Add comment on ffs to say why optimizing it further is a bad idea.

From-SVN: r8564

gcc/config/i386/i386.md

@@ -5439,6 +5439,10 @@
   ""
   "operands[2] = gen_reg_rtx (SImode);")
 
+;; Note, you cannot optimize away the branch following the bsfl by assuming
+;; that the destination is not modified if the input is 0, since not all
+;; x86 implementations do this.
+
 (define_insn ""
   [(set (match_operand:SI 0 "general_operand" "=&r")
 	(plus:SI (ffs:SI (match_operand:SI 1 "general_operand" "rm"))
@@ -5453,7 +5457,6 @@
   xops[0] = operands[0];
   xops[1] = operands[1];
   xops[2] = constm1_rtx;
-  /* Can there be a way to avoid the jump here?  */
   output_asm_insn (AS2 (bsf%L0,%1,%0), xops);
 #ifdef LOCAL_LABEL_PREFIX
   sprintf (buffer, \"jnz %sLFFSSI%d\",