for (long i = 0; i < 16; ++i)
> + c[i] = 40;
> + for (int j = 0; j < 16; j++)
> + {
> + e = (c[j] ? j : d[j]) + d[j];
> + a = f * c[j] ?: ~0;
> + }
> + b = (int) e;
> + if (b != 15)
> + __builtin_abort ();
> +}
> +
> +/* { dg-final { scan-tree-dump-not "FMA" "widening_mul" } } */
> diff --git a/gcc/tree-ssa-math-opts.cc b/gcc/tree-ssa-math-opts.cc
> index e3f88f23628..1655b68cb9c 100644
> --- a/gcc/tree-ssa-math-opts.cc
> +++ b/gcc/tree-ssa-math-opts.cc
> @@ -3481,6 +3481,7 @@ convert_mult_to_fma (gimple *mul_stmt, tree op1, tree
> op2,
> use_operand_p tmp_use_p;
> if (single_imm_use (cast_lhs, &tmp_use_p, &tmp_use))
> use_stmt = tmp_use;
> + result = cast_lhs;
> }
>
> /* For now restrict this operations to single basic blocks. In theory
> @@ -3535,6 +3536,10 @@ convert_mult_to_fma (gimple *mul_stmt, tree op1, tree
> op2,
> &else_value, &len, &bias))
> return false;
>
> + /* The multiplication result must be one of the addition operands. */
> + if (ops[0] != result && ops[1] != result)
> + return false;
> +
> switch (code)
> {
> case MINUS_EXPR:
> --
> 2.52.0
>
