Add integer to f128 conversions

library/compiler-builtins/README.md

@@ -233,12 +233,12 @@ of being added to Rust.
 - [x] fixunstfdi.c
 - [x] fixunstfsi.c
 - [x] fixunstfti.c
-- [ ] floatditf.c
-- [ ] floatsitf.c
-- [ ] floattitf.c
-- [ ] floatunditf.c
-- [ ] floatunsitf.c
-- [ ] floatuntitf.c
+- [x] floatditf.c
+- [x] floatsitf.c
+- [x] floattitf.c
+- [x] floatunditf.c
+- [x] floatunsitf.c
+- [x] floatuntitf.c
 - [x] multf3.c
 - [x] powitf2.c
 - [x] subtf3.c

library/compiler-builtins/build.rs

@@ -532,10 +532,6 @@ mod c {
         if (target.arch == "aarch64" || target.arch == "arm64ec") && consider_float_intrinsics {
             sources.extend(&[
                 ("__comparetf2", "comparetf2.c"),
-                ("__floatditf", "floatditf.c"),
-                ("__floatsitf", "floatsitf.c"),
-                ("__floatunditf", "floatunditf.c"),
-                ("__floatunsitf", "floatunsitf.c"),
                 ("__fe_getround", "fp_mode.c"),
                 ("__fe_raise_inexact", "fp_mode.c"),
             ]);
@@ -550,21 +546,11 @@ mod c {
         }
 
         if target.arch == "mips64" {
-            sources.extend(&[
-                ("__netf2", "comparetf2.c"),
-                ("__floatsitf", "floatsitf.c"),
-                ("__floatunsitf", "floatunsitf.c"),
-                ("__fe_getround", "fp_mode.c"),
-            ]);
+            sources.extend(&[("__netf2", "comparetf2.c"), ("__fe_getround", "fp_mode.c")]);
         }
 
         if target.arch == "loongarch64" {
-            sources.extend(&[
-                ("__netf2", "comparetf2.c"),
-                ("__floatsitf", "floatsitf.c"),
-                ("__floatunsitf", "floatunsitf.c"),
-                ("__fe_getround", "fp_mode.c"),
-            ]);
+            sources.extend(&[("__netf2", "comparetf2.c"), ("__fe_getround", "fp_mode.c")]);
         }
 
         // Remove the assembly implementations that won't compile for the target

library/compiler-builtins/examples/intrinsics.rs

@@ -264,14 +264,18 @@ mod intrinsics {
 
     /* i32 operations */
 
+    // floatsisf
+    pub fn aeabi_i2f(x: i32) -> f32 {
+        x as f32
+    }
+
     // floatsidf
     pub fn aeabi_i2d(x: i32) -> f64 {
         x as f64
     }
 
-    // floatsisf
-    pub fn aeabi_i2f(x: i32) -> f32 {
-        x as f32
+    pub fn floatsitf(x: i32) -> f128 {
+        x as f128
     }
 
     pub fn aeabi_idiv(a: i32, b: i32) -> i32 {
@@ -294,6 +298,10 @@ mod intrinsics {
         x as f64
     }
 
+    pub fn floatditf(x: i64) -> f128 {
+        x as f128
+    }
+
     pub fn mulodi4(a: i64, b: i64) -> i64 {
         a * b
     }
@@ -314,6 +322,18 @@ mod intrinsics {
 
     /* i128 operations */
 
+    pub fn floattisf(x: i128) -> f32 {
+        x as f32
+    }
+
+    pub fn floattidf(x: i128) -> f64 {
+        x as f64
+    }
+
+    pub fn floattitf(x: i128) -> f128 {
+        x as f128
+    }
+
     pub fn lshrti3(a: i128, b: usize) -> i128 {
         a >> b
     }
@@ -328,14 +348,18 @@ mod intrinsics {
 
     /* u32 operations */
 
+    // floatunsisf
+    pub fn aeabi_ui2f(x: u32) -> f32 {
+        x as f32
+    }
+
     // floatunsidf
     pub fn aeabi_ui2d(x: u32) -> f64 {
         x as f64
     }
 
-    // floatunsisf
-    pub fn aeabi_ui2f(x: u32) -> f32 {
-        x as f32
+    pub fn floatunsitf(x: u32) -> f128 {
+        x as f128
     }
 
     pub fn aeabi_uidiv(a: u32, b: u32) -> u32 {
@@ -358,6 +382,10 @@ mod intrinsics {
         x as f64
     }
 
