[LLVMbugs] [Bug 6417] New: inefficient code for inserting i8 into i32

[bugzilla-daemon at llvm.org]

http://llvm.org/bugs/show_bug.cgi?id=6417

           Summary: inefficient code for inserting i8 into i32
           Product: new-bugs
           Version: 2.6
          Platform: PC
        OS/Version: Windows NT
            Status: NEW
          Severity: enhancement
          Priority: P5
         Component: new bugs
        AssignedTo: unassignedbugs at nondot.org
        ReportedBy: pete.cooper at gmail.com
                CC: llvmbugs at cs.uiuc.edu


Created an attachment (id=4304)
 --> (http://llvm.org/bugs/attachment.cgi?id=4304)
original IR

I've been looking into the IR generated for writing a byte of a structure.  The
optimizer reduces it to the code you would use for writing to a bitfield, ie,
masked and with an add for the insertion, but wouldn't it be better to use a
byte move or byte store on x86 or a shuffle on other targets?

I've attached the original IR as well as a modified version to use
insertelement as i think that could generate better code.  Unfortunately, the
asm produced is horrendous with my change.  I understand there are code
generator selection and possible LegalizeDag issues giving me bad asm, but is
there any reason why i shouldn't implement the optimisation for inserts and
just add the code generators we need too?

Note, i've look at 2109 and i think this is similar to that, but its for
doubles, not a vector that fits in a normal x86 register.

-- 
Configure bugmail: http://llvm.org/bugs/userprefs.cgi?tab=email
------- You are receiving this mail because: -------
You are on the CC list for the bug.