// Non-indexed ops generate DS-form loads and stores when the offset fits in 16 bits,
// and on power8 and power9, a multiple of 4 is required for MOVW and MOVD ops.
// On power10, prefixed loads and stores can be used for offsets > 16 bits and <= 32 bits.
// and support for PC relative addressing must be available if relocation is needed.
// On power10, the assembler will determine when to use DS-form or prefixed

		bytesPerLoop := int64(64)
		// This is used when moving more
		// than 8 bytes on power9.  Moves start with
		// as many 8 byte moves as possible, then
		// 4, 2, or 1 byte(s) as remaining.  This will
		// work and be efficient for power8 or later.
		// If there are 64 or more bytes, then a
		// loop is generated to move 32 bytes and
		// update the src and dst addresses on each

		"loong64": {},
		"mips":    {"GOMIPS", "hardfloat", "softfloat"},
		"mips64":  {"GOMIPS64", "hardfloat", "softfloat"},
		"ppc64":   {"GOPPC64", "power8", "power9", "power10"},
		"ppc64le": {"GOPPC64", "power8", "power9", "power10"},
		"ppc64x":  {}, // A pseudo-arch representing both ppc64 and ppc64le
		"s390x":   {},
		"wasm":    {},
		"riscv64": {"GORISCV64", "rva20u64", "rva22u64"},
	}
)

	// Hex constant 0x20004000000
	MOVD $2199090364416, R5         // 60058001
	// Hex constant 0xFFFFFE0004000000
	MOVD $-2198956146688, R5        // 38a08001
	// TODO: On GOPPC64={power8,power9}, this is preprocessed into MOVD $-1, R5; RLDC R5, $33, $63, R5.
	//       This only captures the MOVD. Should the RLDC be appended to the encoding by the test?
	// Hex constant 0xFFFFFFFE00000001

	}
	if goarch == "ppc64" || goarch == "ppc64le" {
		// We treat each powerpc version as a superset of functionality.
		switch goppc64 {
		case "power10":
			asmArgs = append(asmArgs, "-D", "GOPPC64_power10")
			fallthrough
		case "power9":
			asmArgs = append(asmArgs, "-D", "GOPPC64_power9")
			fallthrough
		default: // This should always be power8.
			asmArgs = append(asmArgs, "-D", "GOPPC64_power8")
		}
	}

   * is merely copied. Only as the power set is iterated are the individual subsets created, and
   * these subsets themselves occupy only a small constant amount of memory.
   *
   * @param set the set of elements to construct a power set from
   * @return the power set, as an immutable set of immutable sets

}
func isPowerOfTwo64(n int64) bool {
	return n > 0 && n&(n-1) == 0
}

// isUint64PowerOfTwo reports whether uint64(n) is a power of 2.
func isUint64PowerOfTwo(in int64) bool {
	n := uint64(in)
	return n > 0 && n&(n-1) == 0
}

// isUint32PowerOfTwo reports whether uint32(n) is a power of 2.
func isUint32PowerOfTwo(in int64) bool {
	n := uint64(uint32(in))
	return n > 0 && n&(n-1) == 0
}

   * is merely copied. Only as the power set is iterated are the individual subsets created, and
   * these subsets themselves occupy only a small constant amount of memory.
   *
   * @param set the set of elements to construct a power set from
   * @return the power set, as an immutable set of immutable sets

  auto axis = rewriter.create<ConstOp>(
      loc, GetScalarOfType(rewriter.getIntegerType(64), 0));
  return rewriter.create<ConcatV2Op>(loc, type, ValueRange(vals), axis);
}

// Lowers AddN op to a sequence of AddV2 ops to accumulate operands.
//
// Note that to improve the parallelism, AddN op uses tree-based reduction.
// For example, tf.AddN([0, 1, 2, 3, 4]) behaves as follows:
//

// and verified to be a power of two.
//
// If set, this must be set by the OS init code (typically in osinit) before
// mallocinit. However, setting it at all is optional, and leaving the default
// value is always safe (though potentially less efficient).
//
// Since physHugePageSize is always assumed to be a power of two,