[PATCH] D152677: [InstCombine] If there is a known-bit transform is_pow2 check to just check for any other bits

[Noah Goldstein via Phabricator via llvm-commits]

goldstein.w.n created this revision.
goldstein.w.n added reviewers: nikic, spatel.
Herald added subscribers: StephenFan, hiraditya.
Herald added a project: All.
goldstein.w.n requested review of this revision.
Herald added a project: LLVM.
Herald added a subscriber: llvm-commits.

in `ctpop(X) eq/ne 1` or `ctpop(X) ugt/ule 1`, if there is any
known-bit in `X`, instead of going through `ctpop`, we can just test
if there are any other known bits in `X`. If there are, `X` is not a
power of 2. If there aren't, `X` is a power of 2.

https://alive2.llvm.org/ce/z/T8osF6


Repository:
  rG LLVM Github Monorepo

https://reviews.llvm.org/D152677

Files:
  llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp
  llvm/test/Transforms/InstCombine/ispow2.ll


Index: llvm/test/Transforms/InstCombine/ispow2.ll
===================================================================
--- llvm/test/Transforms/InstCombine/ispow2.ll
+++ llvm/test/Transforms/InstCombine/ispow2.ll
@@ -1213,9 +1213,8 @@
 declare <2 x i32> @llvm.ctpop.2xi32(<2 x i32>)
 define i1 @is_pow2_nz_known_bits(i32 %xin) {
 ; CHECK-LABEL: @is_pow2_nz_known_bits(
-; CHECK-NEXT:    [[X:%.*]] = or i32 [[XIN:%.*]], 64
-; CHECK-NEXT:    [[CNT:%.*]] = call i32 @llvm.ctpop.i32(i32 [[X]]), !range [[RNG3:![0-9]+]]
-; CHECK-NEXT:    [[R:%.*]] = icmp eq i32 [[CNT]], 1
+; CHECK-NEXT:    [[TMP1:%.*]] = and i32 [[XIN:%.*]], -65
+; CHECK-NEXT:    [[R:%.*]] = icmp eq i32 [[TMP1]], 0
 ; CHECK-NEXT:    ret i1 [[R]]
 ;
   %x = or i32 %xin, 64
@@ -1227,7 +1226,7 @@
 define i1 @is_pow2_nz_known_bits_fail_multiuse(i32 %xin) {
 ; CHECK-LABEL: @is_pow2_nz_known_bits_fail_multiuse(
 ; CHECK-NEXT:    [[X:%.*]] = or i32 [[XIN:%.*]], 64
-; CHECK-NEXT:    [[CNT:%.*]] = call i32 @llvm.ctpop.i32(i32 [[X]]), !range [[RNG3]]
+; CHECK-NEXT:    [[CNT:%.*]] = call i32 @llvm.ctpop.i32(i32 [[X]]), !range [[RNG3:![0-9]+]]
 ; CHECK-NEXT:    call void @use.i32(i32 [[CNT]])
 ; CHECK-NEXT:    [[R:%.*]] = icmp eq i32 [[CNT]], 1
 ; CHECK-NEXT:    ret i1 [[R]]
@@ -1241,9 +1240,7 @@
 
 define i1 @not_pow2_nz_known_bits(i32 %xin) {
 ; CHECK-LABEL: @not_pow2_nz_known_bits(
-; CHECK-NEXT:    [[X:%.*]] = or i32 [[XIN:%.*]], 1
-; CHECK-NEXT:    [[CNT:%.*]] = call i32 @llvm.ctpop.i32(i32 [[X]]), !range [[RNG3]]
-; CHECK-NEXT:    [[R:%.*]] = icmp ne i32 [[CNT]], 1
+; CHECK-NEXT:    [[R:%.*]] = icmp ugt i32 [[XIN:%.*]], 1
 ; CHECK-NEXT:    ret i1 [[R]]
 ;
   %x = or i32 %xin, 1
@@ -1267,9 +1264,8 @@
 
