compiler_rt: change the abi of f16 on mac to depend on the other type

This commit is contained in:
Jacob Young 2023-03-27 21:32:37 -04:00 committed by Andrew Kelley
parent 243848167b
commit 5870ffeb82
11 changed files with 43 additions and 29 deletions

View File

@ -81,12 +81,20 @@ pub fn panic(msg: []const u8, error_return_trace: ?*std.builtin.StackTrace, _: ?
/// need for extending them to wider fp types.
/// TODO remove this; do this type selection in the language rather than
/// here in compiler-rt.
pub const F16T = switch (builtin.cpu.arch) {
.aarch64, .aarch64_be, .aarch64_32 => f16,
.riscv64 => if (builtin.zig_backend == .stage1) u16 else f16,
.x86, .x86_64 => if (builtin.target.isDarwin()) u16 else f16,
else => u16,
};
pub fn F16T(comptime other_type: type) type {
return switch (builtin.cpu.arch) {
.aarch64, .aarch64_be, .aarch64_32 => f16,
.riscv64 => if (builtin.zig_backend == .stage1) u16 else f16,
.x86, .x86_64 => if (builtin.target.isDarwin()) switch (other_type) {
// Starting with LLVM 16, Darwin uses different abi for f16
// depending on the type of the other return/argument..???
f32, f64 => u16,
f80, f128 => f16,
else => unreachable,
} else f16,
else => u16,
};
}
pub fn wideMultiply(comptime Z: type, a: Z, b: Z, hi: *Z, lo: *Z) void {
switch (Z) {

View File

@ -45,7 +45,7 @@ fn test__extenddftf2(a: f64, expected_hi: u64, expected_lo: u64) !void {
}
fn test__extendhfsf2(a: u16, expected: u32) !void {
const x = __extendhfsf2(@bitCast(F16T, a));
const x = __extendhfsf2(@bitCast(F16T(f32), a));
const rep = @bitCast(u32, x);
if (rep == expected) {
@ -208,7 +208,7 @@ fn makeInf32() f32 {
}
fn test__extendhftf2(a: u16, expected_hi: u64, expected_lo: u64) !void {
const x = __extendhftf2(@bitCast(F16T, a));
const x = __extendhftf2(@bitCast(F16T(f128), a));
const rep = @bitCast(u128, x);
const hi = @intCast(u64, rep >> 64);

View File

@ -7,6 +7,6 @@ comptime {
@export(__extendhfdf2, .{ .name = "__extendhfdf2", .linkage = common.linkage, .visibility = common.visibility });
}
pub fn __extendhfdf2(a: common.F16T) callconv(.C) f64 {
pub fn __extendhfdf2(a: common.F16T(f64)) callconv(.C) f64 {
return extendf(f64, f16, @bitCast(u16, a));
}

View File

@ -12,11 +12,11 @@ comptime {
@export(__extendhfsf2, .{ .name = "__extendhfsf2", .linkage = common.linkage, .visibility = common.visibility });
}
pub fn __extendhfsf2(a: common.F16T) callconv(.C) f32 {
pub fn __extendhfsf2(a: common.F16T(f32)) callconv(.C) f32 {
return extendf(f32, f16, @bitCast(u16, a));
}
fn __gnu_h2f_ieee(a: common.F16T) callconv(.C) f32 {
fn __gnu_h2f_ieee(a: common.F16T(f32)) callconv(.C) f32 {
return extendf(f32, f16, @bitCast(u16, a));
}

View File

@ -7,6 +7,6 @@ comptime {
@export(__extendhftf2, .{ .name = "__extendhftf2", .linkage = common.linkage, .visibility = common.visibility });
}
pub fn __extendhftf2(a: common.F16T) callconv(.C) f128 {
pub fn __extendhftf2(a: common.F16T(f128)) callconv(.C) f128 {
return extendf(f128, f16, @bitCast(u16, a));
}

View File

@ -7,6 +7,6 @@ comptime {
@export(__extendhfxf2, .{ .name = "__extendhfxf2", .linkage = common.linkage, .visibility = common.visibility });
}
fn __extendhfxf2(a: common.F16T) callconv(.C) f80 {
fn __extendhfxf2(a: common.F16T(f80)) callconv(.C) f80 {
return extend_f80(f16, @bitCast(u16, a));
}

