[LLVMdev] Making Sense of ISel DAG Output

[Dan Gohman]

On Oct 2, 2008, at 2:19 PM, David Greene wrote:

> On Thursday 02 October 2008 12:42, David Greene wrote:
>
>> But let's say you _could_ write such a pattern (because I can).   
>> The input
>> DAG looks like this:
>>
>>            0x391a220: <multiple use>
>>          0x391c970: v2f64 = scalar_to_vector 0x391a220 srcLineNum= 10
>>            0x391ac10: <multiple use>
>>          0x391c8b0: v2f64 = scalar_to_vector 0x391ac10 srcLineNum= 10
>>          0x3927b10: <multiple use>
>>        0x3923100: v2f64 = vector_shuffle 0x391c970, 0x391c8b0,
>> 0x3927b10<0,2> srcLineNum= 10
>>
>> The code that gets produced looks like this:
>>
>> 	%reg1071<def> = MOVSD2PDrm %reg1026, 8, %reg1065, 4294967288,  
>> Mem:LD(8,8)
>> [r66428 + 0]LD(8,8) [r78427 + 0]	 ; srcLine 10
>> 	%reg1072<def> = MOVSD2PDrm %reg1026, 8, %reg1065, 4294967288,  
>> Mem:LD(8,8)
>> [r66428 + 0]LD(8,8) [r78427 + 0]	 ; srcLine 10
>> 	%reg1073<def> = SHUFPDrri %reg1071, %reg1072, 0	 ; srcLine 10
>
> Actrually, it's worse than this.  I wanted to check to make sure  
> something
> else wasn't causing the problem but it appears to come from isel.   
> The full
> output for the DAG looks like this:
>
> 	%reg1059<def> = MOVSX64rm32 %reg1033, 1, %reg0, 4, Mem:LD(4,4)  
> [tmp163 +
> 0]	 ; srcLine 10
> 	%reg1060<def> = MOVSDrm %reg1026, 8, %reg1059, 4294967288,  
> Mem:LD(8,8)
> [r45154 + 0]	 ; srcLine 10
> 	%reg1061<def> = MOVSX64rm32 %reg1033, 1, %reg0, 0, Mem:LD(4,4) [iv. 
> 161162 +
> 0]	 ; srcLine 10
> 	%reg1062<def> = MOVSDrm %reg1026, 8, %reg1061, 4294967288,  
> Mem:LD(8,8)
> [r30158 + 0]	 ; srcLine 10
> 	%reg1063<def> = MOVSD2PDrm %reg1026, 8, %reg1059, 4294967288,  
> Mem:LD(8,8)
> [r30158 + 0]LD(8,8) [r45154 + 0]	 ; srcLine 10
> 	%reg1064<def> = MOVSD2PDrm %reg1026, 8, %reg1059, 4294967288,  
> Mem:LD(8,8)
> [r30158 + 0]LD(8,8) [r45154 + 0]	 ; srcLine 10
> 	%reg1065<def> = SHUFPDrri %reg1063, %reg1064, 0	 ; srcLine 10
>
> Where the <bleep> are these extra dead MOVSDrms coming from?  Note  
> that the
> extra MOVSDrms at least seem to use the correct addresses.

Looking at your dump() output above, it looks like the pre-selection
loads have multiple uses, so even though you've managed to match a
larger pattern that incorporates them, they still need to exist to
satisfy some other users.

Dan