On Mon, 12 Oct 2026, Tamar Christina wrote:
> The example
>
> float
> f (const float *x, int n)
> {
> float acc = 0;
> for (int i = 0; i < n; ++i)
> {
> float d = x[i];
> if (d < 0)
> acc += d;
> acc += d * 3.0f;
> }
> return acc;
> }
>
> compiled at -Ofast fails to vectorize with
>
> note: vect_is_simple_use: operand acc_17 = PHI <acc_13(8), 0.0(18)>,
> type of def: unknown
> missed: Unsupported pattern.
> missed: not vectorized: unsupported use in stmt.
>
> due to
>
> note: Analyze phi: acc_17 = PHI <acc_13(8), 0.0(18)>
> missed: reduction used in loop.
> missed: Unknown def-use cycle pattern.
>
> This is happening because before if-convert we have the following reduction:
>
> acc_1 = PHI <acc_0, acc_2>
> if (cond)
> acc_2 = acc_1 + value;
>
> and so the reduction value is updated only on some paths.
>
> if-cvt changes this into
>
> acc_3 = cond ? acc_2 : acc_1;
>
> but since acc_2 itself is acc_1 + value this is actually
>
> acc_3 = cond ? acc_1 + value : acc_1;
>
> This is semantically sound but hides the recurrence from the vectorizer.
>
> In principle if-cvt creates
>
> acc_1 = PHI <0, acc_3>
> acc_2 = acc_1 + value
> acc_3 = cond ? acc_2 : acc_1
>
> so the loop carried value acc_3 is inside the COND_EXPR and acc_1 is used
> twice
> inside the same reduction. The update chain is
>
> PHI -> PLUS -> COND_EXPR -> PHI
> \--------------^
>
> which is a problem since reduction analysis needs a single statement that
> updates the loop carried accumulator.
>
> Note that if the user supplied this form manually, if-cvt rewrites it to one
> of
> two forms that the vectorizer can handle. It either rewrites it to COND_ADD
> and
> this removing the PLUS + COND_EXPR or it floats out the cond_expr.
>
> This patch does the latter, as picking COND_ADD here is premature, it forces
> selection of a masked target even though the vectorizer these days can convert
> the cond expr to a masked operation on its own based on costing.
>
> Essentially we change
>
> cond ? base + addend : base or cond ? base : base + addend
>
> into
>
> base + (cond ? addend : 0) or base + (cond ? 0 : addend)
>
> this turns the recurrence into a form the vectorizer can handle.
>
> Special care is taken to only do this when the reduction is within the same
> loop and where there is no other uses of the PHI reduction but the PHI node.
>
> i.e. the rewrite is only applied when BASE is the loop-carried reduction PHI
> for
> the current loop and the if-converted PHI feeds the latch update for that same
> reduction. This avoids changing unrelated conditional expressions or cases
> where the apparent recurrence is not the loop reduction being analyzed.
>
> I have not used match.pd's PHI matching machinery here because the issue is
> that ^ can't be used because the PHI in the IL may not correspond to PHI after
> ifcvt lowered it. As an example
>
> int f(const int *x, int n)
> {
> int acc = 0, other = 0;
>
> for (int i = 0; i < n; ++i) {
> int d = x[i];
> int next = acc + d;
>
> if (d < 0) {
> acc = next;
> other += d;
> } else if (d > 100) {
> other += 2 * d;
> } else {
> other += 3 * d;
> }
>
> acc += 4 * d;
> }
> return acc + other;
> }
>
> Has as the incoming PHI
>
> acc_7 = PHI <next_17(4), acc_26(6), acc_26(7)>
>
> If converts notices it has only 2 distinct values and lowers this to
>
> acc_7 = _46 ? next_17 : acc_26
>
> which we can transform whereas match_cond_with_binary_phi
> rejects PHIs that don't have just two values.
>
> As a result it means to even call the (match on statements not in IL
> yet we have to construct a dummy gimple statement
>
> tree res = gimple_phi_result (phi);
> gimple *old_def = SSA_NAME_DEF_STMT (res);
> gassign *match_stmt = gimple_build_assign (res, COND_EXPR, cond, rhs, lhs);
> tree ops[3];
> bool matched = gimple_ifc_cond_reduction (res, ops, follow_all_ssa_edges);
> gimple_assign_set_lhs (match_stmt, NULL_TREE);
> SSA_NAME_DEF_STMT (res) = old_def;
>
> And really doesn't seem worth it for just
>
> (match (ifc_cond_reduction @0 @1 @2)
> (cond @0 (plus:c @1 @2) @1)
> (if (types_match (type, @1, @2))))
> (match (ifc_cond_reduction @0 @1 @2)
> (cond @0 @1 (plus:c @1 @2))
> (if (types_match (type, @1, @2)
> && (@0 = invert_truthvalue (@0), true))))
>
>
> and building the temporary IL I found rather ugly..
