Commit 09bd4a9a86

Andrew Kelley <superjoe30@gmail.com>
2017-08-20 06:33:05
compile-time f32, f64 operations are now correctly lossy
previously we used the bigfloat abstraction to do all compile-time float math. but runtime code and comptime code are supposed to get the same result. so now if you add a f32 to a f32 at compile time it does it with f32 math instead of the bigfloat. float literals still get the bigfloat math. closes #424
1 parent c73a0c9
src/all_types.hpp
@@ -225,6 +225,9 @@ struct ConstExprValue {
         // populated if special == ConstValSpecialStatic
         BigInt x_bigint;
         BigFloat x_bigfloat;
+        float x_f32;
+        double x_f64;
+        __float128 x_f128;
         bool x_bool;
         ConstFn x_fn;
         ConstBoundFnValue x_bound_fn;
src/analyze.cpp
@@ -13,6 +13,7 @@
 #include "ir_print.hpp"
 #include "os.hpp"
 #include "parser.hpp"
+#include "quadmath.hpp"
 #include "zig_llvm.hpp"
 
 static const size_t default_backward_branch_quota = 1000;
@@ -3273,8 +3274,35 @@ static uint32_t hash_const_val(ConstExprValue *const_val) {
                 return result;
             }
         case TypeTableEntryIdFloat:
+            switch (const_val->type->data.floating.bit_count) {
+                case 32:
+                    {
+                        uint32_t result;
+                        memcpy(&result, &const_val->data.x_f32, 4);
+                        return result ^ 4084870010;
+                    }
+                case 64:
+                    {
+                        uint32_t ints[2];
+                        memcpy(&ints[0], &const_val->data.x_f64, 8);
+                        return ints[0] ^ ints[1] ^ 0x22ed43c6;
+                    }
+                case 128:
+                    {
+                        uint32_t ints[4];
+                        memcpy(&ints[0], &const_val->data.x_f128, 16);
+                        return ints[0] ^ ints[1] ^ ints[2] ^ ints[3] ^ 0xb5ffef27;
+                    }
+                default:
+                    zig_unreachable();
+            }
         case TypeTableEntryIdNumLitFloat:
-            return (uint32_t)(const_val->data.x_bigfloat.value * (uint32_t)UINT32_MAX);
+            {
+                __float128 f128 = bigfloat_to_f128(&const_val->data.x_bigfloat);
+                uint32_t ints[4];
+                memcpy(&ints[0], &f128, 16);
+                return ints[0] ^ ints[1] ^ ints[2] ^ ints[3] ^ 0xed8b3dfb;
+            }
         case TypeTableEntryIdArgTuple:
             return (uint32_t)const_val->data.x_arg_tuple.start_index * (uint32_t)281907309 +
                 (uint32_t)const_val->data.x_arg_tuple.end_index * (uint32_t)2290442768;
@@ -3575,7 +3603,25 @@ ConstExprValue *create_const_signed(TypeTableEntry *type, int64_t x) {
 void init_const_float(ConstExprValue *const_val, TypeTableEntry *type, double value) {
     const_val->special = ConstValSpecialStatic;
     const_val->type = type;
-    bigfloat_init_float(&const_val->data.x_bigfloat, value);
+    if (type->id == TypeTableEntryIdNumLitFloat) {
+        bigfloat_init_64(&const_val->data.x_bigfloat, value);
+    } else if (type->id == TypeTableEntryIdFloat) {
+        switch (type->data.floating.bit_count) {
+            case 32:
+                const_val->data.x_f32 = value;
+                break;
+            case 64:
+                const_val->data.x_f64 = value;
+                break;
+            case 128:
+                // if we need this, we should add a function that accepts a __float128 param
+                zig_unreachable();
+            default:
+                zig_unreachable();
+        }
+    } else {
+        zig_unreachable();
+    }
 }
 
 ConstExprValue *create_const_float(TypeTableEntry *type, double value) {
@@ -3816,6 +3862,17 @@ bool const_values_equal(ConstExprValue *a, ConstExprValue *b) {
         case TypeTableEntryIdBool:
             return a->data.x_bool == b->data.x_bool;
         case TypeTableEntryIdFloat:
+            assert(a->type->data.floating.bit_count == b->type->data.floating.bit_count);
+            switch (a->type->data.floating.bit_count) {
+                case 32:
+                    return a->data.x_f32 == b->data.x_f32;
+                case 64:
+                    return a->data.x_f64 == b->data.x_f64;
+                case 128:
+                    return a->data.x_f128 == b->data.x_f128;
+                default:
+                    zig_unreachable();
+            }
         case TypeTableEntryIdNumLitFloat:
             return bigfloat_cmp(&a->data.x_bigfloat, &b->data.x_bigfloat) == CmpEQ;
         case TypeTableEntryIdInt:
@@ -3984,12 +4041,32 @@ void render_const_value(CodeGen *g, Buf *buf, ConstExprValue *const_val) {
             buf_appendf(buf, "{}");
             return;
         case TypeTableEntryIdNumLitFloat:
-        case TypeTableEntryIdFloat:
-            bigfloat_write_buf(buf, &const_val->data.x_bigfloat);
+            bigfloat_append_buf(buf, &const_val->data.x_bigfloat);
             return;
+        case TypeTableEntryIdFloat:
+            switch (type_entry->data.floating.bit_count) {
+                case 32:
+                    buf_appendf(buf, "%f", const_val->data.x_f32);
+                    return;
+                case 64:
+                    buf_appendf(buf, "%f", const_val->data.x_f64);
+                    return;
+                case 128:
+                    {
+                        const size_t extra_len = 100;
+                        size_t old_len = buf_len(buf);
+                        buf_resize(buf, old_len + extra_len);
+                        int len = quadmath_snprintf(buf_ptr(buf) + old_len, extra_len, "%Qf", const_val->data.x_f128);
+                        assert(len > 0);
+                        buf_resize(buf, old_len + len);
+                        return;
+                    }
+                default:
+                    zig_unreachable();
+            }
         case TypeTableEntryIdNumLitInt:
         case TypeTableEntryIdInt:
-            bigint_write_buf(buf, &const_val->data.x_bigint, 10);
+            bigint_append_buf(buf, &const_val->data.x_bigint, 10);
             return;
         case TypeTableEntryIdMetaType:
             buf_appendf(buf, "%s", buf_ptr(&const_val->data.x_type->name));
src/ast_render.cpp
@@ -540,7 +540,7 @@ static void render_node_extra(AstRender *ar, AstNode *node, bool grouped) {
             {
                 Buf rendered_buf = BUF_INIT;
                 buf_resize(&rendered_buf, 0);
-                bigfloat_write_buf(&rendered_buf, node->data.float_literal.bigfloat);
+                bigfloat_append_buf(&rendered_buf, node->data.float_literal.bigfloat);
                 fprintf(ar->f, "%s", buf_ptr(&rendered_buf));
             }
             break;
@@ -548,7 +548,7 @@ static void render_node_extra(AstRender *ar, AstNode *node, bool grouped) {
             {
                 Buf rendered_buf = BUF_INIT;
                 buf_resize(&rendered_buf, 0);
-                bigint_write_buf(&rendered_buf, node->data.int_literal.bigint, 10);
+                bigint_append_buf(&rendered_buf, node->data.int_literal.bigint, 10);
                 fprintf(ar->f, "%s", buf_ptr(&rendered_buf));
             }
             break;
src/bigfloat.cpp
@@ -8,19 +8,19 @@
 #include "bigfloat.hpp"
 #include "bigint.hpp"
 #include "buffer.hpp"
+#include "quadmath.hpp"
 #include <math.h>
 #include <errno.h>
 
