projects / gcc.git / commitdiff
? search:
summary | shortlog | log | commit | commitdiff | tree
raw | patch | inline | side by side (parent: 2be4091)
AVX-512. 56/n. Add plus/minus/abs/neg/andnot insn.
authorAlexander Ivchenko <alexander.ivchenko@intel.com>
Wed, 15 Oct 2014 11:04:03 +0000 (11:04 +0000)
committerKirill Yukhin <kyukhin@gcc.gnu.org>
Wed, 15 Oct 2014 11:04:03 +0000 (11:04 +0000)
gcc/
* config/i386/sse.md (define_mode_iterator VI_AVX2): Extend
to support AVX-512BW.
(define_mode_iterator VI124_AVX2_48_AVX512F): Remove.
(define_expand "<plusminus_insn><mode>3"): Remove masking support.
(define_insn "*<plusminus_insn><mode>3"): Ditto.
(define_expand "<plusminus_insn><VI48_AVX512VL:mode>3_mask"): New.
(define_expand "<plusminus_insn><VI12_AVX512VL:mode>3_mask"): Ditto.
(define_insn "*<plusminus_insn><VI48_AVX512VL:mode>3_mask"): Ditto.
(define_insn "*<plusminus_insn><VI12_AVX512VL:mode>3_mask"): Ditto.
(define_expand "<sse2_avx2>_andnot<mode>3"): Remove masking support.
(define_insn "*andnot<mode>3"): Ditto.
(define_expand "<sse2_avx2>_andnot<VI48_AVX512VL:mode>3_mask"): New.
(define_expand "<sse2_avx2>_andnot<VI12_AVX512VL:mode>3_mask"): Ditto.
(define_insn "*andnot<VI48_AVX512VL:mode>3<mask_name>"): Ditto.
(define_insn "*andnot<VI12_AVX512VL:mode>3<mask_name>"): Ditto.
(define_insn "*abs<mode>2"): Remove masking support.
(define_insn "abs<VI48_AVX512VL:mode>2_mask"): New.
(define_insn "abs<VI12_AVX512VL:mode>2_mask"): Ditto.
(define_expand "abs<mode>2"): Use VI_AVX2 mode iterator.

Co-Authored-By: Andrey Turetskiy <andrey.turetskiy@intel.com>
Co-Authored-By: Anna Tikhonova <anna.tikhonova@intel.com>
Co-Authored-By: Ilya Tocar <ilya.tocar@intel.com>
Co-Authored-By: Ilya Verbin <ilya.verbin@intel.com>
Co-Authored-By: Kirill Yukhin <kirill.yukhin@intel.com>
Co-Authored-By: Maxim Kuznetsov <maxim.kuznetsov@intel.com>
Co-Authored-By: Michael Zolotukhin <michael.v.zolotukhin@intel.com>
From-SVN: r216255

gcc/ChangeLog patch | blob | history
gcc/config/i386/sse.md patch | blob | history

