On Mon, 23 Feb 2026, Tamar Christina wrote:
> > -----Original Message-----
> > From: Richard Biener <[email protected]>
> > Sent: 23 February 2026 12:36
> > To: Tamar Christina <[email protected]>
> > Cc: [email protected]; nd <[email protected]>; [email protected]
> > Subject: RE: [PATCH][vect]: enforce trapping math checks when early breaks
> > [PR124142]
> >
> > On Mon, 23 Feb 2026, Tamar Christina wrote:
> >
> > > > -----Original Message-----
> > > > From: Richard Biener <[email protected]>
> > > > Sent: 23 February 2026 12:07
> > > > To: Tamar Christina <[email protected]>
> > > > Cc: [email protected]; nd <[email protected]>; [email protected]
> > > > Subject: Re: [PATCH][vect]: enforce trapping math checks when early
> > breaks
> > > > [PR124142]
> > > >
> > > > On Mon, 23 Feb 2026, Tamar Christina wrote:
> > > >
> > > > > The loop
> > > > >
> > > > > void
> > > > > __attribute__((noipa))
> > > > > bug (int f, int *w, int l)
> > > > > {
> > > > > int i;
> > > > > for (i = 0; i < l; ++i)
> > > > > while (f % w[i]) w[i]--;
> > > > > }
> > > > >
> > > > > is an uncounted loop which performs a trapping operation during
> > > > vectorization.
> > > > > Normally the vectorizer doesn't stop vectorization if the operation
> > > > > can't
> > be
> > > > > masked (though ifcvt does).
> > > > >
> > > > > For Early breaks however this is unsafe as we are introducing a trap
> > where
> > > > the
> > > > > original scalar code may not have.
> > > > >
> > > > > reductions however don't require masking when the scalar epilogue is
> > used
> > > > to
> > > > > restart the last iteration as we use the previous value of the vector
> > > > > in that
> > > > > case and don't reduce "inactive" lanes. This would need to be
> > > > > adjusted
> > > > when we
> > > > > support staying inside the vector loop directly.
> > > > >
> > > > > Some tests now fail to vectorize (including some tsvc ones) which I
> > > > > could
> > > > have
> > > > > xfail'ed but instead decided to add -fno-trapping-math just to get the
> > > > > additional coverage they provide.
> > > > >
> > > > > Bootstrapped Regtested on aarch64-none-linux-gnu,
> > > > > arm-none-linux-gnueabihf, x86_64-pc-linux-gnu
> > > > > -m32, -m64 and no issues.
> > > > >
> > > > > Any objections?
> > > >
> > > > I was expecting the check to be in vect_analyze_early_break_dependences
> > > > since I think there is a special thing about early break vectorization
> > > > when a trapping operation is before the early exit condition or does
> > > > feed it, like for the testcase at hand. Those operations cannot be
> > > > masked since the very mask is not yet computed.
> > >
> > > I see your point.
> > >
> > > operations before the early break condition but not part of the condition
> > > are also not valid. E.g.
> > >
> > > b = a / d;
> > > if (a > c)
> > > break;
> > >
> > > f[i] = b;
> > >
> > > because we'd be performing the division for iterations that may not have
> > > happened for the scalar loop.
> > >
> > > So the only "safe" trapping operations are those between the last early
> > > exit and the latch exit.
> > >
> > > For them to be safe they would have to be moved similar to stores.
> > >
> > > I can move the check there, but I'd leave the trap stmt motion for now..
> >
> > In particular a possibly trapping operation or call _after_ the
> > early break condition should be OK if we exit the loop and defer
> > the last iteration to the epilog (once we stop that, those need to
> > be masked of course).
>
> Yeah, I was just saying that in the above example, there's no reason we
> couldn't
> move b until after all early break statements and still be able to vectorize
> correctly because b isn't actually needed in that BB and can be sunk.
>
> >
> > Your patch does not catch the case where the trapping stmt is
> > before the early exit but we can use partial vectors - the mask
> > would be wrong (from the main IV) at that point, no?
>
> Definitely, I wasn't arguing that. I was just saying that the above case
> where
> b could be sunk is valid after sinking, that I'd like to punt that particular
> optimization
> to the next release and for now reject the case above.
