[llvm] [X86] Lower `minimum`/`maximum`/`minimumnum`/`maximumnum` using bitwise operations (PR #170069)

[via llvm-commits]

================
@@ -208,12 +198,12 @@ define <32 x half> @test_fminimum_v32f16_szero(<32 x half> %x, <32 x half> %y) "
 define <32 x half> @test_fmaximum_v32f16_nans_szero(<32 x half> %x, <32 x half> %y) {
 ; CHECK-LABEL: test_fmaximum_v32f16_nans_szero:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vpmovw2m %zmm0, %k1
-; CHECK-NEXT:    vpblendmw %zmm1, %zmm0, %zmm2 {%k1}
+; CHECK-NEXT:    vmaxph %zmm1, %zmm0, %zmm2
+; CHECK-NEXT:    vpbroadcastw {{.*#+}} zmm1 = [NaN,NaN,NaN,NaN,NaN,NaN,NaN,NaN,NaN,NaN,NaN,NaN,NaN,NaN,NaN,NaN,NaN,NaN,NaN,NaN,NaN,NaN,NaN,NaN,NaN,NaN,NaN,NaN,NaN,NaN,NaN,NaN]
+; CHECK-NEXT:    vpternlogq {{.*#+}} zmm1 = zmm2 & (zmm1 | zmm0)
----------------
valadaptive wrote:

llvm-mca seems to think the new code is faster despite the memory load, at least in a tight loop. You can try it yourself with:

```asm
# LLVM-MCA-BEGIN old
vpmovw2m %zmm0, %k1
vpblendmw %zmm0, %zmm1, %zmm2 {%k1}
vmovdqu16 %zmm1, %zmm0 {%k1}
vminph %zmm2, %zmm0, %zmm0
# LLVM-MCA-END

# LLVM-MCA-BEGIN new
vminph %zmm1, %zmm0, %zmm1
vpbroadcastw (%rdi), %zmm2
vpternlogq $248, %zmm2, %zmm0, %zmm1
# LLVM-MCA-END
```

Here's what I get with `-mcpu=sapphirerapids`:

```
[0] Code Region - old

Iterations:        100
Instructions:      400
Total Cycles:      1103
Total uOps:        400

Dispatch Width:    6
uOps Per Cycle:    0.36
IPC:               0.36
Block RThroughput: 2.0

[snip]

[1] Code Region - new

Iterations:        100
Instructions:      300
Total Cycles:      607
Total uOps:        400

Dispatch Width:    6
uOps Per Cycle:    0.66
IPC:               0.49
Block RThroughput: 1.0
```

https://github.com/llvm/llvm-project/pull/170069