>
> Bootstrapped Regtested on aarch64-none-linux-gnu,
> arm-none-linux-gnueabihf, x86_64-pc-linux-gnu -m32, -m64 and no issues.
>
> Is this version OK for trunk?
>
> Thanks,
> Tamar
>
> gcc/ChangeLog:
>
> PR tree-optimization/126551
> * tree-if-conv.cc (ifc_temp_var): Support cases with no vUSE.
> (ifc_plus_with_operand_p, ifc_phi_feeds_loop_reduction_p,
> fold_build_cond_reduction_expr): New.
> (predicate_scalar_phi): Use it.
>
> gcc/testsuite/ChangeLog:
>
> PR tree-optimization/126551
> * gcc.dg/vect/vect-switch-ifcvt-2.c: Update testcases.
> * gcc.dg/vect/cond-reduc-ifcvt-1.c: New test.
> * gcc.dg/vect/cond-reduc-ifcvt-2.c: New test.
>
> ---
> diff --git a/gcc/testsuite/gcc.dg/vect/cond-reduc-ifcvt-1.c
> b/gcc/testsuite/gcc.dg/vect/cond-reduc-ifcvt-1.c
> new file mode 100644
> index
> 0000000000000000000000000000000000000000..e4a2e7634a8d6bf02d1d3a6f7f12515e83fb8423
> --- /dev/null
> +++ b/gcc/testsuite/gcc.dg/vect/cond-reduc-ifcvt-1.c
> @@ -0,0 +1,21 @@
> +/* { dg-do compile } */
> +/* { dg-require-effective-target vect_float } */
> +/* { dg-additional-options "-Ofast -fdump-tree-ifcvt-details" } */
> +
> +float
> +f (const float *x, int n)
> +{
> + float acc = 0;
> + for (int i = 0; i < n; ++i)
> + {
> + float d = x[i];
> + if (d < 0)
> + acc += d;
> + acc += d * 3.0f;
> + }
> + return acc;
> +}
> +
> +/* { dg-final { scan-tree-dump "new phi replacement stmt" "ifcvt" } } */
> +/* { dg-final { scan-tree-dump "_ifc_.* = MIN_EXPR" "ifcvt" } } */
> +/* { dg-final { scan-tree-dump " = .* \\+ _ifc_" "ifcvt" } } */
> diff --git a/gcc/testsuite/gcc.dg/vect/cond-reduc-ifcvt-2.c
> b/gcc/testsuite/gcc.dg/vect/cond-reduc-ifcvt-2.c
> new file mode 100644
> index
> 0000000000000000000000000000000000000000..c6015303e5568dd51a7ced1edf33784fdf93f75d
> --- /dev/null
> +++ b/gcc/testsuite/gcc.dg/vect/cond-reduc-ifcvt-2.c
> @@ -0,0 +1,21 @@
> +/* { dg-do compile } */
> +/* { dg-require-effective-target vect_int } */
> +/* { dg-additional-options "-O3 -fdump-tree-ifcvt-details" } */
> +
> +int
> +f (const int *x, int n)
> +{
> + int acc = 0;
> + for (int i = 0; i < n; ++i)
> + {
> + int d = x[i];
> + if (d < 0)
> + acc += d;
> + acc += d * 3;
> + }
> + return acc;
> +}
> +
> +/* { dg-final { scan-tree-dump "new phi replacement stmt" "ifcvt" } } */
> +/* { dg-final { scan-tree-dump "_ifc_.* = MIN_EXPR" "ifcvt" } } */
> +/* { dg-final { scan-tree-dump " = .* \\+ _ifc_" "ifcvt" } } */
> diff --git a/gcc/testsuite/gcc.dg/vect/vect-switch-ifcvt-2.c
> b/gcc/testsuite/gcc.dg/vect/vect-switch-ifcvt-2.c
> index
> f0fcb43e17702492b4fb0e9bc89f0c90b3106fda..03f241aef3750be06e91b7c9c583d4bb091e87e7
> 100644
> --- a/gcc/testsuite/gcc.dg/vect/vect-switch-ifcvt-2.c
> +++ b/gcc/testsuite/gcc.dg/vect/vect-switch-ifcvt-2.c
> @@ -46,4 +46,4 @@ f2 (char *s)
> return c;
> }
>
> -/* { dg-final { scan-tree-dump-times "vectorized 1 loops" 0 "vect" } } */
> +/* { dg-final { scan-tree-dump-times "vectorized 1 loops" 1 "vect" } } */
> diff --git a/gcc/tree-if-conv.cc b/gcc/tree-if-conv.cc
> index
> cc3e1dfb4ef2bbd1edaa5ce188202650f5cb163d..a63e6aacadc7f5df105610fe810aba4372b37ac6
> 100644
> --- a/gcc/tree-if-conv.cc
> +++ b/gcc/tree-if-conv.cc
> @@ -392,7 +392,9 @@ ifc_temp_var (tree type, tree expr, gimple_stmt_iterator
> *gsi)
> {
