| Message ID | BANLkTimb5zs8eCUCQ2eyZFw-d=FerEY9Wg@mail.gmail.com |
|---|---|
| State | New |
| Headers | show |
On 9 June 2011 13:00, Ira Rosen <ira.rosen@linaro.org> wrote: > Hi, > > This patch fixes several vectorizer testsuite failures on ARM: > - vect-16.c checks that the vectorization fails without -ffast-math, > but -ffast-math is a default flag for vector tests on ARM. I renamed > the test to no-fast-math-vect-16.c to avoid the use of the flag for > it. > - vect-peel-3.c and vect-peel-4.c test vectorizer's peeling heuristics > ignoring double-word vectors. This patch changes the misalignment > values of the data-refs. > - bb-slp-10.c contains a misaligned store supported not only on > vect_hw_misalign targets, but also on vect_element_align targets. The > patch fixes that. > > Tested on arm-linux-gnueabi, powerpc64-suse-linux and x86_64-suse-linux. > OK for mainline? > I'll commit this patch tomorrow if nobody objects. Thanks, Ira > Thanks, > Ira > > testsuite/ChangeLog: > > * gcc.dg/vect/vect-16.c: Rename to ... > * gcc.dg/vect/no-fast-math-vect16.c: ... this. > * gcc.dg/vect/vect-peel-3.c: Adjust misalignment values > for double-word vectors. > * gcc.dg/vect/vect-peel-4.c: Likewise. > * gcc.dg/vect/bb-slp-10.c: Replace vect_hw_misalign with > vect_element_align. > * gcc.dg/vect/vect.exp: Run no-fast-math-* tests with > -fno-fast-math. >
> testsuite/ChangeLog: > > * gcc.dg/vect/vect-16.c: Rename to ... > * gcc.dg/vect/no-fast-math-vect16.c: ... this. > * gcc.dg/vect/vect-peel-3.c: Adjust misalignment values > for double-word vectors. > * gcc.dg/vect/vect-peel-4.c: Likewise. > * gcc.dg/vect/bb-slp-10.c: Replace vect_hw_misalign with > vect_element_align. > * gcc.dg/vect/vect.exp: Run no-fast-math-* tests with > -fno-fast-math. Ira, This change broke vect-peel-3.c and vect-peel-4.c on IA64 HP-UX. I get these failures: FAIL: gcc.dg/vect/vect-peel-3.c -flto scan-tree-dump-times vect "Alignment of access forced using peeling" 1 FAIL: gcc.dg/vect/vect-peel-3.c -flto scan-tree-dump-times vect "vectorized 1 loops" 1 FAIL: gcc.dg/vect/vect-peel-3.c scan-tree-dump-times vect "Alignment of access forced using peeling" 1 FAIL: gcc.dg/vect/vect-peel-3.c scan-tree-dump-times vect "vectorized 1 loops" 1 FAIL: gcc.dg/vect/vect-peel-4.c -flto scan-tree-dump-times vect "vectorized 1 loops" 1 FAIL: gcc.dg/vect/vect-peel-4.c scan-tree-dump-times vect "vectorized 1 loops" 1 I think that changing i+6 and i+2 to i+5 and i+1 broke things on HP-UX because we are big-endian and the old values may not have worked on ARM because it is little-endian (just guessing). I am not sure what we want to do to make this test work on both platforms. Do you have any ideas? Steve Ellcey sje@cup.hp.com
Steve Ellcey <sje@cup.hp.com> wrote on 15/06/2011 08:15:27 PM: > > > testsuite/ChangeLog: > > > > * gcc.dg/vect/vect-16.c: Rename to ... > > * gcc.dg/vect/no-fast-math-vect16.c: ... this. > > * gcc.dg/vect/vect-peel-3.c: Adjust misalignment values > > for double-word vectors. > > * gcc.dg/vect/vect-peel-4.c: Likewise. > > * gcc.dg/vect/bb-slp-10.c: Replace vect_hw_misalign with > > vect_element_align. > > * gcc.dg/vect/vect.exp: Run no-fast-math-* tests with > > -fno-fast-math. > > Ira, > > This change broke vect-peel-3.c and vect-peel-4.c on IA64 HP-UX. > I get these failures: > > FAIL: gcc.dg/vect/vect-peel-3.c -flto scan-tree-dump-times vect > "Alignment of access forced using peeling" 1 > FAIL: gcc.dg/vect/vect-peel-3.c -flto scan-tree-dump-times vect > "vectorized 1 loops" 1 > FAIL: gcc.dg/vect/vect-peel-3.c scan-tree-dump-times vect > "Alignment of access forced using peeling" 1 > FAIL: gcc.dg/vect/vect-peel-3.c scan-tree-dump-times vect > "vectorized 1 loops" 1 > FAIL: gcc.dg/vect/vect-peel-4.c -flto scan-tree-dump-times vect > "vectorized 1 loops" 1 > FAIL: gcc.dg/vect/vect-peel-4.c scan-tree-dump-times vect > "vectorized 1 loops" 1 > > I think that changing i+6 and i+2 to i+5 and i+1 broke things on HP-UX > because we are big-endian and the old values may not have worked on ARM > because it is little-endian (just guessing). > > I am not sure what we want to do to make this test work on both platforms. > Do you have any ideas? I don't think it is related to endianess. Could you please send me the dump files (created with -fdump-tree-vect-details)? Thanks, Ira > > Steve Ellcey > sje@cup.hp.com
On Thu, 2011-06-16 at 08:48 +0300, Ira Rosen wrote: > > I am not sure what we want to do to make this test work on both > platforms. > > Do you have any ideas? > > I don't think it is related to endianess. > Could you please send me the dump files (created with > -fdump-tree-vect-details)? > > Thanks, > Ira > > > > > Steve Ellcey > > sje@cup.hp.com I went ahead and created a bug report (PR 49443) and added the dump file to that report as an attachment. Steve Ellcey sje@cup.hp.com
Index: gcc.dg/vect/vect-16.c =================================================================== --- gcc.dg/vect/vect-16.c (revision 174836) +++ gcc.dg/vect/vect-16.c (working copy) @@ -1,38 +0,0 @@ -/* { dg-require-effective-target vect_float } */ - -#include <stdarg.h> -#include "tree-vect.h" - -#define N 16 -#define DIFF 240 - -__attribute__ ((noinline)) -int main1 () -{ - int i; - float b[N] = {0,3,6,9,12,15,18,21,24,27,30,33,36,39,42,45}; - float c[N] = {0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15}; - float diff; - - diff = 0; - for (i = 0; i < N; i++) { - diff += (b[i] - c[i]); - } - - /* check results: */ - if (diff != DIFF) - abort (); - - return 0; -} - -int main (void) -{ - check_vect (); - - return main1 (); -} - -/* Requires fast-math. */ -/* { dg-final { scan-tree-dump-times "vectorized 1 loops" 1 "vect" { xfail *-*-* } } } */ -/* { dg-final { cleanup-tree-dump "vect" } } */ Index: gcc.dg/vect/vect-peel-3.c =================================================================== --- gcc.dg/vect/vect-peel-3.c (revision 174836) +++ gcc.dg/vect/vect-peel-3.c (working copy) @@ -4,9 +4,7 @@ #include "tree-vect.h" #define N 128 -#define RES 21888 - -/* unaligned store. */ +#define RES 21640 int ib[N+10]; int ia[N+10]; @@ -18,11 +16,11 @@ int main1 () int i, suma = 0, sumb = 0, sumc = 0; /* ib and ic have same misalignment, we peel to align them. */ - for (i = 1; i <= N; i++) + for (i = 0; i <= N; i++) { suma += ia[i]; - sumb += ib[i+6]; - sumc += ic[i+2]; + sumb += ib[i+5]; + sumc += ic[i+1]; } /* check results: */ Index: gcc.dg/vect/vect-peel-4.c =================================================================== --- gcc.dg/vect/vect-peel-4.c (revision 174836) +++ gcc.dg/vect/vect-peel-4.c (working copy) @@ -16,13 +16,13 @@ int main1 () /* Don't peel keeping one load and the store aligned. */ for (i = 0; i <= N; i++) { - ia[i] = ib[i] + ib[i+6]; + ia[i] = ib[i] + ib[i+5]; } /* check results: */ for (i = 1; i <= N; i++) { - if (ia[i] != ib[i] + ib[i+6]) + if (ia[i] != ib[i] + ib[i+5]) abort (); } Index: