]> git.ipfire.org Git - thirdparty/gcc.git/commitdiff
i386: Emulate MMX mmx_uavgv8qi3 with SSE
authorH.J. Lu <hongjiu.lu@intel.com>
Wed, 15 May 2019 15:19:19 +0000 (15:19 +0000)
committerH.J. Lu <hjl@gcc.gnu.org>
Wed, 15 May 2019 15:19:19 +0000 (08:19 -0700)
Emulate MMX mmx_uavgv8qi3 with SSE.  Only SSE register source operand is
allowed.

PR target/89021
* config/i386/mmx.md (mmx_uavgv8qi3): Also check TARGET_MMX
and TARGET_MMX_WITH_SSE.
(*mmx_uavgv8qi3): Add SSE emulation.

From-SVN: r271235

gcc/ChangeLog
gcc/config/i386/mmx.md

index 8f3e5f88abc4a091fcea67a11a09fcc1fc8d2fea..0857df4a3a590a25d249660292fd7e4cbb3f2316 100644 (file)
@@ -1,3 +1,10 @@
+2019-05-15  H.J. Lu  <hongjiu.lu@intel.com>
+
+       PR target/89021
+       * config/i386/mmx.md (mmx_uavgv8qi3): Also check TARGET_MMX
+       and TARGET_MMX_WITH_SSE.
+       (*mmx_uavgv8qi3): Add SSE emulation.
+
 2019-05-15  H.J. Lu  <hongjiu.lu@intel.com>
 
        PR target/89021
index e2a1f4c8ace14b940f01346c03127bd065c3b116..9f14d0d015d6e9eae25083111b1c6e266360c3dd 100644 (file)
            (plus:V8HI
              (plus:V8HI
                (zero_extend:V8HI
-                 (match_operand:V8QI 1 "nonimmediate_operand"))
+                 (match_operand:V8QI 1 "register_mmxmem_operand"))
                (zero_extend:V8HI
-                 (match_operand:V8QI 2 "nonimmediate_operand")))
+                 (match_operand:V8QI 2 "register_mmxmem_operand")))
              (const_vector:V8HI [(const_int 1) (const_int 1)
                                  (const_int 1) (const_int 1)
                                  (const_int 1) (const_int 1)
                                  (const_int 1) (const_int 1)]))
            (const_int 1))))]
-  "TARGET_SSE || TARGET_3DNOW"
+  "(TARGET_MMX || TARGET_MMX_WITH_SSE)
+   && (TARGET_SSE || TARGET_3DNOW_A)"
   "ix86_fixup_binary_operands_no_copy (PLUS, V8QImode, operands);")
 
 (define_insn "*mmx_uavgv8qi3"
-  [(set (match_operand:V8QI 0 "register_operand" "=y")
+  [(set (match_operand:V8QI 0 "register_operand" "=y,x,Yv")
        (truncate:V8QI
          (lshiftrt:V8HI
            (plus:V8HI
              (plus:V8HI
                (zero_extend:V8HI
-                 (match_operand:V8QI 1 "nonimmediate_operand" "%0"))
+                 (match_operand:V8QI 1 "register_mmxmem_operand" "%0,0,Yv"))
                (zero_extend:V8HI
-                 (match_operand:V8QI 2 "nonimmediate_operand" "ym")))
+                 (match_operand:V8QI 2 "register_mmxmem_operand" "ym,x,Yv")))
              (const_vector:V8HI [(const_int 1) (const_int 1)
                                  (const_int 1) (const_int 1)
                                  (const_int 1) (const_int 1)
                                  (const_int 1) (const_int 1)]))
            (const_int 1))))]
-  "(TARGET_SSE || TARGET_3DNOW)
+  "(TARGET_MMX || TARGET_MMX_WITH_SSE)
+   && (TARGET_SSE || TARGET_3DNOW_A)
    && ix86_binary_operator_ok (PLUS, V8QImode, operands)"
 {
   /* These two instructions have the same operation, but their encoding
      is different.  Prefer the one that is de facto standard.  */
-  if (TARGET_SSE || TARGET_3DNOW_A)
+  if (TARGET_MMX_WITH_SSE && TARGET_AVX)
+    return "vpavgb\t{%2, %1, %0|%0, %1, %2}";
+  else if (TARGET_SSE || TARGET_3DNOW_A)
     return "pavgb\t{%2, %0|%0, %2}";
   else
     return "pavgusb\t{%2, %0|%0, %2}";
 }
-  [(set_attr "type" "mmxshft")
+  [(set_attr "mmx_isa" "native,x64_noavx,x64_avx")
+   (set_attr "type" "mmxshft,sseiadd,sseiadd")
    (set (attr "prefix_extra")
      (if_then_else
        (not (ior (match_test "TARGET_SSE")
                 (match_test "TARGET_3DNOW_A")))
        (const_string "1")
        (const_string "*")))
-   (set_attr "mode" "DI")])
+   (set_attr "mode" "DI,TI,TI")])
 
 (define_expand "mmx_uavgv4hi3"
   [(set (match_operand:V4HI 0 "register_operand")