> tree new_name = make_temp_ssa_name (type, NULL, "_ifc_");
> gimple *stmt = gimple_build_assign (new_name, expr);
> - gimple_set_vuse (stmt, gimple_vuse (gsi_stmt (*gsi)));
> + gimple *curr_stmt = gsi_stmt (*gsi);
> + if (curr_stmt)
> + gimple_set_vuse (stmt, gimple_vuse (curr_stmt));
> gsi_insert_before (gsi, stmt, GSI_SAME_STMT);
> return new_name;
> }
> @@ -526,6 +528,116 @@ fold_build_cond_expr (tree type, tree cond, tree rhs,
> tree lhs)
> return build3 (COND_EXPR, type, cond, rhs, lhs);
> }
>
> +/* Return true if SUM is an addition of OPERAND and another value. Set
> + ADDEND to that value and OPERAND_FIRST according to the operand order. */
> +
> +static bool
> +ifc_plus_with_operand_p (tree sum, tree operand, tree *addend,
> + bool *operand_first)
> +{
> + if (TREE_CODE (sum) != SSA_NAME)
> + return false;
> +
> + gimple *stmt = SSA_NAME_DEF_STMT (sum);
> + if (!is_gimple_assign (stmt)
> + || gimple_assign_rhs_code (stmt) != PLUS_EXPR)
> + return false;
> +
> + tree op0 = gimple_assign_rhs1 (stmt);
> + tree op1 = gimple_assign_rhs2 (stmt);
> + if (op0 == operand)
> + {
> + *addend = op1;
> + *operand_first = true;
> + }
> + else if (op1 == operand)
> + {
> + *addend = op0;
> + *operand_first = false;
> + }
> + else
> + return false;
> +
> + return types_compatible_p (TREE_TYPE (sum), TREE_TYPE (*addend));
PLUS_EXPR should constrain this to be always true.
> +}
> +
> +/* Return true if BASE is a loop-header PHI whose latch value is defined by
> + the sole use of PHI. */
> +
> +static bool
> +ifc_phi_feeds_loop_reduction_p (gphi *phi, tree base)
> +{
> + use_operand_p use_p;
> + gimple *use_stmt;
> +
> + if (TREE_CODE (base) != SSA_NAME
> + || gimple_code (SSA_NAME_DEF_STMT (base)) != GIMPLE_PHI
> + || !single_imm_use (gimple_phi_result (phi), &use_p, &use_stmt)
> + || !is_gimple_assign (use_stmt))
Could be a .IFN operation like .FMAX? So I think is_gimple_assign
(use_stmt) is premature here.
> + return false;
> +
> + gphi *header_phi = as_a <gphi *> (SSA_NAME_DEF_STMT (base));
a little bit refactoring and using dyn_cast <gphi *> could simplify
the above.
> + class loop *loop = gimple_bb (phi)->loop_father;
> + if (gimple_bb (header_phi) != loop->header)
> + return false;
> +
> + return (PHI_ARG_DEF_FROM_EDGE (header_phi, loop_latch_edge (loop))
> + == gimple_assign_lhs (use_stmt));
so relaxing the is_gimple_assign check would be
PHI_ARG_DEF_FROM_EDGE (..) == SSA_NAME && SSA_NAME_DEF_STMT (...) ==
use_stmt.
> +}
> +
> +/* Return a vectorizer-friendly form of PHI if it selects between BASE +
> ADDEND
> + and BASE. This keeps the reduction value in a single add chain after
> + if-conversion:
> +
> + COND ? BASE + ADDEND : BASE -> BASE + (COND ? ADDEND : 0)
> + COND ? BASE : BASE + ADDEND -> BASE + (!COND ? ADDEND : 0)
> +
> + Preserve the order of the two addition operands if ADDEND comes first.
> +
> + PHI is the if-converted PHI whose result is being replaced. GSI is the
> + insertion point for the temporary conditional addend.
> +
> + When the original expression chooses BASE alone, this version adds zero.
> + For floating-point values, use -0.0 when the sign of zero matters, and
> + check that adding it leaves BASE unchanged. For integers, the chosen
> + value is either ADDEND or zero, so the rewrite cannot cause new overflow.
