OSDN Git Service

* config/i386/sse.md (round<mode>2_sfix): New expander.
authoruros <uros@138bc75d-0d04-0410-961f-82ee72b054a4>
Mon, 14 Nov 2011 19:36:33 +0000 (19:36 +0000)
committeruros <uros@138bc75d-0d04-0410-961f-82ee72b054a4>
Mon, 14 Nov 2011 19:36:33 +0000 (19:36 +0000)
(round<mode>2_vec_pack_sfix): Ditto.
(<sse4_1>_round<ssemodesuffix>_sfix<avxsizesuffix>): Ditto.
(<sse4_1>_round<ssemodesuffix>_vec_pack_sfix<avxsizesuffix>): Ditto.
* config/i386/builtin-types.def (V4SI_FTYPE_V4SF_ROUND,
V8SI_FTYPE_V8SF_ROUND, V4SI_FTYPE_V2DF_V2DF_ROUND,
V8SI_FTYPE_V4DF_V4DF_ROUND): New builtin types.
* config/i386/i386.c (ix86_builtins): Add
IX86_BUILTIN_{FLOORPD,CEILPD,ROUNDPD_AZ}_VEC_PACK_SFIX{,256} and
IX86_BUILTIN_{FLOORPS,CEILPS,ROUNDPS_AZ}_SFIX{,256} defines.
(bdesc_args): Add __builtin_ia32_{floorpd,ceilpd}_vec_pack_sfix{,256},
__builtin_ia32_roundpd_az_vec_pack_sfix{,256},
__builtin_ia32_{floorps,ceilps}_sfix{,256}and
__builtin_ia32_roundps_az_sfix{,256} descriptions.
(ix86_expand_sse_round_vec_pack_sfix): New.
(ix86_expand_args_builtin): Handle V4SI_FTYPE_V4SF_ROUND,
V8SI_FTYPE_V8SF_ROUND, V4SI_FTYPE_V2DF_V2DF_ROUND and
V8SI_FTYPE_V4DF_V4DF_ROUND types.  Check last argument of
CODE_FOR_sse4_1_roundpd_vec_pack_sfix, CODE_FOR_sse4_1_roundps_sfix,
CODE_FOR_avx_roundpd_vec_pack_sfix256 and CODE_FOR_avx_roundps_sfix256.
(ix86_builtin_vectorized_function): Handle
BUILT_IN_{I,L,LL}FLOOR{,F}, BUILT_IN_{I,L,LL}CEIL{,F} and
BUILT_IN_{I,L,LL}ROUND{,F}

testsuite/ChangeLog:

* gcc.target/i386/sse4_1-floor-sfix-vec.c: New test.
* gcc.target/i386/sse4_1-floorf-sfix-vec.c: Ditto.
* gcc.target/i386/avx-floor-sfix-vec.c: Ditto.
* gcc.target/i386/avx-floorf-sfix-vec.c: Ditto.
* gcc.target/i386/sse4_1-ceil-sfix-vec.c: Ditto.
* gcc.target/i386/sse4_1-ceilf-sfix-vec.c: Ditto.
* gcc.target/i386/avx-ceil-sfix-vec.c: Ditto.
* gcc.target/i386/avx-ceilf-sfix-vec.c: Ditto.
* gcc.target/i386/sse4_1-round-sfix-vec.c: Ditto.
* gcc.target/i386/sse4_1-roundf-sfix-vec.c: Ditto.
* gcc.target/i386/avx-round-sfix-vec.c: Ditto.
* gcc.target/i386/avx-roundf-sfix-vec.c: Ditto.

git-svn-id: svn+ssh://gcc.gnu.org/svn/gcc/trunk@181361 138bc75d-0d04-0410-961f-82ee72b054a4

17 files changed:
gcc/ChangeLog
gcc/config/i386/i386-builtin-types.def
gcc/config/i386/i386.c
gcc/config/i386/sse.md
gcc/testsuite/ChangeLog
gcc/testsuite/gcc.target/i386/avx-ceil-sfix-vec.c [new file with mode: 0644]
gcc/testsuite/gcc.target/i386/avx-ceilf-sfix-vec.c [new file with mode: 0644]
gcc/testsuite/gcc.target/i386/avx-floor-sfix-vec.c [new file with mode: 0644]
gcc/testsuite/gcc.target/i386/avx-floorf-sfix-vec.c [new file with mode: 0644]
gcc/testsuite/gcc.target/i386/avx-round-sfix-vec.c [new file with mode: 0644]
gcc/testsuite/gcc.target/i386/avx-roundf-sfix-vec.c [new file with mode: 0644]
gcc/testsuite/gcc.target/i386/sse4_1-ceil-sfix-vec.c [new file with mode: 0644]
gcc/testsuite/gcc.target/i386/sse4_1-ceilf-sfix-vec.c [new file with mode: 0644]
gcc/testsuite/gcc.target/i386/sse4_1-floor-sfix-vec.c [new file with mode: 0644]
gcc/testsuite/gcc.target/i386/sse4_1-floorf-sfix-vec.c [new file with mode: 0644]
gcc/testsuite/gcc.target/i386/sse4_1-round-sfix-vec.c [new file with mode: 0644]
gcc/testsuite/gcc.target/i386/sse4_1-roundf-sfix-vec.c [new file with mode: 0644]

index 51ea394..fabb1da 100644 (file)
@@ -1,3 +1,29 @@
+2011-11-14  Uros Bizjak  <ubizjak@gmail.com>
+
+       * config/i386/sse.md (round<mode>2_sfix): New expander.
+       (round<mode>2_vec_pack_sfix): Ditto.
+       (<sse4_1>_round<ssemodesuffix>_sfix<avxsizesuffix>): Ditto.
+       (<sse4_1>_round<ssemodesuffix>_vec_pack_sfix<avxsizesuffix>): Ditto.
+       * config/i386/builtin-types.def (V4SI_FTYPE_V4SF_ROUND,
+       V8SI_FTYPE_V8SF_ROUND, V4SI_FTYPE_V2DF_V2DF_ROUND,
+       V8SI_FTYPE_V4DF_V4DF_ROUND): New builtin types.
+       * config/i386/i386.c (ix86_builtins): Add
+       IX86_BUILTIN_{FLOORPD,CEILPD,ROUNDPD_AZ}_VEC_PACK_SFIX{,256} and
+       IX86_BUILTIN_{FLOORPS,CEILPS,ROUNDPS_AZ}_SFIX{,256} defines.
+       (bdesc_args): Add __builtin_ia32_{floorpd,ceilpd}_vec_pack_sfix{,256},
+       __builtin_ia32_roundpd_az_vec_pack_sfix{,256},
+       __builtin_ia32_{floorps,ceilps}_sfix{,256}and
+       __builtin_ia32_roundps_az_sfix{,256} descriptions.
+       (ix86_expand_sse_round_vec_pack_sfix): New.
+       (ix86_expand_args_builtin): Handle V4SI_FTYPE_V4SF_ROUND,
+       V8SI_FTYPE_V8SF_ROUND, V4SI_FTYPE_V2DF_V2DF_ROUND and
+       V8SI_FTYPE_V4DF_V4DF_ROUND types.  Check last argument of
+       CODE_FOR_sse4_1_roundpd_vec_pack_sfix, CODE_FOR_sse4_1_roundps_sfix,
+       CODE_FOR_avx_roundpd_vec_pack_sfix256 and CODE_FOR_avx_roundps_sfix256.
+       (ix86_builtin_vectorized_function): Handle
+       BUILT_IN_{I,L,LL}FLOOR{,F}, BUILT_IN_{I,L,LL}CEIL{,F} and
+       BUILT_IN_{I,L,LL}ROUND{,F}
+
 2011-11-14  Jan Hubicka  <jh@suse.cz>
 
