[Bug target/37364] [4.4 Regression] IRA generates inefficient code due to missing regmove pass

[steven at gcc dot gnu dot org]

------- Comment #29 from steven at gcc dot gnu dot org  2008-11-30 21:32 -------
The insns 8 in comment #0 show the regmove transformation that matters here:

With regmove disabled::
(insn:HI 8 7 14 2 ../../include/mmintrin.h:300 (set (reg:V2SI 61)
        (plus:V2SI (reg:V2SI 63 [ x ])
            (mem/c/i:V2SI (symbol_ref:DI ("y") <var_decl 0x7f66abfb5c80 y>) [2
y
+0 S8 A64]))) 992 {*mmx_addv2si3} (expr_list:REG_DEAD (reg:V2SI 63 [ x ])
        (nil)))


vs. with regmove enabled:
(insn:HI 8 7 14 2 ../../include/mmintrin.h:300 (set (reg:V2SI 63 [ x ])
        (plus:V2SI (reg:V2SI 63 [ x ])
            (mem/c/i:V2SI (symbol_ref:DI ("y") <var_decl 0x7fd36e03cc80 y>) [2 
y+0 S8 A64]))) 992 {*mmx_addv2si3} (nil))


This is a "textbook" example of a regmove transformation (where the use of the
word "textbook" is an insult to all textbooks ever written, since nothing in
regmove is "textbook", but oh well...).  It's turned a 3-address instruction
into a 2-address instruction, and this is precisely the transformation that
regmove was originally written for.

In IRA, I would've expected reg 61 and reg 63 to be coalesced to give the same
result.

Vlad already commented on this in comment #6.  It's clear from his description
why IRA did not coalesce reg 61 and reg 63.  

Other compilers do this kind of transformation via reverse copy propagation. 
GCC could perhaps add something like that too, when it transforms a 3-address
insn to a 2-address insn.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37364