Merge pull request #8542 from LemonBoy/floating-point-is-hard-my-dude

Floating point is hard my dude

lib/std/special/compiler_rt/extendXfYf2.zig

@@ -23,6 +23,10 @@ pub fn __extendhfsf2(a: u16) callconv(.C) f32 {
     return @call(.{ .modifier = .always_inline }, extendXfYf2, .{ f32, f16, a });
 }
 
+pub fn __extendhftf2(a: u16) callconv(.C) f128 {
+    return @call(.{ .modifier = .always_inline }, extendXfYf2, .{ f128, f16, a });
+}
+
 pub fn __aeabi_h2f(arg: u16) callconv(.AAPCS) f32 {
     @setRuntimeSafety(false);
     return @call(.{ .modifier = .always_inline }, __extendhfsf2, .{arg});

lib/std/special/compiler_rt/extendXfYf2_test.zig

@@ -4,9 +4,10 @@
 // The MIT license requires this copyright notice to be included in all copies
 // and substantial portions of the software.
 const builtin = @import("builtin");
-const __extenddftf2 = @import("extendXfYf2.zig").__extenddftf2;
 const __extendhfsf2 = @import("extendXfYf2.zig").__extendhfsf2;
+const __extendhftf2 = @import("extendXfYf2.zig").__extendhftf2;
 const __extendsftf2 = @import("extendXfYf2.zig").__extendsftf2;
+const __extenddftf2 = @import("extendXfYf2.zig").__extenddftf2;
 
 fn test__extenddftf2(a: f64, expectedHi: u64, expectedLo: u64) void {
     const x = __extenddftf2(a);
@@ -161,3 +162,49 @@ fn makeNaN32(rand: u32) f32 {
 fn makeInf32() f32 {
     return @bitCast(f32, @as(u32, 0x7f800000));
 }
+
+fn test__extendhftf2(a: u16, expectedHi: u64, expectedLo: u64) void {
+    const x = __extendhftf2(a);
+
+    const rep = @bitCast(u128, x);
+    const hi = @intCast(u64, rep >> 64);
+    const lo = @truncate(u64, rep);
+
+    if (hi == expectedHi and lo == expectedLo)
+        return;
+
+    // test other possible NaN representation(signal NaN)
+    if (expectedHi == 0x7fff800000000000 and expectedLo == 0x0) {
+        if ((hi & 0x7fff000000000000) == 0x7fff000000000000 and
+            ((hi & 0xffffffffffff) > 0 or lo > 0))
+        {
+            return;
+        }
+    }
+
+    @panic("__extendhftf2 test failure");
+}
+
+test "extendhftf2" {
+    // qNaN
+    test__extendhftf2(0x7e00, 0x7fff800000000000, 0x0);
+    // NaN
+    test__extendhftf2(0x7d00, 0x7fff400000000000, 0x0);
+    // inf
+    test__extendhftf2(0x7c00, 0x7fff000000000000, 0x0);
+    test__extendhftf2(0xfc00, 0xffff000000000000, 0x0);
+    // zero
+    test__extendhftf2(0x0000, 0x0000000000000000, 0x0);
+    test__extendhftf2(0x8000, 0x8000000000000000, 0x0);
+    // denormal
+    test__extendhftf2(0x0010, 0x3feb000000000000, 0x0);
+    test__extendhftf2(0x0001, 0x3fe7000000000000, 0x0);
+    test__extendhftf2(0x8001, 0xbfe7000000000000, 0x0);
+
+    // pi
+    test__extendhftf2(0x4248, 0x4000920000000000, 0x0);
+    test__extendhftf2(0xc248, 0xc000920000000000, 0x0);
+
+    test__extendhftf2(0x508c, 0x4004230000000000, 0x0);
+    test__extendhftf2(0x1bb7, 0x3ff6edc000000000, 0x0);
+}

lib/std/special/compiler_rt/truncXfYf2.zig

@@ -13,6 +13,10 @@ pub fn __truncdfhf2(a: f64) callconv(.C) u16 {
     return @bitCast(u16, @call(.{ .modifier = .always_inline }, truncXfYf2, .{ f16, f64, a }));
 }
 
