Go版本

Go1.19+

在你的模块/工作区中go env的输出：

NA

你做了什么？

根据预先计算的p256AffineTable和p256SelectAffine函数的定义：

// p256AffineTable is a table of the first 32 multiples of a point. Points are
// stored at an index offset of -1 like in p256Table, and [0]P is not stored.
type p256AffineTable [32]p256AffinePoint

// p256Precomputed is a series of precomputed multiples of G, the canonical
// generator. The first p256AffineTable contains multiples of G. The second one
// multiples of [2⁶]G, the third one of [2¹²]G, and so on, where each successive
// table is the previous table doubled six times. Six is the width of the
// sliding window used in p256ScalarMult, and having each table already
// pre-doubled lets us avoid the doublings between windows entirely. This table
// MUST NOT be modified, as it aliases into p256PrecomputedEmbed below.
var p256Precomputed *[43]p256AffineTable

// p256SelectAffine sets res to the point at index idx in the table.
// idx must be in [0, 31]. It executes in constant time.
//
//go:noescape
func p256SelectAffine(res *p256AffinePoint, table *p256AffineTable, idx int)

如果每次取一个点，p256SelectAffine中的正确循环次数应该是32。

你看到了什么发生？

1. go/src/crypto/internal/nistec/p256_asm_s390x.s
   36b45bc中的第511行
   | | CMPW COUNT, $65 |

2. go/src/crypto/internal/nistec/p256_asm_ppc64le.s
   36b45bc中的第444行
   | | MOVD$64, COUNT |

3. go/src/crypto/internal/nistec/p256_asm_ppc64le.s
   36b45bc中的第294行
   | | MOVD$17, COUNT |
   这将影响ScalarBaseMult性能。

你期望看到什么？

1. go/src/crypto/internal/nistec/p256_asm_s390x.s
   36b45bc中的第511行
   | | CMPW COUNT, $65 |
   应该为33。

2. go/src/crypto/internal/nistec/p256_asm_ppc64le.s
   36b45bc中的第444行
   | | MOVD$64, COUNT |
   应该为32。

3. go/src/crypto/internal/nistec/p256_asm_ppc64le.s
   36b45bc中的第294行
   | | MOVD$17, COUNT |
   应该为16。
   顺便说一下，对于s390x汇编实现：

4. 尝试使用VLM/VSTM;

5. go/src/crypto/internal/nistec/p256_asm_s390x.s
   36b45bc中的第579行
   | | // VLEF $3, 0(R4), K0 |