>
> That is to say, are you ok with just rejecting the case above (where a
> trapping statement that's not part of the break condition and could be
> sunk and made valid)
Sure, that sounds reasonable (tree sinking should have sunk it anyway,
unless it was loop ifconverted late).
Richard.
>
> Or did you want me to actually try to handle the sinking now.
>
> Thanks,
> Tamar
> >
> > Richard.
> >
> > >
> > > Thanks,
> > > Tamar
> > >
> > > >
> > > > There might be other latent issues with trapping operations but I'd
> > > > rather split those out.
> > > >
> > > > So, am I missing something?
> > > >
> > > > Thanks,
> > > > Richard.
> > > >
> > > > > Thanks,
> > > > > Tamar
> > > > >
> > > > > gcc/ChangeLog:
> > > > >
> > > > > PR tree-optimization/124142
> > > > > * tree-vect-stmts.cc (vectorizable_call,
> > > > > vectorizable_operation):
> > > > > For early break require masking when operation can trap.
> > > > >
> > > > > gcc/testsuite/ChangeLog:
> > > > >
> > > > > PR tree-optimization/124142
> > > > > * gcc.dg/vect/tsvc/vect-tsvc-s481.c: Add -fno-trapping-math.
> > > > > * gcc.dg/vect/tsvc/vect-tsvc-s482.c: Likewise.
> > > > > * gcc.dg/vect/vect-early-break_61.c: Likewise.
> > > > > * gcc.target/aarch64/vect-early-break-cbranch_3.c: Likewise.
> > > > > * gcc.dg/vect/vect-early-break_143-pr124142.c: New test.
> > > > >
> > > > > ---
> > > > > diff --git a/gcc/testsuite/gcc.dg/vect/tsvc/vect-tsvc-s481.c
> > > > b/gcc/testsuite/gcc.dg/vect/tsvc/vect-tsvc-s481.c
> > > > > index
> > > >
> > e4433385d6686806a75fffe22f90e3bfb603564c..23c9961691d2f0c952894
> > > > 9c5c44dae3e264dc3ec 100644
> > > > > --- a/gcc/testsuite/gcc.dg/vect/tsvc/vect-tsvc-s481.c
> > > > > +++ b/gcc/testsuite/gcc.dg/vect/tsvc/vect-tsvc-s481.c
> > > > > @@ -1,7 +1,7 @@
> > > > > /* This file is distributed under the University of Illinois Open
> > > > > Source
> > > > > License. See license.txt for details. */
> > > > >
> > > > > -/* { dg-additional-options "--param vect-epilogues-nomask=0" } */
> > > > > +/* { dg-additional-options "--param vect-epilogues-nomask=0 -fno-
> > > > trapping-math" } */
> > > > > /* { dg-require-effective-target vect_float } */
> > > > > /* { dg-require-effective-target vect_early_break_hw } */
> > > > > /* { dg-add-options vect_early_break } */
> > > > > @@ -41,4 +41,4 @@ int main (int argc, char **argv)
> > > > > return 0;
> > > > > }
> > > > >
> > > > > -/* { dg-final { scan-tree-dump "vectorized 1 loops" "vect" { xfail {
> > > > > !
> > > > vect_early_break} } } } */
> > > > > +/* { dg-final { scan-tree-dump "vectorized 1 loops" "vect" { xfail {
> > > > > !
