[Bug rtl-optimization/17236] inefficient code for long long multiply on x86

[ubizjak at gmail dot com]

------- Comment #18 from ubizjak at gmail dot com  2007-12-19 12:11 -------
Another baby step can be performed by:

Index: optabs.c
===================================================================
--- optabs.c    (revision 131053)
+++ optabs.c    (working copy)
@@ -1245,7 +1245,7 @@ expand_doubleword_mult (enum machine_mod
        return NULL_RTX;
     }

-  adjust = expand_binop (word_mode, smul_optab, op0_high, op1_low,
+  adjust = expand_binop (word_mode, smul_optab, op1_low, op0_high,
                         NULL_RTX, 0, OPTAB_DIRECT);
   if (!adjust)
     return NULL_RTX;
@@ -1274,7 +1274,7 @@ expand_doubleword_mult (enum machine_mod
        return NULL_RTX;
     }

-  temp = expand_binop (word_mode, smul_optab, op1_high, op0_low,
+  temp = expand_binop (word_mode, smul_optab, op0_low, op1_high,
                       NULL_RTX, 0, OPTAB_DIRECT);
   if (!temp)
     return NULL_RTX;

This patch generates (note mem operands to imull):

test:
        subl    $8, %esp
        movl    %ebx, (%esp)
        movl    12(%esp), %ebx
        movl    %esi, 4(%esp)
        movl    20(%esp), %esi
        movl    %ebx, %ecx
        movl    %esi, %eax
        imull   24(%esp), %ecx
        imull   16(%esp), %eax
        addl    %eax, %ecx
        movl    %esi, %eax
        mull    %ebx
        movl    (%esp), %ebx
        leal    (%ecx,%edx), %esi
        movl    %esi, %edx
        movl    4(%esp), %esi
        addl    $8, %esp
        ret

but since mulsi3 is defined as:

(define_insn "*mulsi3_1"
  [(set (match_operand:SI 0 "register_operand" "=r,r,r")
        (mult:SI (match_operand:SI 1 "nonimmediate_operand" "%rm,rm,0")
                 (match_operand:SI 2 "general_operand" "K,i,mr")))
   (clobber (reg:CC FLAGS_REG))]

the question is, why didn't RA perform operand reversal by itself. Currently it
converts _both_ SI multiplies from:

(insn:HI 8 7 9 2 m.c:14 (parallel [
            (set (reg:SI 63)
                (mult:SI (mem/c/i:SI (plus:SI (reg/f:SI 16 argp)
                            (const_int 4 [0x4])) [2 a+4 S4 A32])
                    (reg:SI 66 [ b ])))
            (clobber (reg:CC 17 flags))
        ]) 214 {*mulsi3_1} (expr_list:REG_UNUSED (reg:CC 17 flags)
        (nil)))

into

--cut here--
(insn 35 7 8 2 m.c:14 (set (reg:SI 0 ax [63])
        (mem/c/i:SI (plus:SI (reg/f:SI 7 sp)
                (const_int 16 [0x10])) [2 a+4 S4 A32])) 41 {*movsi_1} (nil))

(insn:HI 8 35 9 2 m.c:14 (parallel [
            (set (reg:SI 0 ax [63])
                (mult:SI (reg:SI 0 ax [63])
                    (reg:SI 4 si [orig:66 b ] [66])))
            (clobber (reg:CC 17 flags))
        ]) 214 {*mulsi3_1} (nil))
--cut here--


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=17236