[Bug middle-end/52890] New: Revision 185336 causes 10% degradation on cpu2000 benchmark 252.eon

[pthaugen at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=52890

             Bug #: 52890
           Summary: Revision 185336 causes 10% degradation on cpu2000
                    benchmark 252.eon
    Classification: Unclassified
           Product: gcc
           Version: 4.8.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: middle-end
        AssignedTo: unassig...@gcc.gnu.org
        ReportedBy: pthau...@gcc.gnu.org
                CC: berg...@gcc.gnu.org, mjam...@suse.cz
              Host: powerpc64-linux
            Target: powerpc64-linux
             Build: powerpc64-linux


Following testcase extracted from benchmark file mrXZRectangle.cc.

$ cat test.cc
// g++ -m64 -O2 -mcpu=power7 -S test.cc
class ggVector3 {
public:
    ggVector3() {e[0] = 1.0; e[1] = e[2] = 0.0; }
    ggVector3(double e0, double e1, double e2) {
                   e[0] = e0; e[1] = e1; e[2] = e2; }
    double e[3];
};

class ggONB3 {
public:
    ggONB3() {  }
    ggONB3(const ggVector3& a, const ggVector3& b, const ggVector3& c) {
         U = a; V = b; W = c; }
private:
    ggVector3 U,V,W;
};

class mrViewingHitRecord {
public:
    ggONB3 UVW;
};

void foo(mrViewingHitRecord& VHR)
{

      VHR.UVW = ggONB3(ggVector3 (1, 0, 0),
                       ggVector3 (0, 0, -1),
                       ggVector3 (0, 1, 0));
}

Revision 185335 produces tight sequence of 8-byte stores:

        stfd 0,8(3)
        stfd 0,16(3)
        stfd 0,24(3)
        stfd 0,32(3)
        ...

Revision 185336, with additional check for SLOW_UNALIGNED_ACCESS now causes
path through store_bit_field() which generates a much larger sequence of byte
stores.

        stb 8,0(3)
        stb 10,1(3)
        stb 9,2(3)
        stb 9,3(3)
        stb 9,4(3)
        stb 9,5(3)
        ...