diff --git a/gcc/ChangeLog b/gcc/ChangeLog
index 283c0d302cda13bde121c98f04bb7a5bca3e0665..efbbe411fd75742adb1ebe69344d216bb277b3f1 100644 (file)
--- a/gcc/ChangeLog
+++ b/gcc/ChangeLog
@@ -1,3 +1,32 @@
+2014-10-15  Alexander Ivchenko  <alexander.ivchenko@intel.com>
+           Maxim Kuznetsov  <maxim.kuznetsov@intel.com>
+           Anna Tikhonova  <anna.tikhonova@intel.com>
+           Ilya Tocar  <ilya.tocar@intel.com>
+           Andrey Turetskiy  <andrey.turetskiy@intel.com>
+           Ilya Verbin  <ilya.verbin@intel.com>
+           Kirill Yukhin  <kirill.yukhin@intel.com>
+           Michael Zolotukhin  <michael.v.zolotukhin@intel.com>
+
+       * config/i386/sse.md (define_mode_iterator VI_AVX2): Extend
+       to support AVX-512BW.
+       (define_mode_iterator VI124_AVX2_48_AVX512F): Remove.
+       (define_expand "<plusminus_insn><mode>3"): Remove masking support.
+       (define_insn "*<plusminus_insn><mode>3"): Ditto.
+       (define_expand "<plusminus_insn><VI48_AVX512VL:mode>3_mask"): New.
+       (define_expand "<plusminus_insn><VI12_AVX512VL:mode>3_mask"): Ditto.
+       (define_insn "*<plusminus_insn><VI48_AVX512VL:mode>3_mask"): Ditto.
+       (define_insn "*<plusminus_insn><VI12_AVX512VL:mode>3_mask"): Ditto.
+       (define_expand "<sse2_avx2>_andnot<mode>3"): Remove masking support.
+       (define_insn "*andnot<mode>3"): Ditto.
+       (define_expand "<sse2_avx2>_andnot<VI48_AVX512VL:mode>3_mask"): New.
+       (define_expand "<sse2_avx2>_andnot<VI12_AVX512VL:mode>3_mask"): Ditto.
+       (define_insn "*andnot<VI48_AVX512VL:mode>3<mask_name>"): Ditto.
+       (define_insn "*andnot<VI12_AVX512VL:mode>3<mask_name>"): Ditto.
+       (define_insn "*abs<mode>2"): Remove masking support.
+       (define_insn "abs<VI48_AVX512VL:mode>2_mask"): New.
+       (define_insn "abs<VI12_AVX512VL:mode>2_mask"): Ditto.
+       (define_expand "abs<mode>2"): Use VI_AVX2 mode iterator.
+
 2014-10-15  Alexander Ivchenko  <alexander.ivchenko@intel.com>
            Maxim Kuznetsov  <maxim.kuznetsov@intel.com>
            Anna Tikhonova  <anna.tikhonova@intel.com>
diff --git a/gcc/config/i386/sse.md b/gcc/config/i386/sse.md
index 018702afca8a166aadf157c27d53f9630201c674..a6cf363957657b121f2594661df5e02711059c57 100644 (file)
--- a/gcc/config/i386/sse.md
+++ b/gcc/config/i386/sse.md
@@ -271,8 +271,8 @@
    (V4DI "TARGET_AVX") V2DI])
 
 (define_mode_iterator VI_AVX2
-  [(V32QI "TARGET_AVX2") V16QI
-   (V16HI "TARGET_AVX2") V8HI
+  [(V64QI "TARGET_AVX512BW") (V32QI "TARGET_AVX2") V16QI
+   (V32HI "TARGET_AVX512BW") (V16HI "TARGET_AVX2") V8HI
    (V16SI "TARGET_AVX512F") (V8SI "TARGET_AVX2") V4SI
    (V8DI "TARGET_AVX512F") (V4DI "TARGET_AVX2") V2DI])
 
@@ -362,12 +362,6 @@
   [(V16HI "TARGET_AVX2") V8HI
    (V8SI "TARGET_AVX2") V4SI])
 
-(define_mode_iterator VI124_AVX2_48_AVX512F
-  [(V32QI "TARGET_AVX2") V16QI
-   (V16HI "TARGET_AVX2") V8HI
-   (V16SI "TARGET_AVX512F") (V8SI "TARGET_AVX2") V4SI
-   (V8DI "TARGET_AVX512F")])
-
 (define_mode_iterator VI124_AVX512F
   [(V32QI "TARGET_AVX2") V16QI
    (V32HI "TARGET_AVX512F") (V16HI "TARGET_AVX2") V8HI
@@ -9143,20 +9137,43 @@
   "TARGET_SSE2"
   "operands[2] = force_reg (<MODE>mode, CONST0_RTX (<MODE>mode));")
 
-(define_expand "<plusminus_insn><mode>3<mask_name>"
+(define_expand "<plusminus_insn><mode>3"
   [(set (match_operand:VI_AVX2 0 "register_operand")
        (plusminus:VI_AVX2
          (match_operand:VI_AVX2 1 "nonimmediate_operand")
          (match_operand:VI_AVX2 2 "nonimmediate_operand")))]
-  "TARGET_SSE2 && <mask_mode512bit_condition>"
+  "TARGET_SSE2"
+  "ix86_fixup_binary_operands_no_copy (<CODE>, <MODE>mode, operands);")
+
+(define_expand "<plusminus_insn><mode>3_mask"
+  [(set (match_operand:VI48_AVX512VL 0 "register_operand")
+       (vec_merge:VI48_AVX512VL
+         (plusminus:VI48_AVX512VL
+           (match_operand:VI48_AVX512VL 1 "nonimmediate_operand")
+           (match_operand:VI48_AVX512VL 2 "nonimmediate_operand"))
+         (match_operand:VI48_AVX512VL 3 "vector_move_operand")
+         (match_operand:<avx512fmaskmode> 4 "register_operand")))]
+  "TARGET_AVX512F"
+  "ix86_fixup_binary_operands_no_copy (<CODE>, <MODE>mode, operands);")
+
+(define_expand "<plusminus_insn><mode>3_mask"
+  [(set (match_operand:VI12_AVX512VL 0 "register_operand")
+       (vec_merge:VI12_AVX512VL
+         (plusminus:VI12_AVX512VL
+           (match_operand:VI12_AVX512VL 1 "nonimmediate_operand")
+           (match_operand:VI12_AVX512VL 2 "nonimmediate_operand"))
+         (match_operand:VI12_AVX512VL 3 "vector_move_operand")
+         (match_operand:<avx512fmaskmode> 4 "register_operand")))]
+  "TARGET_AVX512BW"
   "ix86_fixup_binary_operands_no_copy (<CODE>, <MODE>mode, operands);")
 