-extern "C" {
-    __float128 fmodq(__float128 a, __float128 b);
-    __float128 ceilq(__float128 a);
-    __float128 floorq(__float128 a);
-    __float128 strtoflt128 (const char *s, char **sp);
-    int quadmath_snprintf (char *s, size_t size, const char *format, ...);
+void bigfloat_init_128(BigFloat *dest, __float128 x) {
+    dest->value = x;
 }
 
+void bigfloat_init_32(BigFloat *dest, float x) {
+    dest->value = x;
+}
 
-void bigfloat_init_float(BigFloat *dest, __float128 x) {
+void bigfloat_init_64(BigFloat *dest, double x) {
     dest->value = x;
 }
 
@@ -104,11 +104,13 @@ void bigfloat_mod(BigFloat *dest, const BigFloat *op1, const BigFloat *op2) {
     dest->value = fmodq(fmodq(op1->value, op2->value) + op2->value, op2->value);
 }
 
-void bigfloat_write_buf(Buf *buf, const BigFloat *op) {
-    buf_resize(buf, 256);
-    int len = quadmath_snprintf(buf_ptr(buf), buf_len(buf), "%Qf", op->value);
+void bigfloat_append_buf(Buf *buf, const BigFloat *op) {
+    const size_t extra_len = 100;
+    size_t old_len = buf_len(buf);
+    buf_resize(buf, old_len + extra_len);
+    int len = quadmath_snprintf(buf_ptr(buf) + old_len, extra_len, "%Qf", op->value);
     assert(len > 0);
-    buf_resize(buf, len);
+    buf_resize(buf, old_len + len);
 }
 
 Cmp bigfloat_cmp(const BigFloat *op1, const BigFloat *op2) {
@@ -121,42 +123,15 @@ Cmp bigfloat_cmp(const BigFloat *op1, const BigFloat *op2) {
     }
 }
 
-// TODO this is wrong when compiler running on big endian systems. caught by tests
-void bigfloat_write_ieee597(const BigFloat *op, uint8_t *buf, size_t bit_count, bool is_big_endian) {
-    if (bit_count == 32) {
-        float f32 = op->value;
-        memcpy(buf, &f32, 4);
-    } else if (bit_count == 64) {
-        double f64 = op->value;
-        memcpy(buf, &f64, 8);
-    } else if (bit_count == 128) {
-        __float128 f128 = op->value;
-        memcpy(buf, &f128, 16);
-    } else {
-        zig_unreachable();
-    }
+float bigfloat_to_f32(const BigFloat *bigfloat) {
+    return (float)bigfloat->value;
 }
 
-// TODO this is wrong when compiler running on big endian systems. caught by tests
-void bigfloat_read_ieee597(BigFloat *dest, const uint8_t *buf, size_t bit_count, bool is_big_endian) {
-    if (bit_count == 32) {
-        float f32;
-        memcpy(&f32, buf, 4);
-        dest->value = f32;
-    } else if (bit_count == 64) {
-        double f64;
-        memcpy(&f64, buf, 8);
-        dest->value = f64;
-    } else if (bit_count == 128) {
-        __float128 f128;
-        memcpy(&f128, buf, 16);
-        dest->value = f128;
-    } else {
-        zig_unreachable();
-    }
+double bigfloat_to_f64(const BigFloat *bigfloat) {
+    return (double)bigfloat->value;
 }
 
-double bigfloat_to_double(const BigFloat *bigfloat) {
+__float128 bigfloat_to_f128(const BigFloat *bigfloat) {
     return bigfloat->value;
 }
 
src/bigfloat.hpp
@@ -19,12 +19,16 @@ struct BigFloat {
 
 struct Buf;
 
-void bigfloat_init_float(BigFloat *dest, __float128 x);
+void bigfloat_init_32(BigFloat *dest, float x);
+void bigfloat_init_64(BigFloat *dest, double x);
+void bigfloat_init_128(BigFloat *dest, __float128 x);
 void bigfloat_init_bigfloat(BigFloat *dest, const BigFloat *x);
 void bigfloat_init_bigint(BigFloat *dest, const BigInt *op);
 int bigfloat_init_buf_base10(BigFloat *dest, const uint8_t *buf_ptr, size_t buf_len);
 
-double bigfloat_to_double(const BigFloat *bigfloat);
+float bigfloat_to_f32(const BigFloat *bigfloat);
+double bigfloat_to_f64(const BigFloat *bigfloat);
+__float128 bigfloat_to_f128(const BigFloat *bigfloat);
 
 void bigfloat_add(BigFloat *dest, const BigFloat *op1, const BigFloat *op2);
 void bigfloat_negate(BigFloat *dest, const BigFloat *op);
@@ -35,10 +39,8 @@ void bigfloat_div_trunc(BigFloat *dest, const BigFloat *op1, const BigFloat *op2
 void bigfloat_div_floor(BigFloat *dest, const BigFloat *op1, const BigFloat *op2);
 void bigfloat_rem(BigFloat *dest, const BigFloat *op1, const BigFloat *op2);
 void bigfloat_mod(BigFloat *dest, const BigFloat *op1, const BigFloat *op2);
-void bigfloat_write_buf(Buf *buf, const BigFloat *op);
+void bigfloat_append_buf(Buf *buf, const BigFloat *op);
 Cmp bigfloat_cmp(const BigFloat *op1, const BigFloat *op2);
-void bigfloat_write_ieee597(const BigFloat *op, uint8_t *buf, size_t bit_count, bool is_big_endian);
-void bigfloat_read_ieee597(BigFloat *dest, const uint8_t *buf, size_t bit_count, bool is_big_endian);
 
 
 // convenience functions
src/bigint.cpp
@@ -141,6 +141,21 @@ void bigint_init_unsigned(BigInt *dest, uint64_t x) {
     dest->is_negative = false;
 }
 
+void bigint_init_u128(BigInt *dest, unsigned __int128 x) {
+    uint64_t low = (uint64_t)(x & UINT64_MAX);
+    uint64_t high = (uint64_t)(x >> 64);
+
+    if (high == 0) {
+        return bigint_init_unsigned(dest, low);
+    }
+
+    dest->digit_count = 2;
+    dest->data.digits = allocate_nonzero<uint64_t>(2);
+    dest->data.digits[0] = low;
+    dest->data.digits[1] = high;
+    dest->is_negative = false;
+}
+
 void bigint_init_signed(BigInt *dest, int64_t x) {
     if (x >= 0) {
         return bigint_init_unsigned(dest, x);
@@ -167,9 +182,9 @@ void bigint_init_bigint(BigInt *dest, const BigInt *src) {
 
 void bigint_init_bigfloat(BigInt *dest, const BigFloat *op) {
     if (op->value >= 0) {
-        bigint_init_unsigned(dest, op->value);
+        bigint_init_u128(dest, (unsigned __int128)(op->value));
     } else {
-        bigint_init_unsigned(dest, -op->value);
+        bigint_init_u128(dest, (unsigned __int128)(-op->value));
         dest->is_negative = true;
     }
 }
@@ -1023,7 +1038,7 @@ Cmp bigint_cmp(const BigInt *op1, const BigInt *op2) {
     }
 }
 
