Merge pull request #593 from beetrees/fix-f16-conv-c

Add builtins for `f16`/`f128` float conversions

Author: Amanieu

Cargo.toml

@@ -49,6 +49,10 @@ c = ["cc"]
 # which use inline assembly and fall back to pure Rust versions (if avalible).
 no-asm = []
 
+# Workaround for codegen backends which haven't yet implemented `f16` and
+# `f128` support. Disabled any intrinsics which use those types.
+no-f16-f128 = []
+
 # Flag this library as the unstable compiler-builtins lib
 compiler-builtins = []
 

README.md

@@ -162,7 +162,6 @@ rely on CI.
 - [x] divmodsi4.c
 - [x] divsf3.c
 - [x] divsi3.c
-- [ ] extendhfsf2.c
 - [x] extendsfdf2.c
 - [x] fixdfdi.c
 - [x] fixdfsi.c
@@ -201,9 +200,7 @@ rely on CI.
 - [x] powisf2.c
 - [x] subdf3.c
 - [x] subsf3.c
-- [ ] truncdfhf2.c
 - [x] truncdfsf2.c
-- [ ] truncsfhf2.c
 - [x] udivdi3.c
 - [x] udivmoddi4.c
 - [x] udivmodsi4.c
@@ -233,60 +230,68 @@ These builtins are needed to support 128-bit integers.
 - [x] udivti3.c
 - [x] umodti3.c
 
+These builtins are needed to support `f16` and `f128`, which are in the process of being added to Rust.
+
+- [ ] addtf3.c
+- [ ] comparetf2.c
+- [ ] divtf3.c
+- [x] extenddftf2.c
+- [x] extendhfsf2.c
+- [x] extendhftf2.c
+- [x] extendsftf2.c
+- [ ] fixtfdi.c
+- [ ] fixtfsi.c
+- [ ] fixtfti.c
+- [ ] fixunstfdi.c
+- [ ] fixunstfsi.c
+- [ ] fixunstfti.c
+- [ ] floatditf.c
+- [ ] floatsitf.c
+- [ ] floatunditf.c
+- [ ] floatunsitf.c
+- [ ] multf3.c
+- [ ] powitf2.c
+- [ ] ppc/fixtfdi.c
+- [ ] ppc/fixunstfdi.c
+- [ ] ppc/floatditf.c
+- [ ] ppc/floatunditf.c
+- [ ] subtf3.c
+- [x] truncdfhf2.c
+- [x] truncsfhf2.c
+- [x] trunctfdf2.c
+- [x] trunctfhf2.c
+- [x] trunctfsf2.c
+
 ## Unimplemented functions
 
-These builtins involve floating-point types ("`f128`", "`f80`" and complex numbers) that are not supported by Rust.
+These builtins involve floating-point types ("`f80`" and complex numbers) that are not supported by Rust.
 
-- ~~addtf3.c~~
-- ~~comparetf2.c~~
 - ~~divdc3.c~~
 - ~~divsc3.c~~
 - ~~divtc3.c~~
-- ~~divtf3.c~~
 - ~~divxc3.c~~
-- ~~extenddftf2.c~~
-- ~~extendsftf2.c~~
-- ~~fixtfdi.c~~
-- ~~fixtfsi.c~~
-- ~~fixtfti.c~~
-- ~~fixunstfdi.c~~
-- ~~fixunstfsi.c~~
-- ~~fixunstfti.c~~
 - ~~fixunsxfdi.c~~
 - ~~fixunsxfsi.c~~
 - ~~fixunsxfti.c~~
 - ~~fixxfdi.c~~
 - ~~fixxfti.c~~
-- ~~floatditf.c~~
 - ~~floatdixf.c~~
-- ~~floatsitf.c~~
 - ~~floattixf.c~~
-- ~~floatunditf.c~~
 - ~~floatundixf.c~~
-- ~~floatunsitf.c~~
 - ~~floatuntixf.c~~
 - ~~i386/floatdixf.S~~
 - ~~i386/floatundixf.S~~
 - ~~muldc3.c~~
 - ~~mulsc3.c~~
 - ~~multc3.c~~
-- ~~multf3.c~~
 - ~~mulxc3.c~~
-- ~~powitf2.c~~
 - ~~powixf2.c~~
 - ~~ppc/divtc3.c~~
-- ~~ppc/fixtfdi.c~~
-- ~~ppc/fixunstfdi.c~~
-- ~~ppc/floatditf.c~~
-- ~~ppc/floatunditf.c~~
 - ~~ppc/gcc_qadd.c~~
 - ~~ppc/gcc_qdiv.c~~
 - ~~ppc/gcc_qmul.c~~
 - ~~ppc/gcc_qsub.c~~
 - ~~ppc/multc3.c~~
-- ~~subtf3.c~~
-- ~~trunctfdf2.c~~
-- ~~trunctfsf2.c~~
 - ~~x86_64/floatdixf.c~~
 - ~~x86_64/floatundixf.S~~
 

