simd/archsimd, ppc64le: track SIMD (GOEXPERIMENT=simd) enablement for PowerPC64LE

[jkrishmys]

We are planning SIMD enablement for ppc64le under the accepted
GOEXPERIMENT=simd direction.

This issue tracks PowerPC64LE-specific implementation work, design
considerations, and blockers.

Related umbrella issues:

• simd/archsimd: architecture-specific SIMD intrinsics under a GOEXPERIMENT #73787 — SIMD intrinsics proposal
• simd: rename low-level SIMD package to simd/archsimd (GOEXPERIMENT) #76473 — simd/archsimd package naming

Dependencies / assumptions:

For the PowerPC port to fully utilize all 64 PPC vector registers, regMask needs to be 128 bits in size. After discussion with Junyang Shao, who is expanding regMask for the ARM64 port via https://go-review.googlesource.com/c/go/+/648635, we will develop the code based on the assumption that regMask will be expanded to 128 bits.

We will update this issue as implementation progresses.
cc @golang/ppc64

[phuclv90]

Excuse my PPC SIMD knowledge. I thought it has only 32 vector registers v0-v31?

[jkrishmys]

Excuse my PPC SIMD knowledge. I thought it has only 32 vector registers v0-v31?

Hi @phuclv90, AltiVec/VMX exposes 32 vector registers (v0–v31), which is likely what you're referring to.

With VSX (POWER7 and then on), the architectural register file is extended to 64 VSX (Vector Scalar Extension) registers (vs0–vs63). The lower 32 (vs0–vs31) overlap with the floating-point registers, and the upper 32 (vs32–vs63) overlap with the VMX registers. So for VSX-based SIMD codegen, the compiler effectively has 64 SIMD-capable registers to reason about.

[gabyhelp]

Related Issues

• cmd/asm, cmd/internal/obj/ppc64: ppc64 and ppc64le do not support vector (VMX) registers/instructions #15619 (closed)
• simd/archsimd: architecture-specific SIMD intrinsics under a GOEXPERIMENT #73787
• cmd/asm,cmd/compile,cmd/link,cmd/objdump: ppc64 ISA 3.1 support #44549 (closed)
• x/crypto/internal/chacha20: improve performance for ppc64le #25051 (closed)
• cmd/compile: PPC64 SSA support #16175 (closed)
• cmd/compile,math: improve performance for some math functions for ppc64x #21390 (closed)

_{(Emoji vote if this was helpful or unhelpful; more detailed feedback welcome in this discussion.)}