+    pub fn floatunditf(x: u64) -> f128 {
+        x as f128
+    }
+
     // udivdi3
     pub fn aeabi_uldivmod(a: u64, b: u64) -> u64 {
         a * b
@@ -369,6 +397,18 @@ mod intrinsics {
 
     /* u128 operations */
 
+    pub fn floatuntisf(x: u128) -> f32 {
+        x as f32
+    }
+
+    pub fn floatuntidf(x: u128) -> f64 {
+        x as f64
+    }
+
+    pub fn floatuntitf(x: u128) -> f128 {
+        x as f128
+    }
+
     pub fn muloti4(a: u128, b: u128) -> Option<u128> {
         a.checked_mul(b)
     }
@@ -466,6 +506,16 @@ fn run() {
     bb(fixunstfsi(bb(2.)));
     #[cfg(not(any(target_arch = "powerpc", target_arch = "powerpc64")))]
     bb(fixunstfti(bb(2.)));
+    bb(floatditf(bb(2)));
+    bb(floatsitf(bb(2)));
+    bb(floattidf(bb(2)));
+    bb(floattisf(bb(2)));
+    bb(floattitf(bb(2)));
+    bb(floatunditf(bb(2)));
+    bb(floatunsitf(bb(2)));
+    bb(floatuntidf(bb(2)));
+    bb(floatuntisf(bb(2)));
+    bb(floatuntitf(bb(2)));
     bb(gttf(bb(2.), bb(2.)));
     bb(lshrti3(bb(2), bb(2)));
     bb(lttf(bb(2.), bb(2.)));

library/compiler-builtins/src/float/conv.rs

@@ -104,6 +104,24 @@ mod int_to_float {
         repr::<f64>(e, m)
     }
 
+    #[cfg(f128_enabled)]
+    pub fn u32_to_f128_bits(i: u32) -> u128 {
+        if i == 0 {
+            return 0;
+        }
+        let n = i.leading_zeros();
+
+        // Shift into mantissa position that is correct for the type, but shifted into the lower
+        // 64 bits over so can can avoid 128-bit math.
+        let m = (i as u64) << (shift_f_gt_i::<u32, f128>(n) - 64);
+        let e = exp::<u32, f128>(n) as u64 - 1;
+        // High 64 bits of f128 representation.
+        let h = (e << (f128::SIGNIFICAND_BITS - 64)) + m;
+
+        // Shift back to the high bits, the rest of the mantissa will always be 0.
+        (h as u128) << 64
+    }
+
     pub fn u64_to_f32_bits(i: u64) -> u32 {
         let n = i.leading_zeros();
         let i_m = i.wrapping_shl(n);
@@ -130,6 +148,18 @@ mod int_to_float {
         repr::<f64>(e, m)
     }
 
+    #[cfg(f128_enabled)]
+    pub fn u64_to_f128_bits(i: u64) -> u128 {
+        if i == 0 {
+            return 0;
+        }
+        let n = i.leading_zeros();
+        // Mantissa with implicit bit set
+        let m = (i as u128) << shift_f_gt_i::<u64, f128>(n);
+        let e = exp::<u64, f128>(n) - 1;
+        repr::<f128>(e, m)
+    }
+
     pub fn u128_to_f32_bits(i: u128) -> u32 {
         let n = i.leading_zeros();
         let i_m = i.wrapping_shl(n); // Mantissa, shifted so the first bit is nonzero
@@ -162,6 +192,20 @@ mod int_to_float {
         let e = if i == 0 { 0 } else { exp::<u128, f64>(n) - 1 };
         repr::<f64>(e, m)
     }
+
+    #[cfg(f128_enabled)]
+    pub fn u128_to_f128_bits(i: u128) -> u128 {
+        if i == 0 {
+            return 0;
+        }
+        let n = i.leading_zeros();
+        // Mantissa with implicit bit set
+        let m_base = (i << n) >> f128::EXPONENT_BITS;
+        let adj = (i << n) << (f128::SIGNIFICAND_BITS + 1);
+        let m = m_adj::<f128>(m_base, adj);
+        let e = exp::<u128, f128>(n) - 1;
+        repr::<f128>(e, m)
+    }
 }
 
