[TypePromotion] Add truncate in ConvertTruncs when the original truncate type is not extend type

[bcl5980]

Summary

If the src type is not extend type, after convert the truncate to and we need to truncate the and also to make sure the all user is legal.

The old fix D137613 doesn't work when the truncate convert to and have the other users. So this time I try to add the truncate after and to avoid all these potential issues.

Fix: #59554

[samparker]

Thanks.

llvm/lib/CodeGen/TypePromotion.cpp

for (auto *I : Sinks) {
     }
   }
 }
 
 void IRPromoter::Cleanup() {
   LLVM_DEBUG(dbgs() << "IR Promotion: Cleanup..\n");
   // Some zexts will now have become redundant, along with their trunc
   // operands, so remove them.
-  // Some zexts need to be replaced with truncate if src bitwidth is larger.
   for (auto *V : Visited) {
     if (!isa<ZExtInst>(V))
       continue;
 
     auto ZExt = cast<ZExtInst>(V);
     if (ZExt->getDestTy() != ExtTy)
       continue;
 
     Value *Src = ZExt->getOperand(0);
     if (ZExt->getSrcTy() == ZExt->getDestTy()) {
       LLVM_DEBUG(dbgs() << "IR Promotion: Removing unnecessary cast: " << *ZExt
                         << "\n");
       ReplaceAllUsersOfWith(ZExt, Src);
       continue;
-    } else if (ZExt->getSrcTy()->getScalarSizeInBits() > PromotedWidth) {
-      IRBuilder<> Builder{ZExt};
-      Value *Trunc = Builder.CreateTrunc(Src, ZExt->getDestTy());
-      ReplaceAllUsersOfWith(ZExt, Trunc);
-      continue;
     }
 
     // We've inserted a trunc for a zext sink, but we already know that the
     // input is in range, negating the need for the trunc.
     if (NewInsts.count(Src) && isa<TruncInst>(Src)) {
       auto *Trunc = cast<TruncInst>(Src);
       assert(Trunc->getOperand(0)->getType() == ExtTy &&
              "expected inserted trunc to be operating on i32");
for (auto *V : Visited) {
     Builder.SetInsertPoint(Trunc);
     IntegerType *SrcTy = cast<IntegerType>(Trunc->getOperand(0)->getType());
     IntegerType *DestTy = cast<IntegerType>(TruncTysMap[Trunc][0]);
 
     unsigned NumBits = DestTy->getScalarSizeInBits();
     ConstantInt *Mask =
         ConstantInt::get(SrcTy, APInt::getMaxValue(NumBits).getZExtValue());
     Value *Masked = Builder.CreateAnd(Trunc->getOperand(0), Mask);
+    if (SrcTy != ExtTy)
+      Masked = Builder.CreateTrunc(Masked, ExtTy);
 
     if (auto *I = dyn_cast<Instruction>(Masked))
       NewInsts.insert(I);
 
     ReplaceAllUsersOfWith(Trunc, Masked);
   }
 }
 

llvm/test/Transforms/TypePromotion/AArch64/pr58843.ll

-; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
-; RUN: opt -mtriple=aarch64 -type-promotion -verify -S %s -o - | FileCheck %s
-target datalayout = "e-m:e-i8:8:32-i16:16:32-i64:64-i128:128-n32:64-S128"
-
-; Check the case don't crash due to zext source type bitwidth
-; larger than dest type bitwidth.
-define i1 @test(i8 %arg) {
-; CHECK-LABEL: @test(
-; CHECK-NEXT:    [[EXT1:%.*]] = zext i8 [[ARG:%.*]] to i64
-; CHECK-NEXT:    [[TMP1:%.*]] = and i64 [[EXT1]], 7
-; CHECK-NEXT:    [[TMP2:%.*]] = trunc i64 [[TMP1]] to i32
-; CHECK-NEXT:    [[CMP:%.*]] = icmp ne i32 [[TMP2]], 0
-; CHECK-NEXT:    ret i1 [[CMP]]
-;
-  %ext1 = zext i8 %arg to i64
-  %trunc = trunc i64 %ext1 to i3
-  %ext2 = zext i3 %trunc to i8
-  %cmp = icmp ne i8 %ext2, 0
-  ret i1 %cmp
-}

llvm/test/Transforms/TypePromotion/AArch64/trunc-zext-chain.ll

latch:                                             ; preds = %bb14, %bb9
   %var36 = icmp eq %struct.zot* %var35, null
   %var37 = getelementptr inbounds %struct.zot, %struct.zot* %var35, i64 0, i32 2
   %var38 = load %struct.wobble*, %struct.wobble** %var37, align 8
   br i1 %var36, label %exit, label %header
 
 exit:
   ret i64 %var30
 }
+
+; Check the case don't crash due to zext source type bitwidth
+; larger than dest type bitwidth.
+define i1 @pr58843(i8 %arg) {
+; CHECK-LABEL: @pr58843(
+; CHECK-NEXT:    [[EXT1:%.*]] = zext i8 [[ARG:%.*]] to i64
+; CHECK-NEXT:    [[TMP1:%.*]] = and i64 [[EXT1]], 7
+; CHECK-NEXT:    [[TMP2:%.*]] = trunc i64 [[TMP1]] to i32
+; CHECK-NEXT:    [[CMP:%.*]] = icmp ne i32 [[TMP2]], 0
+; CHECK-NEXT:    ret i1 [[CMP]]
+;
+  %ext1 = zext i8 %arg to i64
+  %trunc = trunc i64 %ext1 to i3
+  %ext2 = zext i3 %trunc to i8
+  %cmp = icmp ne i8 %ext2, 0
+  ret i1 %cmp
+}
+
+; Check the case don't crash due to xor two op have different
+; types
+define i1 @pr59554(i8 %arg) {
+; CHECK-LABEL: @pr59554(
+; CHECK-NEXT:    [[ARG_EXT:%.*]] = zext i8 [[ARG:%.*]] to i64
+; CHECK-NEXT:    [[TMP1:%.*]] = and i64 [[ARG_EXT]], 7
+; CHECK-NEXT:    [[TMP2:%.*]] = trunc i64 [[TMP1]] to i32
+; CHECK-NEXT:    [[SWITCH_TABLEIDX:%.*]] = xor i32 [[TMP2]], 1
+; CHECK-NEXT:    [[SWITCH_LOBIT:%.*]] = icmp ne i32 [[TMP2]], 0
+; CHECK-NEXT:    ret i1 [[SWITCH_LOBIT]]
+;
+  %arg.ext = zext i8 %arg to i64
+  %trunc = trunc i64 %arg.ext to i3
+  %switch.tableidx = xor i3 %trunc, 1
+  %switch.maskindex = zext i3 %trunc to i8
+  %switch.lobit = icmp ne i8 %switch.maskindex, 0
+  ret i1 %switch.lobit
+}