[SelectOpt] Don't treat LogicalAnd/LogicalOr as selects

[dmgreen]

Summary

A select i1 %c, i1 true, i1 %d is just an or and a select i1 %c, i1 %d, i1 false is just an and. There are better treated as such in the logic of SelectOpt, allowing the backend to optimize them to and/or directly.

[davidxl]

This looks good. Sotiris may help with a quick benchmarking for the patch.

[apostolakis]

This looks fine to me as well and it is okay performance-wise (within noise margins for some testing I did).
My only suggestion is to separate this check in a separate function. See inline comment.

llvm/lib/CodeGen/SelectOptimize.cpp

538

I would suggest moving the added check to a new static function maybe called isSpecialSelect and continue if it returns true. It will make the code cleaner and maybe we will add more special select cases that should not be changed to branches.

[dmgreen]

Thanks - that sounds good. I've created a function like you suggest.

llvm/lib/CodeGen/SelectOptimize.cpp

for (auto It = ASI.rbegin(); It != ASI.rend(); ++It) {
       SI->replaceAllUsesWith(PN);
       SI->eraseFromParent();
       INS.erase(SI);
       ++NumSelectsConverted;
     }
   }
 }
 
+static bool isSpecialSelect(SelectInst *SI) {
+  using namespace llvm::PatternMatch;
+
+  // If the select is a logical-and/logical-or then it is better treated as a
+  // and/or by the backend.
+  if (match(SI, m_CombineOr(m_LogicalAnd(m_Value(), m_Value()),
+                            m_LogicalOr(m_Value(), m_Value()))))
+    return true;
+
+  return false;
+}
+
 void SelectOptimize::collectSelectGroups(BasicBlock &BB,
                                          SelectGroups &SIGroups) {
   BasicBlock::iterator BBIt = BB.begin();
   while (BBIt != BB.end()) {
     Instruction *I = &*BBIt++;
     if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
+      if (isSpecialSelect(SI))
+        continue;
+
       SelectGroup SIGroup;
       SIGroup.push_back(SI);
       while (BBIt != BB.end()) {
         Instruction *NI = &*BBIt;
         SelectInst *NSI = dyn_cast<SelectInst>(NI);
         if (NSI && SI->getCondition() == NSI->getCondition()) {
           SIGroup.push_back(NSI);
         } else if (!NI->isDebugOrPseudoInst()) {