        PR middle-end/50598
 2011-11-14  Zolotukhin Michael  <michael.v.zolotukhin@gmail.com>
            Jan Hubicka  <jh@suse.cz>
 
-        * config/i386/i386.h (processor_costs): Add second dimension to
-        stringop_algs array.
-        * config/i386/i386.c (cost models): Initialize second dimension of
-        stringop_algs arrays.
+       * config/i386/i386.h (processor_costs): Add second dimension to
+       stringop_algs array.
+       * config/i386/i386.c (cost models): Initialize second dimension of
+       stringop_algs arrays.
        (core_cost): New costs based on generic64 costs with updated stringop
        values.
-        (promote_duplicated_reg): Add support for vector modes, add
-        declaration.
-        (promote_duplicated_reg_to_size): Likewise.
+       (promote_duplicated_reg): Add support for vector modes, add
+       declaration.
+       (promote_duplicated_reg_to_size): Likewise.
        (processor_target): Set core costs for core variants.
-        (expand_set_or_movmem_via_loop_with_iter): New function.
-        (expand_set_or_movmem_via_loop): Enable reuse of the same iters in
-        different loops, produced by this function.
-        (emit_strset): New function.
-        (expand_movmem_epilogue): Add epilogue generation for bigger sizes,
-        use SSE-moves where possible.
-        (expand_setmem_epilogue): Likewise.
-        (expand_movmem_prologue): Likewise for prologue.
-        (expand_setmem_prologue): Likewise.
-        (expand_constant_movmem_prologue): Likewise.
-        (expand_constant_setmem_prologue): Likewise.
-        (decide_alg): Add new argument align_unknown.  Fix algorithm of
-        strategy selection if TARGET_INLINE_ALL_STRINGOPS is set; Skip sse_loop
-        (decide_alignment): Update desired alignment according to chosen move
-        mode.
-        (ix86_expand_movmem): Change unrolled_loop strategy to use SSE-moves.
-        (ix86_expand_setmem): Likewise.
-        (ix86_slow_unaligned_access): Implementation of new hook
-        slow_unaligned_access.
-        * config/i386/i386.md (strset): Enable half-SSE moves.
-        * config/i386/sse.md (vec_dupv4si): Add expand for vec_dupv4si.
-        (vec_dupv2di): Add expand for vec_dupv2di.
+       (expand_set_or_movmem_via_loop_with_iter): New function.
+       (expand_set_or_movmem_via_loop): Enable reuse of the same iters in
+       different loops, produced by this function.
+       (emit_strset): New function.
+       (expand_movmem_epilogue): Add epilogue generation for bigger sizes,
+       use SSE-moves where possible.
+       (expand_setmem_epilogue): Likewise.
+       (expand_movmem_prologue): Likewise for prologue.
+       (expand_setmem_prologue): Likewise.
+       (expand_constant_movmem_prologue): Likewise.
+       (expand_constant_setmem_prologue): Likewise.
+       (decide_alg): Add new argument align_unknown.  Fix algorithm of
+       strategy selection if TARGET_INLINE_ALL_STRINGOPS is set; Skip sse_loop
+       (decide_alignment): Update desired alignment according to chosen move
+       mode.
+       (ix86_expand_movmem): Change unrolled_loop strategy to use SSE-moves.
+       (ix86_expand_setmem): Likewise.
+       (ix86_slow_unaligned_access): Implementation of new hook
+       slow_unaligned_access.
+       * config/i386/i386.md (strset): Enable half-SSE moves.
+       * config/i386/sse.md (vec_dupv4si): Add expand for vec_dupv4si.
+       (vec_dupv2di): Add expand for vec_dupv2di.
 
 2011-11-14  Dimitrios Apostolou  <jimis@gmx.net>
 
@@ -53,8 +79,7 @@
 
 2011-11-14  Kai Tietz  <ktietz@redhat.com>
 
-       * gcov.c (generate_results): Add missing semicolon and
-       correct indent.
+       * gcov.c (generate_results): Add missing semicolon and correct indent.
 
 2011-11-14  Ira Rosen  <ira.rosen@linaro.org>
 
@@ -71,9 +96,8 @@
        PR target/50694
        * config/sh/sh.h (IS_LITTLE_ENDIAN_OPTION, UNSUPPORTED_SH2A):
        New macros.
-       (DRIVER_SELF_SPECS): Use new macros to filter out
-       unsupported options taking the default configuration into
-       account.
+       (DRIVER_SELF_SPECS): Use new macros to filter out unsupported options
+       taking the default configuration into account.
 
 2011-11-13  Jonathan Wakely  <jwakely.gcc@gmail.com>
 
 
 2011-11-12  Richard Henderson  <rth@redhat.com>
 
-       * config/rs6000/rs6000.md (fix_trunc<SFDF>si2_stfiwx): Use 
+       * config/rs6000/rs6000.md (fix_trunc<SFDF>si2_stfiwx): Use
        nonimmediate_operand for the destination.
        (fixuns_trunc<SFDF>si2_stfiwx): Likewise.
 
index 5dcb68c..79fb142 100644 (file)
@@ -465,6 +465,11 @@ DEF_FUNCTION_TYPE_ALIAS (V4DF_FTYPE_V4DF, ROUND)
 DEF_FUNCTION_TYPE_ALIAS (V4SF_FTYPE_V4SF, ROUND)
 DEF_FUNCTION_TYPE_ALIAS (V8SF_FTYPE_V8SF, ROUND)
 
+DEF_FUNCTION_TYPE_ALIAS (V4SI_FTYPE_V2DF_V2DF, ROUND)
+DEF_FUNCTION_TYPE_ALIAS (V8SI_FTYPE_V4DF_V4DF, ROUND)
+DEF_FUNCTION_TYPE_ALIAS (V4SI_FTYPE_V4SF, ROUND)
+DEF_FUNCTION_TYPE_ALIAS (V8SI_FTYPE_V8SF, ROUND)
+
 DEF_FUNCTION_TYPE_ALIAS (INT_FTYPE_V2DF_V2DF, PTEST)
 DEF_FUNCTION_TYPE_ALIAS (INT_FTYPE_V2DI_V2DI, PTEST)
 DEF_FUNCTION_TYPE_ALIAS (INT_FTYPE_V4DF_V4DF, PTEST)
index a8e8a6a..e307999 100644 (file)
@@ -25457,22 +25457,32 @@ enum ix86_builtins
   IX86_BUILTIN_PMULDQ128,
   IX86_BUILTIN_PMULLD128,
 
