[llvm] [PowerPC] improve performance on the isNan and !isNan function in case of -ffp-model=strict" (PR #215587)

[zhijian lin via llvm-commits]

diggerlin wrote:

> 1. There seems a slightly difference in terms of instruction count between the two PRs in `P8 + VSX`.
> 2. There is also a difference that "reland" used `xscmpudp` and this PR uses `fcmpu`.
> 
> Just an observation from Bob's analysis. Not sure if it is significant or not though.
> 
> ## Instruction count comparison: `isNan` / `!isNan` under `-ffp-model=strict`
> ### `isNan` (i32 3)
> Function	Target	Commit `7e1aba74` (reland)	PR #215587 (fix)	Delta
> `isNan` f64/f32	P8 + VSX	3 (`xscmpudp` + `li` + `iseleq`)	4 (`fcmpu` + 2×`li` + `isel`)	+1
> `isNan` f64/f32	P9	4 (`xststdcdp` + 2×`li` + `iseleq`)	4 (`xststdcdp` + 2×`li` + `iseleq`)	—
> `isNan` f64/f32	P8 no-VSX	3 (`fcmpu` + `li` + `iseleq`)	3 (`fcmpu` + `li` + `iseleq`)	—
> ### `!isNan` (i32 1020)
> Function	Target	Commit `7e1aba74` (reland)	PR #215587 (fix)	Delta
> `!isNan` f64/f32	P8 + VSX	3 (`xscmpudp` + `li` + `isel`)	3 (`fcmpu` + `li` + `isel`)	—
> `!isNan` f64/f32	P9	3 (`xststdcdp` + `li` + `iseleq`)	3 (`xststdcdp` + `li` + `iseleq`)	—
> `!isNan` f64/f32	P8 no-VSX	3 (`fcmpu` + `li` + `isel`)	3 (`fcmpu` + `li` + `isel`)	—

1. we do not re_land the commit 7e1aba74, the implement has some problem(it cause the 32bit crash.  The fix in the current commit is simple and clear.  the title of the patch is  "[improve performance on the isNan and !isNan function in case of -ffp-model=strict"](https://github.com/llvm/llvm-project/pull/215587#top)" not to compare with reland.

2. your compare 
> Function	Target	Commit `7e1aba74` (reland)	PR #215587 (fix)	Delta
> `isNan` f64/f32	P8 + VSX	3 (`xscmpudp` + `li` + `iseleq`)	4 (`fcmpu` + 2×`li` + `isel`)	+1

 in current llvm implement , the SDValue generated by

```
the  ISD::CondCode CC = (Test == fcNan) ? ISD::SETUO : ISD::SETO;
  return DAG.getSelectCC(Dl, Op, Op, DAG.getConstant(1, Dl, ResultVT),
                         DAG.getConstant(0, Dl, ResultVT), CC);
```
do not convert to (`xscmpudp` + `li` + `iseleq`) , instead it convert  4 (`fcmpu` + 2×`li` + `isel`)	for P8+VSX ,  it is common problem for the SDValue , we can have further patch to achieve it.

https://github.com/llvm/llvm-project/pull/215587