build.rs

@@ -217,6 +217,14 @@ mod c {
             }
         }
 
+        // `compiler-rt` requires `COMPILER_RT_HAS_FLOAT16` to be defined to make it use the
+        // `_Float16` type for `f16` intrinsics. This shouldn't matter as all existing `f16`
+        // intrinsics have been ported to Rust in `compiler-builtins` as C compilers don't
+        // support `_Float16` on all targets (whereas Rust does). However, define the macro
+        // anyway to prevent issues like rust#118813 and rust#123885 silently reoccuring if more
+        // `f16` intrinsics get accidentally added here in the future.
+        cfg.define("COMPILER_RT_HAS_FLOAT16", None);
+
         cfg.warnings(false);
 
         if target_env == "msvc" {
@@ -288,13 +296,10 @@ mod c {
             sources.extend(&[
                 ("__divdc3", "divdc3.c"),
                 ("__divsc3", "divsc3.c"),
-                ("__extendhfsf2", "extendhfsf2.c"),
                 ("__muldc3", "muldc3.c"),
                 ("__mulsc3", "mulsc3.c"),
                 ("__negdf2", "negdf2.c"),
                 ("__negsf2", "negsf2.c"),
-                ("__truncdfhf2", "truncdfhf2.c"),
-                ("__truncsfhf2", "truncsfhf2.c"),
             ]);
         }
 
@@ -464,8 +469,6 @@ mod c {
         if (target_arch == "aarch64" || target_arch == "arm64ec") && consider_float_intrinsics {
             sources.extend(&[
                 ("__comparetf2", "comparetf2.c"),
-                ("__extenddftf2", "extenddftf2.c"),
-                ("__extendsftf2", "extendsftf2.c"),
                 ("__fixtfdi", "fixtfdi.c"),
                 ("__fixtfsi", "fixtfsi.c"),
                 ("__fixtfti", "fixtfti.c"),
@@ -476,8 +479,6 @@ mod c {
                 ("__floatsitf", "floatsitf.c"),
                 ("__floatunditf", "floatunditf.c"),
                 ("__floatunsitf", "floatunsitf.c"),
-                ("__trunctfdf2", "trunctfdf2.c"),
-                ("__trunctfsf2", "trunctfsf2.c"),
                 ("__addtf3", "addtf3.c"),
                 ("__multf3", "multf3.c"),
                 ("__subtf3", "subtf3.c"),
@@ -498,7 +499,6 @@ mod c {
 
         if target_arch == "mips64" {
             sources.extend(&[
-                ("__extenddftf2", "extenddftf2.c"),
                 ("__netf2", "comparetf2.c"),
                 ("__addtf3", "addtf3.c"),
                 ("__multf3", "multf3.c"),
@@ -509,14 +509,11 @@ mod c {
                 ("__floatunsitf", "floatunsitf.c"),
                 ("__fe_getround", "fp_mode.c"),
                 ("__divtf3", "divtf3.c"),
-                ("__trunctfdf2", "trunctfdf2.c"),
-                ("__trunctfsf2", "trunctfsf2.c"),
             ]);
         }
 