-  IX86_BUILTIN_ROUNDPD,
-  IX86_BUILTIN_ROUNDPS,
   IX86_BUILTIN_ROUNDSD,
   IX86_BUILTIN_ROUNDSS,
 
+  IX86_BUILTIN_ROUNDPD,
+  IX86_BUILTIN_ROUNDPS,
+
   IX86_BUILTIN_FLOORPD,
   IX86_BUILTIN_CEILPD,
   IX86_BUILTIN_TRUNCPD,
   IX86_BUILTIN_RINTPD,
   IX86_BUILTIN_ROUNDPD_AZ,
+
+  IX86_BUILTIN_FLOORPD_VEC_PACK_SFIX,
+  IX86_BUILTIN_CEILPD_VEC_PACK_SFIX,
+  IX86_BUILTIN_ROUNDPD_AZ_VEC_PACK_SFIX,
+
   IX86_BUILTIN_FLOORPS,
   IX86_BUILTIN_CEILPS,
   IX86_BUILTIN_TRUNCPS,
   IX86_BUILTIN_RINTPS,
   IX86_BUILTIN_ROUNDPS_AZ,
 
+  IX86_BUILTIN_FLOORPS_SFIX,
+  IX86_BUILTIN_CEILPS_SFIX,
+  IX86_BUILTIN_ROUNDPS_AZ_SFIX,
+
   IX86_BUILTIN_PTESTZ,
   IX86_BUILTIN_PTESTC,
   IX86_BUILTIN_PTESTNZC,
@@ -25646,12 +25656,21 @@ enum ix86_builtins
   IX86_BUILTIN_TRUNCPD256,
   IX86_BUILTIN_RINTPD256,
   IX86_BUILTIN_ROUNDPD_AZ256,
+
+  IX86_BUILTIN_FLOORPD_VEC_PACK_SFIX256,
+  IX86_BUILTIN_CEILPD_VEC_PACK_SFIX256,
+  IX86_BUILTIN_ROUNDPD_AZ_VEC_PACK_SFIX256,
+
   IX86_BUILTIN_FLOORPS256,
   IX86_BUILTIN_CEILPS256,
   IX86_BUILTIN_TRUNCPS256,
   IX86_BUILTIN_RINTPS256,
   IX86_BUILTIN_ROUNDPS_AZ256,
 
+  IX86_BUILTIN_FLOORPS_SFIX256,
+  IX86_BUILTIN_CEILPS_SFIX256,
+  IX86_BUILTIN_ROUNDPS_AZ_SFIX256,
+
   IX86_BUILTIN_UNPCKHPD256,
   IX86_BUILTIN_UNPCKLPD256,
   IX86_BUILTIN_UNPCKHPS256,
@@ -26856,14 +26875,22 @@ static const struct builtin_description bdesc_args[] =
   { OPTION_MASK_ISA_ROUND, CODE_FOR_sse4_1_roundpd, "__builtin_ia32_truncpd", IX86_BUILTIN_TRUNCPD, (enum rtx_code) ROUND_TRUNC, (int) V2DF_FTYPE_V2DF_ROUND },
   { OPTION_MASK_ISA_ROUND, CODE_FOR_sse4_1_roundpd, "__builtin_ia32_rintpd", IX86_BUILTIN_RINTPD, (enum rtx_code) ROUND_MXCSR, (int) V2DF_FTYPE_V2DF_ROUND },
 
+  { OPTION_MASK_ISA_ROUND, CODE_FOR_sse4_1_roundpd_vec_pack_sfix, "__builtin_ia32_floorpd_vec_pack_sfix", IX86_BUILTIN_FLOORPD_VEC_PACK_SFIX, (enum rtx_code) ROUND_FLOOR, (int) V4SI_FTYPE_V2DF_V2DF_ROUND },
+  { OPTION_MASK_ISA_ROUND, CODE_FOR_sse4_1_roundpd_vec_pack_sfix, "__builtin_ia32_ceilpd_vec_pack_sfix", IX86_BUILTIN_CEILPD_VEC_PACK_SFIX, (enum rtx_code) ROUND_CEIL, (int) V4SI_FTYPE_V2DF_V2DF_ROUND },
+
   { OPTION_MASK_ISA_ROUND, CODE_FOR_roundv2df2, "__builtin_ia32_roundpd_az", IX86_BUILTIN_ROUNDPD_AZ, UNKNOWN, (int) V2DF_FTYPE_V2DF },
+  { OPTION_MASK_ISA_ROUND, CODE_FOR_roundv2df2_vec_pack_sfix, "__builtin_ia32_roundpd_az_vec_pack_sfix", IX86_BUILTIN_ROUNDPD_AZ_VEC_PACK_SFIX, UNKNOWN, (int) V4SI_FTYPE_V2DF_V2DF },
 
   { OPTION_MASK_ISA_ROUND, CODE_FOR_sse4_1_roundps, "__builtin_ia32_floorps", IX86_BUILTIN_FLOORPS, (enum rtx_code) ROUND_FLOOR, (int) V4SF_FTYPE_V4SF_ROUND },
   { OPTION_MASK_ISA_ROUND, CODE_FOR_sse4_1_roundps, "__builtin_ia32_ceilps", IX86_BUILTIN_CEILPS, (enum rtx_code) ROUND_CEIL, (int) V4SF_FTYPE_V4SF_ROUND },
   { OPTION_MASK_ISA_ROUND, CODE_FOR_sse4_1_roundps, "__builtin_ia32_truncps", IX86_BUILTIN_TRUNCPS, (enum rtx_code) ROUND_TRUNC, (int) V4SF_FTYPE_V4SF_ROUND },
   { OPTION_MASK_ISA_ROUND, CODE_FOR_sse4_1_roundps, "__builtin_ia32_rintps", IX86_BUILTIN_RINTPS, (enum rtx_code) ROUND_MXCSR, (int) V4SF_FTYPE_V4SF_ROUND },
 
+  { OPTION_MASK_ISA_ROUND, CODE_FOR_sse4_1_roundps_sfix, "__builtin_ia32_floorps_sfix", IX86_BUILTIN_FLOORPS_SFIX, (enum rtx_code) ROUND_FLOOR, (int) V4SI_FTYPE_V4SF_ROUND },
+  { OPTION_MASK_ISA_ROUND, CODE_FOR_sse4_1_roundps_sfix, "__builtin_ia32_ceilps_sfix", IX86_BUILTIN_CEILPS_SFIX, (enum rtx_code) ROUND_CEIL, (int) V4SI_FTYPE_V4SF_ROUND },
+
   { OPTION_MASK_ISA_ROUND, CODE_FOR_roundv4sf2, "__builtin_ia32_roundps_az", IX86_BUILTIN_ROUNDPS_AZ, UNKNOWN, (int) V4SF_FTYPE_V4SF },
+  { OPTION_MASK_ISA_ROUND, CODE_FOR_roundv4sf2_sfix, "__builtin_ia32_roundps_az_sfix", IX86_BUILTIN_ROUNDPS_AZ_SFIX, UNKNOWN, (int) V4SI_FTYPE_V4SF },
 