 // Conversions from unsigned integers to floats.
@@ -195,6 +239,24 @@ intrinsics! {
     pub extern "C" fn __floatuntidf(i: u128) -> f64 {
         f64::from_bits(int_to_float::u128_to_f64_bits(i))
     }
+
+    #[ppc_alias = __floatunsikf]
+    #[cfg(f128_enabled)]
+    pub extern "C" fn __floatunsitf(i: u32) -> f128 {
+        f128::from_bits(int_to_float::u32_to_f128_bits(i))
+    }
+
+    #[ppc_alias = __floatundikf]
+    #[cfg(f128_enabled)]
+    pub extern "C" fn __floatunditf(i: u64) -> f128 {
+        f128::from_bits(int_to_float::u64_to_f128_bits(i))
+    }
+
+    #[ppc_alias = __floatuntikf]
+    #[cfg(f128_enabled)]
+    pub extern "C" fn __floatuntitf(i: u128) -> f128 {
+        f128::from_bits(int_to_float::u128_to_f128_bits(i))
+    }
 }
 
 // Conversions from signed integers to floats.
@@ -228,6 +290,24 @@ intrinsics! {
     pub extern "C" fn __floattidf(i: i128) -> f64 {
         int_to_float::signed(i, int_to_float::u128_to_f64_bits)
     }
+
+    #[ppc_alias = __floatsikf]
+    #[cfg(f128_enabled)]
+    pub extern "C" fn __floatsitf(i: i32) -> f128 {
+        int_to_float::signed(i, int_to_float::u32_to_f128_bits)
+    }
+
+    #[ppc_alias = __floatdikf]
+    #[cfg(f128_enabled)]
+    pub extern "C" fn __floatditf(i: i64) -> f128 {
+        int_to_float::signed(i, int_to_float::u64_to_f128_bits)
+    }
+
+    #[ppc_alias = __floattikf]
+    #[cfg(f128_enabled)]
+    pub extern "C" fn __floattitf(i: i128) -> f128 {
+        int_to_float::signed(i, int_to_float::u128_to_f128_bits)
+    }
 }
 
 /// Generic float to unsigned int conversions.

library/compiler-builtins/testcrate/benches/float_conv.rs

@@ -1,7 +1,8 @@
 #![allow(improper_ctypes)]
+#![cfg_attr(f128_enabled, feature(f128))]
 
 use compiler_builtins::float::conv;
-use criterion::{criterion_group, criterion_main, Criterion};
+use criterion::{criterion_main, Criterion};
 use testcrate::float_bench;
 
 /* unsigned int -> float */
@@ -76,6 +77,18 @@ float_bench! {
     ],
 }
 
+#[cfg(f128_enabled)]
+float_bench! {
+    name: conv_u32_f128,
+    sig: (a: u32) -> f128,
+    crate_fn: conv::__floatunsitf,
+    crate_fn_ppc: conv::__floatunsikf,
+    sys_fn: __floatunsitf,
+    sys_fn_ppc: __floatunsikf,
+    sys_available: not(feature = "no-sys-f16-f128-convert"),
+    asm: []
+}
+
 float_bench! {
     name: conv_u64_f32,
     sig: (a: u64) -> f32,
@@ -118,6 +131,18 @@ float_bench! {
     ],
 }
 
+#[cfg(f128_enabled)]
+float_bench! {
+    name: conv_u64_f128,
+    sig: (a: u64) -> f128,
+    crate_fn: conv::__floatunditf,
+    crate_fn_ppc: conv::__floatundikf,
+    sys_fn: __floatunditf,
+    sys_fn_ppc: __floatundikf,
+    sys_available: not(feature = "no-sys-f16-f128-convert"),
+    asm: []
+}
+
 float_bench! {
     name: conv_u128_f32,
     sig: (a: u128) -> f32,
@@ -136,6 +161,18 @@ float_bench! {
     asm: []
 }
 
+#[cfg(f128_enabled)]
+float_bench! {
+    name: conv_u128_f128,
+    sig: (a: u128) -> f128,
+    crate_fn: conv::__floatuntitf,
+    crate_fn_ppc: conv::__floatuntikf,
+    sys_fn: __floatuntitf,
+    sys_fn_ppc: __floatuntikf,
+    sys_available: not(feature = "no-sys-f16-f128-convert"),
+    asm: []
+}
+
 /* signed int -> float */
 
 float_bench! {
@@ -205,6 +242,18 @@ float_bench! {
     ],
 }
 
+#[cfg(f128_enabled)]
+float_bench! {
+    name: conv_i32_f128,
+    sig: (a: i32) -> f128,
+    crate_fn: conv::__floatsitf,
+    crate_fn_ppc: conv::__floatsikf,
+    sys_fn: __floatsitf,
+    sys_fn_ppc: __floatsikf,
+    sys_available: not(feature = "no-sys-f16-f128-convert"),
+    asm: []
+}
+
 float_bench! {
     name: conv_i64_f32,
     sig: (a: i64) -> f32,
@@ -272,6 +321,18 @@ float_bench! {
     ],
 }
 
