[Bug target/65710] [4.9/5 Regression] Thumb1 ICE caused by no register to spill

[vmakarov at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=65710

Vladimir Makarov <vmakarov at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |vmakarov at gcc dot gnu.org

--- Comment #14 from Vladimir Makarov <vmakarov at gcc dot gnu.org> ---
(In reply to Stupachenko Evgeny from comment #13)
> The issue reproduced with -march=corei7 as well.
> 
> Currently the hottest loop looks as following:
> L1.
> mov    (%esp),%ebp /* potentially redundant.  */
> and    $0x7fff,%ebp                                          
> movzwl 0x400(%esi,%ebp,2),%ecx                               
> mov    %ecx,(%esp) 
> cmp    %ecx,0x4(%esp) /*  Why not reuse ebp?  */        
> jae    380                                                   
> subl   $0x1,0x8(%esp)                                        
> je     380                                                   
> mov    (%esp),%ebp /* ebp could contain correct value (if reused above).  */
> add    %edi,%ebp                                             
> cmp    0x0(%ebp,%ebx,1),%al                                  
> jne    L1
> movzbl 0xf(%esp),%edx                                       
> cmp    -0x1(%ebp,%ebx,1),%dl                               
> jne    L1
> 
> Unnecessary fills and spills of ebp added.
> Before the patch the loop was 2 instruction shorter.
> 
> L1.
> and    $0x7fff,%edx                                             
> movzwl 0x400(%esi,%edx,2),%edx                                  
> cmp    %edx,0x4(%esp)                                           
> jae    3c0                                                      
> sub    $0x1,%ebp                                                
> je     3c0                                                      
> lea    (%edi,%edx,1),%ecx                                       
> movzbl (%esp),%eax                                              
> cmp    (%ecx,%ebx,1),%al                                        
> jne    L1
> movzbl 0xb(%esp),%eax                                           
> cmp    -0x1(%ecx,%ebx,1),%al                                    
> jne    L1.
> 
> Please let me know if you need dumps.

Could you add a pre-processed file with the hot loop.