   { OPTION_MASK_ISA_ROUND, CODE_FOR_sse4_1_ptest, "__builtin_ia32_ptestz128", IX86_BUILTIN_PTESTZ, EQ, (int) INT_FTYPE_V2DI_V2DI_PTEST },
   { OPTION_MASK_ISA_ROUND, CODE_FOR_sse4_1_ptest, "__builtin_ia32_ptestc128", IX86_BUILTIN_PTESTC, LTU, (int) INT_FTYPE_V2DI_V2DI_PTEST },
@@ -26983,13 +27010,21 @@ static const struct builtin_description bdesc_args[] =
   { OPTION_MASK_ISA_AVX, CODE_FOR_avx_roundpd256, "__builtin_ia32_rintpd256", IX86_BUILTIN_RINTPD256, (enum rtx_code) ROUND_MXCSR, (int) V4DF_FTYPE_V4DF_ROUND },
 
   { OPTION_MASK_ISA_AVX, CODE_FOR_roundv4df2, "__builtin_ia32_roundpd_az256", IX86_BUILTIN_ROUNDPD_AZ256, UNKNOWN, (int) V4DF_FTYPE_V4DF },
+  { OPTION_MASK_ISA_AVX, CODE_FOR_roundv4df2_vec_pack_sfix, "__builtin_ia32_roundpd_az_vec_pack_sfix256", IX86_BUILTIN_ROUNDPD_AZ_VEC_PACK_SFIX256, UNKNOWN, (int) V8SI_FTYPE_V4DF_V4DF },
+
+  { OPTION_MASK_ISA_AVX, CODE_FOR_avx_roundpd_vec_pack_sfix256, "__builtin_ia32_floorpd_vec_pack_sfix256", IX86_BUILTIN_FLOORPD_VEC_PACK_SFIX256, (enum rtx_code) ROUND_FLOOR, (int) V8SI_FTYPE_V4DF_V4DF_ROUND },
+  { OPTION_MASK_ISA_AVX, CODE_FOR_avx_roundpd_vec_pack_sfix256, "__builtin_ia32_ceilpd_vec_pack_sfix256", IX86_BUILTIN_CEILPD_VEC_PACK_SFIX256, (enum rtx_code) ROUND_CEIL, (int) V8SI_FTYPE_V4DF_V4DF_ROUND },
 
   { OPTION_MASK_ISA_AVX, CODE_FOR_avx_roundps256, "__builtin_ia32_floorps256", IX86_BUILTIN_FLOORPS256, (enum rtx_code) ROUND_FLOOR, (int) V8SF_FTYPE_V8SF_ROUND },
   { OPTION_MASK_ISA_AVX, CODE_FOR_avx_roundps256, "__builtin_ia32_ceilps256", IX86_BUILTIN_CEILPS256, (enum rtx_code) ROUND_CEIL, (int) V8SF_FTYPE_V8SF_ROUND },
   { OPTION_MASK_ISA_AVX, CODE_FOR_avx_roundps256, "__builtin_ia32_truncps256", IX86_BUILTIN_TRUNCPS256, (enum rtx_code) ROUND_TRUNC, (int) V8SF_FTYPE_V8SF_ROUND },
   { OPTION_MASK_ISA_AVX, CODE_FOR_avx_roundps256, "__builtin_ia32_rintps256", IX86_BUILTIN_RINTPS256, (enum rtx_code) ROUND_MXCSR, (int) V8SF_FTYPE_V8SF_ROUND },
 
+  { OPTION_MASK_ISA_AVX, CODE_FOR_avx_roundps_sfix256, "__builtin_ia32_floorps_sfix256", IX86_BUILTIN_FLOORPS_SFIX256, (enum rtx_code) ROUND_FLOOR, (int) V8SI_FTYPE_V8SF_ROUND },
+  { OPTION_MASK_ISA_AVX, CODE_FOR_avx_roundps_sfix256, "__builtin_ia32_ceilps_sfix256", IX86_BUILTIN_CEILPS_SFIX256, (enum rtx_code) ROUND_CEIL, (int) V8SI_FTYPE_V8SF_ROUND },
+
   { OPTION_MASK_ISA_AVX, CODE_FOR_roundv8sf2, "__builtin_ia32_roundps_az256", IX86_BUILTIN_ROUNDPS_AZ256, UNKNOWN, (int) V8SF_FTYPE_V8SF },
+  { OPTION_MASK_ISA_AVX, CODE_FOR_roundv8sf2_sfix, "__builtin_ia32_roundps_az_sfix256", IX86_BUILTIN_ROUNDPS_AZ_SFIX256, UNKNOWN, (int) V8SI_FTYPE_V8SF },
 
   { OPTION_MASK_ISA_AVX, CODE_FOR_avx_unpckhpd256,  "__builtin_ia32_unpckhpd256", IX86_BUILTIN_UNPCKHPD256, UNKNOWN, (int) V4DF_FTYPE_V4DF_V4DF },
   { OPTION_MASK_ISA_AVX, CODE_FOR_avx_unpcklpd256,  "__builtin_ia32_unpcklpd256", IX86_BUILTIN_UNPCKLPD256, UNKNOWN, (int) V4DF_FTYPE_V4DF_V4DF },
@@ -28430,7 +28465,7 @@ ix86_expand_sse_comi (const struct builtin_description *d, tree exp,
   return SUBREG_REG (target);
 }
 
-/* Subroutine of ix86_expand_args_builtin to take care of round insns.  */
+/* Subroutines of ix86_expand_args_builtin to take care of round insns.  */
 
 static rtx
 ix86_expand_sse_round (const struct builtin_description *d, tree exp,
@@ -28463,6 +28498,44 @@ ix86_expand_sse_round (const struct builtin_description *d, tree exp,
   return target;
 }
 
+static rtx
+ix86_expand_sse_round_vec_pack_sfix (const struct builtin_description *d,
+                                    tree exp, rtx target)
+{
+  rtx pat;
+  tree arg0 = CALL_EXPR_ARG (exp, 0);
+  tree arg1 = CALL_EXPR_ARG (exp, 1);
+  rtx op0 = expand_normal (arg0);
+  rtx op1 = expand_normal (arg1);
+  rtx op2;
+  enum machine_mode tmode = insn_data[d->icode].operand[0].mode;
+  enum machine_mode mode0 = insn_data[d->icode].operand[1].mode;
+  enum machine_mode mode1 = insn_data[d->icode].operand[2].mode;
+
+  if (optimize || target == 0
+      || GET_MODE (target) != tmode
+      || !insn_data[d->icode].operand[0].predicate (target, tmode))
+    target = gen_reg_rtx (tmode);
+
+  op0 = safe_vector_operand (op0, mode0);
+  op1 = safe_vector_operand (op1, mode1);
+
+  if ((optimize && !register_operand (op0, mode0))
+      || !insn_data[d->icode].operand[0].predicate (op0, mode0))
+    op0 = copy_to_mode_reg (mode0, op0);
+  if ((optimize && !register_operand (op1, mode1))
+      || !insn_data[d->icode].operand[1].predicate (op1, mode1))
+    op1 = copy_to_mode_reg (mode1, op1);
+
+  op2 = GEN_INT (d->comparison);
+
+  pat = GEN_FCN (d->icode) (target, op0, op1, op2);
+  if (! pat)
+    return 0;
+  emit_insn (pat);
+  return target;
+}
+
 /* Subroutine of ix86_expand_builtin to take care of ptest insns.  */
 