+#[cfg(f128_enabled)]
+float_bench! {
+    name: conv_i64_f128,
+    sig: (a: i64) -> f128,
+    crate_fn: conv::__floatditf,
+    crate_fn_ppc: conv::__floatdikf,
+    sys_fn: __floatditf,
+    sys_fn_ppc: __floatdikf,
+    sys_available: not(feature = "no-sys-f16-f128-convert"),
+    asm: []
+}
+
 float_bench! {
     name: conv_i128_f32,
     sig: (a: i128) -> f32,
@@ -290,6 +351,18 @@ float_bench! {
     asm: []
 }
 
+#[cfg(f128_enabled)]
+float_bench! {
+    name: conv_i128_f128,
+    sig: (a: i128) -> f128,
+    crate_fn: conv::__floattitf,
+    crate_fn_ppc: conv::__floattikf,
+    sys_fn: __floattitf,
+    sys_fn_ppc: __floattikf,
+    sys_available: not(feature = "no-sys-f16-f128-convert"),
+    asm: []
+}
+
 /* float -> unsigned int */
 
 #[cfg(not(all(target_arch = "powerpc64", target_endian = "little")))]
@@ -397,6 +470,39 @@ float_bench! {
     asm: []
 }
 
+#[cfg(f128_enabled)]
+float_bench! {
+    name: conv_f128_u32,
+    sig: (a: f128) -> u32,
+    crate_fn: conv::__fixunstfsi,
+    crate_fn_ppc: conv::__fixunskfsi,
+    sys_fn: __fixunstfsi,
+    sys_available: not(feature = "no-sys-f16-f128-convert"),
+    asm: []
+}
+
+#[cfg(f128_enabled)]
+float_bench! {
+    name: conv_f128_u64,
+    sig: (a: f128) -> u64,
+    crate_fn: conv::__fixunstfdi,
+    crate_fn_ppc: conv::__fixunskfdi,
+    sys_fn: __fixunstfdi,
+    sys_available: not(feature = "no-sys-f16-f128-convert"),
+    asm: []
+}
+
+#[cfg(f128_enabled)]
+float_bench! {
+    name: conv_f128_u128,
+    sig: (a: f128) -> u128,
+    crate_fn: conv::__fixunstfti,
+    crate_fn_ppc: conv::__fixunskfti,
+    sys_fn: __fixunstfti,
+    sys_available: not(feature = "no-sys-f16-f128-convert"),
+    asm: []
+}
+
 /* float -> signed int */
 
 #[cfg(not(all(target_arch = "powerpc64", target_endian = "little")))]
@@ -504,43 +610,79 @@ float_bench! {
     asm: []
 }
 
-criterion_group!(
-    float_conv,
-    conv_u32_f32,
-    conv_u32_f64,
-    conv_u64_f32,
-    conv_u64_f64,
-    conv_u128_f32,
-    conv_u128_f64,
-    conv_i32_f32,
-    conv_i32_f64,
-    conv_i64_f32,
-    conv_i64_f64,
-    conv_i128_f32,
-    conv_i128_f64,
-    conv_f64_u32,
-    conv_f64_u64,
-    conv_f64_u128,
-    conv_f64_i32,
-    conv_f64_i64,
-    conv_f64_i128,
-);
+#[cfg(f128_enabled)]
+float_bench! {
+    name: conv_f128_i32,
+    sig: (a: f128) -> i32,
+    crate_fn: conv::__fixtfsi,
+    crate_fn_ppc: conv::__fixkfsi,
+    sys_fn: __fixtfsi,
+    sys_available: not(feature = "no-sys-f16-f128-convert"),
+    asm: []
+}
 