llvm/test/CodeGen/AArch64/selectopt-logical.ll

+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt -select-optimize -mtriple=aarch64-linux-gnu -mcpu=neoverse-v2 -S < %s | FileCheck %s
+
+define i32 @test(ptr nocapture noundef readnone %x, i32 noundef %iters) {
+; CHECK-LABEL: @test(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[CMP3:%.*]] = icmp sgt i32 [[ITERS:%.*]], 0
+; CHECK-NEXT:    br i1 [[CMP3]], label [[FOR_BODY_I_PREHEADER:%.*]], label [[FOR_COND_CLEANUP:%.*]]
+; CHECK:       for.body.i.preheader:
+; CHECK-NEXT:    [[I_05:%.*]] = phi i32 [ [[INC:%.*]], [[INNER_LOOP_010_EXIT:%.*]] ], [ 0, [[ENTRY:%.*]] ]
+; CHECK-NEXT:    [[RES_04:%.*]] = phi i32 [ [[OR:%.*]], [[INNER_LOOP_010_EXIT]] ], [ 0, [[ENTRY]] ]
+; CHECK-NEXT:    br label [[FOR_BODY_I:%.*]]
+; CHECK:       for.cond.cleanup:
+; CHECK-NEXT:    [[RES_0_LCSSA:%.*]] = phi i32 [ 0, [[ENTRY]] ], [ [[OR]], [[INNER_LOOP_010_EXIT]] ]
+; CHECK-NEXT:    ret i32 [[RES_0_LCSSA]]
+; CHECK:       for.body.i:
+; CHECK-NEXT:    [[INDVARS_IV_I:%.*]] = phi i64 [ 0, [[FOR_BODY_I_PREHEADER]] ], [ [[INDVARS_IV_NEXT_I_4:%.*]], [[FOR_BODY_I]] ]
+; CHECK-NEXT:    [[ALL_0_OFF010_I:%.*]] = phi i1 [ true, [[FOR_BODY_I_PREHEADER]] ], [ [[ALL_0_OFF0__I_4:%.*]], [[FOR_BODY_I]] ]
+; CHECK-NEXT:    [[ANY_0_OFF09_I:%.*]] = phi i1 [ false, [[FOR_BODY_I_PREHEADER]] ], [ [[DOTANY_0_OFF0_I_4:%.*]], [[FOR_BODY_I]] ]
+; CHECK-NEXT:    [[ARRAYIDX_I:%.*]] = getelementptr inbounds float, ptr [[X:%.*]], i64 [[INDVARS_IV_I]]
+; CHECK-NEXT:    [[TMP0:%.*]] = load float, ptr [[ARRAYIDX_I]], align 4
+; CHECK-NEXT:    [[CMP1_I:%.*]] = fcmp fast olt float [[TMP0]], 0.000000e+00
+; CHECK-NEXT:    [[INDVARS_IV_NEXT_I:%.*]] = add nuw nsw i64 [[INDVARS_IV_I]], 1
+; CHECK-NEXT:    [[ARRAYIDX_I_1:%.*]] = getelementptr inbounds float, ptr [[X]], i64 [[INDVARS_IV_NEXT_I]]
+; CHECK-NEXT:    [[TMP1:%.*]] = load float, ptr [[ARRAYIDX_I_1]], align 4
+; CHECK-NEXT:    [[CMP1_I_1:%.*]] = fcmp fast olt float [[TMP1]], 0.000000e+00
+; CHECK-NEXT:    [[INDVARS_IV_NEXT_I_1:%.*]] = add nuw nsw i64 [[INDVARS_IV_I]], 2
+; CHECK-NEXT:    [[ARRAYIDX_I_2:%.*]] = getelementptr inbounds float, ptr [[X]], i64 [[INDVARS_IV_NEXT_I_1]]
+; CHECK-NEXT:    [[TMP2:%.*]] = load float, ptr [[ARRAYIDX_I_2]], align 4
+; CHECK-NEXT:    [[CMP1_I_2:%.*]] = fcmp fast olt float [[TMP2]], 0.000000e+00
+; CHECK-NEXT:    [[INDVARS_IV_NEXT_I_2:%.*]] = add nuw nsw i64 [[INDVARS_IV_I]], 3
+; CHECK-NEXT:    [[ARRAYIDX_I_3:%.*]] = getelementptr inbounds float, ptr [[X]], i64 [[INDVARS_IV_NEXT_I_2]]
+; CHECK-NEXT:    [[TMP3:%.*]] = load float, ptr [[ARRAYIDX_I_3]], align 4
+; CHECK-NEXT:    [[CMP1_I_3:%.*]] = fcmp fast olt float [[TMP3]], 0.000000e+00
+; CHECK-NEXT:    [[INDVARS_IV_NEXT_I_3:%.*]] = add nuw nsw i64 [[INDVARS_IV_I]], 4
+; CHECK-NEXT:    [[ARRAYIDX_I_4:%.*]] = getelementptr inbounds float, ptr [[X]], i64 [[INDVARS_IV_NEXT_I_3]]
+; CHECK-NEXT:    [[TMP4:%.*]] = load float, ptr [[ARRAYIDX_I_4]], align 4
+; CHECK-NEXT:    [[CMP1_I_4:%.*]] = fcmp fast olt float [[TMP4]], 0.000000e+00
+; CHECK-NEXT:    [[TMP5:%.*]] = select i1 [[CMP1_I_4]], i1 true, i1 [[CMP1_I_3]]
+; CHECK-NEXT:    [[TMP6:%.*]] = select i1 [[TMP5]], i1 true, i1 [[CMP1_I_2]]
+; CHECK-NEXT:    [[TMP7:%.*]] = select i1 [[TMP6]], i1 true, i1 [[CMP1_I_1]]
+; CHECK-NEXT:    [[TMP8:%.*]] = select i1 [[TMP7]], i1 true, i1 [[CMP1_I]]
+; CHECK-NEXT:    [[DOTANY_0_OFF0_I_4]] = select i1 [[TMP8]], i1 true, i1 [[ANY_0_OFF09_I]]
+; CHECK-NEXT:    [[TMP9:%.*]] = select i1 [[CMP1_I_4]], i1 [[CMP1_I_3]], i1 false
+; CHECK-NEXT:    [[TMP10:%.*]] = select i1 [[TMP9]], i1 [[CMP1_I_2]], i1 false
+; CHECK-NEXT:    [[TMP11:%.*]] = select i1 [[TMP10]], i1 [[CMP1_I_1]], i1 false
+; CHECK-NEXT:    [[TMP12:%.*]] = select i1 [[TMP11]], i1 [[CMP1_I]], i1 false
+; CHECK-NEXT:    [[ALL_0_OFF0__I_4]] = select i1 [[TMP12]], i1 [[ALL_0_OFF010_I]], i1 false
