[llvm] ValueTracking: Check if x is undef for fma(x, x, y) analysis (PR #174763)

[Sergey Shcherbinin via llvm-commits]

SergeyShch01 wrote:

Hello @arsenm, 
This commit caused a performance regression of the benchmark we are tracking. Could you please have a look?

The reproducer is as follows (should be compiled w/ -O3):

```
#include <cmath>
double kernel(int64_t ibegin, int64_t iend,
                       double x0, double h,
                       double y, double xp, double yp)
{
    double sumx = 0.0;
    for (int64_t i = ibegin; i < iend; ++i) {
        double x = x0 + i * h;
        double denom = (x - xp)*(x - xp) + (y - yp)*(y - yp);
        denom = 1.0 / sqrt(denom);
        sumx += denom;
    }    
    return sumx;
}
```

This commit leads to the vectorization not beeing applied. Debug logs show the following message: 

```
LV: Not vectorizing: Found a non-intrinsic callsite 
```
On previous revision (12c89e40) this loop was successfully vectorized.

https://github.com/llvm/llvm-project/pull/174763