 define i1 @is_pow2_or_z_known_bits(i32 %xin) {
 ; CHECK-LABEL: @is_pow2_or_z_known_bits(
-; CHECK-NEXT:    [[X:%.*]] = or i32 [[XIN:%.*]], -2147483648
-; CHECK-NEXT:    [[CNT:%.*]] = call i32 @llvm.ctpop.i32(i32 [[X]]), !range [[RNG3]]
-; CHECK-NEXT:    [[R:%.*]] = icmp ult i32 [[CNT]], 2
+; CHECK-NEXT:    [[TMP1:%.*]] = and i32 [[XIN:%.*]], 2147483647
+; CHECK-NEXT:    [[R:%.*]] = icmp eq i32 [[TMP1]], 0
 ; CHECK-NEXT:    ret i1 [[R]]
 ;
   %x = or i32 %xin, 2147483648
@@ -1280,9 +1276,8 @@
 
 define <2 x i1> @not_pow2_or_z_known_bits(<2 x i32> %xin) {
 ; CHECK-LABEL: @not_pow2_or_z_known_bits(
-; CHECK-NEXT:    [[X:%.*]] = or <2 x i32> [[XIN:%.*]], <i32 64, i32 64>
-; CHECK-NEXT:    [[CNT:%.*]] = call <2 x i32> @llvm.ctpop.v2i32(<2 x i32> [[X]]), !range [[RNG3]]
-; CHECK-NEXT:    [[R:%.*]] = icmp ugt <2 x i32> [[CNT]], <i32 1, i32 1>
+; CHECK-NEXT:    [[TMP1:%.*]] = and <2 x i32> [[XIN:%.*]], <i32 -65, i32 -65>
+; CHECK-NEXT:    [[R:%.*]] = icmp ne <2 x i32> [[TMP1]], zeroinitializer
 ; CHECK-NEXT:    ret <2 x i1> [[R]]
 ;
   %x = or <2 x i32> %xin, <i32 64, i32 64>
Index: llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp
===================================================================
--- llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp
+++ llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp
@@ -6769,6 +6769,35 @@
   if (Instruction *Res = foldICmpPow2Test(I, Builder))
     return Res;
 
+  {
+    // Power of 2 test:
+    //    isPow2OrZero : ctpop(X) u< 2
+    //    isPow2       : ctpop(X) == 1
+    //    NotPow2OrZero: ctpop(X) u> 1
+    //    NotPow2      : ctpop(X) != 1
+    // If we know any bit of X can be folded to:
+    //    IsPow2       : X & (~Bit) == 0
+    //    NotPow2      : X & (~Bit) != 0
+    Value *A;
+    const APInt *C;
+    if (match(Op0, m_OneUse(m_Intrinsic<Intrinsic::ctpop>(m_Value(A)))) &&
+        match(Op1, m_APInt(C)) &&
+        (((I.isEquality() || Pred == ICmpInst::ICMP_UGT) && *C == 1) ||
+         (Pred == ICmpInst::ICMP_ULT && *C == 2))) {
+      KnownBits AKnown = computeKnownBits(A, /*Depth*/ 0, &I);
+      // No need to check for count > 1, that should be already constant folded.
+      if (AKnown.countMinPopulation() == 1) {
+        Value *And = Builder.CreateAnd(
+            A, Constant::getIntegerValue(A->getType(), ~(AKnown.One)));
+        return new ICmpInst(
+            (Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_ULT)
+                ? ICmpInst::ICMP_EQ
+                : ICmpInst::ICMP_NE,
+            And, Constant::getNullValue(A->getType()));
+      }
+    }
+  }
+
   if (Instruction *Res = foldICmpOfUAddOv(I))
     return Res;
 


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D152677.530400.patch
Type: text/x-patch
Size: 4341 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20230612/9ad8ed04/attachment.bin>