Merge branch 'master' into windows-ci

Author: tgross35

.release-plz.toml

@@ -7,6 +7,7 @@ publish_allow_dirty = true
 [[package]]
 name = "compiler_builtins"
 semver_check = false
+changelog_include = ["libm"] # libm is included as part of builtins
 
 [[package]]
 name = "libm"

Cargo.toml

@@ -6,6 +6,7 @@ members = [
     "crates/libm-macros",
     "crates/musl-math-sys",
     "crates/panic-handler",
+    "crates/symbol-check",
     "crates/util",
     "libm",
     "libm-test",

README.md

@@ -5,7 +5,7 @@ This repository contains two main crates:
 * `compiler-builtins`: symbols that the compiler expects to be available at
   link time
 * `libm`: a Rust implementation of C math libraries, used to provide
-  implementations in `ocre`.
+  implementations in `core`.
 
 More details are at [compiler-builtins/README.md](compiler-builtins/README.md)
 and [libm/README.md](libm/README.md).

builtins-test-intrinsics/Cargo.toml

@@ -1,7 +1,7 @@
 [package]
 name = "builtins-test-intrinsics"
 version = "0.1.0"
-edition = "2021"
+edition = "2024"
 publish = false
 license = "MIT OR Apache-2.0"
 

builtins-test-intrinsics/src/main.rs

@@ -13,11 +13,14 @@
 #![no_std]
 #![no_main]
 
+// Ensure this `compiler_builtins` gets used, rather than the version injected from the sysroot.
+extern crate compiler_builtins;
 extern crate panic_handler;
 
+// SAFETY: no definitions, only used for linking
 #[cfg(all(not(thumb), not(windows), not(target_arch = "wasm32")))]
 #[link(name = "c")]
-extern "C" {}
+unsafe extern "C" {}
 
 // Every function in this module maps will be lowered to an intrinsic by LLVM, if the platform
 // doesn't have native support for the operation used in the function. ARM has a naming convention
@@ -651,22 +654,23 @@ fn something_with_a_dtor(f: &dyn Fn()) {
 
 #[unsafe(no_mangle)]
 #[cfg(not(thumb))]
-fn main(_argc: core::ffi::c_int, _argv: *const *const u8) -> core::ffi::c_int {
+extern "C" fn main(_argc: core::ffi::c_int, _argv: *const *const u8) -> core::ffi::c_int {
     run();
     0
 }
 
 #[unsafe(no_mangle)]
 #[cfg(thumb)]
-pub fn _start() -> ! {
+extern "C" fn _start() -> ! {
     run();
     loop {}
 }
 
+// SAFETY: no definitions, only used for linking
 #[cfg(windows)]
 #[link(name = "kernel32")]
 #[link(name = "msvcrt")]
-extern "C" {}
+unsafe extern "C" {}
 
 // ARM targets need these symbols
 #[unsafe(no_mangle)]

builtins-test/benches/float_cmp.rs

@@ -1,27 +1,38 @@
 #![cfg_attr(f128_enabled, feature(f128))]
 
 use builtins_test::float_bench;
-use compiler_builtins::float::cmp;
+use compiler_builtins::float::cmp::{self, CmpResult};
 use criterion::{Criterion, criterion_main};
 
 /// `gt` symbols are allowed to return differing results, they just get compared
 /// to 0.
-fn gt_res_eq(a: i32, b: i32) -> bool {
+fn gt_res_eq(mut a: CmpResult, mut b: CmpResult) -> bool {
+    // FIXME: Our CmpResult used to be `i32`, but GCC/LLVM expect `isize`. on 64-bit platforms,
+    // this means the top half of the word may be garbage if built with an old version of
+    // `compiler-builtins`, so add a hack around this.
+    //
+    // This can be removed once a version of `compiler-builtins` with the return type fix makes
+    // it upstream.
+    if size_of::<CmpResult>() == 8 {
+        a = a as i32 as CmpResult;
+        b = b as i32 as CmpResult;
+    }
+
     let a_lt_0 = a <= 0;
     let b_lt_0 = b <= 0;
     (a_lt_0 && b_lt_0) || (!a_lt_0 && !b_lt_0)
 }
 
 float_bench! {
     name: cmp_f32_gt,
-    sig: (a: f32, b: f32) -> i32,
+    sig: (a: f32, b: f32) -> CmpResult,
     crate_fn: cmp::__gtsf2,
     sys_fn: __gtsf2,
     sys_available: all(),
     output_eq: gt_res_eq,
     asm: [
         #[cfg(target_arch = "x86_64")] {
-            let ret: i32;
+            let ret: CmpResult;
             asm!(
                 "xor     {ret:e}, {ret:e}",
                 "ucomiss {a}, {b}",
@@ -36,7 +47,7 @@ float_bench! {
         };
 