+pub fn __trunctfhf2(a: f128) callconv(.C) u16 {
+    return @bitCast(u16, @call(.{ .modifier = .always_inline }, truncXfYf2, .{ f16, f128, a }));
+}
+
 pub fn __trunctfsf2(a: f128) callconv(.C) f32 {
     return @call(.{ .modifier = .always_inline }, truncXfYf2, .{ f32, f128, a });
 }
@@ -122,7 +126,7 @@ fn truncXfYf2(comptime dst_t: type, comptime src_t: type, a: src_t) dst_t {
         if (shift > srcSigBits) {
             absResult = 0;
         } else {
-            const sticky: src_rep_t = significand << @intCast(SrcShift, srcBits - shift);
+            const sticky: src_rep_t = @boolToInt(significand << @intCast(SrcShift, srcBits - shift) != 0);
             const denormalizedSignificand: src_rep_t = significand >> @intCast(SrcShift, shift) | sticky;
             absResult = @intCast(dst_rep_t, denormalizedSignificand >> (srcSigBits - dstSigBits));
             const roundBits: src_rep_t = denormalizedSignificand & roundMask;

lib/std/special/compiler_rt/truncXfYf2_test.zig

@@ -242,3 +242,59 @@ test "truncdfsf2" {
     // huge number becomes inf
     test__truncdfsf2(340282366920938463463374607431768211456.0, 0x7f800000);
 }
+
+const __trunctfhf2 = @import("truncXfYf2.zig").__trunctfhf2;
+
+fn test__trunctfhf2(a: f128, expected: u16) void {
+    const x = __trunctfhf2(a);
+
+    const rep = @bitCast(u16, x);
+    if (rep == expected) {
+        return;
+    }
+
+    @import("std").debug.warn("got 0x{x} wanted 0x{x}\n", .{ rep, expected });
+
+    @panic("__trunctfhf2 test failure");
+}
+
+test "trunctfhf2" {
+    // qNaN
+    test__trunctfhf2(@bitCast(f128, @as(u128, 0x7fff8000000000000000000000000000)), 0x7e00);
+    // NaN
+    test__trunctfhf2(@bitCast(f128, @as(u128, 0x7fff0000000000000000000000000001)), 0x7e00);
+    // inf
+    test__trunctfhf2(@bitCast(f128, @as(u128, 0x7fff0000000000000000000000000000)), 0x7c00);
+    test__trunctfhf2(-@bitCast(f128, @as(u128, 0x7fff0000000000000000000000000000)), 0xfc00);
+    // zero
+    test__trunctfhf2(0.0, 0x0);
+    test__trunctfhf2(-0.0, 0x8000);
+
+    test__trunctfhf2(3.1415926535, 0x4248);
+    test__trunctfhf2(-3.1415926535, 0xc248);
+    test__trunctfhf2(0x1.987124876876324p+100, 0x7c00);
+    test__trunctfhf2(0x1.987124876876324p+12, 0x6e62);
+    test__trunctfhf2(0x1.0p+0, 0x3c00);
+    test__trunctfhf2(0x1.0p-14, 0x0400);
+    // denormal
+    test__trunctfhf2(0x1.0p-20, 0x0010);
+    test__trunctfhf2(0x1.0p-24, 0x0001);
+    test__trunctfhf2(-0x1.0p-24, 0x8001);
+    test__trunctfhf2(0x1.5p-25, 0x0001);
+    // and back to zero
+    test__trunctfhf2(0x1.0p-25, 0x0000);
+    test__trunctfhf2(-0x1.0p-25, 0x8000);
+    // max (precise)
+    test__trunctfhf2(65504.0, 0x7bff);
+    // max (rounded)
+    test__trunctfhf2(65519.0, 0x7bff);
+    // max (to +inf)
+    test__trunctfhf2(65520.0, 0x7c00);
+    test__trunctfhf2(65536.0, 0x7c00);
+    test__trunctfhf2(-65520.0, 0xfc00);
+
+    test__trunctfhf2(0x1.23a2abb4a2ddee355f36789abcdep+5, 0x508f);
+    test__trunctfhf2(0x1.e3d3c45bd3abfd98b76a54cc321fp-9, 0x1b8f);
+    test__trunctfhf2(0x1.234eebb5faa678f4488693abcdefp+453, 0x7c00);
+    test__trunctfhf2(0x1.edcba9bb8c76a5a43dd21f334634p-43, 0x0);
+}

src/stage1/bigfloat.cpp

@@ -9,6 +9,7 @@
 #include "bigint.hpp"
 #include "buffer.hpp"
 #include "softfloat.hpp"
+#include "softfloat_ext.hpp"
 #include "parse_f128.h"
 #include <stdio.h>
 #include <math.h>
@@ -60,9 +61,7 @@ void bigfloat_init_bigint(BigFloat *dest, const BigInt *op) {
 
         if (i == 0) {
             if (op->is_negative) {
-                float128_t zero_f128;
-                ui32_to_f128M(0, &zero_f128);
-                f128M_sub(&zero_f128, &dest->value, &dest->value);
+                f128M_neg(&dest->value, &dest->value);
             }
             return;
         }
@@ -89,9 +88,7 @@ void bigfloat_add(BigFloat *dest, const BigFloat *op1, const BigFloat *op2) {
 }
 
