projects / gcc.git / commitdiff
? search:
summary | shortlog | log | commit | commitdiff | tree
raw | patch | inline | side by side (parent: 6edf4f2)
AVX-512. Add byte/word plus/minus/avg.
authorAlexander Ivchenko <alexander.ivchenko@intel.com>
Thu, 28 Aug 2014 06:37:12 +0000 (06:37 +0000)
committerKirill Yukhin <kyukhin@gcc.gnu.org>
Thu, 28 Aug 2014 06:37:12 +0000 (06:37 +0000)
gcc/
(define_mode_iterator VI12_AVX2): Add V64QI and V32HI modes.
(define_expand "<sse2_avx2>_<plusminus_insn><mode>3<mask_name>"): Add
masking.
(define_insn "*<sse2_avx2>_<plusminus_insn><mode>3<mask_name>"): Ditto.
(define_expand "<sse2_avx2>_uavg<mode>3<mask_name>"): Ditto.
(define_insn "*<sse2_avx2>_uavg<mode>3<mask_name>"): Ditto.
(define_insn "*mul<mode>3"): Add EVEX version.

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: r214672

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

diff --git a/gcc/ChangeLog b/gcc/ChangeLog
index 9d545feaa6e3b323bc323ff8aa9489489f3f984b..10fa23bf498a8e88ed35af26cc98908d29915c97 100644 (file)
--- a/gcc/ChangeLog
+++ b/gcc/ChangeLog
@@ -1,3 +1,20 @@
+2014-08-28  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>
+
+       (define_mode_iterator VI12_AVX2): Add V64QI and V32HI modes.
+       (define_expand "<sse2_avx2>_<plusminus_insn><mode>3<mask_name>"): Add
+       masking.
+       (define_insn "*<sse2_avx2>_<plusminus_insn><mode>3<mask_name>"): Ditto.
+       (define_expand "<sse2_avx2>_uavg<mode>3<mask_name>"): Ditto.
+       (define_insn "*<sse2_avx2>_uavg<mode>3<mask_name>"): Ditto.
+       (define_insn "*mul<mode>3"): Add EVEX version.
+
 2014-08-28  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 03512a5c55a1ef5c4ed8da20c46b1464c01b476b..c602eebf2ff3ded25d9e1132ec448c9865097d57 100644 (file)
--- a/gcc/config/i386/sse.md
+++ b/gcc/config/i386/sse.md
@@ -338,8 +338,8 @@
   [(V2TI "TARGET_AVX2") TI])
 
 (define_mode_iterator VI12_AVX2
-  [(V32QI "TARGET_AVX2") V16QI
-   (V16HI "TARGET_AVX2") V8HI])
+  [(V64QI "TARGET_AVX512BW") (V32QI "TARGET_AVX2") V16QI
+   (V32HI "TARGET_AVX512BW") (V16HI "TARGET_AVX2") V8HI])
 
 (define_mode_iterator VI24_AVX2
   [(V16HI "TARGET_AVX2") V8HI
@@ -8109,27 +8109,28 @@
    (set_attr "prefix" "<mask_prefix3>")
    (set_attr "mode" "<sseinsnmode>")])
 
-(define_expand "<sse2_avx2>_<plusminus_insn><mode>3"
+(define_expand "<sse2_avx2>_<plusminus_insn><mode>3<mask_name>"
   [(set (match_operand:VI12_AVX2 0 "register_operand")
        (sat_plusminus:VI12_AVX2
          (match_operand:VI12_AVX2 1 "nonimmediate_operand")
          (match_operand:VI12_AVX2 2 "nonimmediate_operand")))]
-  "TARGET_SSE2"
+  "TARGET_SSE2 && <mask_mode512bit_condition> && <mask_avx512bw_condition>"
   "ix86_fixup_binary_operands_no_copy (<CODE>, <MODE>mode, operands);")
 
-(define_insn "*<sse2_avx2>_<plusminus_insn><mode>3"
+(define_insn "*<sse2_avx2>_<plusminus_insn><mode>3<mask_name>"
   [(set (match_operand:VI12_AVX2 0 "register_operand" "=x,v")
        (sat_plusminus:VI12_AVX2
          (match_operand:VI12_AVX2 1 "nonimmediate_operand" "<comm>0,v")
          (match_operand:VI12_AVX2 2 "nonimmediate_operand" "xm,vm")))]
-  "TARGET_SSE2 && ix86_binary_operator_ok (<CODE>, <MODE>mode, operands)"
+  "TARGET_SSE2 && <mask_mode512bit_condition> && <mask_avx512bw_condition>
+   && 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|%0, %1, %2}"
+   vp<plusminus_mnemonic><ssemodesuffix>\t{%2, %1, %0<mask_operand3>|%0<mask_operand3>, %1, %2}"
   [(set_attr "isa" "noavx,avx")
    (set_attr "type" "sseiadd")
    (set_attr "prefix_data16" "1,*")
-   (set_attr "prefix" "orig,vex")
+   (set_attr "prefix" "orig,maybe_evex")
    (set_attr "mode" "TI")])
 
