SIMDMask<>.any() and .all() dramatically slower than scalar boolean operations

[S1D1T1]

Description

Converting arrays in an existing App to SIMD4 created noticeable slowdowns. Profile shows the bottlenecks are calls to
any(SIMDMask<SIMD4<Float>>) or all(...)
calculating the same values with alternate simple methods resulted in a dramatic speedup:

Reproduction

Code listing with CPU profiler data. Note Bottleneck: 82% of function time attributed to any()

        // sum the acceleration vectors acting on a given point, by a set of masses
       func gravAccelerationSum(_ posX:SIMD4<Float>,_ posY:SIMD4<Float>) -> (SIMD4<Float>,SIMD4<Float>) {
         var accelerationX:SIMD4<Float> = .zero
         var accelerationY:SIMD4<Float> = .zero
     
5.5%         for grav in self.gravs {
2.3%           let xDiffs = posX - grav.position.x
2.3%           let yDiffs = posY - grav.position.y
     
3.5%           let distanceSquared = xDiffs * xDiffs + yDiffs * yDiffs
     		// compare to range threshold. ignore if ALL 4 locations beyond a certain range
           let inRange = distanceSquared .< grav.influenceRadius4 // creates SIMDMask
     
82%           if any(inRange) {  // ** Bottleneck
    // also tried:   if !all(inRange), with comparable results
     //      if inRange[0] || inRange[1] || inRange[2] || inRange[3] { // ** WORKAROUND
0.8%           let n = distanceSquared + grav.softeningSquared
0.7%             let gravDenominator = n * n.squareRoot()
1.5%             accelerationX +=   (grav.attractionTimesTimeTimesBaseG) * xDiffs / gravDenominator
1.5%             accelerationY +=   (grav.attractionTimesTimeTimesBaseG) * yDiffs / gravDenominator
           }
         }
         return (accelerationX,accelerationY)
       }

The bottleneck is particularly apparent given the heavy computation in the rest of the function - calculating multi body gravitational attractions - dwarfed by checking 4 bits? : )
switching to the workaround code drops the boolean check from 82% to <1% and the other values bubble up appropriately. real world speed shows large increase. Functionality is identical aside from the performance change.

measuring execution time using osSignPoster

      let signpost = physicsSP.beginInterval("gravAccelerationSum",id:physicsSP.makeSignpostID())

      for i in 0..<SIMD4Count { // SIMD4Count = 6250 in this test case
        (gravInfluences.x[i], gravInfluences.y[i]) = gravAccelerationSum(particleLocations.x[i],particleLocations.y[i])
      }
      physicsSP.endInterval("gravAccelerationSum", signpost)

these durations are outside a loop which call the function 6250x in the test runs

Average duration x 6250. using any(): 31.85ms
Average duration x 6250. using workaround:: 9.98ms

Expected behavior

expected SIMDMask.any() and .all() to execute at least as efficiently as individual boolean comparisons

Environment

swift-driver version: 1.115 Apple Swift version 6.0.2 (swiftlang-6.0.2.1.2 clang-1600.0.26.4)
Target: arm64-apple-macosx14.0

• Xcode 16.1 and 16.2 - tested vs both
• macOS 14.7.1
• Apple M2 Max cpu

Additional information

related bug (yet to be filed): similar bottleneck in SIMD4.max(), manual comparison is order of magnitude faster.

this is my first compiler bug report. if i’m doing it wrong, tell me how to do it right.

[S1D1T1]

this is related to, if not a duplicate of #72413