gcc.dg/vect/no-fast-math-vect16.c =================================================================== --- gcc.dg/vect/no-fast-math-vect16.c (revision 0) +++ gcc.dg/vect/no-fast-math-vect16.c (revision 0) @@ -0,0 +1,38 @@ +/* { dg-require-effective-target vect_float } */ + +#include <stdarg.h> +#include "tree-vect.h" + +#define N 16 +#define DIFF 240 + +__attribute__ ((noinline)) +int main1 () +{ + int i; + float b[N] = {0,3,6,9,12,15,18,21,24,27,30,33,36,39,42,45}; + float c[N] = {0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15}; + float diff; + + diff = 0; + for (i = 0; i < N; i++) { + diff += (b[i] - c[i]); + } + + /* check results: */ + if (diff != DIFF) + abort (); + + return 0; +} + +int main (void) +{ + check_vect (); + + return main1 (); +} + +/* Requires fast-math. */ +/* { dg-final { scan-tree-dump-times "vectorized 1 loops" 1 "vect" { xfail *-*-* } } } */ +/* { dg-final { cleanup-tree-dump "vect" } } */ Index: gcc.dg/vect/bb-slp-10.c =================================================================== --- gcc.dg/vect/bb-slp-10.c (revision 174836) +++ gcc.dg/vect/bb-slp-10.c (working copy) @@ -49,7 +49,7 @@ int main (void) return 0; } -/* { dg-final { scan-tree-dump-times "unsupported alignment in basic block." 1 "slp" { xfail vect_hw_misalign } } } */ -/* { dg-final { scan-tree-dump-times "basic block vectorized using SLP" 1 "slp" { target vect_hw_misalign } } } */ +/* { dg-final { scan-tree-dump-times "unsupported alignment in basic block." 1 "slp" { xfail vect_element_align } } } */ +/* { dg-final { scan-tree-dump-times "basic block vectorized using SLP" 1 "slp" { target vect_element_align } } } */ /* { dg-final { cleanup-tree-dump "slp" } } */ Index: gcc.dg/vect/vect.exp =================================================================== --- gcc.dg/vect/vect.exp (revision 174836) +++ gcc.dg/vect/vect.exp (working copy) @@ -107,6 +107,12 @@ lappend DEFAULT_VECTCFLAGS "-ffast-math" dg-runtest [lsort [glob -nocomplain $srcdir/$subdir/fast-math-*.\[cS\]]] \ "" $DEFAULT_VECTCFLAGS +# -fno-fast-math tests +set DEFAULT_VECTCFLAGS $SAVED_DEFAULT_VECTCFLAGS +lappend DEFAULT_VECTCFLAGS "-fno-fast-math" +dg-runtest [lsort [glob -nocomplain $srcdir/$subdir/no-fast-math-*.\[cS\]]] \ + "" $DEFAULT_VECTCFLAGS + # -fno-math-errno tests set DEFAULT_VECTCFLAGS $SAVED_DEFAULT_VECTCFLAGS lappend DEFAULT_VECTCFLAGS "-fno-math-errno"
Hi, This patch fixes several vectorizer testsuite failures on ARM: - vect-16.c checks that the vectorization fails without -ffast-math, but -ffast-math is a default flag for vector tests on ARM. I renamed the test to no-fast-math-vect-16.c to avoid the use of the flag for it. - vect-peel-3.c and vect-peel-4.c test vectorizer's peeling heuristics ignoring double-word vectors. This patch changes the misalignment values of the data-refs. - bb-slp-10.c contains a misaligned store supported not only on vect_hw_misalign targets, but also on vect_element_align targets. The patch fixes that. Tested on arm-linux-gnueabi, powerpc64-suse-linux and x86_64-suse-linux. OK for mainline? Thanks, Ira testsuite/ChangeLog: * gcc.dg/vect/vect-16.c: Rename to ... * gcc.dg/vect/no-fast-math-vect16.c: ... this. * gcc.dg/vect/vect-peel-3.c: Adjust misalignment values for double-word vectors. * gcc.dg/vect/vect-peel-4.c: Likewise. * gcc.dg/vect/bb-slp-10.c: Replace vect_hw_misalign with vect_element_align. * gcc.dg/vect/vect.exp: Run no-fast-math-* tests with -fno-fast-math.