https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116979
Jakub Jelinek <jakub at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |uros at gcc dot gnu.org
--- Comment #11 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
C testcase (-O2 -mfma):
struct S { __complex__ float f; };
struct S
foo (const struct S *a, const struct S *b)
{
struct S r;
r.f = a->f * b->f;
return r;
}
I've tried to catch this up during widening_mul with
--- gcc/tree-ssa-math-opts.cc.jj 2024-11-23 13:00:31.445983255 +0100
+++ gcc/tree-ssa-math-opts.cc 2024-12-12 18:22:42.078324724 +0100
@@ -3092,7 +3092,17 @@ convert_mult_to_fma_1 (tree mul_result,
if (!can_interpret_as_conditional_op_p (use_stmt, &cond, &code,
ops, &else_value,
&len, &bias))
- gcc_unreachable ();
+ {
+ gcc_assert (gimple_call_internal_p (use_stmt, IFN_VEC_ADDSUB));
+ cond = NULL_TREE;
+ code = ERROR_MARK;
+ ops[0] = gimple_call_arg (use_stmt, 0);
+ ops[1] = gimple_call_arg (use_stmt, 1);
+ ops[2] = NULL_TREE;
+ else_value = NULL_TREE;
+ len = NULL_TREE;
+ bias = NULL_TREE;
+ }
addop = ops[0] == result ? ops[1] : ops[0];
if (code == MINUS_EXPR)
@@ -3118,6 +3128,9 @@ convert_mult_to_fma_1 (tree mul_result,
else if (cond)
fma_stmt = gimple_build_call_internal (IFN_COND_FMA, 5, cond, mulop1,
op2, addop, else_value);
+ else if (code == ERROR_MARK)
+ fma_stmt = gimple_build_call_internal (IFN_VEC_FMADDSUB, 3, mulop1,
+ op2, addop);
else
fma_stmt = gimple_build_call_internal (IFN_FMA, 3, mulop1, op2, addop);
gimple_set_lhs (fma_stmt, gimple_get_lhs (use_stmt));
@@ -3419,22 +3432,38 @@ convert_mult_to_fma (gimple *mul_stmt, t
tree cond, else_value, ops[3], len, bias;
tree_code code;
- if (!can_interpret_as_conditional_op_p (use_stmt, &cond, &code, ops,
- &else_value, &len, &bias))
- return false;
-
- switch (code)
+ if (can_interpret_as_conditional_op_p (use_stmt, &cond, &code, ops,
+ &else_value, &len, &bias))
+ switch (code)
+ {
+ case MINUS_EXPR:
+ if (ops[1] == result)
+ negate_p = !negate_p;
+ break;
+ case PLUS_EXPR:
+ break;
+ default:
+ /* FMA can only be formed from PLUS and MINUS. */
+ return false;
+ }
+ else if (gimple_call_internal_p (use_stmt, IFN_VEC_ADDSUB))
{
- case MINUS_EXPR:
- if (ops[1] == result)
- negate_p = !negate_p;
- break;
- case PLUS_EXPR:
- break;
- default:
- /* FMA can only be formed from PLUS and MINUS. */
- return false;
+ cond = NULL_TREE;
+ code = ERROR_MARK;
+ ops[0] = gimple_call_arg (use_stmt, 0);
+ ops[1] = gimple_call_arg (use_stmt, 1);
+ ops[2] = NULL_TREE;
+ else_value = NULL_TREE;
+ len = NULL_TREE;
+ bias = NULL_TREE;
+ if (negate_p
+ || ops[0] != result
+ || !direct_internal_fn_supported_p (IFN_VEC_FMADDSUB, type,
+ opt_type))
+ return false;
}
+ else
+ return false;
if (len)
{
but it doesn't work anyway, because vec_fmaddsubv2sf4 is not supported.
And if I change the testcase s/float/double/, then it uses vfmaddsub132pd even
without my patch.
So maybe the fix would be to introduce vec_fmaddsubv2sf4 and maybe
vec_fmsubaddv2sf4
patterns for (TARGET_FMA || TARGET_FMA4 || TARGET_AVX512VL) &&
TARGET_MMX_WITH_SSE && ix86_partial_vec_fp_math ?
CCing Uros, I really have limited experience with the mmx with sse stuff.
