On 7/23/2024 11:20 AM, Richard Sandiford wrote:
Edwin Lu <e...@rivosinc.com> writes:
On 7/23/2024 4:56 AM, Richard Biener wrote:
On Tue, Jul 23, 2024 at 1:03 AM Edwin Lu <e...@rivosinc.com> wrote:
Hi Richard,
On 5/31/2024 1:48 AM, Richard Biener wrote:
On Thu, May 30, 2024 at 2:11 AM Patrick O'Neill <patr...@rivosinc.com> wrote:
From: Greg McGary <g...@rivosinc.com>
Still a NACK. If remain ends up zero then
/* Try to use a single smaller load when we are about
to load excess elements compared to the unrolled
scalar loop. */
if (known_gt ((vec_num * j + i + 1) * nunits,
(group_size * vf - gap)))
{
poly_uint64 remain = ((group_size * vf - gap)
- (vec_num * j + i) * nunits);
if (known_ge ((vec_num * j + i + 1) * nunits
- (group_size * vf - gap), nunits))
/* DR will be unused. */
ltype = NULL_TREE;
needs to be re-formulated so that the combined conditions make sure
this doesn't happen. The outer known_gt should already ensure that
remain > 0. For correctness that should possibly be maybe_gt though.
Yeah. FWIW, I mentioned the maybe_gt thing in
https://gcc.gnu.org/pipermail/gcc-patches/2024-May/653013.html:
Pre-existing, but shouldn't this be maybe_gt rather than known_gt?
We can only skip doing it if we know for sure that the load won't cross
the gap. (Not sure whether the difference can trigger in practice.)
But AFAICT, the known_gt doesn't inherently prove that remain is known
to be nonzero. It just proves that the gap between the end of the scalar
accesses and the end of this vector is known to be nonzero.
Putting the list back in the loop and CCing Richard S.
I'm currently looking into this patch and am trying to figure out what
is going on. Stepping through gdb, I see that remain == {coeffs = {0,
2}} and nunits == {coeffs = {2, 2}} (the outer known_gt returned true
with known_gt({coeffs = {8, 8}}, {coeffs = {6, 8}})).
From what I understand, this falls under the umbrella of 0 <= remain <
nunits. The divide by zero error is because of the 0 <= remain which is
coming from the constant_multiple_p function in poly-int.h where it
performs the modulus NCa(a.coeffs[0]) % NCb(b.coeffs[0]).
(https://github.com/gcc-mirror/gcc/blob/master/gcc/poly-int.h#L1970-L1971)
> if (known_ge ((vec_num * j + i + 1) * nunits
> - (group_size * vf - gap),
nunits))
> /* DR will be unused. */
> ltype = NULL_TREE;
This if condition is a bit suspicious to me though. I'm seeing that it's
evaluating known_ge({coeffs = {2, 0}}, {coeffs = {2, 2}}) which is
returning false. Should it be maybe_ge instead?
No, we can only not emit a load if we know it won't be used, not if
it eventually cannot be used.
Agreed.
[switching round for easier reply]
After running some
tests, to me it looks like it doesn't vectorize quite as often; however,
I'm not fully sure what else to do when the coeffs can potentially be
equal to 0.
Should it even be possible for there to be a {coeffs = {0, n}}
situation? My understanding of how poly_ints are used for representing
vectorization is that the first coefficient is the number of elements
needed to make the minimum supported vector size. That is, if vector
lengths are 128 bits, element size is 32 bits, coeff[0] should be
minimum of 4. Is this understanding correct?
I was told n can be negative, but nunits.coeff[0] should be non-zero.
What would it mean for the coeffs[0] to be 0? Would that mean the vector length
supports 0 bits?
coeffs = {A,B} just means A+B*X, where X is the number of vector
"chunks" beyond the minimum length. It's certainly valid for a poly_int
to have a zero coeffs[0] (i.e. zero A). For example, (the length of a
vector) - (the minimum length) would have this property.
Thanks for the explanation! I have a few clarification questions about this.
If I understand correctly, B would represent the number of elements the vector
can have (for 128b vector operating on 32b elements, B == 4, but if operating
on 64b elements B == 2); however, I'm not too sure what A represents.
On the poly_int docs, it says
An indeterminate value of 0 should usually represent the minimum possible
runtime value, with c0 specifying the value in that case.
"minimum possible runtime value" doesn't make sense to me. Does it mean the
potential minimum bound of elements it will operate on?
What is j and i when the divisor is zero?