-void bigint_write_buf(Buf *buf, const BigInt *op, uint64_t base) {
+void bigint_append_buf(Buf *buf, const BigInt *op, uint64_t base) {
     if (op->digit_count == 0) {
         buf_append_char(buf, '0');
         return;
src/bigint.hpp
@@ -30,6 +30,7 @@ enum Cmp {
 };
 
 void bigint_init_unsigned(BigInt *dest, uint64_t x);
+void bigint_init_u128(BigInt *dest, unsigned __int128 x);
 void bigint_init_signed(BigInt *dest, int64_t x);
 void bigint_init_bigint(BigInt *dest, const BigInt *src);
 void bigint_init_bigfloat(BigInt *dest, const BigFloat *op);
@@ -76,7 +77,7 @@ void bigint_truncate(BigInt *dest, const BigInt *op, size_t bit_count, bool is_s
 
 Cmp bigint_cmp(const BigInt *op1, const BigInt *op2);
 
-void bigint_write_buf(Buf *buf, const BigInt *op, uint64_t base);
+void bigint_append_buf(Buf *buf, const BigInt *op, uint64_t base);
 
 size_t bigint_ctz(const BigInt *bi, size_t bit_count);
 size_t bigint_clz(const BigInt *bi, size_t bit_count);
src/codegen.cpp
@@ -1243,10 +1243,6 @@ static LLVMValueRef bigint_to_llvm_const(LLVMTypeRef type_ref, BigInt *bigint) {
     }
 }
 
-static LLVMValueRef bigfloat_to_llvm_const(LLVMTypeRef type_ref, BigFloat *bigfloat) {
-    return LLVMConstReal(type_ref, bigfloat_to_double(bigfloat));
-}
-
 static LLVMValueRef gen_div(CodeGen *g, bool want_debug_safety, bool want_fast_math,
         LLVMValueRef val1, LLVMValueRef val2,
         TypeTableEntry *type_entry, DivKind div_kind)
@@ -3455,7 +3451,23 @@ static LLVMValueRef gen_const_val(CodeGen *g, ConstExprValue *const_val) {
             return LLVMConstInt(g->builtin_types.entry_pure_error->type_ref,
                     const_val->data.x_pure_err->value, false);
         case TypeTableEntryIdFloat:
-            return bigfloat_to_llvm_const(type_entry->type_ref, &const_val->data.x_bigfloat);
+            switch (type_entry->data.floating.bit_count) {
+                case 32:
+                    return LLVMConstReal(type_entry->type_ref, const_val->data.x_f32);
+                case 64:
+                    return LLVMConstReal(type_entry->type_ref, const_val->data.x_f64);
+                case 128:
+                    {
+                        // TODO make sure this is correct on big endian targets too
+                        uint8_t buf[16];
+                        memcpy(buf, &const_val->data.x_f128, 16);
+                        LLVMValueRef as_int = LLVMConstIntOfArbitraryPrecision(LLVMInt128Type(), 2,
+                                (uint64_t*)buf);
+                        return LLVMConstBitCast(as_int, type_entry->type_ref);
+                    }
+                default:
+                    zig_unreachable();
+            }
         case TypeTableEntryIdBool:
             if (const_val->data.x_bool) {
                 return LLVMConstAllOnes(LLVMInt1Type());
@@ -3937,8 +3949,11 @@ static void do_code_gen(CodeGen *g) {
             // Generate debug info for it but that's it.
             ConstExprValue *const_val = var->value;
             assert(const_val->special != ConstValSpecialRuntime);
-            TypeTableEntry *var_type = g->builtin_types.entry_f64;
-            LLVMValueRef init_val = bigfloat_to_llvm_const(var_type->type_ref, &const_val->data.x_bigfloat);
+            TypeTableEntry *var_type = g->builtin_types.entry_f128;
+            ConstExprValue coerced_value;
+            coerced_value.type = var_type;
+            coerced_value.data.x_f128 = bigfloat_to_f128(&const_val->data.x_bigfloat);
+            LLVMValueRef init_val = gen_const_val(g, &coerced_value);
             gen_global_var(g, var, init_val, var_type);
             continue;
         }
src/ir.cpp
@@ -12,6 +12,7 @@
 #include "ir_print.hpp"
 #include "os.hpp"
 #include "parseh.hpp"
+#include "quadmath.hpp"
 #include "range_set.hpp"
 
 struct IrExecContext {
@@ -6272,6 +6273,546 @@ static bool const_val_fits_in_num_lit(ConstExprValue *const_val, TypeTableEntry
         (const_val->type->id == TypeTableEntryIdInt || const_val->type->id == TypeTableEntryIdNumLitInt)));
 }
 