 static rtx
@@ -28736,7 +28809,12 @@ ix86_expand_args_builtin (const struct builtin_description *d,
     case V4DF_FTYPE_V4DF_ROUND:
     case V4SF_FTYPE_V4SF_ROUND:
     case V8SF_FTYPE_V8SF_ROUND:
+    case V4SI_FTYPE_V4SF_ROUND:
+    case V8SI_FTYPE_V8SF_ROUND:
       return ix86_expand_sse_round (d, exp, target);
+    case V4SI_FTYPE_V2DF_V2DF_ROUND:
+    case V8SI_FTYPE_V4DF_V4DF_ROUND:
+      return ix86_expand_sse_round_vec_pack_sfix (d, exp, target);
     case INT_FTYPE_V8SF_V8SF_PTEST:
     case INT_FTYPE_V4DI_V4DI_PTEST:
     case INT_FTYPE_V4DF_V4DF_PTEST:
@@ -29053,15 +29131,22 @@ ix86_expand_args_builtin (const struct builtin_description *d,
                error ("the last argument must be an 1-bit immediate");
                return const0_rtx;
 
-             case CODE_FOR_sse4_1_roundpd:
-             case CODE_FOR_sse4_1_roundps:
              case CODE_FOR_sse4_1_roundsd:
              case CODE_FOR_sse4_1_roundss:
+
+             case CODE_FOR_sse4_1_roundpd:
+             case CODE_FOR_sse4_1_roundps:
+             case CODE_FOR_avx_roundpd256:
+             case CODE_FOR_avx_roundps256:
+
+             case CODE_FOR_sse4_1_roundpd_vec_pack_sfix:
+             case CODE_FOR_sse4_1_roundps_sfix:
+             case CODE_FOR_avx_roundpd_vec_pack_sfix256:
+             case CODE_FOR_avx_roundps_sfix256:
+
              case CODE_FOR_sse4_1_blendps:
              case CODE_FOR_avx_blendpd256:
              case CODE_FOR_avx_vpermilv4df:
-             case CODE_FOR_avx_roundpd256:
-             case CODE_FOR_avx_roundps256:
                error ("the last argument must be a 4-bit immediate");
                return const0_rtx;
 
@@ -30087,6 +30172,70 @@ ix86_builtin_vectorized_function (tree fndecl, tree type_out,
        }
       break;
 
+    case BUILT_IN_IFLOOR:
+    case BUILT_IN_LFLOOR:
+    case BUILT_IN_LLFLOOR:
+      /* The round insn does not trap on denormals.  */
+      if (flag_trapping_math || !TARGET_ROUND)
+       break;
+
+      if (out_mode == SImode && in_mode == DFmode)
+       {
+         if (out_n == 4 && in_n == 2)
+           return ix86_builtins[IX86_BUILTIN_FLOORPD_VEC_PACK_SFIX];
+         else if (out_n == 8 && in_n == 4)
+           return ix86_builtins[IX86_BUILTIN_FLOORPD_VEC_PACK_SFIX256];
+       }
+      break;
+
+    case BUILT_IN_IFLOORF:
+    case BUILT_IN_LFLOORF:
+    case BUILT_IN_LLFLOORF:
+      /* The round insn does not trap on denormals.  */
+      if (flag_trapping_math || !TARGET_ROUND)
+       break;
+
+      if (out_mode == SImode && in_mode == SFmode)
+       {
+         if (out_n == 4 && in_n == 4)
+           return ix86_builtins[IX86_BUILTIN_FLOORPS_SFIX];
+         else if (out_n == 8 && in_n == 8)
+           return ix86_builtins[IX86_BUILTIN_FLOORPS_SFIX256];
+       }
+      break;
+
+    case BUILT_IN_ICEIL:
+    case BUILT_IN_LCEIL:
+    case BUILT_IN_LLCEIL:
+      /* The round insn does not trap on denormals.  */
+      if (flag_trapping_math || !TARGET_ROUND)
+       break;
+
+      if (out_mode == SImode && in_mode == DFmode)
+       {
+         if (out_n == 4 && in_n == 2)
+           return ix86_builtins[IX86_BUILTIN_CEILPD_VEC_PACK_SFIX];
+         else if (out_n == 8 && in_n == 4)
+           return ix86_builtins[IX86_BUILTIN_CEILPD_VEC_PACK_SFIX256];
+       }
+      break;
+
+    case BUILT_IN_ICEILF:
+    case BUILT_IN_LCEILF:
+    case BUILT_IN_LLCEILF:
+      /* The round insn does not trap on denormals.  */
+      if (flag_trapping_math || !TARGET_ROUND)
+       break;
+
+      if (out_mode == SImode && in_mode == SFmode)
+       {
+         if (out_n == 4 && in_n == 4)
+           return ix86_builtins[IX86_BUILTIN_CEILPS_SFIX];
+         else if (out_n == 8 && in_n == 8)
+           return ix86_builtins[IX86_BUILTIN_CEILPS_SFIX256];
+       }
+      break;
+
     case BUILT_IN_IRINT:
     case BUILT_IN_LRINT:
     case BUILT_IN_LLRINT:
@@ -30111,6 +30260,38 @@ ix86_builtin_vectorized_function (tree fndecl, tree type_out,
        }
       break;
 
+    case BUILT_IN_IROUND:
+    case BUILT_IN_LROUND:
+    case BUILT_IN_LLROUND:
+      /* The round insn does not trap on denormals.  */
+      if (flag_trapping_math || !TARGET_ROUND)
+       break;
+
+      if (out_mode == SImode && in_mode == DFmode)
+       {
+         if (out_n == 4 && in_n == 2)
+           return ix86_builtins[IX86_BUILTIN_ROUNDPD_AZ_VEC_PACK_SFIX];
+         else if (out_n == 8 && in_n == 4)
+           return ix86_builtins[IX86_BUILTIN_ROUNDPD_AZ_VEC_PACK_SFIX256];
+       }
+      break;
+
+    case BUILT_IN_IROUNDF:
+    case BUILT_IN_LROUNDF:
+    case BUILT_IN_LLROUNDF:
+      /* The round insn does not trap on denormals.  */
+      if (flag_trapping_math || !TARGET_ROUND)
+       break;
+
+      if (out_mode == SImode && in_mode == SFmode)
+       {
+         if (out_n == 4 && in_n == 4)
+           return ix86_builtins[IX86_BUILTIN_ROUNDPS_AZ_SFIX];
+         else if (out_n == 8 && in_n == 8)
+           return ix86_builtins[IX86_BUILTIN_ROUNDPS_AZ_SFIX256];
+       }
+      break;
+
     case BUILT_IN_COPYSIGN:
       if (out_mode == DFmode && in_mode == DFmode)
        {
index 4938abf..147646b 100644 (file)
    (set_attr "prefix" "maybe_vex")
    (set_attr "mode" "<MODE>")])
 
