Fix legalization of SIMD fneg (#1286)

Previously `fsub` was used but this fails when negating -0.0 and +0.0 in the SIMD spec tests; using more instructions, this change uses shifts to create a constant for flipping the most significant bit of each lane with `bxor`.

cranelift/codegen/meta/src/isa/x86/legalize.rs

@@ -40,7 +40,6 @@ pub(crate) fn define(shared: &mut SharedDefinitions, x86_instructions: &Instruct
     let fmax = insts.by_name("fmax");
     let fmin = insts.by_name("fmin");
     let fneg = insts.by_name("fneg");
-    let fsub = insts.by_name("fsub");
     let iadd = insts.by_name("iadd");
     let icmp = insts.by_name("icmp");
     let iconst = insts.by_name("iconst");
@@ -48,6 +47,7 @@ pub(crate) fn define(shared: &mut SharedDefinitions, x86_instructions: &Instruct
     let ineg = insts.by_name("ineg");
     let insertlane = insts.by_name("insertlane");
     let ishl = insts.by_name("ishl");
+    let ishl_imm = insts.by_name("ishl_imm");
     let isub = insts.by_name("isub");
     let popcnt = insts.by_name("popcnt");
     let raw_bitcast = insts.by_name("raw_bitcast");
@@ -550,9 +550,18 @@ pub(crate) fn define(shared: &mut SharedDefinitions, x86_instructions: &Instruct
 
     for ty in &[F32, F64] {
         let fneg = fneg.bind(vector(*ty, sse_vector_size));
+        let lane_type_as_int = LaneType::int_from_bits(LaneType::from(*ty).lane_bits() as u16);
+        let uimm8_shift = Literal::constant(&imm.uimm8, lane_type_as_int.lane_bits() as i64 - 1);
+        let vconst = vconst.bind(vector(lane_type_as_int, sse_vector_size));
+        let bitcast_to_float = raw_bitcast.bind(vector(*ty, sse_vector_size));
         narrow.legalize(
             def!(b = fneg(a)),
-            vec![def!(c = vconst(u128_zeroes)), def!(b = fsub(c, a))],
+            vec![
+                def!(c = vconst(ones)),
+                def!(d = ishl_imm(c, uimm8_shift)), // Create a mask of all 0s except the MSB.
+                def!(e = bitcast_to_float(d)),      // Cast mask to the floating-point type.
+                def!(b = bxor(a, e)),               // Flip the MSB.
+            ],
         );
     }
 

cranelift/filetests/filetests/isa/x86/simd-arithmetic-legalize.clif

@@ -39,13 +39,17 @@ function %fneg_legalized() {
 ebb0:
     v0 = vconst.f32x4 [0x1.0 0x2.0 0x3.0 0x4.0]
     v1 = fneg v0
-    ; check: v4 = vconst.f32x4 0x00
-    ; nextln: v1 = fsub v4, v0
+    ; check: v4 = vconst.i32x4 0xffffffffffffffffffffffffffffffff
+    ; nextln: v5 = ishl_imm v4, 31
+    ; nextln: v6 = raw_bitcast.f32x4 v5
+    ; nextln: v1 = bxor v0, v6
 
     v2 = vconst.f64x2 [0x1.0 0x2.0]
     v3 = fneg v2
-    ; check: v5 = vconst.f64x2 0x00
-    ; nextln: v3 = fsub v5, v2
+    ; check: v7 = vconst.i64x2 0xffffffffffffffffffffffffffffffff
+    ; nextln: v8 = ishl_imm v7, 63
+    ; nextln: v9 = raw_bitcast.f64x2 v8
+    ; nextln: v3 = bxor v2, v9
 
     return
 }

cranelift/filetests/filetests/isa/x86/simd-arithmetic-run.clif

@@ -240,6 +240,19 @@ ebb0:
 }
 ; run
 
+function %fneg_f32x4() -> b1 {
+ebb0:
+    v0 = vconst.f32x4 [0x0.0 -0x0.0 -Inf Inf]
+    v1 = fneg v0
+
+    v2 = vconst.f32x4 [-0x0.0 0x0.0 Inf -Inf]
+    v3 = fcmp eq v1, v2
+    v4 = vall_true v3
+
+    return v4
+}
+; run
+
 function %fabs_f32x4() -> b1 {
 ebb0:
     v0 = vconst.f32x4 [0x0.0 -0x1.0 0x2.0 -0x3.0]