-(define_insn "*<plusminus_insn><mode>3<mask_name>"
+(define_insn "*<plusminus_insn><mode>3"
   [(set (match_operand:VI_AVX2 0 "register_operand" "=x,v")
        (plusminus:VI_AVX2
          (match_operand:VI_AVX2 1 "nonimmediate_operand" "<comm>0,v")
          (match_operand:VI_AVX2 2 "nonimmediate_operand" "xm,vm")))]
-  "TARGET_SSE2 && ix86_binary_operator_ok (<CODE>, <MODE>mode, operands) && <mask_mode512bit_condition>"
+  "TARGET_SSE2
+   && ix86_binary_operator_ok (<CODE>, <MODE>mode, operands)"
   "@
    p<plusminus_mnemonic><ssemodesuffix>\t{%2, %0|%0, %2}
    vp<plusminus_mnemonic><ssemodesuffix>\t{%2, %1, %0<mask_operand3>|%0<mask_operand3>, %1, %2}"
@@ -9166,6 +9183,35 @@
    (set_attr "prefix" "<mask_prefix3>")
    (set_attr "mode" "<sseinsnmode>")])
 
+(define_insn "*<plusminus_insn><mode>3_mask"
+  [(set (match_operand:VI48_AVX512VL 0 "register_operand" "=v")
+       (vec_merge:VI48_AVX512VL
+         (plusminus:VI48_AVX512VL
+           (match_operand:VI48_AVX512VL 1 "nonimmediate_operand" "<comm>v")
+           (match_operand:VI48_AVX512VL 2 "nonimmediate_operand" "vm"))
+         (match_operand:VI48_AVX512VL 3 "vector_move_operand" "0C")
+         (match_operand:<avx512fmaskmode> 4 "register_operand" "Yk")))]
+  "TARGET_AVX512F
+   && ix86_binary_operator_ok (<CODE>, <MODE>mode, operands)"
+  "vp<plusminus_mnemonic><ssemodesuffix>\t{%2, %1, %0%{%4%}%N3|%0%{%4%}%N3, %1, %2}"
+  [(set_attr "type" "sseiadd")
+   (set_attr "prefix" "evex")
+   (set_attr "mode" "<sseinsnmode>")])
+
+(define_insn "*<plusminus_insn><mode>3_mask"
+  [(set (match_operand:VI12_AVX512VL 0 "register_operand" "=v")
+       (vec_merge:VI12_AVX512VL
+         (plusminus:VI12_AVX512VL
+           (match_operand:VI12_AVX512VL 1 "nonimmediate_operand" "<comm>v")
+           (match_operand:VI12_AVX512VL 2 "nonimmediate_operand" "vm"))
+         (match_operand:VI12_AVX512VL 3 "vector_move_operand" "0C")
+         (match_operand:<avx512fmaskmode> 4 "register_operand" "Yk")))]
+  "TARGET_AVX512BW && ix86_binary_operator_ok (<CODE>, <MODE>mode, operands)"
+  "vp<plusminus_mnemonic><ssemodesuffix>\t{%2, %1, %0%{%4%}%N3|%0%{%4%}%N3, %1, %2}"
+  [(set_attr "type" "sseiadd")
+   (set_attr "prefix" "evex")
+   (set_attr "mode" "<sseinsnmode>")])
+
 (define_expand "<sse2_avx2>_<plusminus_insn><mode>3<mask_name>"
   [(set (match_operand:VI12_AVX2 0 "register_operand")
        (sat_plusminus:VI12_AVX2
@@ -10654,19 +10700,41 @@
   operands[2] = force_reg (<MODE>mode, gen_rtx_CONST_VECTOR (<MODE>mode, v));
 })
 
