[AtomicOp](fix) Extend atomic min/max canonicalization

[CHNJZ]

Background

PR [#1460](#1460) introduced AtomicMaxMinCanonicalizer to fuse the integer atomic operations expanded from floating-point atomic_min/atomic_max back into a single floating-point atomic operation before discrete-mask conversion.

However, the canonicalizer did not cover the following two IR forms.

Changes

1. Support scalar floating-point atomic min/max

   Floating-point atomics whose value is produced by a scalar reduction use scalar types instead of TensorType. Their positive sign mask is also represented as:

   %sign = arith.shrui %value_bits, %c31_i32 : i32
   %positive = arith.cmpi eq, %sign, %c0_i32 : i32

   Extend the canonicalizer to:

   • Obtain the element type from both scalar and tensor values.
   • Recognize the scalar cmpi eq 0 sign-mask form.
   • Create an all-true mask for both scalar and shaped atomic operations.

2. Support pointers represented as splat(bitcast(ptr))

   For tensor atomics constructed from a scalar base pointer, the pointer conversion may be represented as:

   %int_ptr = tt.bitcast %float_ptr : !tt.ptr<f32> -> !tt.ptr<i32>
   %int_ptrs = tt.splat %int_ptr : !tt.ptr<i32> -> tensor<...x!tt.ptr<i32>>

   PR [AtomicOp](fix) Fuse expanded floating-point atomic min/max #1460 only recognized atomics whose pointer was directly defined by tt.bitcast.

   Extend the canonicalizer to recognize both direct bitcast and splat(bitcast(ptr)) forms, rebuild the corresponding floating-point pointer tensor, and remove dead pointer splat/bitcast operations after fusion.

Test

• Add regression coverage for scalar floating-point atomic_max.
• Add regression coverage for 4D floating-point atomic_min using a splatted scalar base pointer.

[github-actions]

✅ DCO Check Passed

Commits

Identity	Commits
chnjz233 <ji…@huawei.com>	1 (signed off)

[CHNJZ]

/retry