[LLVMbugs] [Bug 702] NEW: [codegen] Compile copysign more optimally

[bugzilla-daemon at cs.uiuc.edu]

http://llvm.cs.uiuc.edu/bugs/show_bug.cgi?id=702

           Summary: [codegen] Compile copysign more optimally
           Product: libraries
           Version: 1.0
          Platform: All
        OS/Version: All
            Status: NEW
          Severity: enhancement
          Priority: P2
         Component: Common Code Generator Code
        AssignedTo: unassignedbugs at nondot.org
        ReportedBy: sabre at nondot.org
                CC: natebegeman at mac.com


It would be nice to add:

1. an llvm intrinsic (llvm.copysign.f32/f64), that represents copysign[f] when negative zeros
    and SNANs are not important (basically -ffast-math), the CFE could generate this.
2. a selectiondag node to represent the same condition, also when UnsafeFPMath is enabled.

Then the legalizer can turn copysign into either 1) stores to mem followed by bitfield stuff and a load, 
or 2) fabs/fneg/setcc/select if the target has them.

With this, the PPC backend would codegen copysign as:

copysign(x, y) {
   fabs   xPos, x
   fnabs   xNeg, x
   fsel   x, y, xPos, xNeg
   return x;
}

which is nice and fast.  X86/SSE has a similar lowering, but it would be custom lowered.

This is important for codes that use the FORTRAN copysign intrinsic or the C99 copysign functions 
heavily.

-Chris



------- You are receiving this mail because: -------
You are on the CC list for the bug, or are watching someone who is.