[AArch64] Sink mismatching wide extends to mul (#164986)

If we have v4i64 mul(zext(v4i16), sext(v4i16)), we can code-generate
that as v4i64 smull(v4i32 zext(v4i16), sext(v4i16), as
zext(x)==sext(zext(x)). This teaches the part of CGP that sinks operands
to uses about that, so that it can treat a zext that is more than twice
the width as a sext.

Author: davemgreen

llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp

@@ -6657,10 +6657,15 @@ bool AArch64TTIImpl::isProfitableToSinkOperands(
           Ops.push_back(&Ext->getOperandUse(0));
         Ops.push_back(&Op);
 
-        if (isa<SExtInst>(Ext))
+        if (isa<SExtInst>(Ext)) {
           NumSExts++;
-        else
+        } else {
           NumZExts++;
+          // A zext(a) is also a sext(zext(a)), if we take more than 2 steps.
+          if (Ext->getOperand(0)->getType()->getScalarSizeInBits() * 2 <
+              I->getType()->getScalarSizeInBits())
+            NumSExts++;
+        }
 
         continue;
       }