         #[cfg(target_arch = "aarch64")] {
-            let ret: i32;
+            let ret: CmpResult;
             asm!(
                 "fcmp    {a:s}, {b:s}",
                 "cset    {ret:w}, gt",
@@ -53,13 +64,13 @@ float_bench! {
 
 float_bench! {
     name: cmp_f32_unord,
-    sig: (a: f32, b: f32) -> i32,
+    sig: (a: f32, b: f32) -> CmpResult,
     crate_fn: cmp::__unordsf2,
     sys_fn: __unordsf2,
     sys_available: all(),
     asm: [
         #[cfg(target_arch = "x86_64")] {
-            let ret: i32;
+            let ret: CmpResult;
             asm!(
                 "xor     {ret:e}, {ret:e}",
                 "ucomiss {a}, {b}",
@@ -74,7 +85,7 @@ float_bench! {
         };
 
         #[cfg(target_arch = "aarch64")] {
-            let ret: i32;
+            let ret: CmpResult;
             asm!(
                 "fcmp    {a:s}, {b:s}",
                 "cset    {ret:w}, vs",
@@ -91,14 +102,14 @@ float_bench! {
 
 float_bench! {
     name: cmp_f64_gt,
-    sig: (a: f64, b: f64) -> i32,
+    sig: (a: f64, b: f64) -> CmpResult,
     crate_fn: cmp::__gtdf2,
     sys_fn: __gtdf2,
     sys_available: all(),
     output_eq: gt_res_eq,
     asm: [
         #[cfg(target_arch = "x86_64")] {
-            let ret: i32;
+            let ret: CmpResult;
             asm!(
                 "xor     {ret:e}, {ret:e}",
                 "ucomisd {a}, {b}",
@@ -113,7 +124,7 @@ float_bench! {
         };
 
         #[cfg(target_arch = "aarch64")] {
-            let ret: i32;
+            let ret: CmpResult;
             asm!(
                 "fcmp    {a:d}, {b:d}",
                 "cset {ret:w}, gt",
@@ -130,13 +141,13 @@ float_bench! {
 
 float_bench! {
     name: cmp_f64_unord,
-    sig: (a: f64, b: f64) -> i32,
+    sig: (a: f64, b: f64) -> CmpResult,
     crate_fn: cmp::__unorddf2,
     sys_fn: __unorddf2,
     sys_available: all(),
     asm: [
         #[cfg(target_arch = "x86_64")] {
-            let ret: i32;
+            let ret: CmpResult;
             asm!(
                 "xor     {ret:e}, {ret:e}",
                 "ucomisd {a}, {b}",
@@ -151,7 +162,7 @@ float_bench! {
         };
 
         #[cfg(target_arch = "aarch64")] {
-            let ret: i32;
+            let ret: CmpResult;
             asm!(
                 "fcmp    {a:d}, {b:d}",
                 "cset    {ret:w}, vs",
@@ -168,7 +179,7 @@ float_bench! {
 
 float_bench! {
     name: cmp_f128_gt,
-    sig: (a: f128, b: f128) -> i32,
+    sig: (a: f128, b: f128) -> CmpResult,
     crate_fn: cmp::__gttf2,
     crate_fn_ppc: cmp::__gtkf2,
     sys_fn: __gttf2,
@@ -180,7 +191,7 @@ float_bench! {
 
 float_bench! {
     name: cmp_f128_unord,
-    sig: (a: f128, b: f128) -> i32,
+    sig: (a: f128, b: f128) -> CmpResult,
     crate_fn: cmp::__unordtf2,
     crate_fn_ppc: cmp::__unordkf2,
     sys_fn: __unordtf2,

builtins-test/src/bench.rs

@@ -358,8 +358,8 @@ impl_testio!(float f16);
 impl_testio!(float f32, f64);
 #[cfg(f128_enabled)]
 impl_testio!(float f128);
-impl_testio!(int i16, i32, i64, i128);
-impl_testio!(int u16, u32, u64, u128);
+impl_testio!(int i8, i16, i32, i64, i128, isize);
+impl_testio!(int u8, u16, u32, u64, u128, usize);
 impl_testio!((float, int)(f32, i32));
 impl_testio!((float, int)(f64, i32));
 #[cfg(f128_enabled)]

builtins-test/src/lib.rs

@@ -40,6 +40,75 @@ pub const N: u32 = if cfg!(target_arch = "x86_64") && !cfg!(debug_assertions) {
     10_000
 };
 
