[Instcombine] Avoid widening trunc+sext to trunc+shl+ashr when not profitable

Skip the transform that replaces:
  %a = trunc i64 %x to i16
  %b = sext i16 %a to i32
with
  %a = trunc i64 %x to i32
  %b = shl i32 %a, 16
  %c = ashr exact i32 %b, 16
when (pre-trunc) source type is wider than the final destination type.
Modern architectures typically have efficient sign-extend instruction.
It is preferable to preserve the sext for this case.

Resolves #116019

Author: wenju-he

llvm/lib/Transforms/InstCombine/InstCombineCasts.cpp

@@ -1528,7 +1528,17 @@ Instruction *InstCombinerImpl::visitSExt(SExtInst &Sext) {
   }
 
   // Try to extend the entire expression tree to the wide destination type.
-  if (shouldChangeType(SrcTy, DestTy) && canEvaluateSExtd(Src, DestTy)) {
+  bool shouldExtendExpression = true;
+  Value *X = nullptr;
+  // Do not extend expression in the trunc + sext pattern when destination type
+  // is narrower than original (pre-trunc) type: modern architectures typically
+  // provide efficient sign-extend instruction, so preserving the sext is
+  // preferable here.
+  if (match(Src, m_Trunc(m_Value(X))))
+    if (X->getType()->getScalarSizeInBits() > DestBitSize)
+      shouldExtendExpression = false;
+  if (shouldExtendExpression && shouldChangeType(SrcTy, DestTy) &&
+      canEvaluateSExtd(Src, DestTy)) {
     // Okay, we can transform this!  Insert the new expression now.
     LLVM_DEBUG(
         dbgs() << "ICE: EvaluateInDifferentType converting expression type"
@@ -1548,8 +1558,7 @@ Instruction *InstCombinerImpl::visitSExt(SExtInst &Sext) {
                                       ShAmt);
   }
 
-  Value *X;
-  if (match(Src, m_Trunc(m_Value(X)))) {
+  if (X) {
     // If the input has more sign bits than bits truncated, then convert
     // directly to final type.
     unsigned XBitSize = X->getType()->getScalarSizeInBits();

llvm/test/Transforms/InstCombine/sext.ll

@@ -320,6 +320,17 @@ define i10 @test19(i10 %i) {
   ret i10 %d
 }
 
+define i32 @test20(i64 %i) {
+; CHECK-LABEL: @test20(
+; CHECK-NEXT:    [[A:%.*]] = trunc i64 [[I:%.*]] to i16
+; CHECK-NEXT:    [[B:%.*]] = sext i16 [[A]] to i32
+; CHECK-NEXT:    ret i32 [[B]]
+;
+  %a = trunc i64 %i to i16
+  %b = sext i16 %a to i32
+  ret i32 %b
+}
+
 define i32 @smear_set_bit(i32 %x) {
 ; CHECK-LABEL: @smear_set_bit(
 ; CHECK-NEXT:    [[TMP1:%.*]] = shl i32 [[X:%.*]], 24