-// FIXME: ppc64le has a sporadic overflow panic in the crate functions
-// <https://github.com/rust-lang/compiler-builtins/issues/617#issuecomment-2125914639>
-#[cfg(not(all(target_arch = "powerpc64", target_endian = "little")))]
-criterion_group!(
-    float_conv_not_ppc64le,
-    conv_f32_u32,
-    conv_f32_u64,
-    conv_f32_u128,
-    conv_f32_i32,
-    conv_f32_i64,
-    conv_f32_i128,
-);
+#[cfg(f128_enabled)]
+float_bench! {
+    name: conv_f128_i64,
+    sig: (a: f128) -> i64,
+    crate_fn: conv::__fixtfdi,
+    crate_fn_ppc: conv::__fixkfdi,
+    sys_fn: __fixtfdi,
+    sys_available: not(feature = "no-sys-f16-f128-convert"),
+    asm: []
+}
+
+#[cfg(f128_enabled)]
+float_bench! {
+    name: conv_f128_i128,
+    sig: (a: f128) -> i128,
+    crate_fn: conv::__fixtfti,
+    crate_fn_ppc: conv::__fixkfti,
+    sys_fn: __fixtfti,
+    sys_available: not(feature = "no-sys-f16-f128-convert"),
+    asm: []
+}
+
+pub fn float_conv() {
+    let mut criterion = Criterion::default().configure_from_args();
+
+    conv_u32_f32(&mut criterion);
+    conv_u32_f64(&mut criterion);
+    conv_u64_f32(&mut criterion);
+    conv_u64_f64(&mut criterion);
+    conv_u128_f32(&mut criterion);
+    conv_u128_f64(&mut criterion);
+    conv_i32_f32(&mut criterion);
+    conv_i32_f64(&mut criterion);
+    conv_i64_f32(&mut criterion);
+    conv_i64_f64(&mut criterion);
+    conv_i128_f32(&mut criterion);
+    conv_i128_f64(&mut criterion);
+    conv_f64_u32(&mut criterion);
+    conv_f64_u64(&mut criterion);
+    conv_f64_u128(&mut criterion);
+    conv_f64_i32(&mut criterion);
+    conv_f64_i64(&mut criterion);
+    conv_f64_i128(&mut criterion);
+
+    #[cfg(all(f128_enabled))]
+    // FIXME: ppc64le has a sporadic overflow panic in the crate functions
+    // <https://github.com/rust-lang/compiler-builtins/issues/617#issuecomment-2125914639>
+    #[cfg(not(all(target_arch = "powerpc64", target_endian = "little")))]
+    {
+        conv_u32_f128(&mut criterion);
+        conv_u64_f128(&mut criterion);
+        conv_u128_f128(&mut criterion);
+        conv_i32_f128(&mut criterion);
+        conv_i64_f128(&mut criterion);
+        conv_i128_f128(&mut criterion);
+        conv_f128_u32(&mut criterion);
+        conv_f128_u64(&mut criterion);
+        conv_f128_u128(&mut criterion);
+        conv_f128_i32(&mut criterion);
+        conv_f128_i64(&mut criterion);
+        conv_f128_i128(&mut criterion);
+    }
+}
 
-#[cfg(all(target_arch = "powerpc64", target_endian = "little"))]
 criterion_main!(float_conv);
-
-#[cfg(not(all(target_arch = "powerpc64", target_endian = "little")))]
-criterion_main!(float_conv, float_conv_not_ppc64le);

library/compiler-builtins/testcrate/tests/conv.rs

@@ -117,6 +117,28 @@ mod i_to_f {
         u128, __floatuntidf;
         i128, __floattidf;
     }
+
+    #[cfg(not(feature = "no-f16-f128"))]
+    #[cfg(not(any(target_arch = "powerpc", target_arch = "powerpc64")))]
+    i_to_f! { f128, Quad, not(feature = "no-sys-f128-int-convert"),
+        u32, __floatunsitf;
+        i32, __floatsitf;
+        u64, __floatunditf;
+        i64, __floatditf;
+        u128, __floatuntitf;
+        i128, __floattitf;
+    }
+
+    #[cfg(not(feature = "no-f16-f128"))]
+    #[cfg(any(target_arch = "powerpc", target_arch = "powerpc64"))]
+    i_to_f! { f128, Quad, not(feature = "no-sys-f128-int-convert"),
+        u32, __floatunsikf;
+        i32, __floatsikf;
+        u64, __floatundikf;
+        i64, __floatdikf;
+        u128, __floatuntikf;
+        i128, __floattikf;
+    }
 }
 
 // PowerPC tests are failing on LLVM 13: https://github.com/rust-lang/rust/issues/88520