+(define_expand "<sse4_1>_round<ssemodesuffix>_sfix<avxsizesuffix>"
+  [(match_operand:<sseintvecmode> 0 "register_operand" "")
+   (match_operand:VF1 1 "nonimmediate_operand" "")
+   (match_operand:SI 2 "const_0_to_15_operand" "")]
+  "TARGET_ROUND"
+{
+  rtx tmp = gen_reg_rtx (<MODE>mode);
+
+  emit_insn
+    (gen_<sse4_1>_round<ssemodesuffix><avxsizesuffix> (tmp, operands[1],
+                                                      operands[2]));
+  emit_insn
+    (gen_fix_trunc<mode><sseintvecmodelower>2 (operands[0], tmp));
+  DONE;
+})
+
+(define_expand "<sse4_1>_round<ssemodesuffix>_vec_pack_sfix<avxsizesuffix>"
+  [(match_operand:<ssepackfltmode> 0 "register_operand" "")
+   (match_operand:VF2 1 "nonimmediate_operand" "")
+   (match_operand:VF2 2 "nonimmediate_operand" "")
+   (match_operand:SI 3 "const_0_to_15_operand" "")]
+  "TARGET_ROUND"
+{
+  rtx tmp0, tmp1;
+
+  tmp0 = gen_reg_rtx (<MODE>mode);
+  tmp1 = gen_reg_rtx (<MODE>mode);
+
+  emit_insn
+    (gen_<sse4_1>_round<ssemodesuffix><avxsizesuffix> (tmp0, operands[1],
+                                                      operands[3]));
+  emit_insn
+    (gen_<sse4_1>_round<ssemodesuffix><avxsizesuffix> (tmp1, operands[2],
+                                                      operands[3]));
+  emit_insn
+    (gen_vec_pack_sfix_trunc_<mode> (operands[0], tmp0, tmp1));
+  DONE;
+})
+
 (define_insn "sse4_1_round<ssescalarmodesuffix>"
   [(set (match_operand:VF_128 0 "register_operand" "=x,x")
        (vec_merge:VF_128
   operands[5] = GEN_INT (ROUND_TRUNC);
 })
 
+(define_expand "round<mode>2_sfix"
+  [(match_operand:<sseintvecmode> 0 "register_operand" "")
+   (match_operand:VF1 1 "nonimmediate_operand" "")]
+  "TARGET_ROUND && !flag_trapping_math"
+{
+  rtx tmp = gen_reg_rtx (<MODE>mode);
+
+  emit_insn (gen_round<mode>2 (tmp, operands[1]));
+
+  emit_insn
+    (gen_fix_trunc<mode><sseintvecmodelower>2 (operands[0], tmp));
+  DONE;
+})
+
+(define_expand "round<mode>2_vec_pack_sfix"
+  [(match_operand:<ssepackfltmode> 0 "register_operand" "")
+   (match_operand:VF2 1 "nonimmediate_operand" "")
+   (match_operand:VF2 2 "nonimmediate_operand" "")]
+  "TARGET_ROUND && !flag_trapping_math"
+{
+  rtx tmp0, tmp1;
+
+  tmp0 = gen_reg_rtx (<MODE>mode);
+  tmp1 = gen_reg_rtx (<MODE>mode);
+
+  emit_insn (gen_round<mode>2 (tmp0, operands[1]));
+  emit_insn (gen_round<mode>2 (tmp1, operands[2]));
+
+  emit_insn
+    (gen_vec_pack_sfix_trunc_<mode> (operands[0], tmp0, tmp1));
+  DONE;
+})
+
 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 ;;
 ;; Intel SSE4.2 string/text processing instructions
index 7a7445e..e371383 100644 (file)
@@ -1,3 +1,18 @@
+2011-11-14  Uros Bizjak  <ubizjak@gmail.com>
+
+       * gcc.target/i386/sse4_1-floor-sfix-vec.c: New test.
+       * gcc.target/i386/sse4_1-floorf-sfix-vec.c: Ditto.
+       * gcc.target/i386/avx-floor-sfix-vec.c: Ditto.
+       * gcc.target/i386/avx-floorf-sfix-vec.c: Ditto.
+       * gcc.target/i386/sse4_1-ceil-sfix-vec.c: Ditto.
+       * gcc.target/i386/sse4_1-ceilf-sfix-vec.c: Ditto.
+       * gcc.target/i386/avx-ceil-sfix-vec.c: Ditto.
+       * gcc.target/i386/avx-ceilf-sfix-vec.c: Ditto.
+       * gcc.target/i386/sse4_1-round-sfix-vec.c: Ditto.
+       * gcc.target/i386/sse4_1-roundf-sfix-vec.c: Ditto.
+       * gcc.target/i386/avx-round-sfix-vec.c: Ditto.
+       * gcc.target/i386/avx-roundf-sfix-vec.c: Ditto.
+
 2011-11-14  Fabien ChĂȘne  <fabien@gcc.gnu.org>
 
        PR c++/6936
 
 2011-11-09  Janne Blomqvist  <jb@gcc.gnu.org>
 
-        PR libfortran/50016
-        * gfortran.dg/inquire_size.f90: Don't flush the unit.
+       PR libfortran/50016
+       * gfortran.dg/inquire_size.f90: Don't flush the unit.
 
 2011-11-09  Richard Guenther  <rguenther@suse.de>
 
 
 2011-11-07  Janne Blomqvist  <jb@gcc.gnu.org>
 
-        PR libfortran/45723
-        * gfortran.dg/open_dev_null.f90: Remove testcase.
+       PR libfortran/45723
+       * gfortran.dg/open_dev_null.f90: Remove testcase.
 
