[PATCH] D32352: Go to eleven

[Simon Pilgrim via Phabricator via llvm-commits]

RKSimon added inline comments.


================
Comment at: test/CodeGen/X86/mul-constant-i16.ll:263
+; X64-NEXT:    addl %edi, %eax
 ; X64-NEXT:    # kill: %AX<def> %AX<kill> %EAX<kill>
 ; X64-NEXT:    retq
----------------
lsaba wrote:
> need to be careful when replacing one mul instruction with 2 lea instructions + other instructions, in the general case this should be faster, but if the RA chooses RBP/R13/EBP as the base register of the lea instruction  (instead of rdi in this case), the lea will be a slow lea so this sequence: 
> leal (%r13,%r13,2), %eax
> leal (%r13,%rax,4),%eax 
> addl %r13d, %eax
> 
> will be slower than the mul instruction. 
> 
> patch https://reviews.llvm.org/D32277 will attempt to replace this sequence with 
> leal (,%r13,2), %eax
> add %r13,%eax
> leal (,%rax,4),%eax 
> add %r13,%eax
> addl %r13d, %eax
> 
> which will cost ~the same as the mul instruction but with increase in code size.
So do you want to wait until D32277 has landed and only replace the 2*lea+other cases if !Subtarget.Slow3OpsLEA() ?


https://reviews.llvm.org/D32352