> + Return null if the rewrite is unsafe. */
> +
> +static tree
> +fold_build_cond_reduction_expr (gphi *phi, tree type, tree cond, tree rhs,
> + tree lhs, gimple_stmt_iterator *gsi)
> +{
> + if (!SCALAR_FLOAT_TYPE_P (type)
> + && (!ANY_INTEGRAL_TYPE_P (type) || TYPE_OVERFLOW_TRAPS (type)))
why ANY_? do we really want to allow vectors?
> + return NULL_TREE;
> +
> + tree base, addend;
> + bool base_first;
> + if (ifc_plus_with_operand_p (rhs, lhs, &addend, &base_first))
> + base = lhs;
> + else if (ifc_plus_with_operand_p (lhs, rhs, &addend, &base_first))
> + {
> + base = rhs;
> + cond = invert_truthvalue (cond);
> + }
> + else
> + return NULL_TREE;
> +
> + if (!ifc_phi_feeds_loop_reduction_p (phi, base))
> + return NULL_TREE;
> +
> + tree zero = SCALAR_FLOAT_TYPE_P (type) && HONOR_SIGNED_ZEROS (type)
> + ? build_real (type, dconstm0) : build_zero_cst (type);
> + if (SCALAR_FLOAT_TYPE_P (type)
> + && !fold_real_zero_addition_p (type, base, zero, 0))
> + return NULL_TREE;
Didn't we say using -0.0 is OK when doing in-order reductions? So
why the extra check? Adding a comment is worthwhile though.
> +
> + tree expr = fold_build_cond_expr (type, cond, addend, zero);
> + tree tmp = ifc_temp_var (type, expr, gsi);
> + return base_first ? build2 (PLUS_EXPR, type, base, tmp)
> + : build2 (PLUS_EXPR, type, tmp, base);
> +}
> +
> /* Add condition NC to the predicate list of basic block BB. LOOP is
> the loop to be if-converted. Use predicate of cd-equivalent block
> for join bb if it exists: we call basic blocks bb1 and bb2
> @@ -2488,8 +2600,14 @@ predicate_scalar_phi (gphi *phi, gimple_stmt_iterator
> *gsi)
> }
> else
> /* Build new RHS using selected condition and arguments. */
> - rhs = fold_build_cond_expr (TREE_TYPE (res), unshare_expr (cond),
> - arg0, arg1);
> + {
> + tree type = TREE_TYPE (res);
> + rhs = fold_build_cond_reduction_expr (phi, type, cond, arg0,
> + arg1, gsi);
> + if (!rhs)
> + rhs = fold_build_cond_expr (type, unshare_expr (cond), arg0,
> + arg1);
> + }
> new_stmt = gimple_build_assign (res, rhs);
> gsi_insert_before (gsi, new_stmt, GSI_SAME_STMT);
> gimple_stmt_iterator new_gsi = gsi_for_stmt (new_stmt);
> @@ -2574,9 +2692,16 @@ predicate_scalar_phi (gphi *phi, gimple_stmt_iterator
> *gsi)
> NULL_TREE, true, GSI_SAME_STMT);
> if (!(is_cond_scalar_reduction (phi, &reduc, arg0 , arg1,
> &op0, &op1, true, &has_nop, &nop_reduc)))
> - rhs = fold_build_cond_expr (TREE_TYPE (res), unshare_expr (cond),
> - swap ? arg1 : arg0,
> - swap ? arg0 : arg1);
> + {
> + tree type = TREE_TYPE (res);
> + tree rhs_arg = swap ? arg1 : arg0;
> + tree lhs_arg = swap ? arg0 : arg1;
> + rhs = fold_build_cond_reduction_expr (phi, type, cond, rhs_arg,
> + lhs_arg, gsi);
> + if (!rhs)
> + rhs = fold_build_cond_expr (type, unshare_expr (cond), rhs_arg,
> + lhs_arg);
> + }
So seeing the above two hunks I wonder whether it would be better
structured to enhance
is_cond_scalar_reduction and convert_scalar_cond_reduction?
I didn't really get what it is that makes the new case differ?
That said, I'm OK keeping them separate but the APIs differing
(query + transform vs. try-transform) is a bit off-putting.
Thanks,
Richard.
> else
> {
> /* Convert reduction stmt into vectorizable form. */
>
>
>
--
Richard Biener <[email protected]>
SUSE Software Solutions Germany GmbH,
Frankenstrasse 146, 90461 Nuernberg, Germany;
GF: Stefan Gaiser, Jochen Jaser, Abhinav Puri; (HRB 36809, AG Nuernberg)
