[PATCH] D48223: DAG combine "and|or (select c, -1, 0), x" -> "select c, x, 0|-1"

[Stanislav Mekhanoshin via Phabricator via llvm-commits]

rampitec marked an inline comment as done.
rampitec added inline comments.


================
Comment at: test/CodeGen/X86/dagcombine-select.ll:1
+; RUN: llc -march=x86-64 -verify-machineinstrs < %s | FileCheck -enable-var-scope %s
+
----------------
lebedev.ri wrote:
> lebedev.ri wrote:
> > Most tests (and practically all new x86 tests) use `utils/update_llc_test_checks.py` script to auto-generate these check lines.
> (not actually done, that does not look like the utility's output, and the first line does not say the script was used)
Oops, sorry. Clicked wrong checkbox. Now updated.


================
Comment at: test/CodeGen/X86/dagcombine-select.ll:3-17
+; CHECK-LABEL: {{^}}select_and1:
+; CHECK:      cmpl    $11, %edi
+; CHECK-NEXT: cmovgel %esi, %eax
+; CHECK-NEXT: retq
+define i32 @select_and1(i32 %x, i32 %y) {
+  %c = icmp slt i32 %x, 11
+  %s = select i1 %c, i32 0, i32 -1
----------------
lebedev.ri wrote:
> rampitec wrote:
> > rampitec wrote:
> > > lebedev.ri wrote:
> > > > Hm, is there some omitted instruction, or is this actually better than what we currently normally do?
> > > > https://godbolt.org/g/7ULPfH
> > > Yes. I have updated the test to contain the full ISA. First xor to zero out eax was omitted.
> > > I am not sure what compiler explorer does, but that is what trunk llc has produced:
> > > 
> > > 
> > > ```
> > >         xorl    %eax, %eax
> > >         cmpl    $11, %edi
> > >         setl    %al
> > >         decl    %eax
> > >         andl    %esi, %eax
> > >         retq
> > > 
> > > ```
> > > 
> > > I assume difference comes from running or not running opt.
> > E.g. compare w/o opt: https://godbolt.org/g/NMZ9he
> Ok, so the only difference is that the strictness of the comparison is inverted (`cmovg` vs `cmovge` and the other way around).
Do you see "and eax, esi" instruction on the second from the right tab using my link? This one was eliminated.

In general if you run opt that is not an issue. If you run llc only (as in the test) you can see it. Normally that shall not happen, as I wrote in the description InstCombine can get rid of it. However, amdgcn can produce it during lowering and InstCombine does not play.


https://reviews.llvm.org/D48223