fix: missing neg in rsqrt derivative

[avik-pal]

No description provided.

[Pangoraw]

The corresponding test needs to be updated:

Enzyme-JAX/test/lit_tests/diffrules/stablehlo/rsqrt.mlir

Lines 19 to 29 in 4e28cbc

	// REVERSE: func.func @main(%arg0: tensor<2xf32>, %arg1: tensor<2xf32>) -> tensor<2xf32> {
	// REVERSE-NEXT: %cst = stablehlo.constant dense<2.000000e+00> : tensor<2xf32>
	// REVERSE-NEXT: %cst_0 = arith.constant dense<0.000000e+00> : tensor<2xf32>
	// REVERSE-NEXT: %0 = arith.addf %arg1, %cst_0 : tensor<2xf32>
	// REVERSE-NEXT: %1 = stablehlo.sqrt %arg0 : tensor<2xf32>
	// REVERSE-NEXT: %2 = stablehlo.multiply %arg0, %1 : tensor<2xf32>
	// REVERSE-NEXT: %3 = stablehlo.multiply %cst, %2 : tensor<2xf32>
	// REVERSE-NEXT: %4 = stablehlo.divide %0, %3 : tensor<2xf32>
	// REVERSE-NEXT: %5 = arith.addf %4, %cst_0 : tensor<2xf32>
	// REVERSE-NEXT: return %5 : tensor<2xf32>
	// REVERSE-NEXT: }

[wsmoses]

Minor but can you put the neg on the mul and not on the div? This will enable better constant propagation (since often times differrt will be 1 or 0 or a phi thereof so then it simplifies nicely)

[wsmoses]

Also actually why not just make the constant 2 into -2?

[wsmoses]

This also only changes one AD mode, presumably you’ll need to change the 2 to be a -2 in the line above as well