View File

@ -11,10 +11,10 @@ comptime {
}
}
pub fn __truncdfhf2(a: f64) callconv(.C) common.F16T {
return @bitCast(common.F16T, truncf(f16, f64, a));
pub fn __truncdfhf2(a: f64) callconv(.C) common.F16T(f64) {
return @bitCast(common.F16T(f64), truncf(f16, f64, a));
}
fn __aeabi_d2h(a: f64) callconv(.AAPCS) u16 {
return @bitCast(common.F16T, truncf(f16, f64, a));
return @bitCast(common.F16T(f64), truncf(f16, f64, a));
}

View File

@ -12,14 +12,14 @@ comptime {
@export(__truncsfhf2, .{ .name = "__truncsfhf2", .linkage = common.linkage, .visibility = common.visibility });
}
pub fn __truncsfhf2(a: f32) callconv(.C) common.F16T {
return @bitCast(common.F16T, truncf(f16, f32, a));
pub fn __truncsfhf2(a: f32) callconv(.C) common.F16T(f32) {
return @bitCast(common.F16T(f32), truncf(f16, f32, a));
}
fn __gnu_f2h_ieee(a: f32) callconv(.C) common.F16T {
return @bitCast(common.F16T, truncf(f16, f32, a));
fn __gnu_f2h_ieee(a: f32) callconv(.C) common.F16T(f32) {
return @bitCast(common.F16T(f32), truncf(f16, f32, a));
}
fn __aeabi_f2h(a: f32) callconv(.AAPCS) u16 {
return @bitCast(common.F16T, truncf(f16, f32, a));
return @bitCast(common.F16T(f32), truncf(f16, f32, a));
}

View File

@ -7,6 +7,6 @@ comptime {
@export(__trunctfhf2, .{ .name = "__trunctfhf2", .linkage = common.linkage, .visibility = common.visibility });
}
pub fn __trunctfhf2(a: f128) callconv(.C) common.F16T {
return @bitCast(common.F16T, truncf(f16, f128, a));
pub fn __trunctfhf2(a: f128) callconv(.C) common.F16T(f128) {
return @bitCast(common.F16T(f128), truncf(f16, f128, a));
}

View File

@ -7,6 +7,6 @@ comptime {
@export(__truncxfhf2, .{ .name = "__truncxfhf2", .linkage = common.linkage, .visibility = common.visibility });
}
fn __truncxfhf2(a: f80) callconv(.C) common.F16T {
return @bitCast(common.F16T, trunc_f80(f16, a));
fn __truncxfhf2(a: f80) callconv(.C) common.F16T(f80) {
return @bitCast(common.F16T(f80), trunc_f80(f16, a));
}

View File

@ -2873,6 +2873,12 @@ typedef zig_repr_f16 zig_f16;
#undef zig_init_special_f16
#define zig_init_special_f16(sign, name, arg, repr) repr
#endif
#if __APPLE__
typedef zig_repr_f16 zig_compiler_rt_f16;
#else
typedef zig_f16 zig_compiler_rt_f16;
#endif
#define zig_compiler_rt_abbrev_zig_compiler_rt_f16 zig_compiler_rt_abbrev_zig_f16
#define zig_has_f32 1
#define zig_bitSizeOf_f32 32
@ -3088,15 +3094,15 @@ zig_float_from_repr(f128)
#define zig_convert_builtin(ResType, operation, ArgType, version) \
zig_extern ResType zig_expand_concat(zig_expand_concat(zig_expand_concat(__##operation, \
zig_compiler_rt_abbrev_##ArgType), zig_compiler_rt_abbrev_##ResType), version)(ArgType);
zig_convert_builtin(zig_f16, trunc, zig_f32, 2)
zig_convert_builtin(zig_f16, trunc, zig_f64, 2)
zig_convert_builtin(zig_compiler_rt_f16, trunc, zig_f32, 2)
zig_convert_builtin(zig_compiler_rt_f16, trunc, zig_f64, 2)
zig_convert_builtin(zig_f16, trunc, zig_f80, 2)
zig_convert_builtin(zig_f16, trunc, zig_f128, 2)
zig_convert_builtin(zig_f32, extend, zig_f16, 2)
zig_convert_builtin(zig_f32, extend, zig_compiler_rt_f16, 2)
zig_convert_builtin(zig_f32, trunc, zig_f64, 2)
zig_convert_builtin(zig_f32, trunc, zig_f80, 2)
zig_convert_builtin(zig_f32, trunc, zig_f128, 2)
zig_convert_builtin(zig_f64, extend, zig_f16, 2)
zig_convert_builtin(zig_f64, extend, zig_compiler_rt_f16, 2)
zig_convert_builtin(zig_f64, extend, zig_f32, 2)
zig_convert_builtin(zig_f64, trunc, zig_f80, 2)
zig_convert_builtin(zig_f64, trunc, zig_f128, 2)