Add egraph optimization for fneg's cancelling out (bytecodealliance#5910)

This implements comments from bytecodealliance#5895 to cancel out `fneg` operations in
`fma` instructions. Additional support for `fmul` is added as well.

Author: alexcrichton

cranelift/codegen/src/opts/algebraic.isle

@@ -527,3 +527,13 @@
 (rule (simplify
        (vselect ty (fcmp _ (FloatCC.GreaterThan) x y) x y))
       (fmax_pseudo ty x y))
+
+;; If both of the multiplied arguments to an `fma` are negated then remove
+;; both of them since they cancel out.
+(rule (simplify (fma ty (fneg ty x) (fneg ty y) z))
+      (fma ty x y z))
+
+;; If both of the multiplied arguments to an `fmul` are negated then remove
+;; both of them since they cancel out.
+(rule (simplify (fmul ty (fneg ty x) (fneg ty y)))
+      (fmul ty x y))

cranelift/filetests/filetests/egraph/algebraic.clif

@@ -483,3 +483,25 @@ block0(v1: i16):
 
 ; check: v4 = sextend.i64 v1
 ; check: return v4
+
+function %fma_double_fneg(f32, f32, f32) -> f32 {
+block0(v1: f32, v2: f32, v3: f32):
+    v4 = fneg v1
+    v5 = fneg v2
+    v6 = fma v4, v5, v3
+    return v6
+}
+
+; check: v7 = fma v1, v2, v3
+; check: return v7
+
+function %fmul_double_fneg(f32, f32) -> f32 {
+block0(v1: f32, v2: f32):
+    v3 = fneg v1
+    v4 = fneg v2
+    v5 = fmul v3, v4
+    return v5
+}
+
+; check: v6 = fmul v1, v2
+; check: return v6