mx.compile: assigning an elementwise expression to a negative-strided slice writes only one element

[cetagostini]

Describe the bug

Under mx.compile, assigning to a negative-strided slice (a[::-1] = expr or a[::-1] += expr) produces wrong results when the right-hand side is a non-trivial elementwise expression of a negative-strided view. Only a single element ends up written; the rest stay zero. The same function is correct in eager mode, and is correct under compile when the slice is positive-strided or when the RHS is a bare view.

To Reproduce

import mlx.core as mx

def f(x):
    base = mx.zeros_like(x)
    base[::-1] += 2.0 * x[::-1]      # negative-strided scatter; RHS is an elementwise expr
    return base

x = mx.arange(6, dtype=mx.float32)
print("eager   :", f(x).tolist())              # [0.0, 2.0, 4.0, 6.0, 8.0, 10.0]   (correct)
print("compiled:", mx.compile(f)(x).tolist())  # [0.0, 0.0, 0.0, 0.0, 0.0, 10.0]   (WRONG)

The set form fails identically:

def g(x):
    base = mx.zeros_like(x)
    base[::-1] = 2.0 * x[::-1]
    return base

mx.compile(g)(x).tolist()   # [0.0, 0.0, 0.0, 0.0, 0.0, 10.0]  (WRONG; eager is [0,2,4,6,8,10])

Characterization (what does / doesn't trigger it)

All in mx.compile, x = mx.arange(6, dtype=mx.float32):

Case	Compiled result	Correct?
base[::-1] += 2.0 * x[::-1]	[0,0,0,0,0,10]	❌
base[::-1] = 2.0 * x[::-1]	[0,0,0,0,0,10]	❌
base[::-1] += x[::-1] + x[::-1]	[0,0,0,0,0,10]	❌
base[::-1] += x[::-1] (bare view RHS)	[0,1,2,3,4,5]	✅
base[::-1] += x (non-strided RHS)	[5,4,3,2,1,0]	✅
base[::2] += 2.0 * x[::2] (positive step)	[0,0,4,0,8,0]	✅
any of the above in eager mode	—	✅

So the bug needs the combination of (a) a negative-strided destination slice, (b) an RHS that is an elementwise expression (not a bare view), and (c) mx.compile.

Expected behavior

mx.compile(f)(x) should equal f(x) (eager), i.e. [0, 2, 4, 6, 8, 10].

Desktop (please complete the following information):

• OS: macOS 26.5.1 (arm64, Apple Silicon)
• MLX version: 0.31.2