Dorit Nuzman - Re: [PATCH, i386]: Vectorize conversions for i386 (original) (raw)

This is the mail archive of the gcc-patches@gcc.gnu.orgmailing list for the GCC project.

• From: Dorit Nuzman
• To: Uros Bizjak
• Cc: GCC Patches , Richard Guenther , Tehila Meyzels
• Date: Sun, 11 Feb 2007 13:51:27 +0200
• Subject: Re: [PATCH, i386]: Vectorize conversions for i386

Uros Bizjak ubizjak@gmail.com wrote on 08/02/2007 16:50:23:

Hello!

This patch implements conversions for i386 and x86_64 targets. Patch

Cool, thanks!

builds on (yet uncommitted) patch that implements vectorized conversions infrastructure by Tehila Meyzels (http://gcc.gnu.org/ml/gcc-patches/2007-02/msg00494.html).

I plan to commit it soonish. The vectorization-of-induction patch that I committed a couple of days ago actually exposed a couple of problems in this patch - I'm fixing it now - will commit the fixed version (and submit the fixes to the list) this weekend.

Here is the patch I ended up committing.

The diffs relative to Tehila's patch are as follows:

1. Mark the arguments of the call-expr for renaming:

•  FOR_EACH_SSA_TREE_OPERAND (sym, new_stmt, iter, SSA_OP_ALL_VIRTUALS)

•    {

•      if (TREE_CODE (sym) == SSA_NAME)

•        sym = SSA_NAME_VAR (sym);

•      mark_sym_for_renaming (sym);

•    }

this is necessary because on some targets the builtins we call are not marked as readonly (e.g. Altivec), which IIUC causes the compiler to create vdefs/vuses, which resulted in an ICE for some of there not getting renamed.

1. This bit was missing in the patch I was testing, which caused most of the failures I saw:

• /* Supportable by target? */
• if (!targetm.vectorize.builtin_conversion (code, vectype_in))
• {

•  if (vect_print_dump_info (REPORT_DETAILS))

•    fprintf (vect_dump, "op not supported by target.");

•  return false;

• }

1. Updates to the vectorizer testsuite to reflect that more loops get vectorized.

Index: testsuite/ChangeLog

--- testsuite/ChangeLog (revision 121815) +++ testsuite/ChangeLog (working copy) @@ -1,3 +1,13 @@ +2007-02-11 Tehila Meyzels tehila@il.ibm.com

•     Dorit Nuzman  <dorit@il.ibm.com>

• * gcc.dg/vect/vect-intfloat-conversion.c-1:  New test.

• * gcc.dg/vect/vect-intfloat-conversion.c-2:  New test.

• * gcc.dg/vect/vect-93.c: Another loop gets vectorized  on powerpc.

• * gcc.dg/vect/vect-113.c: Likewise.

• * gcc.dg/vect/vect-iv-11.c: A loop gets vectorized.

• 2007-02-10 Richard Henderson rth@redhat.com

   * lib/target-supports.exp (check_effective_target_tls): Redefine

Index: testsuite/gcc.dg/vect/vect-93.c

--- testsuite/gcc.dg/vect/vect-93.c (revision 121815) +++ testsuite/gcc.dg/vect/vect-93.c (working copy) @@ -65,12 +65,21 @@ return 0; }

-/* in main1 / -/ { dg-final { scan-tree-dump-times "vectorized 2 loops" 1 "vect" } } / +/ 2 loops vectorized in main1, 2 loops vectorized in main:

• the first loop in main requires vectorization of conversions,
• the second loop in main requires vectorization of misaliged load: */
• +/* main && main1 together: / +/ { dg-final { scan-tree-dump-times "vectorized 2 loops" 2 "vect" { target powerpc*-- } } } / / { dg-final { scan-tree-dump-times "Alignment of access forced using peeling" 2 "vect" { target vect_no_align } } } / / { dg-final { scan-tree-dump-times "Alignment of access forced using peeling" 3 "vect" { xfail vect_no_align } } } */

-/* in main / -/ { dg-final { scan-tree-dump-times "vectorized 1 loops" 1 "vect" { xfail vect_no_align } } } / +/ in main1: / +/ { dg-final { scan-tree-dump-times "vectorized 2 loops" 1 "vect" { target {! powerpc*-- } } } } / +/ { dg-final { scan-tree-dump-times "vectorized 2 loops" 1 "vect" { target vect_no_align } } } / + +/ in main: / +/ { dg-final { scan-tree-dump-times "vectorized 0 loops" 1 "vect" { target vect_no_align } } } / / { dg-final { scan-tree-dump-times "Vectorizing an unaligned access" 1 "vect" { xfail vect_no_align } } } / + / { dg-final { cleanup-tree-dump "vect" } } */ Index: testsuite/gcc.dg/vect/vect-iv-11.c