The values I see in gdb are: vec_num = 4 j = 0 i = 3 vf = {coeffs = {2,
2}} nunits = {coeffs = {2, 2}} group_size = 4 gap = 2 vect_align = 2
remain = {coeffs = {0, 2}}
OK, so let's use D to mean "data" and G to mean "gap". Then, for the
minimum vector length of 2 elements, we have:
DD GG DD GG
The last load will read beyond the scalar loop if the vector loop happens
to handle all elements of the scalar loop.
For a vector length of 4 elements, we have:
DDGG DDGG DDGG DDGG
where every load contains both data and gaps. The same will be true
for larger vectors.
That's where remain={0,2} is coming from. The last load is fully redundant
for the minimum VL but not for larger VL.
I think I understand the example but just want to see if I fully understand.
If we have 7 data elements with minimum vector length of 2 elements, the vector
loop would essentially run 3 times on
DD GG DD GG DD GG
There is an extra Data element which hasn't been processed. Since there is one
element left over and is less than the minimum vector length, we can either
1. run scalar execution of the number of remaining elements (finishing scalar
loop) or
2. see if a smaller vector size can process it nicely (i.e. minimum vector
length of 1 element in this case. remain now is {1, 1}) or
3. use the same vector size but mask out the remaining element
Is this a correct?
Based on that, the patch below looks correct to me, but I might have
misunderstood the intent.
As an alternative to the original patch, would this also make sense?
diff --git a/gcc/tree-vect-stmts.cc b/gcc/tree-vect-stmts.cc
index aab3aa59962..cd657ac63af 100644
--- a/gcc/tree-vect-stmts.cc
+++ b/gcc/tree-vect-stmts.cc
@@ -11479,7 +11479,7 @@ vectorizable_load (vec_info *vinfo,
/* Try to use a single smaller load when we are about
to load excess elements compared to the unrolled
scalar loop. */
- if (known_gt ((vec_num * j + i + 1) * nunits,
+ if (maybe_gt ((vec_num * j + i + 1) * nunits,
(group_size * vf - gap)))
{
poly_uint64 remain = ((group_size * vf - gap)
@@ -11502,6 +11502,10 @@ vectorizable_load (vec_info *vinfo,
/* Aligned access to the gap area when there's
at least one element in it is OK. */
;
+ else if (maybe_eq (remain, 0))
+ /* Handle remain.coeffs[0] == 0 case. Number of
+ elements is an exact multiple of vector length. */
+ ;
else
{
/* remain should now be > 0 and < nunits. */
Thanks,
Richard
Edwin
gcc/ChangeLog:
* gcc/tree-vect-stmts.cc (gcc/tree-vect-stmts.cc): Prevent
divide-by-zero.
* testsuite/gcc.target/riscv/rvv/autovec/no-segment.c: Remove dg-ice.
---
No changes in v3. Depends on the risc-v backend option added in patch 1 to
trigger the ICE.
---
gcc/testsuite/gcc.target/riscv/rvv/autovec/no-segment.c | 1 -
gcc/tree-vect-stmts.cc | 3 ++-
2 files changed, 2 insertions(+), 2 deletions(-)
diff --git a/gcc/testsuite/gcc.target/riscv/rvv/autovec/no-segment.c
b/gcc/testsuite/gcc.target/riscv/rvv/autovec/no-segment.c
index dfbe09f01a1..79d03612a22 100644
--- a/gcc/testsuite/gcc.target/riscv/rvv/autovec/no-segment.c
+++ b/gcc/testsuite/gcc.target/riscv/rvv/autovec/no-segment.c
@@ -1,6 +1,5 @@
/* { dg-do compile } */
/* { dg-options "-march=rv64gcv -mabi=lp64d -mrvv-vector-bits=scalable -O3
-mno-autovec-segment" } */
-/* { dg-ice "Floating point exception" } */
enum e { c, d };
enum g { f };
diff --git a/gcc/tree-vect-stmts.cc b/gcc/tree-vect-stmts.cc
index 4219ad832db..34f5736ba00 100644
--- a/gcc/tree-vect-stmts.cc
+++ b/gcc/tree-vect-stmts.cc
@@ -11558,7 +11558,8 @@ vectorizable_load (vec_info *vinfo,
- (vec_num * j + i) * nunits);
/* remain should now be > 0 and < nunits. */
unsigned num;
- if (constant_multiple_p (nunits, remain, &num))
+ if (known_gt (remain, 0)
+ && constant_multiple_p (nunits, remain, &num))
{
tree ptype;
new_vtype
--
2.43.2
Edwin