+static bool float_has_fraction(ConstExprValue *const_val) {
+    if (const_val->type->id == TypeTableEntryIdNumLitFloat) {
+        return bigfloat_has_fraction(&const_val->data.x_bigfloat);
+    } else if (const_val->type->id == TypeTableEntryIdFloat) {
+        switch (const_val->type->data.floating.bit_count) {
+            case 32:
+                return floorf(const_val->data.x_f32) != const_val->data.x_f32;
+            case 64:
+                return floor(const_val->data.x_f64) != const_val->data.x_f64;
+            case 128:
+                return floorq(const_val->data.x_f128) != const_val->data.x_f128;
+            default:
+                zig_unreachable();
+        }
+    } else {
+        zig_unreachable();
+    }
+}
+
+static void float_append_buf(Buf *buf, ConstExprValue *const_val) {
+    if (const_val->type->id == TypeTableEntryIdNumLitFloat) {
+        bigfloat_append_buf(buf, &const_val->data.x_bigfloat);
+    } else if (const_val->type->id == TypeTableEntryIdFloat) {
+        switch (const_val->type->data.floating.bit_count) {
+            case 32:
+                buf_appendf(buf, "%f", const_val->data.x_f32);
+                break;
+            case 64:
+                buf_appendf(buf, "%f", const_val->data.x_f64);
+                break;
+            case 128:
+                {
+                    const size_t extra_len = 100;
+                    size_t old_len = buf_len(buf);
+                    buf_resize(buf, old_len + extra_len);
+                    int len = quadmath_snprintf(buf_ptr(buf) + old_len, extra_len, "%Qf", const_val->data.x_f128);
+                    assert(len > 0);
+                    buf_resize(buf, old_len + len);
+                    break;
+                }
+            default:
+                zig_unreachable();
+        }
+    } else {
+        zig_unreachable();
+    }
+}
+
+static void float_init_bigint(BigInt *bigint, ConstExprValue *const_val) {
+    if (const_val->type->id == TypeTableEntryIdNumLitFloat) {
+        bigint_init_bigfloat(bigint, &const_val->data.x_bigfloat);
+    } else if (const_val->type->id == TypeTableEntryIdFloat) {
+        switch (const_val->type->data.floating.bit_count) {
+            case 32:
+                if (const_val->data.x_f32 >= 0) {
+                    bigint_init_unsigned(bigint, (uint64_t)(const_val->data.x_f32));
+                } else {
+                    bigint_init_unsigned(bigint, (uint64_t)(-const_val->data.x_f32));
+                    bigint->is_negative = true;
+                }
+                break;
+            case 64:
+                if (const_val->data.x_f64 >= 0) {
+                    bigint_init_unsigned(bigint, (uint64_t)(const_val->data.x_f64));
+                } else {
+                    bigint_init_unsigned(bigint, (uint64_t)(-const_val->data.x_f64));
+                    bigint->is_negative = true;
+                }
+                break;
+            case 128:
+                if (const_val->data.x_f128 >= 0) {
+                    bigint_init_u128(bigint, (unsigned __int128)(const_val->data.x_f128));
+                } else {
+                    bigint_init_u128(bigint, (unsigned __int128)(-const_val->data.x_f128));
+                    bigint->is_negative = true;
+                }
+                break;
+            default:
+                zig_unreachable();
+        }
+    } else {
+        zig_unreachable();
+    }
+}
+
+static void float_init_bigfloat(ConstExprValue *dest_val, BigFloat *bigfloat) {
+    if (dest_val->type->id == TypeTableEntryIdNumLitFloat) {
+        bigfloat_init_bigfloat(&dest_val->data.x_bigfloat, bigfloat);
+    } else if (dest_val->type->id == TypeTableEntryIdFloat) {
+        switch (dest_val->type->data.floating.bit_count) {
+            case 32:
+                dest_val->data.x_f32 = bigfloat_to_f32(bigfloat);
+                break;
+            case 64:
+                dest_val->data.x_f64 = bigfloat_to_f64(bigfloat);
+                break;
+            case 128:
+                dest_val->data.x_f128 = bigfloat_to_f128(bigfloat);
+                break;
+            default:
+                zig_unreachable();
+        }
+    } else {
+        zig_unreachable();
+    }
+}
+
+static void float_init_f32(ConstExprValue *dest_val, float x) {
+    if (dest_val->type->id == TypeTableEntryIdNumLitFloat) {
+        bigfloat_init_32(&dest_val->data.x_bigfloat, x);
+    } else if (dest_val->type->id == TypeTableEntryIdFloat) {
+        switch (dest_val->type->data.floating.bit_count) {
+            case 32:
+                dest_val->data.x_f32 = x;
+                break;
+            case 64:
+                dest_val->data.x_f64 = x;
+                break;
+            case 128:
+                dest_val->data.x_f128 = x;
+                break;
+            default:
+                zig_unreachable();
+        }
+    } else {
+        zig_unreachable();
+    }
+}
+
+static void float_init_f64(ConstExprValue *dest_val, double x) {
+    if (dest_val->type->id == TypeTableEntryIdNumLitFloat) {
+        bigfloat_init_64(&dest_val->data.x_bigfloat, x);
+    } else if (dest_val->type->id == TypeTableEntryIdFloat) {
+        switch (dest_val->type->data.floating.bit_count) {
+            case 32:
+                dest_val->data.x_f32 = x;
+                break;
+            case 64:
+                dest_val->data.x_f64 = x;
+                break;
+            case 128:
+                dest_val->data.x_f128 = x;
+                break;
+            default:
+                zig_unreachable();
+        }
+    } else {
+        zig_unreachable();
+    }
+}
+
+static void float_init_f128(ConstExprValue *dest_val, __float128 x) {
+    if (dest_val->type->id == TypeTableEntryIdNumLitFloat) {
+        bigfloat_init_128(&dest_val->data.x_bigfloat, x);
+    } else if (dest_val->type->id == TypeTableEntryIdFloat) {
+        switch (dest_val->type->data.floating.bit_count) {
+            case 32:
+                dest_val->data.x_f32 = x;
+                break;
+            case 64:
+                dest_val->data.x_f64 = x;
+                break;
+            case 128:
+                dest_val->data.x_f128 = x;
+                break;
+            default:
+                zig_unreachable();
+        }
+    } else {
+        zig_unreachable();
+    }
+}
+
+static void float_init_float(ConstExprValue *dest_val, ConstExprValue *src_val) {
+    if (src_val->type->id == TypeTableEntryIdNumLitFloat) {
+        float_init_bigfloat(dest_val, &src_val->data.x_bigfloat);
+    } else if (src_val->type->id == TypeTableEntryIdFloat) {
+        switch (src_val->type->data.floating.bit_count) {
+            case 32:
+                float_init_f32(dest_val, src_val->data.x_f32);
+                break;
+            case 64:
+                float_init_f64(dest_val, src_val->data.x_f64);
+                break;
+            case 128:
+                float_init_f128(dest_val, src_val->data.x_f128);
+                break;
+            default:
+                zig_unreachable();
+        }
+    } else {
+        zig_unreachable();
+    }
+}
+
+static Cmp float_cmp(ConstExprValue *op1, ConstExprValue *op2) {
+    assert(op1->type == op2->type);
+    if (op1->type->id == TypeTableEntryIdNumLitFloat) {
+        return bigfloat_cmp(&op1->data.x_bigfloat, &op2->data.x_bigfloat);
+    } else if (op1->type->id == TypeTableEntryIdFloat) {
+        switch (op1->type->data.floating.bit_count) {
+            case 32:
+                if (op1->data.x_f32 > op2->data.x_f32) {
+                    return CmpGT;
+                } else if (op1->data.x_f32 < op2->data.x_f32) {
+                    return CmpLT;
+                } else {
+                    return CmpEQ;
+                }
+            case 64:
+                if (op1->data.x_f64 > op2->data.x_f64) {
+                    return CmpGT;
+                } else if (op1->data.x_f64 < op2->data.x_f64) {
+                    return CmpLT;
+                } else {
+                    return CmpEQ;
+                }
+            case 128:
+                if (op1->data.x_f128 > op2->data.x_f128) {
+                    return CmpGT;
+                } else if (op1->data.x_f128 < op2->data.x_f128) {
+                    return CmpLT;
+                } else {
+                    return CmpEQ;
+                }
+            default:
+                zig_unreachable();
+        }
+    } else {
+        zig_unreachable();
+    }
+}
+
+static Cmp float_cmp_zero(ConstExprValue *op) {
+    if (op->type->id == TypeTableEntryIdNumLitFloat) {
+        return bigfloat_cmp_zero(&op->data.x_bigfloat);
+    } else if (op->type->id == TypeTableEntryIdFloat) {
+        switch (op->type->data.floating.bit_count) {
+            case 32:
+                if (op->data.x_f32 < 0.0) {
+                    return CmpLT;
+                } else if (op->data.x_f32 > 0.0) {
+                    return CmpGT;
+                } else {
+                    return CmpEQ;
+                }
+            case 64:
+                if (op->data.x_f64 < 0.0) {
+                    return CmpLT;
+                } else if (op->data.x_f64 > 0.0) {
+                    return CmpGT;
+                } else {
+                    return CmpEQ;
+                }
+            case 128:
+                if (op->data.x_f128 < 0.0) {
+                    return CmpLT;
+                } else if (op->data.x_f128 > 0.0) {
+                    return CmpGT;
+                } else {
+                    return CmpEQ;
+                }
+            default:
+                zig_unreachable();
+        }
+    } else {
+        zig_unreachable();
+    }
+}
+
+static void float_add(ConstExprValue *out_val, ConstExprValue *op1, ConstExprValue *op2) {
+    assert(op1->type == op2->type);
+    out_val->type = op1->type;
+    if (op1->type->id == TypeTableEntryIdNumLitFloat) {
+        bigfloat_add(&out_val->data.x_bigfloat, &op1->data.x_bigfloat, &op2->data.x_bigfloat);
+    } else if (op1->type->id == TypeTableEntryIdFloat) {
+        switch (op1->type->data.floating.bit_count) {
+            case 32:
+                out_val->data.x_f32 =  op1->data.x_f32 + op2->data.x_f32;
+                return;
+            case 64:
+                out_val->data.x_f64 =  op1->data.x_f64 + op2->data.x_f64;
+                return;
+            case 128:
+                out_val->data.x_f128 =  op1->data.x_f128 + op2->data.x_f128;
+                return;
+            default:
+                zig_unreachable();
+        }
+    } else {
+        zig_unreachable();
+    }
+}
+
+static void float_sub(ConstExprValue *out_val, ConstExprValue *op1, ConstExprValue *op2) {
+    assert(op1->type == op2->type);
+    out_val->type = op1->type;
+    if (op1->type->id == TypeTableEntryIdNumLitFloat) {
+        bigfloat_sub(&out_val->data.x_bigfloat, &op1->data.x_bigfloat, &op2->data.x_bigfloat);
+    } else if (op1->type->id == TypeTableEntryIdFloat) {
+        switch (op1->type->data.floating.bit_count) {
+            case 32:
+                out_val->data.x_f32 = op1->data.x_f32 - op2->data.x_f32;
+                return;
+            case 64:
+                out_val->data.x_f64 = op1->data.x_f64 - op2->data.x_f64;
+                return;
+            case 128:
+                out_val->data.x_f128 = op1->data.x_f128 - op2->data.x_f128;
+                return;
+            default:
+                zig_unreachable();
+        }
+    } else {
+        zig_unreachable();
+    }
+}
+
+static void float_mul(ConstExprValue *out_val, ConstExprValue *op1, ConstExprValue *op2) {
+    assert(op1->type == op2->type);
+    out_val->type = op1->type;
+    if (op1->type->id == TypeTableEntryIdNumLitFloat) {
+        bigfloat_mul(&out_val->data.x_bigfloat, &op1->data.x_bigfloat, &op2->data.x_bigfloat);
+    } else if (op1->type->id == TypeTableEntryIdFloat) {
+        switch (op1->type->data.floating.bit_count) {
+            case 32:
+                out_val->data.x_f32 = op1->data.x_f32 * op2->data.x_f32;
+                return;
+            case 64:
+                out_val->data.x_f64 = op1->data.x_f64 * op2->data.x_f64;
+                return;
+            case 128:
+                out_val->data.x_f128 = op1->data.x_f128 * op2->data.x_f128;
+                return;
+            default:
+                zig_unreachable();
+        }
+    } else {
+        zig_unreachable();
+    }
+}
+
+static void float_div(ConstExprValue *out_val, ConstExprValue *op1, ConstExprValue *op2) {
+    assert(op1->type == op2->type);
+    out_val->type = op1->type;
+    if (op1->type->id == TypeTableEntryIdNumLitFloat) {
+        bigfloat_div(&out_val->data.x_bigfloat, &op1->data.x_bigfloat, &op2->data.x_bigfloat);
+    } else if (op1->type->id == TypeTableEntryIdFloat) {
+        switch (op1->type->data.floating.bit_count) {
+            case 32:
+                out_val->data.x_f32 = op1->data.x_f32 / op2->data.x_f32;
+                return;
+            case 64:
+                out_val->data.x_f64 = op1->data.x_f64 / op2->data.x_f64;
+                return;
+            case 128:
+                out_val->data.x_f128 = op1->data.x_f128 / op2->data.x_f128;
+                return;
+            default:
+                zig_unreachable();
+        }
+    } else {
+        zig_unreachable();
+    }
+}
+
+static void float_div_trunc(ConstExprValue *out_val, ConstExprValue *op1, ConstExprValue *op2) {
+    assert(op1->type == op2->type);
+    out_val->type = op1->type;
+    if (op1->type->id == TypeTableEntryIdNumLitFloat) {
+        bigfloat_div_trunc(&out_val->data.x_bigfloat, &op1->data.x_bigfloat, &op2->data.x_bigfloat);
+    } else if (op1->type->id == TypeTableEntryIdFloat) {
+        switch (op1->type->data.floating.bit_count) {
+            case 32:
+                out_val->data.x_f32 = op1->data.x_f32 / op2->data.x_f32;
+                if (out_val->data.x_f32 >= 0.0) {
+                    out_val->data.x_f32 = floorf(out_val->data.x_f32);
+                } else {
+                    out_val->data.x_f32 = ceilf(out_val->data.x_f32);
+                }
+                return;
+            case 64:
+                out_val->data.x_f64 = op1->data.x_f64 / op2->data.x_f64;
+                if (out_val->data.x_f64 >= 0.0) {
+                    out_val->data.x_f64 = floor(out_val->data.x_f64);
+                } else {
+                    out_val->data.x_f64 = ceil(out_val->data.x_f64);
+                }
+                return;
+            case 128:
+                out_val->data.x_f128 = op1->data.x_f128 / op2->data.x_f128;
+                if (out_val->data.x_f128 >= 0.0) {
+                    out_val->data.x_f128 = floorq(out_val->data.x_f128);
+                } else {
+                    out_val->data.x_f128 = ceilq(out_val->data.x_f128);
+                }
+                return;
+            default:
+                zig_unreachable();
+        }
+    } else {
+        zig_unreachable();
+    }
+}
+
+static void float_div_floor(ConstExprValue *out_val, ConstExprValue *op1, ConstExprValue *op2) {
+    assert(op1->type == op2->type);
+    out_val->type = op1->type;
+    if (op1->type->id == TypeTableEntryIdNumLitFloat) {
+        bigfloat_div_floor(&out_val->data.x_bigfloat, &op1->data.x_bigfloat, &op2->data.x_bigfloat);
+    } else if (op1->type->id == TypeTableEntryIdFloat) {
+        switch (op1->type->data.floating.bit_count) {
+            case 32:
+                out_val->data.x_f32 = floorf(op1->data.x_f32 / op2->data.x_f32);
+                return;
+            case 64:
+                out_val->data.x_f64 = floor(op1->data.x_f64 / op2->data.x_f64);
+                return;
+            case 128:
+                out_val->data.x_f128 = floorq(op1->data.x_f128 / op2->data.x_f128);
+                return;
+            default:
+                zig_unreachable();
+        }
+    } else {
+        zig_unreachable();
+    }
+}
+
+static void float_rem(ConstExprValue *out_val, ConstExprValue *op1, ConstExprValue *op2) {
+    assert(op1->type == op2->type);
+    out_val->type = op1->type;
+    if (op1->type->id == TypeTableEntryIdNumLitFloat) {
+        bigfloat_rem(&out_val->data.x_bigfloat, &op1->data.x_bigfloat, &op2->data.x_bigfloat);
+    } else if (op1->type->id == TypeTableEntryIdFloat) {
+        switch (op1->type->data.floating.bit_count) {
+            case 32:
+                out_val->data.x_f32 = fmodf(op1->data.x_f32, op2->data.x_f32);
+                return;
+            case 64:
+                out_val->data.x_f64 = fmod(op1->data.x_f64, op2->data.x_f64);
+                return;
+            case 128:
+                out_val->data.x_f128 = fmodq(op1->data.x_f128, op2->data.x_f128);
+                return;
+            default:
+                zig_unreachable();
+        }
+    } else {
+        zig_unreachable();
+    }
+}
+
+static void float_mod(ConstExprValue *out_val, ConstExprValue *op1, ConstExprValue *op2) {
+    assert(op1->type == op2->type);
+    out_val->type = op1->type;
+    if (op1->type->id == TypeTableEntryIdNumLitFloat) {
+        bigfloat_mod(&out_val->data.x_bigfloat, &op1->data.x_bigfloat, &op2->data.x_bigfloat);
+    } else if (op1->type->id == TypeTableEntryIdFloat) {
+        switch (op1->type->data.floating.bit_count) {
+            case 32:
+                out_val->data.x_f32 = fmodf(fmodf(op1->data.x_f32, op2->data.x_f32) + op2->data.x_f32, op2->data.x_f32);
+                return;
+            case 64:
+                out_val->data.x_f64 = fmod(fmod(op1->data.x_f64, op2->data.x_f64) + op2->data.x_f64, op2->data.x_f64);
+                return;
+            case 128:
+                out_val->data.x_f128 = fmodq(fmodq(op1->data.x_f128, op2->data.x_f128) + op2->data.x_f128, op2->data.x_f128);
+                return;
+            default:
+                zig_unreachable();
+        }
+    } else {
+        zig_unreachable();
+    }
+}
+
+static void float_negate(ConstExprValue *out_val, ConstExprValue *op) {
+    out_val->type = op->type;
+    if (op->type->id == TypeTableEntryIdNumLitFloat) {
+        bigfloat_negate(&out_val->data.x_bigfloat, &op->data.x_bigfloat);
+    } else if (op->type->id == TypeTableEntryIdFloat) {
+        switch (op->type->data.floating.bit_count) {
+            case 32:
+                out_val->data.x_f32 = -op->data.x_f32;
+                return;
+            case 64:
+                out_val->data.x_f64 = -op->data.x_f64;
+                return;
+            case 128:
+                out_val->data.x_f128 = -op->data.x_f128;
+                return;
+            default:
+                zig_unreachable();
+        }
+    } else {
+        zig_unreachable();
+    }
+}
+
+void float_write_ieee597(ConstExprValue *op, uint8_t *buf, bool is_big_endian) {
+    if (op->type->id == TypeTableEntryIdFloat) {
+        switch (op->type->data.floating.bit_count) {
+            case 32:
+                memcpy(buf, &op->data.x_f32, 4); // TODO wrong when compiler is big endian
+                return;
+            case 64:
+                memcpy(buf, &op->data.x_f64, 8); // TODO wrong when compiler is big endian
+                return;
+            case 128:
+                memcpy(buf, &op->data.x_f128, 16); // TODO wrong when compiler is big endian
+                return;
+            default:
+                zig_unreachable();
+        }
+    } else {
+        zig_unreachable();
+    }
+}
+
+void float_read_ieee597(ConstExprValue *val, uint8_t *buf, bool is_big_endian) {
+    if (val->type->id == TypeTableEntryIdFloat) {
+        switch (val->type->data.floating.bit_count) {
+            case 32:
+                memcpy(&val->data.x_f32, buf, 4); // TODO wrong when compiler is big endian
+                return;
+            case 64:
+                memcpy(&val->data.x_f64, buf, 8); // TODO wrong when compiler is big endian
+                return;
+            case 128:
+                memcpy(&val->data.x_f128, buf, 16); // TODO wrong when compiler is big endian
+                return;
+            default:
+                zig_unreachable();
+        }
+    } else {
+        zig_unreachable();
+    }
+}
+
 static bool ir_num_lit_fits_in_other_type(IrAnalyze *ira, IrInstruction *instruction, TypeTableEntry *other_type,
         bool explicit_cast)
 {
@@ -6314,9 +6855,9 @@ static bool ir_num_lit_fits_in_other_type(IrAnalyze *ira, IrInstruction *instruc
     if (explicit_cast && (other_type->id == TypeTableEntryIdInt || other_type->id == TypeTableEntryIdNumLitInt) &&
         const_val_is_float)
     {
-        if (bigfloat_has_fraction(&const_val->data.x_bigfloat)) {
+        if (float_has_fraction(const_val)) {
             Buf *val_buf = buf_alloc();
-            bigfloat_write_buf(val_buf, &const_val->data.x_bigfloat);
+            float_append_buf(val_buf, const_val);
 
