Remove redundant folding of bswap(bswap(x))

[arsenm]

Summary

Idempotent intrinsics are already generally handled, so no need to special case bswap.

[• rafael]

// bswap(trunc(bswap(x))) -> trunc(lshr(x, c))
if (match(IIOperand, m_Trunc(m_BSwap(m_Value(X))))) {
  unsigned C = X->getType()->getPrimitiveSizeInBits() -

Index: test/Transforms/InstSimplify/call.ll

• test/Transforms/InstSimplify/call.ll

+++ test/Transforms/InstSimplify/call.ll
@@ -102,6 +102,17 @@

ret float %r4

}

+declare i32 @llvm.bswap.i32(i32) nounwind readnone
+
+; CHECK-LABEL: @test_idempotence_2(
+; CHECK: bswap

What does this match? Didn't we remove the bswap?

In general, it is probably better to match the code that is generated
in a bit more detail. In particular, this function is so small that
you should be able to just use a few CHECK-NEXT until the ret.

+; CHECK-NOT: bswap
+define i32 @test_idempotence_2(i32 %a) {
+ %a0 = call i32 @llvm.bswap.i32(i32 %a)
+ %a1 = call i32 @llvm.bswap.i32(i32 %a0)
+ ret i32 %a1
+}
+

[arsenm]

This is just a thinko. I was misreading this as bswap(bswap(x)) -> bswap(x), but it actually removes the bswap which makes more sense

lib/Analysis/InstructionSimplify.cpp

   // Unary idempotent: f(f(x)) = f(x)
   case Intrinsic::fabs:
   case Intrinsic::floor:
   case Intrinsic::ceil:
   case Intrinsic::trunc:
   case Intrinsic::rint:
   case Intrinsic::nearbyint:
   case Intrinsic::round:
+  case Intrinsic::bswap:
     return true;
   }
 }
 
 template <typename IterTy>
 static Value *SimplifyIntrinsic(Intrinsic::ID IID, IterTy ArgBegin, IterTy ArgEnd,
                                 const Query &Q, unsigned MaxRecurse) {
   // Perform idempotent optimizations

lib/Transforms/InstCombine/InstCombineCalls.cpp

case Intrinsic::objectsize: {
     if (getObjectSize(II->getArgOperand(0), Size, DL, TLI))
       return ReplaceInstUsesWith(CI, ConstantInt::get(CI.getType(), Size));
     return nullptr;
   }
   case Intrinsic::bswap: {
     Value *IIOperand = II->getArgOperand(0);
     Value *X = nullptr;
 
-    // bswap(bswap(x)) -> x
-    if (match(IIOperand, m_BSwap(m_Value(X))))
-        return ReplaceInstUsesWith(CI, X);
-
     // bswap(trunc(bswap(x))) -> trunc(lshr(x, c))
     if (match(IIOperand, m_Trunc(m_BSwap(m_Value(X))))) {
       unsigned C = X->getType()->getPrimitiveSizeInBits() -
         IIOperand->getType()->getPrimitiveSizeInBits();
       Value *CV = ConstantInt::get(X->getType(), C);
       Value *V = Builder->CreateLShr(X, CV);
       return new TruncInst(V, IIOperand->getType());
     }

test/Transforms/InstSimplify/call.ll

; CHECK-NOT: nearbyint
   %r1 = fadd float %r0, %c1
   %r2 = fadd float %r1, %d1
   %r3 = fadd float %r2, %e1
   %r4 = fadd float %r3, %f1
 
   ret float %r4
 }
 
+declare i32 @llvm.bswap.i32(i32) nounwind readnone
+
+; CHECK-LABEL: @test_idempotence_2(
+; CHECK: bswap
+; CHECK-NOT: bswap
+define i32 @test_idempotence_2(i32 %a) {
+  %a0 = call i32 @llvm.bswap.i32(i32 %a)
+  %a1 = call i32 @llvm.bswap.i32(i32 %a0)
+  ret i32 %a1
+}
+
 define i8* @operator_new() {
 entry:
   %call = tail call noalias i8* @_Znwm(i64 8)
   %cmp = icmp eq i8* %call, null
   br i1 %cmp, label %cast.end, label %cast.notnull
 
 cast.notnull:                                     ; preds = %entry
   %add.ptr = getelementptr inbounds i8* %call, i64 4