--- testsuite/gcc.dg/vect/vect-iv-11.c (revision 121815) +++ testsuite/gcc.dg/vect/vect-iv-11.c (working copy) @@ -28,5 +28,5 @@ return 0; }

-/* { dg-final { scan-tree-dump-times "vectorized 1 loops" 1 "vect" { xfail --* } } } / +/ { dg-final { scan-tree-dump-times "vectorized 1 loops" 1 "vect" } } / / { dg-final { cleanup-tree-dump "vect" } } */ Index: testsuite/gcc.dg/vect/vect-113.c

--- testsuite/gcc.dg/vect/vect-113.c (revision 121815) +++ testsuite/gcc.dg/vect/vect-113.c (working copy) @@ -11,7 +11,7 @@ int i; float a[N];

• /* Induction. */

• /* Induction and type conversion. */ for ( i = 0; i < N; i++) { a[i] = i; @@ -32,5 +32,5 @@ return main1 (); }

-/* { dg-final { scan-tree-dump-times "vectorized 0 loops" 1 "vect" } } / +/ { dg-final { scan-tree-dump-times "vectorized 1 loops" 1 "vect" { target powerpc*-- } } } / / { dg-final { cleanup-tree-dump "vect" } } */

Bootstraped on powerpc-linux, and on i386-linux.

dorit

(See attached file: induc.feb11.txt)

dorit

Patch was bootstrapped on x86_64-pc-linux-gnu and regression tested for c,c++ and gfortran.

This patch also includes two testcases from original patch, changed to account for i386 and x86_64 targets and a testcase that checks float-int conversions.

2007-02-08 Uros Bizjak ubizjak@gmail.com

    * config/i386/i386.c (TARGET_VECTORIZE_BUILTIN_CONVERSION):

Define. (ix86_builtin_conversion): New function.

testsuite/ChangeLog:

    * gcc.dg/vect/vect-intfloat-conversion-1.c: Scan for vectorized

loop also for i?86-- and x86_64-- targets. * gcc.dg/vect/vect-intfloat-conversion-2.c: Ditto. * gcc.dg/vect/vect-floatint-conversion-1.c: New.

Uros.

Index: config/i386/i386.c

--- config/i386/i386.c (revision 121711) +++ config/i386/i386.c (working copy) @@ -1516,6 +1516,7 @@ static void ix86_init_builtins (void); static rtx ix86_expand_builtin (tree, rtx, rtx, enum machine_mode, int); static tree ix86_builtin_vectorized_function (enum built_in_function, tree, tree); +static tree ix86_builtin_conversion (enum tree_code, tree); static const char *ix86_mangle_fundamental_type (tree); static tree ix86_stack_protect_fail (void); static rtx ix86_internal_arg_pointer (void); @@ -1580,8 +1581,11 @@ #define TARGET_INIT_BUILTINS ix86_init_builtins #undef TARGET_EXPAND_BUILTIN #define TARGET_EXPAND_BUILTIN ix86_expand_builtin + #undef TARGET_VECTORIZE_BUILTIN_VECTORIZED_FUNCTION #define TARGET_VECTORIZE_BUILTIN_VECTORIZED_FUNCTION ix86_builtin_vectorized_function +#undef TARGET_VECTORIZE_BUILTIN_CONVERSION +#define TARGET_VECTORIZE_BUILTIN_CONVERSION ix86_builtin_conversion

#undef TARGET_ASM_FUNCTION_EPILOGUE #define TARGET_ASM_FUNCTION_EPILOGUE ix86_output_function_epilogue @@ -18056,6 +18060,40 @@ return NULL_TREE; }

+/* Returns a decl of a function that implements conversion of the + input vector of type TYPE, or NULL_TREE if it is not available. / + +static tree +ix86_builtin_conversion (enum tree_code code, tree type) +{ + if (TREE_CODE (type) != VECTOR_TYPE) + return NULL_TREE; + + switch (code) + { + case FLOAT_EXPR: + switch (TYPE_MODE (type)) + { + case V4SImode: + return ix86_builtins[IX86_BUILTIN_CVTDQ2PS]; + default: + return NULL_TREE; + } + + case FIX_TRUNC_EXPR: + switch (TYPE_MODE (type)) + { + case V4SFmode: + return ix86_builtins[IX86_BUILTIN_CVTTPS2DQ]; + default: + return NULL_TREE; + } + default: + return NULL_TREE; + + } +} + / Store OPERAND to the memory after reload is completed. This means that we can't easily use assign_stack_local. */ rtx Index: testsuite/gcc.dg/vect/vect-intfloat-conversion-1.c

