Re: [PATCH] Tweak gcc.target/i386/avx2-vp{add,sub}q-3.c (PR target/82851)
On Thu, Feb 22, 2018 at 7:16 PM, Jakub Jelinek wrote:
> Hi!
>
> These tests FAIL the vp.*q.*ymm insn scan with some tunings, e.g.
> -mtune=silvermont or -mtune=atom, because vectorizing it using AVX2
> is based on costs considered too expensive.
> E.g. for -mtune=silvermont I see for VF 4:
>
> avx2-vpop-check.h:16:3: note: Cost model analysis:
> Vector inside of loop cost: 330
> Vector prologue cost: 0
> Vector epilogue cost: 0
> Scalar iteration cost: 48
> Scalar outside cost: 0
> Vector outside cost: 0
> prologue iterations: 0
> epilogue iterations: 0
> avx2-vpop-check.h:16:3: note: cost model: the vector iteration cost = 330
> divided by the scalar iteration cost = 48 is greater or equal to the vect
> orization factor = 4.
> avx2-vpop-check.h:16:3: note: not vectorized: vectorization not profitable.
> avx2-vpop-check.h:16:3: note: not vectorized: vector version will never be
> profitable.
> avx2-vpop-check.h:16:3: note: * Re-trying analysis with vector size 16
>
> while with -mtune=generic I see:
>
> avx2-vpop-check.h:16:3: note: Cost model analysis:
> Vector inside of loop cost: 64
> Vector prologue cost: 0
> Vector epilogue cost: 0
> Scalar iteration cost: 40
> Scalar outside cost: 0
> Vector outside cost: 0
> prologue iterations: 0
> epilogue iterations: 0
> Calculated minimum iters for profitability: 0
>
> I think we can either use -mtune=generic as in the patch below,
> or add -fno-vect-cost-model.
>
> Tested on x86_64-linux with:
> make check-gcc
> RUNTESTFLAGS='--target_board=unix\{-m32,-m32/-mtune=silvermont,-m32/-mtune=atom,-m64,-m64/-mtune=silvermont,-m64/-mtune=atom\}
> i386.exp=avx2-vp*q-3.c'
> Ok for trunk?
>
> 2018-02-22 Jakub Jelinek
>
> PR target/82851
> * gcc.target/i386/avx2-vpaddq-3.c: Add -mtune=generic to dg-options.
> * gcc.target/i386/avx2-vpsubq-3.c: Likewise.
OK.
Thanks,
Uros.
> --- gcc/testsuite/gcc.target/i386/avx2-vpaddq-3.c.jj2015-05-29
> 15:02:59.475249515 +0200
> +++ gcc/testsuite/gcc.target/i386/avx2-vpaddq-3.c 2018-02-22
> 18:44:45.353178253 +0100
> @@ -1,5 +1,5 @@
> /* { dg-do run } */
> -/* { dg-options "-mavx2 -mno-prefer-avx128 -O2 -ftree-vectorize -save-temps"
> } */
> +/* { dg-options "-mavx2 -mno-prefer-avx128 -O2 -ftree-vectorize
> -mtune=generic -save-temps" } */
> /* { dg-require-effective-target avx2 } */
>
>
> --- gcc/testsuite/gcc.target/i386/avx2-vpsubq-3.c.jj2015-05-29
> 15:03:00.572232560 +0200
> +++ gcc/testsuite/gcc.target/i386/avx2-vpsubq-3.c 2018-02-22
> 18:45:02.131168864 +0100
> @@ -1,5 +1,5 @@
> /* { dg-do run } */
> -/* { dg-options "-mavx2 -mno-prefer-avx128 -O2 -ftree-vectorize -save-temps"
> } */
> +/* { dg-options "-mavx2 -mno-prefer-avx128 -O2 -ftree-vectorize
> -mtune=generic -save-temps" } */
> /* { dg-require-effective-target avx2 } */
>
>
>
> Jakub
[PATCH] Tweak gcc.target/i386/avx2-vp{add,sub}q-3.c (PR target/82851)
Hi!
These tests FAIL the vp.*q.*ymm insn scan with some tunings, e.g.
-mtune=silvermont or -mtune=atom, because vectorizing it using AVX2
is based on costs considered too expensive.
E.g. for -mtune=silvermont I see for VF 4:
avx2-vpop-check.h:16:3: note: Cost model analysis:
Vector inside of loop cost: 330
Vector prologue cost: 0
Vector epilogue cost: 0
Scalar iteration cost: 48
Scalar outside cost: 0
Vector outside cost: 0
prologue iterations: 0
epilogue iterations: 0
avx2-vpop-check.h:16:3: note: cost model: the vector iteration cost = 330
divided by the scalar iteration cost = 48 is greater or equal to the vect
orization factor = 4.
avx2-vpop-check.h:16:3: note: not vectorized: vectorization not profitable.
avx2-vpop-check.h:16:3: note: not vectorized: vector version will never be
profitable.
avx2-vpop-check.h:16:3: note: * Re-trying analysis with vector size 16
while with -mtune=generic I see:
avx2-vpop-check.h:16:3: note: Cost model analysis:
Vector inside of loop cost: 64
Vector prologue cost: 0
Vector epilogue cost: 0
Scalar iteration cost: 40
Scalar outside cost: 0
Vector outside cost: 0
prologue iterations: 0
epilogue iterations: 0
Calculated minimum iters for profitability: 0
I think we can either use -mtune=generic as in the patch below,
or add -fno-vect-cost-model.
Tested on x86_64-linux with:
make check-gcc
RUNTESTFLAGS='--target_board=unix\{-m32,-m32/-mtune=silvermont,-m32/-mtune=atom,-m64,-m64/-mtune=silvermont,-m64/-mtune=atom\}
i386.exp=avx2-vp*q-3.c'
Ok for trunk?
2018-02-22 Jakub Jelinek
PR target/82851
* gcc.target/i386/avx2-vpaddq-3.c: Add -mtune=generic to dg-options.
* gcc.target/i386/avx2-vpsubq-3.c: Likewise.
--- gcc/testsuite/gcc.target/i386/avx2-vpaddq-3.c.jj2015-05-29
15:02:59.475249515 +0200
+++ gcc/testsuite/gcc.target/i386/avx2-vpaddq-3.c 2018-02-22
18:44:45.353178253 +0100
@@ -1,5 +1,5 @@
/* { dg-do run } */
-/* { dg-options "-mavx2 -mno-prefer-avx128 -O2 -ftree-vectorize -save-temps" }
*/
+/* { dg-options "-mavx2 -mno-prefer-avx128 -O2 -ftree-vectorize -mtune=generic
-save-temps" } */
/* { dg-require-effective-target avx2 } */
--- gcc/testsuite/gcc.target/i386/avx2-vpsubq-3.c.jj2015-05-29
15:03:00.572232560 +0200
+++ gcc/testsuite/gcc.target/i386/avx2-vpsubq-3.c 2018-02-22
18:45:02.131168864 +0100
@@ -1,5 +1,5 @@
/* { dg-do run } */
-/* { dg-options "-mavx2 -mno-prefer-avx128 -O2 -ftree-vectorize -save-temps" }
*/
+/* { dg-options "-mavx2 -mno-prefer-avx128 -O2 -ftree-vectorize -mtune=generic
-save-temps" } */
/* { dg-require-effective-target avx2 } */
Jakub
