[InstCombine] Eliminate icmp+zext pairs over phis more aggressively (#121767)

When folding icmp over phi, add a special case for `icmp eq (zext(bool),
0)`, which is known to fold to `!bool` and thus won't increase the
instruction count. This helps convert more phis to i1, esp. in loops.

This is based on existing logic we have to support this for icmp of
ucmp/scmp.

Author: nikic

llvm/lib/Transforms/InstCombine/InstructionCombining.cpp

@@ -1822,12 +1822,29 @@ Instruction *InstCombinerImpl::foldOpIntoPhi(Instruction &I, PHINode *PN,
       continue;
     }
 
-    // If the only use of phi is comparing it with a constant then we can
-    // put this comparison in the incoming BB directly after a ucmp/scmp call
-    // because we know that it will simplify to a single icmp.
-    const APInt *Ignored;
-    if (isa<CmpIntrinsic>(InVal) && InVal->hasOneUser() &&
-        match(&I, m_ICmp(m_Specific(PN), m_APInt(Ignored)))) {
+    // Handle some cases that can't be fully simplified, but where we know that
+    // the two instructions will fold into one.
+    auto WillFold = [&]() {
+      if (!InVal->hasOneUser())
+        return false;
+
+      // icmp of ucmp/scmp with constant will fold to icmp.
+      const APInt *Ignored;
+      if (isa<CmpIntrinsic>(InVal) &&
+          match(&I, m_ICmp(m_Specific(PN), m_APInt(Ignored))))
+        return true;
+
+      // icmp eq zext(bool), 0 will fold to !bool.
+      if (isa<ZExtInst>(InVal) &&
+          cast<ZExtInst>(InVal)->getSrcTy()->isIntOrIntVectorTy(1) &&
+          match(&I,
+                m_SpecificICmp(ICmpInst::ICMP_EQ, m_Specific(PN), m_Zero())))
+        return true;
+
+      return false;
+    };
+
+    if (WillFold()) {
       OpsToMoveUseToIncomingBB.push_back(i);
       NewPhiValues.push_back(nullptr);
       continue;

llvm/test/Transforms/InstCombine/phi.ll

@@ -2828,13 +2828,13 @@ define i1 @test_zext_icmp_eq_0(i1 %a, i1 %b, i32 %c) {
 ; CHECK-NEXT:  entry:
 ; CHECK-NEXT:    br i1 [[A:%.*]], label [[IF:%.*]], label [[ELSE:%.*]]
 ; CHECK:       if:
-; CHECK-NEXT:    [[B_EXT:%.*]] = zext i1 [[B:%.*]] to i32
+; CHECK-NEXT:    [[TMP0:%.*]] = xor i1 [[B:%.*]], true
 ; CHECK-NEXT:    br label [[JOIN:%.*]]
 ; CHECK:       else:
+; CHECK-NEXT:    [[TMP1:%.*]] = icmp eq i32 [[C:%.*]], 0
 ; CHECK-NEXT:    br label [[JOIN]]
 ; CHECK:       join:
-; CHECK-NEXT:    [[PHI:%.*]] = phi i32 [ [[B_EXT]], [[IF]] ], [ [[C:%.*]], [[ELSE]] ]
-; CHECK-NEXT:    [[CMP:%.*]] = icmp eq i32 [[PHI]], 0
+; CHECK-NEXT:    [[CMP:%.*]] = phi i1 [ [[TMP0]], [[IF]] ], [ [[TMP1]], [[ELSE]] ]
 ; CHECK-NEXT:    ret i1 [[CMP]]
 ;
 entry:
@@ -2916,10 +2916,9 @@ define i1 @test_zext_icmp_eq_0_loop(i1 %c, i1 %b) {
 ; CHECK-NEXT:  entry:
 ; CHECK-NEXT:    br label [[LOOP:%.*]]
 ; CHECK:       loop:
-; CHECK-NEXT:    [[PHI:%.*]] = phi i32 [ 1, [[ENTRY:%.*]] ], [ [[EXT:%.*]], [[LOOP]] ]
-; CHECK-NEXT:    [[X:%.*]] = icmp eq i32 [[PHI]], 0
+; CHECK-NEXT:    [[X:%.*]] = phi i1 [ false, [[ENTRY:%.*]] ], [ [[TMP0:%.*]], [[LOOP]] ]
 ; CHECK-NEXT:    [[Y:%.*]] = and i1 [[X]], [[B:%.*]]
-; CHECK-NEXT:    [[EXT]] = zext i1 [[Y]] to i32
+; CHECK-NEXT:    [[TMP0]] = xor i1 [[Y]], true
 ; CHECK-NEXT:    br i1 [[C:%.*]], label [[LOOP]], label [[EXIT:%.*]]
 ; CHECK:       exit:
 ; CHECK-NEXT:    ret i1 [[X]]