             ir_add_error(ira, instruction,
                 buf_sprintf("fractional component prevents float value %s from being casted to type '%s'",
@@ -6324,11 +6865,11 @@ static bool ir_num_lit_fits_in_other_type(IrAnalyze *ira, IrInstruction *instruc
                     buf_ptr(&other_type->name)));
             return false;
         } else {
-            BigInt bigint;
-            bigint_init_bigfloat(&bigint, &const_val->data.x_bigfloat);
             if (other_type->id == TypeTableEntryIdNumLitInt) {
                 return true;
             } else {
+                BigInt bigint;
+                float_init_bigint(&bigint, const_val);
                 if (bigint_fits_in_bits(&bigint, other_type->data.integral.bit_count,
                     other_type->data.integral.is_signed))
                 {
@@ -6342,10 +6883,10 @@ static bool ir_num_lit_fits_in_other_type(IrAnalyze *ira, IrInstruction *instruc
     Buf *val_buf = buf_alloc();
     if (const_val_is_float) {
         num_lit_str = "float";
-        bigfloat_write_buf(val_buf, &const_val->data.x_bigfloat);
+        float_append_buf(val_buf, const_val);
     } else {
         num_lit_str = "integer";
-        bigint_write_buf(val_buf, &const_val->data.x_bigint, 10);
+        bigint_append_buf(val_buf, &const_val->data.x_bigint, 10);
     }
 
