Alias tracking fails for basic views like reshape and diagonal causing silent data staleness upon base tensor mutation

[rookieLiu2018]

bug描述 Describe the Bug

Description:
When compiling models in static mode, view tensors are not correctly synchronized if their original base tensor is modified in-place.

In Eager execution, modifying a base tensor automatically updates its active views, whether they are shape-changing views, subset views, or offset views. However, in static mode, this alias relationship is broken. The view tensor retains a stale, pre-mutation snapshot of the data, which leads to a critical silent value mismatch.

Minimal Reproducer:

import os
import paddle

def kernel_contiguous_views(x):
    # 1. Shape-changing View
    v_resh = paddle.reshape(x, [4])
    
    # 2. Subset View
    v_slic = x[1, :]
    
    # 3. Offset View
    v_diag = paddle.diagonal(x)
    
    # CORE MUTATION: Mutate the base tensor in-place
    x[1, 1] = 99.0
    
    return paddle.sum(v_resh), paddle.sum(v_slic), paddle.sum(v_diag)

# Initialize data: [[0.0, 1.0], [2.0, 3.0]]
x_init = paddle.arange(4, dtype="float32").reshape([2, 2])

# Eager Execution
e_resh, e_slic, e_diag = kernel_contiguous_views(x_init.clone())

# Static PIR Execution
static_fn = paddle.jit.to_static(kernel_contiguous_views, full_graph=True)
s_resh, s_slic, s_diag = static_fn(x_init.clone())

print("=== Basic View Alias Tracking Failure ===")
print(f"[reshape]  Eager Sum: {e_resh.item():.1f} | Static Sum: {s_resh.item():.1f}")
print(f"[slice]    Eager Sum: {e_slic.item():.1f} | Static Sum: {s_slic.item():.1f}")
print(f"[diagonal] Eager Sum: {e_diag.item():.1f} | Static Sum: {s_diag.item():.1f}")

Output:

=== Basic View Alias Tracking Failure ===
[reshape]  Eager Sum: 102.0 | Static Sum: 6.0
[slice]    Eager Sum: 101.0 | Static Sum: 5.0
[diagonal] Eager Sum: 99.0 | Static Sum: 3.0

其他补充信息 Additional Supplementary Information

Paddle Version: 3.3.0