Lots of spills generated

[llvmbot]

Bugzilla Link	10089
Resolution	FIXED
Resolved on	Jun 30, 2011 20:15
Version	trunk
OS	All
Attachments	jsinterp.ii
Reporter	LLVM Bugzilla Contributor
CC	@stoklund

Extended Description

The included testcase compiles to 23572 bytes with clang and 4940 bytes with gcc 4.2. The clang produced code has lots of:

    movq    %rax, %rcx
                                    ## implicit-def: EAX
    movl    %eax, -92(%rbp)         ## 4-byte Spill
                                    ## implicit-def: EAX
    movl    %eax, -68(%rbp)         ## 4-byte Spill
                                    ## implicit-def: EAX
    movq    %rax, -88(%rbp)         ## 8-byte Spill
                                    ## implicit-def: EAX
    movl    %eax, -496(%rbp)        ## 4-byte Spill
                                    ## implicit-def: EAX
    movl    %eax, -492(%rbp)        ## 4-byte Spill
    movq    %rcx, -488(%rbp)        ## 8-byte Spill
                                    ## implicit-def: EAX

[stoklund]

This file uses indirectbr, and it has a handful of phis with many predecessors in indirectgoto.backedge.

It looks like phi lowering or taildup goes crazy, and the coalescer is unable to clean up, creating lots of implicit_defs.

Linear scan and greedy produce similar results.

[llvmbot]

Fixed in 134216.