 2011-11-07  Uros Bizjak  <ubizjak@gmail.com>
 
diff --git a/gcc/testsuite/gcc.target/i386/avx-ceil-sfix-vec.c b/gcc/testsuite/gcc.target/i386/avx-ceil-sfix-vec.c
new file mode 100644 (file)
index 0000000..ac0911f
--- /dev/null
@@ -0,0 +1,9 @@
+/* { dg-do run } */
+/* { dg-options "-O2 -ffast-math -ftree-vectorize -mavx" } */
+/* { dg-require-effective-target avx } */
+/* { dg-skip-if "no M_PI" { vxworks_kernel } } */
+
+#define CHECK_H "avx-check.h"
+#define TEST avx_test
+
+#include "sse4_1-ceil-sfix-vec.c"
diff --git a/gcc/testsuite/gcc.target/i386/avx-ceilf-sfix-vec.c b/gcc/testsuite/gcc.target/i386/avx-ceilf-sfix-vec.c
new file mode 100644 (file)
index 0000000..789b78e
--- /dev/null
@@ -0,0 +1,9 @@
+/* { dg-do run } */
+/* { dg-options "-O2 -ffast-math -ftree-vectorize -mavx" } */
+/* { dg-require-effective-target avx } */
+/* { dg-skip-if "no M_PI" { vxworks_kernel } } */
+
+#define CHECK_H "avx-check.h"
+#define TEST avx_test
+
+#include "sse4_1-ceilf-sfix-vec.c"
diff --git a/gcc/testsuite/gcc.target/i386/avx-floor-sfix-vec.c b/gcc/testsuite/gcc.target/i386/avx-floor-sfix-vec.c
new file mode 100644 (file)
index 0000000..efa557c
--- /dev/null
@@ -0,0 +1,9 @@
+/* { dg-do run } */
+/* { dg-options "-O2 -ffast-math -ftree-vectorize -mavx" } */
+/* { dg-require-effective-target avx } */
+/* { dg-skip-if "no M_PI" { vxworks_kernel } } */
+
+#define CHECK_H "avx-check.h"
+#define TEST avx_test
+
+#include "sse4_1-floor-sfix-vec.c"
diff --git a/gcc/testsuite/gcc.target/i386/avx-floorf-sfix-vec.c b/gcc/testsuite/gcc.target/i386/avx-floorf-sfix-vec.c
new file mode 100644 (file)
index 0000000..0c1587a
--- /dev/null
@@ -0,0 +1,9 @@
+/* { dg-do run } */
+/* { dg-options "-O2 -ffast-math -ftree-vectorize -mavx" } */
+/* { dg-require-effective-target avx } */
+/* { dg-skip-if "no M_PI" { vxworks_kernel } } */
+
+#define CHECK_H "avx-check.h"
+#define TEST avx_test
+
+#include "sse4_1-floorf-sfix-vec.c"
diff --git a/gcc/testsuite/gcc.target/i386/avx-round-sfix-vec.c b/gcc/testsuite/gcc.target/i386/avx-round-sfix-vec.c
new file mode 100644 (file)
index 0000000..5adfffa
--- /dev/null
@@ -0,0 +1,9 @@
+/* { dg-do run } */
+/* { dg-options "-O2 -ffast-math -ftree-vectorize -mavx" } */
+/* { dg-require-effective-target avx } */
+/* { dg-skip-if "no M_PI" { vxworks_kernel } } */
+
+#define CHECK_H "avx-check.h"
+#define TEST avx_test
+
+#include "sse4_1-round-sfix-vec.c"
diff --git a/gcc/testsuite/gcc.target/i386/avx-roundf-sfix-vec.c b/gcc/testsuite/gcc.target/i386/avx-roundf-sfix-vec.c
new file mode 100644 (file)
index 0000000..1fd4591
--- /dev/null
@@ -0,0 +1,9 @@
+/* { dg-do run } */
+/* { dg-options "-O2 -ffast-math -ftree-vectorize -mavx" } */
+/* { dg-require-effective-target avx } */
+/* { dg-skip-if "no M_PI" { vxworks_kernel } } */
+
+#define CHECK_H "avx-check.h"
+#define TEST avx_test
+
+#include "sse4_1-roundf-sfix-vec.c"
diff --git a/gcc/testsuite/gcc.target/i386/sse4_1-ceil-sfix-vec.c b/gcc/testsuite/gcc.target/i386/sse4_1-ceil-sfix-vec.c
new file mode 100644 (file)
index 0000000..ca07d9c
--- /dev/null
@@ -0,0 +1,62 @@
+/* { dg-do run } */
+/* { dg-options "-O2 -ffast-math -ftree-vectorize -msse4.1" } */
+/* { dg-require-effective-target sse4 } */
+/* { dg-skip-if "no M_PI" { vxworks_kernel } } */
+
+#ifndef CHECK_H
+#define CHECK_H "sse4_1-check.h"
+#endif
+
+#ifndef TEST
+#define TEST sse4_1_test
+#endif
+
+#include CHECK_H
+
+#include <math.h>
+
+extern double ceil (double);
+
+#define NUM 64
+
+static void
+__attribute__((__target__("fpmath=sse")))
+init_src (double *src)
+{
+  int i, sign = 1;
+  double f = rand ();
+
+  for (i = 0; i < NUM; i++)
+    {
+      src[i] = (i + 1) * f * M_PI * sign;
+      if (i < (NUM / 2))
+       {
+          if ((i % 6) == 0)
+           f = f * src[i];
+        }
+      else if (i == (NUM / 2))
+       f = rand ();
+      else if ((i % 6) == 0)
+       f = 1 / (f * (i + 1) * src[i] * M_PI * sign);
+      sign = -sign;
+    }
+}
+
+static void
+__attribute__((__target__("fpmath=387")))
+TEST (void)
+{
+  double a[NUM];
+  int r[NUM];
+  int i;
+
+  init_src (a);
+
+  for (i = 0; i < NUM; i++)
+    r[i] = (int) ceil (a[i]);
+
+  /* check results:  */
+  for (i = 0; i < NUM; i++)
+    if (r[i] != (int) ceil (a[i]))
+      abort();
+}
diff --git a/gcc/testsuite/gcc.target/i386/sse4_1-ceilf-sfix-vec.c b/gcc/testsuite/gcc.target/i386/sse4_1-ceilf-sfix-vec.c
new file mode 100644 (file)
index 0000000..b0559bf
--- /dev/null
@@ -0,0 +1,62 @@
+/* { dg-do run } */
+/* { dg-options "-O2 -ffast-math -ftree-vectorize -msse4.1" } */
+/* { dg-require-effective-target sse4 } */
+/* { dg-skip-if "no M_PI" { vxworks_kernel } } */
+
+#ifndef CHECK_H
+#define CHECK_H "sse4_1-check.h"
+#endif
+
+#ifndef TEST
+#define TEST sse4_1_test
+#endif
+
+#include CHECK_H
+
+#include <math.h>
+
+extern float ceilf (float);
+
+#define NUM 64
+
+static void
+__attribute__((__target__("fpmath=sse")))
+init_src (float *src)
+{
+  int i, sign = 1;
+  float f = rand ();
+
+  for (i = 0; i < NUM; i++)
+    {
+      src[i] = (i + 1) * f * M_PI * sign;
+      if (i < (NUM / 2))
+       {
+          if ((i % 6) == 0)
+           f = f * src[i];
+        }
+      else if (i == (NUM / 2))
+       f = rand ();
+      else if ((i % 6) == 0)
+       f = 1 / (f * (i + 1) * src[i] * M_PI * sign);
+      sign = -sign;
+    }
+}
+
+static void
+__attribute__((__target__("fpmath=387")))
+TEST (void)
+{
+  float a[NUM];
+  int r[NUM];
+  int i;
+
+  init_src (a);
+
+  for (i = 0; i < NUM; i++)
+    r[i] = (int) ceilf (a[i]);
+
+  /* check results:  */
+  for (i = 0; i < NUM; i++)