         if target_arch == "loongarch64" {
             sources.extend(&[
-                ("__extenddftf2", "extenddftf2.c"),
                 ("__netf2", "comparetf2.c"),
                 ("__addtf3", "addtf3.c"),
                 ("__multf3", "multf3.c"),
@@ -527,8 +524,6 @@ mod c {
                 ("__floatunsitf", "floatunsitf.c"),
                 ("__fe_getround", "fp_mode.c"),
                 ("__divtf3", "divtf3.c"),
-                ("__trunctfdf2", "trunctfdf2.c"),
-                ("__trunctfsf2", "trunctfsf2.c"),
             ]);
         }
 

ci/run.sh

@@ -28,6 +28,8 @@ else
     $run --features c --release
     $run --features no-asm
     $run --features no-asm --release
+    $run --features no-f16-f128
+    $run --features no-f16-f128 --release
 fi
 
 if [ -d /builtins-target ]; then
@@ -45,6 +47,8 @@ cargo build --target "$target" --features c
 cargo build --target "$target" --release --features c
 cargo build --target "$target" --features no-asm
 cargo build --target "$target" --release --features no-asm
+cargo build --target "$target" --features no-f16-f128
+cargo build --target "$target" --release --features no-f16-f128
 
 PREFIX=${target//unknown-/}-
 case "$target" in

src/float/extend.rs

@@ -82,3 +82,37 @@ intrinsics! {
         a as f64 // LLVM generate 'fcvtds'
     }
 }
+
+#[cfg(not(feature = "no-f16-f128"))]
+intrinsics! {
+    #[avr_skip]
+    #[aapcs_on_arm]
+    #[arm_aeabi_alias = __aeabi_h2f]
+    pub extern "C" fn __extendhfsf2(a: f16) -> f32 {
+        extend(a)
+    }
+
+    #[avr_skip]
+    #[aapcs_on_arm]
+    pub extern "C" fn __gnu_h2f_ieee(a: f16) -> f32 {
+        extend(a)
+    }
+
+    #[avr_skip]
+    #[aapcs_on_arm]
+    pub extern "C" fn __extendhftf2(a: f16) -> f128 {
+        extend(a)
+    }
+
+    #[avr_skip]
+    #[aapcs_on_arm]
+    pub extern "C" fn __extendsftf2(a: f32) -> f128 {
+        extend(a)
+    }
+
+    #[avr_skip]
+    #[aapcs_on_arm]
+    pub extern "C" fn __extenddftf2(a: f64) -> f128 {
+        extend(a)
+    }
+}

src/float/mod.rs

@@ -127,7 +127,20 @@ macro_rules! float_impl {
                 self.to_bits() as Self::SignedInt
             }
             fn eq_repr(self, rhs: Self) -> bool {
-                if self.is_nan() && rhs.is_nan() {
+                #[cfg(feature = "mangled-names")]
+                fn is_nan(x: $ty) -> bool {
+                    // When using mangled-names, the "real" compiler-builtins might not have the
+                    // necessary builtin (__unordtf2) to test whether `f128` is NaN.
+                    // FIXME(f16_f128): Remove once the nightly toolchain has the __unordtf2 builtin
+                    // x is NaN if all the bits of the exponent are set and the significand is non-0
+                    x.repr() & $ty::EXPONENT_MASK == $ty::EXPONENT_MASK
+                        && x.repr() & $ty::SIGNIFICAND_MASK != 0
+                }
+                #[cfg(not(feature = "mangled-names"))]
+                fn is_nan(x: $ty) -> bool {
+                    x.is_nan()
+                }
+                if is_nan(self) && is_nan(rhs) {
                     true
                 } else {
                     self.repr() == rhs.repr()
@@ -171,5 +184,9 @@ macro_rules! float_impl {
     };
 }
 