--- testsuite/gcc.dg/vect/vect-intfloat-conversion-1.c (revision 0) +++ testsuite/gcc.dg/vect/vect-intfloat-conversion-1.c (revision 0) @@ -0,0 +1,38 @@ +/* { dg-require-effective-target vect_int } / + +#include <stdarg.h> +#include "tree-vect.h" + +#define N 32 + +int main1 () +{ + int i; + int ib[N] = {0,3,6,9,12,15,18,21,24,27,30,33,36,39,42,45,0,3,6,9, 12,15,18,21,24,27,30,33,36,39,42,45}; + float fa[N]; + + / int -> float / + for (i = 0; i < N; i++) + { + fa[i] = (float) ib[i]; + } + + / check results: / + for (i = 0; i < N; i++) + { + if (fa[i] != (float) ib[i]) + abort (); + } + + return 0; +} + +int main (void) +{ + check_vect (); + + return main1 (); +} + +/ { dg-final { scan-tree-dump-times "vectorized 1 loops" 1 "vect" { target powerpc*-- i?86-- x86_64-- } } } / +/ { dg-final { cleanup-tree-dump "vect" } } */ Index: testsuite/gcc.dg/vect/vect-intfloat-conversion-2.c

--- testsuite/gcc.dg/vect/vect-intfloat-conversion-2.c (revision 0) +++ testsuite/gcc.dg/vect/vect-intfloat-conversion-2.c (revision 0) @@ -0,0 +1,40 @@ +/* { dg-require-effective-target vect_int } / + +#include <stdarg.h> +#include "tree-vect.h" + +#define N 32 + +int main1 () +{ + int i; + int int_arr[N] = {0,3,6,9,12,15,18,21,24,27,30,33,36,39,42,45,0, 3,6,9,12,15,18,21,24,27,30,33,36,39,42,45}; + float float_arr[N]; + char char_arr[N]; + + for (i = 0; i < N; i++){ + float_arr[i] = (float) int_arr[i]; + char_arr[i] = 0; + } + + / check results: / + for (i = 0; i < N; i++) + { + if (float_arr[i] != (float) int_arr[i]) + abort (); + if (char_arr[i] != 0) + abort (); + } + + return 0; +} + +int main (void) +{ + check_vect (); + + return main1 (); +} + +/ { dg-final { scan-tree-dump-times "vectorized 1 loops" 1 "vect" { target powerpc*-- i?86-- x86_64-- } } } / +/ { dg-final { cleanup-tree-dump "vect" } } */ Index: testsuite/gcc.dg/vect/vect-floatint-conversion-1.c

--- testsuite/gcc.dg/vect/vect-floatint-conversion-1.c (revision 0) +++ testsuite/gcc.dg/vect/vect-floatint-conversion-1.c (revision 0) @@ -0,0 +1,40 @@ +/* { dg-require-effective-target vect_float } */ + +#include <stdarg.h> +#include "tree-vect.h" + +#define N 32 + +int +main1 () +{

• int i;
• float fb[N] = {0.4,3.5,6.6,9.4,12.5,15.6,18.4,21.5,24.6,27.4,30. 5,33.6,36.4,39.5,42.6,45.4,0.5,3.6,6.4,9.5,12.6,15.4,18.5,21.6,24.4, 27.5,30.6,33.4,36.5,39.6,42.4,45.5};
• int ia[N];
• /* float -> int */
• for (i = 0; i < N; i++)
• {

•  ia[i] = (int) fb[i];

• }
• /* check results: */
• for (i = 0; i < N; i++)
• {

•  if (ia[i] != (int) fb[i])

• abort ();
• }
• return 0; +}
• +int +main (void) +{
• check_vect ();
• return main1 (); +}
• +/* { dg-final { scan-tree-dump-times "vectorized 1 loops" 1 "vect" { target i?86-- x86_64-- } } } / +/ { dg-final { cleanup-tree-dump "vect" } } */

Attachment:induc.feb11.txt
Description: Text document

• Follow-Ups:
  • Re: [PATCH, i386]: Vectorize conversions for i386
    * From: Richard Guenther
• References:
  • Re: [PATCH, i386]: Vectorize conversions for i386
    * From: Dorit Nuzman