Branch-Prediction and Value-Prediction are two similar fields but they are really different fields.
When a branch prediction is wrong we have to change the execution path and therefore flush the executed instructions (on the wrong path).
With a value prediction the execution path is correct so you don't have to flush the instructions, but you have to replay them with the right value.
These two techniques involve very different hardware.
Here in the article it is about value prediction emulated in software using branch prediction, it uses the hardware of the branch prediction therefore mispredicted instructions must be flushed.
The only valid example of value prediction you gave here is for prediction of the stack address for memory renaming. In all other cases you have to flush the instructions, while for memory renaming you can just replay it.
And this is a very restricted form of value prediction. There is no real complete value prediction mechanism in modern processors yet.
When a branch prediction is wrong we have to change the execution path and therefore flush the executed instructions (on the wrong path).
With a value prediction the execution path is correct so you don't have to flush the instructions, but you have to replay them with the right value. These two techniques involve very different hardware.
Here in the article it is about value prediction emulated in software using branch prediction, it uses the hardware of the branch prediction therefore mispredicted instructions must be flushed.
The only valid example of value prediction you gave here is for prediction of the stack address for memory renaming. In all other cases you have to flush the instructions, while for memory renaming you can just replay it.
And this is a very restricted form of value prediction. There is no real complete value prediction mechanism in modern processors yet.