-(define_expand "<sse2_avx2>_andnot<mode>3<mask_name>"
+(define_expand "<sse2_avx2>_andnot<mode>3"
   [(set (match_operand:VI_AVX2 0 "register_operand")
        (and:VI_AVX2
          (not:VI_AVX2 (match_operand:VI_AVX2 1 "register_operand"))
          (match_operand:VI_AVX2 2 "nonimmediate_operand")))]
-  "TARGET_SSE2 && <mask_mode512bit_condition>")
+  "TARGET_SSE2")
 
-(define_insn "*andnot<mode>3<mask_name>"
+(define_expand "<sse2_avx2>_andnot<mode>3_mask"
+  [(set (match_operand:VI48_AVX512VL 0 "register_operand")
+       (vec_merge:VI48_AVX512VL
+         (and:VI48_AVX512VL
+           (not:VI48_AVX512VL
+             (match_operand:VI48_AVX512VL 1 "register_operand"))
+           (match_operand:VI48_AVX512VL 2 "nonimmediate_operand"))
+         (match_operand:VI48_AVX512VL 3 "vector_move_operand")
+         (match_operand:<avx512fmaskmode> 4 "register_operand")))]
+  "TARGET_AVX512F")
+
+(define_expand "<sse2_avx2>_andnot<mode>3_mask"
+  [(set (match_operand:VI12_AVX512VL 0 "register_operand")
+       (vec_merge:VI12_AVX512VL
+         (and:VI12_AVX512VL
+           (not:VI12_AVX512VL
+             (match_operand:VI12_AVX512VL 1 "register_operand"))
+           (match_operand:VI12_AVX512VL 2 "nonimmediate_operand"))
+         (match_operand:VI12_AVX512VL 3 "vector_move_operand")
+         (match_operand:<avx512fmaskmode> 4 "register_operand")))]
+  "TARGET_AVX512BW")
+
+(define_insn "*andnot<mode>3"
   [(set (match_operand:VI 0 "register_operand" "=x,v")
        (and:VI
          (not:VI (match_operand:VI 1 "register_operand" "0,v"))
          (match_operand:VI 2 "nonimmediate_operand" "xm,vm")))]
-  "TARGET_SSE && <mask_mode512bit_condition>"
+  "TARGET_SSE"
 {
   static char buf[64];
   const char *ops;
@@ -10740,7 +10808,7 @@
            (eq_attr "mode" "TI"))
        (const_string "1")
        (const_string "*")))
-   (set_attr "prefix" "<mask_prefix3>")
+   (set_attr "prefix" "orig,vex")
    (set (attr "mode")
        (cond [(and (match_test "<MODE_SIZE> == 16")
                    (match_test "TARGET_SSE_PACKED_SINGLE_INSN_OPTIMAL"))
@@ -10758,6 +10826,36 @@
              ]
              (const_string "<sseinsnmode>")))])
 
+(define_insn "*andnot<mode>3_mask"
+  [(set (match_operand:VI48_AVX512VL 0 "register_operand" "=v")
+       (vec_merge:VI48_AVX512VL
+         (and:VI48_AVX512VL
+           (not:VI48_AVX512VL
+             (match_operand:VI48_AVX512VL 1 "register_operand" "v"))
+           (match_operand:VI48_AVX512VL 2 "nonimmediate_operand" "vm"))
+         (match_operand:VI48_AVX512VL 3 "vector_move_operand" "0C")
+         (match_operand:<avx512fmaskmode> 4 "register_operand" "Yk")))]
+  "TARGET_AVX512F"
+  "vpandn<ssemodesuffix>\t{%2, %1, %0%{%4%}%N3|%0%{%4%}%N3, %1, %2}";
+  [(set_attr "type" "sselog")
+   (set_attr "prefix" "evex")
+   (set_attr "mode" "<sseinsnmode>")])
+
+(define_insn "*andnot<mode>3_mask"
+  [(set (match_operand:VI12_AVX512VL 0 "register_operand" "=v")
+       (vec_merge:VI12_AVX512VL
+         (and:VI12_AVX512VL
+           (not:VI12_AVX512VL
+             (match_operand:VI12_AVX512VL 1 "register_operand" "v"))
+           (match_operand:VI12_AVX512VL 2 "nonimmediate_operand" "vm"))
+         (match_operand:VI12_AVX512VL 3 "vector_move_operand" "0C")
+         (match_operand:<avx512fmaskmode> 4 "register_operand" "Yk")))]
+  "TARGET_AVX512BW"
+  "vpandn<ssemodesuffix>\t{%2, %1, %0%{%4%}%N3|%0%{%4%}%N3, %1, %2}";
+  [(set_attr "type" "sselog")
+   (set_attr "prefix" "evex")
+   (set_attr "mode" "<sseinsnmode>")])
+
 (define_expand "<code><mode>3"
   [(set (match_operand:VI 0 "register_operand")
        (any_logic:VI
@@ -13673,22 +13771,48 @@
    (set (attr "prefix_rex") (symbol_ref "x86_extended_reg_mentioned_p (insn)"))
    (set_attr "mode" "DI")])
 