 (define_expand "mul<mode>3"
@@ -8151,8 +8152,8 @@
 
 (define_insn "*mul<mode>3"
   [(set (match_operand:VI2_AVX2 0 "register_operand" "=x,x")
-       (mult:VI2_AVX2 (match_operand:VI2_AVX2 1 "nonimmediate_operand" "%0,x")
-                      (match_operand:VI2_AVX2 2 "nonimmediate_operand" "xm,xm")))]
+        (mult:VI2_AVX2 (match_operand:VI2_AVX2 1 "nonimmediate_operand" "%0,v")
+                       (match_operand:VI2_AVX2 2 "nonimmediate_operand" "xm,vm")))]
   "TARGET_SSE2 && ix86_binary_operator_ok (MULT, <MODE>mode, operands)"
   "@
    pmullw\t{%2, %0|%0, %2}
@@ -11318,7 +11319,7 @@
 ;;
 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 
-(define_expand "<sse2_avx2>_uavg<mode>3"
+(define_expand "<sse2_avx2>_uavg<mode>3<mask_name>"
   [(set (match_operand:VI12_AVX2 0 "register_operand")
        (truncate:VI12_AVX2
          (lshiftrt:<ssedoublemode>
@@ -11328,34 +11329,44 @@
                  (match_operand:VI12_AVX2 1 "nonimmediate_operand"))
                (zero_extend:<ssedoublemode>
                  (match_operand:VI12_AVX2 2 "nonimmediate_operand")))
-             (match_dup 3))
+             (match_dup <mask_expand_op3>))
            (const_int 1))))]
-  "TARGET_SSE2"
+  "TARGET_SSE2 && <mask_mode512bit_condition> && <mask_avx512bw_condition>"
 {
+  rtx tmp;
+  if (<mask_applied>)
+    tmp = operands[3];
   operands[3] = CONST1_RTX(<MODE>mode);
   ix86_fixup_binary_operands_no_copy (PLUS, <MODE>mode, operands);
+
+  if (<mask_applied>)
+    {
+      operands[5] = operands[3];
+      operands[3] = tmp;
+    }
 })
 
-(define_insn "*<sse2_avx2>_uavg<mode>3"
-  [(set (match_operand:VI12_AVX2 0 "register_operand" "=x,x")
+(define_insn "*<sse2_avx2>_uavg<mode>3<mask_name>"
+  [(set (match_operand:VI12_AVX2 0 "register_operand" "=x,v")
        (truncate:VI12_AVX2
          (lshiftrt:<ssedoublemode>
            (plus:<ssedoublemode>
              (plus:<ssedoublemode>
                (zero_extend:<ssedoublemode>
-                 (match_operand:VI12_AVX2 1 "nonimmediate_operand" "%0,x"))
+                 (match_operand:VI12_AVX2 1 "nonimmediate_operand" "%0,v"))
                (zero_extend:<ssedoublemode>
-                 (match_operand:VI12_AVX2 2 "nonimmediate_operand" "xm,xm")))
-             (match_operand:VI12_AVX2 3 "const1_operand"))
+                 (match_operand:VI12_AVX2 2 "nonimmediate_operand" "xm,vm")))
+             (match_operand:VI12_AVX2 <mask_expand_op3> "const1_operand"))
            (const_int 1))))]
-  "TARGET_SSE2 && ix86_binary_operator_ok (PLUS, <MODE>mode, operands)"
+  "TARGET_SSE2 && <mask_mode512bit_condition> && <mask_avx512bw_condition>
+   && ix86_binary_operator_ok (PLUS, <MODE>mode, operands)"
   "@
    pavg<ssemodesuffix>\t{%2, %0|%0, %2}
-   vpavg<ssemodesuffix>\t{%2, %1, %0|%0, %1, %2}"
+   vpavg<ssemodesuffix>\t{%2, %1, %0<mask_operand3>|%0<mask_operand3>, %1, %2}"
   [(set_attr "isa" "noavx,avx")
    (set_attr "type" "sseiadd")
    (set_attr "prefix_data16" "1,*")
-   (set_attr "prefix" "orig,vex")
+   (set_attr "prefix" "orig,<mask_prefix>")
    (set_attr "mode" "<sseinsnmode>")])
 
 ;; The correct representation for this is absolutely enormous, and