[DAG] select Cond, -1, C --> or (sext Cond), C if Cond is MVT::i1

[deadalnix]

Summary

This seems to be beneficial overall, except for midpoint-int.ll .

The X86 backend seems to generate zeroing that are not necesary.

[deadalnix]

llvm/test/CodeGen/X86/memcmp-more-load-pairs-x32.ll

163–166

This seems to be unnecessary.

[pengfei]

llvm/test/CodeGen/X86/memcmp-more-load-pairs-x32.ll

163–166

Looks like dependency breaking for sbb:
https://github.com/llvm/llvm-project/blob/main/llvm/lib/Target/X86/X86ISelDAGToDAG.cpp#L5874-L5876

[deadalnix]

llvm/test/CodeGen/X86/memcmp-more-load-pairs-x32.ll

163–166

Good catch, thanks.

[deadalnix]

Reabse and variosu tweaks.

[deadalnix]

llvm/test/CodeGen/X86/select_const.ll

290

TODO: (sext Cond) | (Pow2 - 1) -> Pow2 - (zext (not Cond)) ?

Maybe if inverting Cond if for free? Is there a way to check for this?

Or maybe this is best reserved for X86DAGToDAGISel::matchAddressRecursively?

[deadalnix]

llvm/test/CodeGen/X86/select_const.ll

290

I meant (sext Cond) | (Pow2 - 1) -> (zext (not Cond)) * Pow2 - 1 ?

[deadalnix]

llvm/test/CodeGen/X86/select_const.ll

290

Doing it in selectLEAAddr doesn't work as the score doesn't get high enough.

[deadalnix]

Rebase on top of D131358

[deadalnix]

llvm/test/CodeGen/X86/memcmp.ll

371–372

In both cases here, cmpl %edx, %ecx seems to be recomputed for no reason (?) It seems to be able to reuse cmpl %ecx, %eax in length5 and cmpw %dx, %cx in length3, at least before this diff.

Is there a reason why this isn't the case here? I'm not quite sure how memcmp is being lowered, and -debug really isn't useful here. @gchatelet , you were looking for ways to improve memcmp, I think there is a low hanging fruit somewhere in there.

[pengfei]

llvm/test/CodeGen/X86/memcmp.ll

371–372

xor will change EFLAGS, so we need to recomputed to get the correct EFLAGS.

[deadalnix]

llvm/test/CodeGen/X86/memcmp.ll

371–372

Sure, but the xor itself doesn't seem necessary. For instance, length5 does:

setae %al
movzbl %al, %eax

Which does away witth he need for the xor, with in turn does away with the need to recompute the cmpl. Both function do the same thing in that block, there is no reason to have different codegen.

[RKSimon]

please can you rebase this against trunk latest?

[deadalnix]

rebase

[RKSimon]

llvm/test/CodeGen/X86/select_const.ll

204

Comment doesn't match select_lea_2 (but the others are still OK).