<charconv>: simplify _Assemble_floating_point_value and optimize _Right_shift_with_rounding

[statementreply]

• Modify the behavior of _Assemble_floating_point_value_no_shift (renamed to avoid potential ODR issue) to gracefully handle the special cases below, and remove special case handling code from _Assemble_floating_point_value.

  • When the significand carries over to a higher bit after rounding up, we need to renormalize the significand and increase the exponent to keep the significand within the normalized range.
    (example: 0x1.fffffffffffff8p+0 rounds to 0x1.0000000000000p+1)
  • In some cases, the new exponent becomes greater than the maximum exponent of the floating point format, so the result overflows.
    (example: 0x1.fffffffffffff8p+1023 overflows to ∞)
  • In some cases, the mantissa of a subnormal value becomes normalized after rounding up, so the result becomes a normal value.
    (example: 0x0.fffffffffffff8p-1022 rounds to 0x1.0000000000000-p1022)

• Optimize _Right_shift_with_rounding with the branchless rounding technique in hex to_chars.

[StephanTLavavej]

Thanks! I think I understand the overall approach here, and the new assembly process makes more sense than the previous control flow. I'm marking this as Request Changes for the "tail bits" bug that I believe I found (plus testing). There's an additional question about space/time improvements for the rounding technique.

[statementreply]

Here are my benchmark results (Intel Core i5-8400, fixed CPU clock speed at 2.7 GHz, VS 2019 16.8 Preview 2, Clang/LLVM 11.0.0-rc1). The measured times are average nanoseconds per floating-point string.

Times with default dynamic CPU clock speed setting are around 0.7x the values below.

Scenario	MSVC Baseline	MSVC +Assemble	MSVC +Rounding	MSVC +Both	LLVM Baseline	LLVM +Assemble	LLVM +Rounding	LLVM +Both
x64 float hex 5 (exact)	49.0	49.4	48.7	49.2	51.9	51.7	51.5	51.4
x64 float hex 6 (rounding)	61.2	61.2	52.2	💚 53.1	55.5	56.4	54.7	55.1
x64 double hex 13 (exact)	64.6	62.3	64.4	✔️ 62.4	66.8	66.9	66.7	66.9
x64 double hex 14 (rounding)	75.4	74.2	67.8	💚 66.7	70.9	71.9	70.3	70.9
x64 float plain shortest roundtrip	159.9	162.3	151.2	✔️ 153.4	144.4	146.4	145.8	143.9
x64 double plain shortest roundtrip	247.2	244.9	237.3	✔️ 232.8	222.4	225.6	224.5	224.0
x86 float hex 5 (exact)	62.0	61.1	62.0	61.1	61.3	60.4	61.5	60.3
x86 float hex 6 (rounding)	86.3	85.4	71.6	💚 71.0	70.8	71.8	68.2	68.8
x86 double hex 13 (exact)	80.6	83.6	80.7	❌ 83.6	78.2	78.7	78.9	79.8
x86 double hex 14 (rounding)	106.7	106.9	91.2	💚 91.2	91.6	89.6	87.9	89.5
x86 float plain shortest roundtrip	205.2	204.5	186.5	💚 185.9	175.0	175.4	169.9	✔️ 169.9
x86 double plain shortest roundtrip	311.0	316.4	293.2	✔️ 297.3	267.5	271.2	265.7	266.2

[StephanTLavavej]

Thanks for the detailed perf data - nice improvements for MSVC! I'll push a one-line change to use logical AND with bool. FYI @cbezault as you had previously approved.

[StephanTLavavej]

Thanks again for improving <charconv> again! 🚀 😸