+/// Additional constants that determine how the integer gets fuzzed.
+trait FuzzInt: MinInt {
+    /// LUT used for maximizing the space covered and minimizing the computational cost of fuzzing
+    /// in `builtins-test`. For example, Self = u128 produces [0,1,2,7,8,15,16,31,32,63,64,95,96,
+    /// 111,112,119,120,125,126,127].
+    const FUZZ_LENGTHS: [u8; 20] = make_fuzz_lengths(Self::BITS);
+
+    /// The number of entries of `FUZZ_LENGTHS` actually used. The maximum is 20 for u128.
+    const FUZZ_NUM: usize = {
+        let log2 = Self::BITS.ilog2() as usize;
+        if log2 == 3 {
+            // case for u8
+            6
+        } else {
+            // 3 entries on each extreme, 2 in the middle, and 4 for each scale of intermediate
+            // boundaries.
+            8 + (4 * (log2 - 4))
+        }
+    };
+}
+
+impl<I> FuzzInt for I where I: MinInt {}
+
+const fn make_fuzz_lengths(bits: u32) -> [u8; 20] {
+    let mut v = [0u8; 20];
+    v[0] = 0;
+    v[1] = 1;
+    v[2] = 2; // important for parity and the iX::MIN case when reversed
+    let mut i = 3;
+
+    // No need for any more until the byte boundary, because there should be no algorithms
+    // that are sensitive to anything not next to byte boundaries after 2. We also scale
+    // in powers of two, which is important to prevent u128 corner tests from getting too
+    // big.
+    let mut l = 8;
+    loop {
+        if l >= ((bits / 2) as u8) {
+            break;
+        }
+        // get both sides of the byte boundary
+        v[i] = l - 1;
+        i += 1;
+        v[i] = l;
+        i += 1;
+        l *= 2;
+    }
+
+    if bits != 8 {
+        // add the lower side of the middle boundary
+        v[i] = ((bits / 2) - 1) as u8;
+        i += 1;
+    }
+
+    // We do not want to jump directly from the Self::BITS/2 boundary to the Self::BITS
+    // boundary because of algorithms that split the high part up. We reverse the scaling
+    // as we go to Self::BITS.
+    let mid = i;
+    let mut j = 1;
+    loop {
+        v[i] = (bits as u8) - (v[mid - j]) - 1;
+        if j == mid {
+            break;
+        }
+        i += 1;
+        j += 1;
+    }
+    v
+}
+
 /// Random fuzzing step. When run several times, it results in excellent fuzzing entropy such as:
 /// 11110101010101011110111110011111
 /// 10110101010100001011101011001010
@@ -92,10 +161,9 @@ fn fuzz_step<I: Int>(rng: &mut Xoshiro128StarStar, x: &mut I) {
 macro_rules! edge_cases {
     ($I:ident, $case:ident, $inner:block) => {
         for i0 in 0..$I::FUZZ_NUM {
-            let mask_lo = (!$I::UnsignedInt::ZERO).wrapping_shr($I::FUZZ_LENGTHS[i0] as u32);
+            let mask_lo = (!$I::Unsigned::ZERO).wrapping_shr($I::FUZZ_LENGTHS[i0] as u32);
             for i1 in i0..I::FUZZ_NUM {
-                let mask_hi =
-                    (!$I::UnsignedInt::ZERO).wrapping_shl($I::FUZZ_LENGTHS[i1 - i0] as u32);
+                let mask_hi = (!$I::Unsigned::ZERO).wrapping_shl($I::FUZZ_LENGTHS[i1 - i0] as u32);
                 let $case = I::from_unsigned(mask_lo & mask_hi);
                 $inner
             }
@@ -107,7 +175,7 @@ macro_rules! edge_cases {
 /// edge cases, followed by a more random fuzzer that runs `n` times.
 pub fn fuzz<I: Int, F: FnMut(I)>(n: u32, mut f: F)
 where
-    <I as MinInt>::UnsignedInt: Int,
+    <I as MinInt>::Unsigned: Int,
 {
     // edge case tester. Calls `f` 210 times for u128.
     // zero gets skipped by the loop
@@ -128,7 +196,7 @@ where
 /// The same as `fuzz`, except `f` has two inputs.
 pub fn fuzz_2<I: Int, F: Fn(I, I)>(n: u32, f: F)
 where
-    <I as MinInt>::UnsignedInt: Int,
+    <I as MinInt>::Unsigned: Int,
 {
     // Check cases where the first and second inputs are zero. Both call `f` 210 times for `u128`.
     edge_cases!(I, case, {

builtins-test/tests/aeabi_memclr.rs

@@ -24,7 +24,8 @@ macro_rules! panic {
     };
 }
 
-extern "C" {
+// SAFETY: defined in  compiler-builtins
+unsafe extern "aapcs" {
     fn __aeabi_memclr4(dest: *mut u8, n: usize);
     fn __aeabi_memset4(dest: *mut u8, n: usize, c: u32);
 }

builtins-test/tests/aeabi_memcpy.rs

@@ -22,7 +22,8 @@ macro_rules! panic {
     };
 }
 
-extern "C" {
+// SAFETY: defined in  compiler-builtins
+unsafe extern "aapcs" {
     fn __aeabi_memcpy(dest: *mut u8, src: *const u8, n: usize);
     fn __aeabi_memcpy4(dest: *mut u8, src: *const u8, n: usize);
 }