+#[cfg(not(feature = "no-f16-f128"))]
+float_impl!(f16, u16, i16, i8, 16, 10);
 float_impl!(f32, u32, i32, i16, 32, 23);
 float_impl!(f64, u64, i64, i16, 64, 52);
+#[cfg(not(feature = "no-f16-f128"))]
+float_impl!(f128, u128, i128, i16, 128, 112);

src/float/trunc.rs

@@ -52,8 +52,10 @@ where
         // destination format.  We can convert by simply right-shifting with
         // rounding and adjusting the exponent.
         abs_result = (a_abs >> sign_bits_delta).cast();
-        let tmp = src_exp_bias.wrapping_sub(dst_exp_bias) << R::SIGNIFICAND_BITS;
-        abs_result = abs_result.wrapping_sub(tmp.cast());
+        // Cast before shifting to prevent overflow.
+        let bias_diff: R::Int = src_exp_bias.wrapping_sub(dst_exp_bias).cast();
+        let tmp = bias_diff << R::SIGNIFICAND_BITS;
+        abs_result = abs_result.wrapping_sub(tmp);
 
         let round_bits = a_abs & round_mask;
         if round_bits > halfway {
@@ -67,13 +69,17 @@ where
         // a is NaN.
         // Conjure the result by beginning with infinity, setting the qNaN
         // bit and inserting the (truncated) trailing NaN field.
-        abs_result = (dst_inf_exp << R::SIGNIFICAND_BITS).cast();
+        // Cast before shifting to prevent overflow.
+        let dst_inf_exp: R::Int = dst_inf_exp.cast();
+        abs_result = dst_inf_exp << R::SIGNIFICAND_BITS;
         abs_result |= dst_qnan;
         abs_result |= dst_nan_code
             & ((a_abs & src_nan_code) >> (F::SIGNIFICAND_BITS - R::SIGNIFICAND_BITS)).cast();
     } else if a_abs >= overflow {
         // a overflows to infinity.
-        abs_result = (dst_inf_exp << R::SIGNIFICAND_BITS).cast();
+        // Cast before shifting to prevent overflow.
+        let dst_inf_exp: R::Int = dst_inf_exp.cast();
+        abs_result = dst_inf_exp << R::SIGNIFICAND_BITS;
     } else {
         // a underflows on conversion to the destination type or is an exact
         // zero.  The result may be a denormal or zero.  Extract the exponent
@@ -124,3 +130,44 @@ intrinsics! {
         a as f32
     }
 }
+
+#[cfg(not(feature = "no-f16-f128"))]
+intrinsics! {
+    #[avr_skip]
+    #[aapcs_on_arm]
+    #[arm_aeabi_alias = __aeabi_f2h]
+    pub extern "C" fn __truncsfhf2(a: f32) -> f16 {
+        trunc(a)
+    }
+
+    #[avr_skip]
+    #[aapcs_on_arm]
+    pub extern "C" fn __gnu_f2h_ieee(a: f32) -> f16 {
+        trunc(a)
+    }
+
+    #[avr_skip]
+    #[aapcs_on_arm]
+    #[arm_aeabi_alias = __aeabi_d2h]
+    pub extern "C" fn __truncdfhf2(a: f64) -> f16 {
+        trunc(a)
+    }
+
+    #[avr_skip]
+    #[aapcs_on_arm]
+    pub extern "C" fn __trunctfhf2(a: f128) -> f16 {
+        trunc(a)
+    }
+
+    #[avr_skip]
+    #[aapcs_on_arm]
+    pub extern "C" fn __trunctfsf2(a: f128) -> f32 {
+        trunc(a)
+    }
+
+    #[avr_skip]
+    #[aapcs_on_arm]
+    pub extern "C" fn __trunctfdf2(a: f128) -> f64 {
+        trunc(a)
+    }
+}

src/lib.rs

@@ -13,6 +13,8 @@
 #![feature(naked_functions)]
 #![feature(repr_simd)]
 #![feature(c_unwind)]
+#![cfg_attr(not(feature = "no-f16-f128"), feature(f16))]
+#![cfg_attr(not(feature = "no-f16-f128"), feature(f128))]
 #![no_builtins]
 #![no_std]
 #![allow(unused_features)]