     ir_add_error(ira, instruction,
@@ -6767,7 +7308,20 @@ static void eval_const_expr_implicit_cast(CastOp cast_op,
             }
         case CastOpNumLitToConcrete:
             if (other_val->type->id == TypeTableEntryIdNumLitFloat) {
-                bigfloat_init_bigfloat(&const_val->data.x_bigfloat, &other_val->data.x_bigfloat);
+                assert(new_type->id == TypeTableEntryIdFloat);
+                switch (new_type->data.floating.bit_count) {
+                    case 32:
+                        const_val->data.x_f32 = bigfloat_to_f32(&other_val->data.x_bigfloat);
+                        break;
+                    case 64:
+                        const_val->data.x_f64 = bigfloat_to_f64(&other_val->data.x_bigfloat);
+                        break;
+                    case 128:
+                        const_val->data.x_f128 = bigfloat_to_f128(&other_val->data.x_bigfloat);
+                        break;
+                    default:
+                        zig_unreachable();
+                }
             } else if (other_val->type->id == TypeTableEntryIdNumLitInt) {
                 bigint_init_bigint(&const_val->data.x_bigint, &other_val->data.x_bigint);
             } else {
@@ -6780,11 +7334,29 @@ static void eval_const_expr_implicit_cast(CastOp cast_op,
             // can't do it
             break;
         case CastOpIntToFloat:
-            bigfloat_init_bigint(&const_val->data.x_bigfloat, &other_val->data.x_bigint);
-            const_val->special = ConstValSpecialStatic;
-            break;
+            {
+                assert(new_type->id == TypeTableEntryIdFloat);
+
+                BigFloat bigfloat;
+                bigfloat_init_bigint(&bigfloat, &other_val->data.x_bigint);
+                switch (new_type->data.floating.bit_count) {
+                    case 32:
+                        const_val->data.x_f32 = bigfloat_to_f32(&bigfloat);
+                        break;
+                    case 64:
+                        const_val->data.x_f64 = bigfloat_to_f64(&bigfloat);
+                        break;
+                    case 128:
+                        const_val->data.x_f128 = bigfloat_to_f128(&bigfloat);
+                        break;
+                    default:
+                        zig_unreachable();
+                }
+                const_val->special = ConstValSpecialStatic;
+                break;
+            }
         case CastOpFloatToInt:
-            bigint_init_bigfloat(&const_val->data.x_bigint, &other_val->data.x_bigfloat);
+            float_init_bigint(&const_val->data.x_bigint, other_val);
             const_val->special = ConstValSpecialStatic;
             break;
         case CastOpBoolToInt:
@@ -7387,12 +7959,12 @@ static IrInstruction *ir_analyze_widen_or_shorten(IrAnalyze *ira, IrInstruction
         }
         IrInstruction *result = ir_create_const(&ira->new_irb, source_instr->scope,
                 source_instr->source_node, wanted_type);
+        result->value.type = wanted_type;
         if (wanted_type->id == TypeTableEntryIdInt) {
             bigint_init_bigint(&result->value.data.x_bigint, &val->data.x_bigint);
         } else {
-            bigfloat_init_bigfloat(&result->value.data.x_bigfloat, &val->data.x_bigfloat);
+            float_init_float(&result->value, val);
         }
-        result->value.type = wanted_type;
         return result;
     }
 
