convert bf16 to u16

Author: gwenzek

src/MacroTranslator.zig

@@ -355,8 +355,17 @@ fn parseCNumLit(mt: *MacroTranslator) ParseError!ZigNode {
             .L => "c_longdouble",
             .W, .F64x => "f80",
             .Q, .F128 => "f128",
-            .BF16 => "u16",
-
+            .BF16 => {
+                // Zig doesn't support BFloat16, so the best thing we can do is convert it to u16,
+                // so that the value is bit-to-bit identic between the emitted Zig and original C code.
+                // https://github.com/ziglang/zig/issues/3148
+                const value_u16 = parseBfloat16AsU16(bytes.items);
+                std.debug.assert(bytes.capacity >= 6); // guaranteed since we allocated literal len + 3
+                const bf16_hex = std.fmt.bufPrint(bytes.allocatedSlice(), "0x{x}", .{value_u16}) catch unreachable;
+                const rhs = try ZigTag.integer_literal.create(arena, bf16_hex);
+                const type_node = try ZigTag.type.create(arena, "u16");
+                return ZigTag.as.create(arena, .{ .lhs = type_node, .rhs = rhs });
+            },
             else => {
                 try mt.fail("TODO: float literal suffix: '{s}'", .{suffix_str});
                 return error.ParseError;
@@ -1359,3 +1368,11 @@ fn parseCUnaryExpr(mt: *MacroTranslator, scope: *Scope) ParseError!ZigNode {
 
     return try mt.parseCPostfixExpr(scope, null);
 }
+
+fn parseBfloat16AsU16(bytes: []const u8) u16 {
+    const value_f32 = std.fmt.parseFloat(f32, bytes) catch math.nan(f32);
+    var int: u32 = @bitCast(value_f32);
+    // Round up if needed.
+    int += 0x8000;
+    return @truncate(int >> 16);
+}

test/cases/translate/bf16_suffix_after_float_literal.c

@@ -0,0 +1,5 @@
+#define EXACT_BF16 0.007873535BF16
+
+// translate
+//
+// pub const EXACT_BF16 = @as(u16, 0x3c01);