+    if (r[i] != (int) ceilf (a[i]))
+      abort();
+}
diff --git a/gcc/testsuite/gcc.target/i386/sse4_1-floor-sfix-vec.c b/gcc/testsuite/gcc.target/i386/sse4_1-floor-sfix-vec.c
new file mode 100644 (file)
index 0000000..2083a60
--- /dev/null
@@ -0,0 +1,62 @@
+/* { dg-do run } */
+/* { dg-options "-O2 -ffast-math -ftree-vectorize -msse4.1" } */
+/* { dg-require-effective-target sse4 } */
+/* { dg-skip-if "no M_PI" { vxworks_kernel } } */
+
+#ifndef CHECK_H
+#define CHECK_H "sse4_1-check.h"
+#endif
+
+#ifndef TEST
+#define TEST sse4_1_test
+#endif
+
+#include CHECK_H
+
+#include <math.h>
+
+extern double floor (double);
+
+#define NUM 64
+
+static void
+__attribute__((__target__("fpmath=sse")))
+init_src (double *src)
+{
+  int i, sign = 1;
+  double f = rand ();
+
+  for (i = 0; i < NUM; i++)
+    {
+      src[i] = (i + 1) * f * M_PI * sign;
+      if (i < (NUM / 2))
+       {
+          if ((i % 6) == 0)
+           f = f * src[i];
+        }
+      else if (i == (NUM / 2))
+       f = rand ();
+      else if ((i % 6) == 0)
+       f = 1 / (f * (i + 1) * src[i] * M_PI * sign);
+      sign = -sign;
+    }
+}
+
+static void
+__attribute__((__target__("fpmath=387")))
+TEST (void)
+{
+  double a[NUM];
+  int r[NUM];
+  int i;
+
+  init_src (a);
+
+  for (i = 0; i < NUM; i++)
+    r[i] = (int) floor (a[i]);
+
+  /* check results:  */
+  for (i = 0; i < NUM; i++)
+    if (r[i] != (int) floor (a[i]))
+      abort();
+}
diff --git a/gcc/testsuite/gcc.target/i386/sse4_1-floorf-sfix-vec.c b/gcc/testsuite/gcc.target/i386/sse4_1-floorf-sfix-vec.c
new file mode 100644 (file)
index 0000000..aa2976d
--- /dev/null
@@ -0,0 +1,62 @@
+/* { dg-do run } */
+/* { dg-options "-O2 -ffast-math -ftree-vectorize -msse4.1" } */
+/* { dg-require-effective-target sse4 } */
+/* { dg-skip-if "no M_PI" { vxworks_kernel } } */
+
+#ifndef CHECK_H
+#define CHECK_H "sse4_1-check.h"
+#endif
+
+#ifndef TEST
+#define TEST sse4_1_test
+#endif
+
+#include CHECK_H
+
+#include <math.h>
+
+extern float floorf (float);
+
+#define NUM 64
+
+static void
+__attribute__((__target__("fpmath=sse")))
+init_src (float *src)
+{
+  int i, sign = 1;
+  float f = rand ();
+
+  for (i = 0; i < NUM; i++)
+    {
+      src[i] = (i + 1) * f * M_PI * sign;
+      if (i < (NUM / 2))
+       {
+          if ((i % 6) == 0)
+           f = f * src[i];
+        }
+      else if (i == (NUM / 2))
+       f = rand ();
+      else if ((i % 6) == 0)
+       f = 1 / (f * (i + 1) * src[i] * M_PI * sign);
+      sign = -sign;
+    }
+}
+
+static void
+__attribute__((__target__("fpmath=387")))
+TEST (void)
+{
+  float a[NUM];
+  int r[NUM];
+  int i;
+
+  init_src (a);
+
+  for (i = 0; i < NUM; i++)
+    r[i] = (int) floorf (a[i]);
+
+  /* check results:  */
+  for (i = 0; i < NUM; i++)
+    if (r[i] != (int) floorf (a[i]))
+      abort();
+}
diff --git a/gcc/testsuite/gcc.target/i386/sse4_1-round-sfix-vec.c b/gcc/testsuite/gcc.target/i386/sse4_1-round-sfix-vec.c
new file mode 100644 (file)
index 0000000..9abbe55
--- /dev/null
@@ -0,0 +1,62 @@
+/* { dg-do run } */
+/* { dg-options "-O2 -ffast-math -ftree-vectorize -msse4.1" } */
+/* { dg-require-effective-target sse4 } */
+/* { dg-skip-if "no M_PI" { vxworks_kernel } } */
+
+#ifndef CHECK_H
+#define CHECK_H "sse4_1-check.h"
+#endif
+
+#ifndef TEST
+#define TEST sse4_1_test
+#endif
+
+#include CHECK_H
+
+#include <math.h>
+
+extern double round (double);
+
+#define NUM 64
+
+static void
+__attribute__((__target__("fpmath=sse")))
+init_src (double *src)
+{
+  int i, sign = 1;
+  double f = rand ();
+
+  for (i = 0; i < NUM; i++)
+    {
+      src[i] = (i + 1) * f * M_PI * sign;
+      if (i < (NUM / 2))
+       {
+          if ((i % 6) == 0)
+           f = f * src[i];
+        }
+      else if (i == (NUM / 2))
+       f = rand ();
+      else if ((i % 6) == 0)
+       f = 1 / (f * (i + 1) * src[i] * M_PI * sign);
+      sign = -sign;
+    }
+}
+
+static void
+__attribute__((__target__("fpmath=387")))
+TEST (void)
+{
+  double a[NUM];
+  int r[NUM];
+  int i;
+
+  init_src (a);
+
+  for (i = 0; i < NUM; i++)
+    r[i] = (int) round (a[i]);
+
+  /* check results:  */
+  for (i = 0; i < NUM; i++)
+    if (r[i] != (int) round (a[i]))
+      abort();
+}
diff --git a/gcc/testsuite/gcc.target/i386/sse4_1-roundf-sfix-vec.c b/gcc/testsuite/gcc.target/i386/sse4_1-roundf-sfix-vec.c
new file mode 100644 (file)
index 0000000..5384e5c
--- /dev/null
@@ -0,0 +1,62 @@
+/* { dg-do run } */
+/* { dg-options "-O2 -ffast-math -ftree-vectorize -msse4.1" } */
+/* { dg-require-effective-target sse4 } */
+/* { dg-skip-if "no M_PI" { vxworks_kernel } } */
+
+#ifndef CHECK_H
+#define CHECK_H "sse4_1-check.h"
+#endif
+
+#ifndef TEST
+#define TEST sse4_1_test
+#endif
+
+#include CHECK_H
+
+#include <math.h>
+
+extern float roundf (float);
+
+#define NUM 64
+
+static void
+__attribute__((__target__("fpmath=sse")))
+init_src (float *src)
+{
+  int i, sign = 1;
+  float f = rand ();
+
+  for (i = 0; i < NUM; i++)
+    {
+      src[i] = (i + 1) * f * M_PI * sign;
+      if (i < (NUM / 2))
+       {
+          if ((i % 6) == 0)
+           f = f * src[i];
+        }
+      else if (i == (NUM / 2))
+       f = rand ();
+      else if ((i % 6) == 0)
+       f = 1 / (f * (i + 1) * src[i] * M_PI * sign);
+      sign = -sign;
+    }
+}
+
+static void
+__attribute__((__target__("fpmath=387")))
+TEST (void)
+{
+  float a[NUM];
+  int r[NUM];
+  int i;
+
+  init_src (a);
+
+  for (i = 0; i < NUM; i++)
+    r[i] = (int) roundf (a[i]);
+
+  /* check results:  */
+  for (i = 0; i < NUM; i++)
+    if (r[i] != (int) roundf (a[i]))
+      abort();
+}