+; CHECK-NEXT:    [[INDVARS_IV_NEXT_I_4]] = add nuw nsw i64 [[INDVARS_IV_I]], 5
+; CHECK-NEXT:    [[EXITCOND_NOT_I_4:%.*]] = icmp eq i64 [[INDVARS_IV_NEXT_I_4]], 10000
+; CHECK-NEXT:    br i1 [[EXITCOND_NOT_I_4]], label [[INNER_LOOP_010_EXIT]], label [[FOR_BODY_I]]
+; CHECK:       inner_loop_010.exit:
+; CHECK-NEXT:    [[TMP13:%.*]] = select i1 [[DOTANY_0_OFF0_I_4]], i32 2, i32 3
+; CHECK-NEXT:    [[TMP14:%.*]] = select i1 [[ALL_0_OFF0__I_4]], i32 1, i32 [[TMP13]]
+; CHECK-NEXT:    [[OR]] = or i32 [[TMP14]], [[RES_04]]
+; CHECK-NEXT:    [[INC]] = add nuw nsw i32 [[I_05]], 1
+; CHECK-NEXT:    [[EXITCOND_NOT:%.*]] = icmp eq i32 [[INC]], [[ITERS]]
+; CHECK-NEXT:    br i1 [[EXITCOND_NOT]], label [[FOR_COND_CLEANUP]], label [[FOR_BODY_I_PREHEADER]]
+;
+entry:
+  %cmp3 = icmp sgt i32 %iters, 0
+  br i1 %cmp3, label %for.body.i.preheader, label %for.cond.cleanup
+
+for.body.i.preheader:                             ; preds = %entry, %inner_loop_010.exit
+  %i.05 = phi i32 [ %inc, %inner_loop_010.exit ], [ 0, %entry ]
+  %res.04 = phi i32 [ %or, %inner_loop_010.exit ], [ 0, %entry ]
+  br label %for.body.i
+
+for.cond.cleanup:                                 ; preds = %inner_loop_010.exit, %entry
+  %res.0.lcssa = phi i32 [ 0, %entry ], [ %or, %inner_loop_010.exit ]
+  ret i32 %res.0.lcssa
+
+for.body.i:                                       ; preds = %for.body.i, %for.body.i.preheader
+  %indvars.iv.i = phi i64 [ 0, %for.body.i.preheader ], [ %indvars.iv.next.i.4, %for.body.i ]
+  %all.0.off010.i = phi i1 [ true, %for.body.i.preheader ], [ %all.0.off0..i.4, %for.body.i ]
+  %any.0.off09.i = phi i1 [ false, %for.body.i.preheader ], [ %.any.0.off0.i.4, %for.body.i ]
+  %arrayidx.i = getelementptr inbounds float, ptr %x, i64 %indvars.iv.i
+  %0 = load float, ptr %arrayidx.i, align 4
+  %cmp1.i = fcmp fast olt float %0, 0.000000e+00
+  %indvars.iv.next.i = add nuw nsw i64 %indvars.iv.i, 1
+  %arrayidx.i.1 = getelementptr inbounds float, ptr %x, i64 %indvars.iv.next.i
+  %1 = load float, ptr %arrayidx.i.1, align 4
+  %cmp1.i.1 = fcmp fast olt float %1, 0.000000e+00
+  %indvars.iv.next.i.1 = add nuw nsw i64 %indvars.iv.i, 2
+  %arrayidx.i.2 = getelementptr inbounds float, ptr %x, i64 %indvars.iv.next.i.1
+  %2 = load float, ptr %arrayidx.i.2, align 4
+  %cmp1.i.2 = fcmp fast olt float %2, 0.000000e+00
+  %indvars.iv.next.i.2 = add nuw nsw i64 %indvars.iv.i, 3
+  %arrayidx.i.3 = getelementptr inbounds float, ptr %x, i64 %indvars.iv.next.i.2
+  %3 = load float, ptr %arrayidx.i.3, align 4
+  %cmp1.i.3 = fcmp fast olt float %3, 0.000000e+00
+  %indvars.iv.next.i.3 = add nuw nsw i64 %indvars.iv.i, 4
+  %arrayidx.i.4 = getelementptr inbounds float, ptr %x, i64 %indvars.iv.next.i.3
+  %4 = load float, ptr %arrayidx.i.4, align 4
+  %cmp1.i.4 = fcmp fast olt float %4, 0.000000e+00
+  %5 = select i1 %cmp1.i.4, i1 true, i1 %cmp1.i.3
+  %6 = select i1 %5, i1 true, i1 %cmp1.i.2
+  %7 = select i1 %6, i1 true, i1 %cmp1.i.1
+  %8 = select i1 %7, i1 true, i1 %cmp1.i
+  %.any.0.off0.i.4 = select i1 %8, i1 true, i1 %any.0.off09.i
+  %9 = select i1 %cmp1.i.4, i1 %cmp1.i.3, i1 false
+  %10 = select i1 %9, i1 %cmp1.i.2, i1 false
+  %11 = select i1 %10, i1 %cmp1.i.1, i1 false
+  %12 = select i1 %11, i1 %cmp1.i, i1 false
+  %all.0.off0..i.4 = select i1 %12, i1 %all.0.off010.i, i1 false
+  %indvars.iv.next.i.4 = add nuw nsw i64 %indvars.iv.i, 5
+  %exitcond.not.i.4 = icmp eq i64 %indvars.iv.next.i.4, 10000
+  br i1 %exitcond.not.i.4, label %inner_loop_010.exit, label %for.body.i
+
+inner_loop_010.exit:                              ; preds = %for.body.i
+  %13 = select i1 %.any.0.off0.i.4, i32 2, i32 3
+  %14 = select i1 %all.0.off0..i.4, i32 1, i32 %13
+  %or = or i32 %14, %res.04
+  %inc = add nuw nsw i32 %i.05, 1
+  %exitcond.not = icmp eq i32 %inc, %iters
+  br i1 %exitcond.not, label %for.cond.cleanup, label %for.body.i.preheader
+}