> > > > vect_early_break } } } } */
> > > > > diff --git a/gcc/testsuite/gcc.dg/vect/tsvc/vect-tsvc-s482.c
> > > > b/gcc/testsuite/gcc.dg/vect/tsvc/vect-tsvc-s482.c
> > > > > index
> > > >
> > 146df409ecc64e9535583c0d0083469d7aa24031..38426845a3d115432bfc
> > > > 68d9e1e7539fe10e27e2 100644
> > > > > --- a/gcc/testsuite/gcc.dg/vect/tsvc/vect-tsvc-s482.c
> > > > > +++ b/gcc/testsuite/gcc.dg/vect/tsvc/vect-tsvc-s482.c
> > > > > @@ -1,7 +1,7 @@
> > > > > /* This file is distributed under the University of Illinois Open
> > > > > Source
> > > > > License. See license.txt for details. */
> > > > >
> > > > > -/* { dg-additional-options "--param vect-epilogues-nomask=0" } */
> > > > > +/* { dg-additional-options "--param vect-epilogues-nomask=0 -fno-
> > > > trapping-math" } */
> > > > > /* { dg-require-effective-target vect_float } */
> > > > > /* { dg-require-effective-target vect_early_break_hw } */
> > > > > /* { dg-add-options vect_early_break } */
> > > > > diff --git a/gcc/testsuite/gcc.dg/vect/vect-early-break_143-pr124142.c
> > > > b/gcc/testsuite/gcc.dg/vect/vect-early-break_143-pr124142.c
> > > > > new file mode 100644
> > > > > index
> > > >
> > 0000000000000000000000000000000000000000..b9141e3f15a9d15211
> > > > 8c2d8e84fc45d5e83066cb
> > > > > --- /dev/null
> > > > > +++ b/gcc/testsuite/gcc.dg/vect/vect-early-break_143-pr124142.c
> > > > > @@ -0,0 +1,16 @@
> > > > > +/* { dg-do compile } */
> > > > > +/* { dg-add-options vect_early_break } */
> > > > > +/* { dg-require-effective-target vect_early_break } */
> > > > > +/* { dg-require-effective-target vect_int } */
> > > > > +
> > > > > +/* { dg-final { scan-tree-dump-not "LOOP VECTORIZED" "vect" } } */
> > > > > +
> > > > > +void
> > > > > +__attribute__((noipa))
> > > > > +bug (int f, int *w, int l)
> > > > > +{
> > > > > + int i;
> > > > > + for (i = 0; i < l; ++i)
> > > > > + while (f % w[i]) w[i]--;
> > > > > +}
> > > > > +
> > > > > diff --git a/gcc/testsuite/gcc.dg/vect/vect-early-break_61.c
> > > > b/gcc/testsuite/gcc.dg/vect/vect-early-break_61.c
> > > > > index
> > > >
> > c789ec01f32c6b958c6a3663531a7b7517b94477..0cd20549cdc656baad2b
> > > > 5561e39ca999bca98bb0 100644
> > > > > --- a/gcc/testsuite/gcc.dg/vect/vect-early-break_61.c
> > > > > +++ b/gcc/testsuite/gcc.dg/vect/vect-early-break_61.c
> > > > > @@ -2,6 +2,7 @@
> > > > > /* { dg-do compile } */
> > > > > /* { dg-require-effective-target vect_early_break } */
> > > > > /* { dg-require-effective-target vect_float } */
> > > > > +/* { dg-additional-options "-fno-trapping-math" } */
> > > > >
> > > > > typedef float real_t;
> > > > > __attribute__((aligned(64))) real_t a[32000], b[32000], c[32000];
> > > > > diff --git a/gcc/testsuite/gcc.target/aarch64/vect-early-break-
> > cbranch_3.c
> > > > b/gcc/testsuite/gcc.target/aarch64/vect-early-break-cbranch_3.c
> > > > > index
> > > >
> > 8980b9f04f9a15ded9e90954e25b2c0578681761..b252b80ef291efd4f6fc9
> > > > 8224a44ef7fadc68d6c 100644
> > > > > --- a/gcc/testsuite/gcc.target/aarch64/vect-early-break-cbranch_3.c
> > > > > +++ b/gcc/testsuite/gcc.target/aarch64/vect-early-break-cbranch_3.c
> > > > > @@ -1,5 +1,5 @@
> > > > > /* { dg-do compile } */
> > > > > -/* { dg-options "-O3 -fno-schedule-insns -fno-reorder-blocks -fno-
> > > > schedule-insns2 --param aarch64-autovec-preference=asimd-only" } */
> > > > > +/* { dg-options "-O3 -fno-trapping-math -fno-schedule-insns -fno-
> > reorder-
> > > > blocks -fno-schedule-insns2 --param aarch64-autovec-preference=asimd-
> > > > only" } */
> > > > > /* { dg-final { check-function-bodies "**" "" "" { target lp64 } } }
> > > > > */
> > > > >