-(define_insn "<mask_codefor>abs<mode>2<mask_name>"
-  [(set (match_operand:VI124_AVX2_48_AVX512F 0 "register_operand" "=v")
-       (abs:VI124_AVX2_48_AVX512F
-         (match_operand:VI124_AVX2_48_AVX512F 1 "nonimmediate_operand" "vm")))]
-  "TARGET_SSSE3 && <mask_mode512bit_condition>"
-  "%vpabs<ssemodesuffix>\t{%1, %0<mask_operand2>|%0<mask_operand2>, %1}"
+(define_insn "*abs<mode>2"
+  [(set (match_operand:VI_AVX2 0 "register_operand" "=v")
+       (abs:VI_AVX2
+         (match_operand:VI_AVX2 1 "nonimmediate_operand" "vm")))]
+  "TARGET_SSSE3"
+  "%vpabs<ssemodesuffix>\t{%1, %0|%0, %1}"
   [(set_attr "type" "sselog1")
    (set_attr "prefix_data16" "1")
    (set_attr "prefix_extra" "1")
    (set_attr "prefix" "maybe_vex")
    (set_attr "mode" "<sseinsnmode>")])
 
+(define_insn "abs<mode>2_mask"
+  [(set (match_operand:VI48_AVX512VL 0 "register_operand" "=v")
+       (vec_merge:VI48_AVX512VL
+         (abs:VI48_AVX512VL
+           (match_operand:VI48_AVX512VL 1 "nonimmediate_operand" "vm"))
+         (match_operand:VI48_AVX512VL 2 "vector_move_operand" "0C")
+         (match_operand:<avx512fmaskmode> 3 "register_operand" "Yk")))]
+  "TARGET_AVX512F"
+  "vpabs<ssemodesuffix>\t{%1, %0%{%3%}%N2|%0%{%3%}%N2, %1}"
+  [(set_attr "type" "sselog1")
+   (set_attr "prefix" "evex")
+   (set_attr "mode" "<sseinsnmode>")])
+
+(define_insn "abs<mode>2_mask"
+  [(set (match_operand:VI12_AVX512VL 0 "register_operand" "=v")
+       (vec_merge:VI12_AVX512VL
+         (abs:VI12_AVX512VL
+           (match_operand:VI12_AVX512VL 1 "nonimmediate_operand" "vm"))
+         (match_operand:VI12_AVX512VL 2 "vector_move_operand" "0C")
+         (match_operand:<avx512fmaskmode> 3 "register_operand" "Yk")))]
+  "TARGET_AVX512BW"
+  "vpabs<ssemodesuffix>\t{%1, %0%{%3%}%N2|%0%{%3%}%N2, %1}"
+  [(set_attr "type" "sselog1")
+   (set_attr "prefix" "evex")
+   (set_attr "mode" "<sseinsnmode>")])
+
 (define_expand "abs<mode>2"
-  [(set (match_operand:VI124_AVX2_48_AVX512F 0 "register_operand")
-       (abs:VI124_AVX2_48_AVX512F
-         (match_operand:VI124_AVX2_48_AVX512F 1 "nonimmediate_operand")))]
+  [(set (match_operand:VI_AVX2 0 "register_operand")
+       (abs:VI_AVX2
+         (match_operand:VI_AVX2 1 "nonimmediate_operand")))]
   "TARGET_SSE2"
 {
   if (!TARGET_SSSE3)