@@ -7415,7 +7987,7 @@ static IrInstruction *ir_analyze_int_to_enum(IrAnalyze *ira, IrInstruction *sour
         bigint_init_unsigned(&enum_member_count, wanted_type->data.enumeration.src_field_count);
         if (bigint_cmp(&val->data.x_bigint, &enum_member_count) != CmpLT) {
             Buf *val_buf = buf_alloc();
-            bigint_write_buf(val_buf, &val->data.x_bigint, 10);
+            bigint_append_buf(val_buf, &val->data.x_bigint, 10);
             ir_add_error(ira, source_instr,
                 buf_sprintf("integer value %s too big for enum '%s' which has %" PRIu32 " fields",
                     buf_ptr(val_buf), buf_ptr(&wanted_type->name), wanted_type->data.enumeration.src_field_count));
@@ -7444,7 +8016,7 @@ static IrInstruction *ir_analyze_number_to_literal(IrAnalyze *ira, IrInstruction
     IrInstruction *result = ir_create_const(&ira->new_irb, source_instr->scope,
             source_instr->source_node, wanted_type);
     if (wanted_type->id == TypeTableEntryIdNumLitFloat) {
-        bigfloat_init_bigfloat(&result->value.data.x_bigfloat, &val->data.x_bigfloat);
+        float_init_float(&result->value, val);
     } else if (wanted_type->id == TypeTableEntryIdNumLitInt) {
         bigint_init_bigint(&result->value.data.x_bigint, &val->data.x_bigint);
     } else {
@@ -7469,7 +8041,7 @@ static IrInstruction *ir_analyze_int_to_err(IrAnalyze *ira, IrInstruction *sourc
         bigint_init_unsigned(&err_count, ira->codegen->error_decls.length);
         if (bigint_cmp_zero(&val->data.x_bigint) == CmpEQ || bigint_cmp(&val->data.x_bigint, &err_count) != CmpLT) {
             Buf *val_buf = buf_alloc();
-            bigint_write_buf(val_buf, &val->data.x_bigint, 10);
+            bigint_append_buf(val_buf, &val->data.x_bigint, 10);
             ir_add_error(ira, source_instr,
                 buf_sprintf("integer value %s represents no error", buf_ptr(val_buf)));
             return ira->codegen->invalid_instruction;
@@ -8304,7 +8876,7 @@ static TypeTableEntry *ir_analyze_bin_op_cmp(IrAnalyze *ira, IrInstructionBinOp
     if ((value_is_comptime(op1_val) && value_is_comptime(op2_val)) || resolved_type->id == TypeTableEntryIdVoid) {
         bool answer;
         if (resolved_type->id == TypeTableEntryIdNumLitFloat || resolved_type->id == TypeTableEntryIdFloat) {
-            Cmp cmp_result = bigfloat_cmp(&op1_val->data.x_bigfloat, &op2_val->data.x_bigfloat);
+            Cmp cmp_result = float_cmp(op1_val, op2_val);
             answer = resolve_cmp_op_id(op_id, cmp_result);
         } else if (resolved_type->id == TypeTableEntryIdNumLitInt || resolved_type->id == TypeTableEntryIdInt) {
             Cmp cmp_result = bigint_cmp(&op1_val->data.x_bigint, &op2_val->data.x_bigint);
@@ -8379,7 +8951,7 @@ static int ir_eval_math_op(TypeTableEntry *type_entry, ConstExprValue *op1_val,
     {
         is_int = false;
         is_float = true;
-        op2_zcmp = bigfloat_cmp_zero(&op2_val->data.x_bigfloat);
+        op2_zcmp = float_cmp_zero(op2_val);
     } else {
         zig_unreachable();
     }
@@ -8447,7 +9019,7 @@ static int ir_eval_math_op(TypeTableEntry *type_entry, ConstExprValue *op1_val,
             if (is_int) {
                 bigint_add(&out_val->data.x_bigint, &op1_val->data.x_bigint, &op2_val->data.x_bigint);
             } else {
-                bigfloat_add(&out_val->data.x_bigfloat, &op1_val->data.x_bigfloat, &op2_val->data.x_bigfloat);
+                float_add(out_val, op1_val, op2_val);
             }
             break;
         case IrBinOpAddWrap:
@@ -8459,7 +9031,7 @@ static int ir_eval_math_op(TypeTableEntry *type_entry, ConstExprValue *op1_val,
             if (is_int) {
                 bigint_sub(&out_val->data.x_bigint, &op1_val->data.x_bigint, &op2_val->data.x_bigint);
             } else {
-                bigfloat_sub(&out_val->data.x_bigfloat, &op1_val->data.x_bigfloat, &op2_val->data.x_bigfloat);
+                float_sub(out_val, op1_val, op2_val);
             }
             break;
         case IrBinOpSubWrap:
@@ -8471,7 +9043,7 @@ static int ir_eval_math_op(TypeTableEntry *type_entry, ConstExprValue *op1_val,
             if (is_int) {
                 bigint_mul(&out_val->data.x_bigint, &op1_val->data.x_bigint, &op2_val->data.x_bigint);
             } else {
-                bigfloat_mul(&out_val->data.x_bigfloat, &op1_val->data.x_bigfloat, &op2_val->data.x_bigfloat);
+                float_mul(out_val, op1_val, op2_val);
             }
             break;
         case IrBinOpMultWrap:
@@ -8481,20 +9053,20 @@ static int ir_eval_math_op(TypeTableEntry *type_entry, ConstExprValue *op1_val,
             break;
         case IrBinOpDivUnspecified:
             assert(is_float);
-            bigfloat_div(&out_val->data.x_bigfloat, &op1_val->data.x_bigfloat, &op2_val->data.x_bigfloat);
+            float_div(out_val, op1_val, op2_val);
             break;
         case IrBinOpDivTrunc:
             if (is_int) {
                 bigint_div_trunc(&out_val->data.x_bigint, &op1_val->data.x_bigint, &op2_val->data.x_bigint);
             } else {
-                bigfloat_div_trunc(&out_val->data.x_bigfloat, &op1_val->data.x_bigfloat, &op2_val->data.x_bigfloat);
+                float_div_trunc(out_val, op1_val, op2_val);
             }
             break;
         case IrBinOpDivFloor:
             if (is_int) {
                 bigint_div_floor(&out_val->data.x_bigint, &op1_val->data.x_bigint, &op2_val->data.x_bigint);
             } else {
-                bigfloat_div_floor(&out_val->data.x_bigfloat, &op1_val->data.x_bigfloat, &op2_val->data.x_bigfloat);
+                float_div_floor(out_val, op1_val, op2_val);
             }
             break;
         case IrBinOpDivExact:
@@ -8506,10 +9078,10 @@ static int ir_eval_math_op(TypeTableEntry *type_entry, ConstExprValue *op1_val,
                     return ErrorExactDivRemainder;
                 }
             } else {
-                bigfloat_div_trunc(&out_val->data.x_bigfloat, &op1_val->data.x_bigfloat, &op2_val->data.x_bigfloat);
-                BigFloat remainder;
-                bigfloat_rem(&remainder, &op1_val->data.x_bigfloat, &op2_val->data.x_bigfloat);
-                if (bigfloat_cmp_zero(&remainder) != CmpEQ) {
+                float_div_trunc(out_val, op1_val, op2_val);
+                ConstExprValue remainder;
+                float_rem(&remainder, op1_val, op2_val);
+                if (float_cmp_zero(&remainder) != CmpEQ) {
                     return ErrorExactDivRemainder;
                 }
             }
@@ -8518,14 +9090,14 @@ static int ir_eval_math_op(TypeTableEntry *type_entry, ConstExprValue *op1_val,
             if (is_int) {
                 bigint_rem(&out_val->data.x_bigint, &op1_val->data.x_bigint, &op2_val->data.x_bigint);
             } else {
-                bigfloat_rem(&out_val->data.x_bigfloat, &op1_val->data.x_bigfloat, &op2_val->data.x_bigfloat);
+                float_rem(out_val, op1_val, op2_val);
             }
             break;
         case IrBinOpRemMod:
             if (is_int) {
                 bigint_mod(&out_val->data.x_bigint, &op1_val->data.x_bigint, &op2_val->data.x_bigint);
             } else {
-                bigfloat_mod(&out_val->data.x_bigfloat, &op1_val->data.x_bigfloat, &op2_val->data.x_bigfloat);
+                float_mod(out_val, op1_val, op2_val);
             }
             break;
     }
@@ -8678,16 +9250,16 @@ static TypeTableEntry *ir_analyze_bin_op_math(IrAnalyze *ira, IrInstructionBinOp
                         ok = bigint_cmp(&rem_result, &mod_result) == CmpEQ;
                     }
                 } else {
-                    if (bigfloat_cmp_zero(&op2->value.data.x_bigfloat) == CmpEQ) {
+                    if (float_cmp_zero(&op2->value) == CmpEQ) {
                         // the division by zero error will be caught later, but we don't
                         // have a remainder function ambiguity problem
                         ok = true;
                     } else {
-                        BigFloat rem_result;
-                        BigFloat mod_result;
-                        bigfloat_rem(&rem_result, &op1->value.data.x_bigfloat, &op2->value.data.x_bigfloat);
-                        bigfloat_mod(&mod_result, &op1->value.data.x_bigfloat, &op2->value.data.x_bigfloat);
-                        ok = bigfloat_cmp(&rem_result, &mod_result) == CmpEQ;
+                        ConstExprValue rem_result;
+                        ConstExprValue mod_result;
+                        float_rem(&rem_result, &op1->value, &op2->value);
+                        float_mod(&mod_result, &op1->value, &op2->value);
+                        ok = float_cmp(&rem_result, &mod_result) == CmpEQ;
                     }
                 }
             }
@@ -9835,7 +10407,7 @@ static TypeTableEntry *ir_analyze_negation(IrAnalyze *ira, IrInstructionUnOp *un
 