> > > > > #pragma GCC target "+sve"
> > > > > diff --git a/gcc/tree-vect-stmts.cc b/gcc/tree-vect-stmts.cc
> > > > > index
> > > >
> > 22285250aa8d4f600721c647a83d1b2bafe7ef2a..cb7217ad6f92ad4b2bd54
> > > > 3733c802606f3fba03e 100644
> > > > > --- a/gcc/tree-vect-stmts.cc
> > > > > +++ b/gcc/tree-vect-stmts.cc
> > > > > @@ -3650,6 +3650,18 @@ vectorizable_call (vec_info *vinfo,
> > > > > LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo) = false;
> > > > > }
> > > > > }
> > > > > +
> > > > > + /* If the operation traps, and we're an early break loop then
> > > > > don't
> > allow
> > > > > + vectorization if masking isn't supported. */
> > > > > + if (LOOP_VINFO_EARLY_BREAKS (loop_vinfo))
> > > > > + {
> > > > > + if (dump_enabled_p ())
> > > > > + dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
> > > > > + "can't use a fully-masked loop and loop
> > > > > has "
> > > > > + "multiple exits. Cannot vectorize as
> > > > > operation "
> > > > > + "may trap.\n");
> > > > > + return false;
> > > > > + }
> > > > > }
> > > > >
> > > > > /* If that fails, try asking for a target-specific built-in
> > > > > function. */
> > > > > @@ -6792,7 +6804,8 @@ vectorizable_operation (vec_info *vinfo,
> > > > > Similarly, if this operation is part of a reduction, a
> > > > > fully-masked
> > > > > loop should only change the active lanes of the reduction chain,
> > > > > keeping the inactive lanes as-is. */
> > > > > - bool mask_out_inactive = ((!is_invariant && gimple_could_trap_p
> > (stmt))
> > > > > + bool could_trap_p = gimple_could_trap_p (stmt);
> > > > > + bool mask_out_inactive = ((!is_invariant && could_trap_p)
> > > > > || reduc_idx >= 0);
> > > > >
> > > > > if (cost_vec) /* transformation not required. */
> > > > > @@ -6821,6 +6834,23 @@ vectorizable_operation (vec_info *vinfo,
> > > > > }
> > > > > }
> > > > >
> > > > > + /* If the operation traps, and we're an early break loop then
> > > > > don't
> > allow
> > > > > + vectorization if masking isn't supported. Reductions are OK
> > > > > because if
> > > > > + we take an early exit, the epilog will get the value of the
> > > > > previous
> > > > > + iterations and we recompute the remainder. */
> > > > > + if (loop_vinfo
> > > > > + && could_trap_p
> > > > > + && !LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo)
> > > > > + && LOOP_VINFO_EARLY_BREAKS (loop_vinfo))
> > > > > + {
> > > > > + if (dump_enabled_p ())
> > > > > + dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
> > > > > + "can't use a fully-masked loop and loop
> > > > > has "
> > > > > + "multiple exits. Cannot vectorize as
> > > > > operation "
> > > > > + "may trap.\n");
> > > > > + return false;
> > > > > + }
> > > > > +
> > > > > /* Put types on constant and invariant SLP children. */
> > > > > if (!vect_maybe_update_slp_op_vectype (slp_op0, vectype)
> > > > > || !vect_maybe_update_slp_op_vectype (slp_op1, vectype)
> > > > >
> > > > >
> > > > >
> > > >
> > > > --
> > > > Richard Biener <[email protected]>
> > > > SUSE Software Solutions Germany GmbH,
> > > > Frankenstrasse 146, 90461 Nuernberg, Germany;
> > > > GF: Jochen Jaser, Andrew McDonald, Werner Knoblich; (HRB 36809, AG
> > > > Nuernberg)
> > >
> >
> > --
> > Richard Biener <[email protected]>
> > SUSE Software Solutions Germany GmbH,
> > Frankenstrasse 146, 90461 Nuernberg, Germany;
> > GF: Jochen Jaser, Andrew McDonald, Werner Knoblich; (HRB 36809, AG
> > Nuernberg)
>
--
Richard Biener <[email protected]>
SUSE Software Solutions Germany GmbH,
Frankenstrasse 146, 90461 Nuernberg, Germany;
GF: Jochen Jaser, Andrew McDonald, Werner Knoblich; (HRB 36809, AG Nuernberg)