 void bigfloat_negate(BigFloat *dest, const BigFloat *op) {
-    float128_t zero_f128;
-    ui32_to_f128M(0, &zero_f128);
-    f128M_sub(&zero_f128, &op->value, &dest->value);
+    f128M_neg(&op->value, &dest->value);
 }
 
 void bigfloat_sub(BigFloat *dest, const BigFloat *op1, const BigFloat *op2) {

src/stage1/codegen.cpp

@@ -7436,7 +7436,10 @@ static LLVMValueRef gen_const_val(CodeGen *g, ZigValue *const_val, const char *n
         case ZigTypeIdFloat:
             switch (type_entry->data.floating.bit_count) {
                 case 16:
-                    return LLVMConstReal(get_llvm_type(g, type_entry), zig_f16_to_double(const_val->data.x_f16));
+                    {
+                        LLVMValueRef as_int = LLVMConstInt(LLVMInt16Type(), const_val->data.x_f16.v, false);
+                        return LLVMConstBitCast(as_int, get_llvm_type(g, type_entry));
+                    }
                 case 32:
                     return LLVMConstReal(get_llvm_type(g, type_entry), const_val->data.x_f32);
                 case 64:

src/stage1/ir.cpp

@@ -11363,11 +11363,8 @@ static void float_negate(ZigValue *out_val, ZigValue *op) {
     } else if (op->type->id == ZigTypeIdFloat) {
         switch (op->type->data.floating.bit_count) {
             case 16:
-                {
-                    const float16_t zero = zig_double_to_f16(0);
-                    out_val->data.x_f16 = f16_sub(zero, op->data.x_f16);
-                    return;
-                }
+                out_val->data.x_f16 = f16_neg(op->data.x_f16);
+                return;
             case 32:
                 out_val->data.x_f32 = -op->data.x_f32;
                 return;
@@ -11375,9 +11372,7 @@ static void float_negate(ZigValue *out_val, ZigValue *op) {
                 out_val->data.x_f64 = -op->data.x_f64;
                 return;
             case 128:
-                float128_t zero_f128;
-                ui32_to_f128M(0, &zero_f128);
-                f128M_sub(&zero_f128, &op->data.x_f128, &out_val->data.x_f128);
+                f128M_neg(&op->data.x_f128, &out_val->data.x_f128);
                 return;
             default:
                 zig_unreachable();

src/stage1/softfloat_ext.cpp

@@ -1,6 +1,8 @@
 #include "softfloat_ext.hpp"
 
 extern "C" {
+    #include "platform.h"
+    #include "internals.h"
     #include "softfloat.h"
 }
 
@@ -22,4 +24,15 @@ void f128M_trunc(const float128_t *aPtr, float128_t *zPtr) {
     } else {
         f128M_roundToInt(aPtr, softfloat_round_min, false, zPtr);
     } 
+}
+
+float16_t f16_neg(const float16_t a) {
+    union ui16_f16 uZ;
+    uZ.ui = a.v ^ (UINT16_C(1) << 15);
+    return uZ.f;
+}
+
+void f128M_neg(const float128_t *aPtr, float128_t *zPtr) {
+    zPtr->v[indexWord(2,1)] = aPtr->v[indexWord(2,1)] ^ (UINT64_C(1) << 63);
+    zPtr->v[indexWord(2,0)] = aPtr->v[indexWord(2,0)];
 }

src/stage1/softfloat_ext.hpp

@@ -5,5 +5,8 @@
 
 void f128M_abs(const float128_t *aPtr, float128_t *zPtr);
 void f128M_trunc(const float128_t *aPtr, float128_t *zPtr);
+void f128M_neg(const float128_t *aPtr, float128_t *zPtr);
+
+float16_t f16_neg(const float16_t a);
 
 #endif

test/stage1/behavior/math.zig

@@ -843,3 +843,20 @@ test "compare undefined literal with comptime_int" {
     x = true;
     expect(x);
 }
+
+test "signed zeros are represented properly" {
+    const S = struct {
+        fn doTheTest() void {
+            inline for ([_]type{ f16, f32, f64, f128 }) |T| {
+                const ST = std.meta.Int(.unsigned, @typeInfo(T).Float.bits);
+                var as_fp_val = -@as(T, 0.0);
+                var as_uint_val = @bitCast(ST, as_fp_val);
+                // Ensure the sign bit is set.
+                expect(as_uint_val >> (@typeInfo(T).Float.bits - 1) == 1);
+            }
+        }
+    };
+
+    S.doTheTest();
+    comptime S.doTheTest();
+}