[Bug rtl-optimization/37571] Performance regression with -mtune=core2

["Joey dot ye at intel dot com" <gcc-bugzilla at gcc dot gnu dot org>]

------- Comment #1 from Joey dot ye at intel dot com  2008-09-18 16:01 -------
Root cause is that instruction length of fused jcc is set to 16, which prevent
the block from merging and copying. For some reason Core2 runs poorly with a
unmerged branch block under certain circonstances.

Following patch fixes it:

Index: i386.md
===================================================================
--- i386.md     (revision 3923)
+++ i386.md     (working copy)
@@ -421,6 +421,9 @@
         ]
         (const_int 1)))

+(define_attr "length_jcc_fuse" ""
+  (const_int 0))
+
 ;; The (bounding maximum) length of an instruction in bytes.
 ;; ??? fistp and frndint are in fact fldcw/{fistp,frndint}/fldcw sequences.
 ;; Later we may want to split them and compute proper length as for
@@ -442,7 +445,8 @@
               (plus (attr "prefix_rep")
                     (plus (attr "prefix_data16")
                           (plus (attr "length_immediate")
-                                (attr "length_address")))))))
+                                (plus (attr "length_address")
+                                       (attr "length_jcc_fuse"))))))))

 ;; The `memory' attribute is `none' if no memory is referenced, `load' or
 ;; `store' if there is a simple memory reference therein, or `unknown'
@@ -645,7 +649,7 @@
 (include "k6.md")
 (include "athlon.md")
 (include "geode.md")
-;;(include "core2.md")
+(include "core2.md")


 ;; Operand and operator predicates and constraints
@@ -14033,7 +14037,8 @@
   return "test{<imodesuffix>}\t%2, %2\n\t"
         "%+j%E1\t%l0\t" ASM_COMMENT_START " fused";
 }
-  [(set_attr "type" "multi")
+  [(set_attr "type" "icmp")
+   (set_attr "length_jcc_fuse" "2")
    (set_attr "mode" "<MODE>")])

 (define_insn "*jcc_fused_2"
@@ -14048,7 +14053,8 @@
   return "test{<imodesuffix>}\t%2, %2\n\t"
         "%+j%e1\t%l0\t" ASM_COMMENT_START " fused";
 }
-  [(set_attr "type" "multi")
+  [(set_attr "type" "icmp")
+   (set_attr "length_jcc_fuse" "2")
    (set_attr "mode" "<MODE>")])

 (define_insn "*jcc_fused_3"
@@ -14066,7 +14072,8 @@
   return "cmp{<imodesuffix>}\t{%3, %2|%2, %3}\n\t"
         "%+j%E1\t%l0\t" ASM_COMMENT_START " fused";
 }
-  [(set_attr "type" "multi")
+  [(set_attr "type" "icmp")
+   (set_attr "length_jcc_fuse" "2")
    (set_attr "mode" "<MODE>")])

 (define_insn "*jcc_fused_4"
@@ -14084,7 +14091,8 @@
   return "cmp{<imodesuffix>}\t{%3, %2|%2, %3}\n\t"
         "%+j%e1\t%l0\t" ASM_COMMENT_START " fused";
 }
-  [(set_attr "type" "multi")
+  [(set_attr "type" "icmp")
+   (set_attr "length_jcc_fuse" "2")
    (set_attr "mode" "<MODE>")])

 ;; In general it is not safe to assume too much about CCmode registers,


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37571