             ConstExprValue *out_val = ir_build_const_from(ira, &un_op_instruction->base);
             if (is_float) {
-                bigfloat_negate(&out_val->data.x_bigfloat, &target_const_val->data.x_bigfloat);
+                float_negate(out_val, target_const_val);
             } else if (is_wrap_op) {
                 bigint_negate_wrap(&out_val->data.x_bigint, &target_const_val->data.x_bigint,
                         expr_type->data.integral.bit_count);
@@ -13780,8 +14352,7 @@ static void buf_write_value_bytes(CodeGen *codegen, uint8_t *buf, ConstExprValue
                     codegen->is_big_endian);
             return;
         case TypeTableEntryIdFloat:
-            bigfloat_write_ieee597(&val->data.x_bigfloat, buf, val->type->data.floating.bit_count,
-                    codegen->is_big_endian);
+            float_write_ieee597(val, buf, codegen->is_big_endian);
             return;
         case TypeTableEntryIdPointer:
             if (val->data.x_ptr.special == ConstPtrSpecialHardCodedAddr) {
@@ -13841,8 +14412,7 @@ static void buf_read_value_bytes(CodeGen *codegen, uint8_t *buf, ConstExprValue
                     codegen->is_big_endian, val->type->data.integral.is_signed);
             return;
         case TypeTableEntryIdFloat:
-            bigfloat_read_ieee597(&val->data.x_bigfloat, buf, val->type->data.floating.bit_count,
-                    codegen->is_big_endian);
+            float_read_ieee597(val, buf, codegen->is_big_endian);
             return;
         case TypeTableEntryIdPointer:
             {
src/quadmath.hpp
@@ -0,0 +1,19 @@
+/*
+ * Copyright (c) 2017 Andrew Kelley
+ *
+ * This file is part of zig, which is MIT licensed.
+ * See http://opensource.org/licenses/MIT
+ */
+
+#ifndef ZIG_QUADMATH_HPP
+#define ZIG_QUADMATH_HPP
+
+extern "C" {
+    __float128 fmodq(__float128 a, __float128 b);
+    __float128 ceilq(__float128 a);
+    __float128 floorq(__float128 a);
+    __float128 strtoflt128 (const char *s, char **sp);
+    int quadmath_snprintf (char *s, size_t size, const char *format, ...);
+}
+
+#endif
src/tokenizer.cpp
@@ -257,7 +257,7 @@ static void set_token_id(Tokenize *t, Token *token, TokenId id) {
     if (id == TokenIdIntLiteral) {
         bigint_init_unsigned(&token->data.int_lit.bigint, 0);
     } else if (id == TokenIdFloatLiteral) {
-        bigfloat_init_float(&token->data.float_lit.bigfloat, 0.0);
+        bigfloat_init_32(&token->data.float_lit.bigfloat, 0.0f);
         token->data.float_lit.overflow = false;
     } else if (id == TokenIdStringLiteral || id == TokenIdSymbol) {
         memset(&token->data.str_lit.str, 0, sizeof(Buf));
@@ -345,7 +345,7 @@ static void end_float_token(Tokenize *t) {
     uint64_t double_bits = (exponent_bits << 52) | significand_bits;
     double dbl_value;
     safe_memcpy(&dbl_value, (double *)&double_bits, 1);
-    bigfloat_init_float(&t->cur_tok->data.float_lit.bigfloat, dbl_value);
+    bigfloat_init_64(&t->cur_tok->data.float_lit.bigfloat, dbl_value);
 }
 
 static void end_token(Tokenize *t) {
std/math/atan2.zig
@@ -24,7 +24,7 @@ const assert = @import("../debug.zig").assert;
 pub const atan2 = atan2_workaround;
 
 // TODO issue #393
-pub fn atan2_workaround(comptime T: type, x: T, y: T) -> T {
+fn atan2_workaround(comptime T: type, x: T, y: T) -> T {
     switch (T) {
         f32 => @inlineCall(atan2_32, x, y),
         f64 => @inlineCall(atan2_64, x, y),
std/math/sin.zig
@@ -147,6 +147,7 @@ fn sin64(x_: f64) -> f64 {
 test "math.sin" {
     assert(sin(f32(0.0)) == sin32(0.0));
     assert(sin(f64(0.0)) == sin64(0.0));
+    assert(comptime {math.sin(f64(2))} == math.sin(f64(2)));
 }
 
 test "math.sin32" {
test/cases/eval.zig
@@ -355,3 +355,23 @@ test "@setEvalBranchQuota" {
         assert(sum == 500500);
     }
 }
+
+test "float literal at compile time not lossy" {
+    assert(16777216.0 + 1.0 == 16777217.0);
+    assert(9007199254740992.0 + 1.0 == 9007199254740993.0);
+}
+
+test "f32 at compile time is lossy" {
+    assert(f32(1 << 24) + 1 == 1 << 24);
+}
+
+test "f64 at compile time is lossy" {
+    assert(f64(1 << 53) + 1 == 1 << 53);
+}
+
+test "f128 at compile time is lossy" {
+    assert(f128(10384593717069655257060992658440192.0) + 1 == 10384593717069655257060992658440192.0);
+}
+
+// TODO need a better implementation of bigfloat_init_bigint
+// assert(f128(1 << 113) == 10384593717069655257060992658440192);