Merge branch 'master' into sync_from_rust2

.github/workflows/ci.yml

@@ -11,7 +11,7 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        libgccjit_version: ["libgccjit.so", "libgccjit_without_int128.so"]
+        libgccjit_version: ["libgccjit.so", "libgccjit_without_int128.so", "libgccjit12.so"]
 
     steps:
     - uses: actions/checkout@v2
@@ -78,12 +78,21 @@ jobs:
         key: ${{ runner.os }}-cargo-build-target-${{ hashFiles('rust-toolchain') }}
 
     - name: Build
+      if: matrix.libgccjit_version != 'libgccjit12.so'
       run: |
         ./prepare_build.sh
         ./build.sh
         cargo test
         ./clean_all.sh
 
+    - name: Build
+      if: matrix.libgccjit_version == 'libgccjit12.so'
+      run: |
+        ./prepare_build.sh
+        ./build.sh --no-default-features
+        cargo test --no-default-features
+        ./clean_all.sh
+
     - name: Prepare dependencies
       run: |
         git config --global user.email "user@example.com"
@@ -98,6 +107,7 @@ jobs:
         args: --release
 
     - name: Test
+      if: matrix.libgccjit_version != 'libgccjit12.so'
       run: |
         # Enable backtraces for easier debugging
         export RUST_BACKTRACE=1
@@ -107,3 +117,15 @@ jobs:
         export RUN_RUNS=2
 
         ./test.sh --release
+
+    - name: Test
+      if: matrix.libgccjit_version == 'libgccjit12.so'
+      run: |
+        # Enable backtraces for easier debugging
+        export RUST_BACKTRACE=1
+
+        # Reduce amount of benchmark runs as they are slow
+        export COMPILE_RUNS=2
+        export RUN_RUNS=2
+
+        ./test.sh --release --no-default-features

.gitignore

@@ -13,9 +13,13 @@ perf.data.old
 /rust
 /simple-raytracer
 /regex
+/rand
 gimple*
 *asm
 res
 test-backend
 gcc_path
 benchmarks
+tools/llvm-project
+tools/llvmint
+tools/llvmint-2

.rustfmt.toml

@@ -0,0 +1 @@
+disable_all_formatting = true

Cargo.lock

@@ -41,7 +41,7 @@ dependencies = [
 [[package]]
 name = "gccjit"
 version = "1.0.0"
-source = "git+https://github.com/antoyo/gccjit.rs#bdecdecfb8a02ec861a39a350f990faa33bd31c3"
+source = "git+https://github.com/antoyo/gccjit.rs#bdb86fb5092895ff5589726b33250010c64d93f6"
 dependencies = [
  "gccjit_sys",
 ]
@@ -49,7 +49,7 @@ dependencies = [
 [[package]]
 name = "gccjit_sys"
 version = "0.0.1"
-source = "git+https://github.com/antoyo/gccjit.rs#bdecdecfb8a02ec861a39a350f990faa33bd31c3"
+source = "git+https://github.com/antoyo/gccjit.rs#bdb86fb5092895ff5589726b33250010c64d93f6"
 dependencies = [
  "libc 0.1.12",
 ]

Cargo.toml

@@ -9,9 +9,17 @@ license = "MIT OR Apache-2.0"
 crate-type = ["dylib"]
 
 [[test]]
-name = "lang_tests"
+name = "lang_tests_debug"
-path = "tests/lib.rs"
+path = "tests/lang_tests_debug.rs"
 harness = false
+[[test]]
+name = "lang_tests_release"
+path = "tests/lang_tests_release.rs"
+harness = false
+
+[features]
+default = ["master"]
+master = ["gccjit/master"]
 
 [dependencies]
 gccjit = { git = "https://github.com/antoyo/gccjit.rs" }

build.sh

@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash
 
 #set -x
 set -e
@@ -6,6 +6,8 @@ set -e
 codegen_channel=debug
 sysroot_channel=debug
 
+flags=
+
 while [[ $# -gt 0 ]]; do
     case $1 in
         --release)
@@ -16,6 +18,15 @@ while [[ $# -gt 0 ]]; do
             sysroot_channel=release
             shift
             ;;
+        --no-default-features)
+            flags="$flags --no-default-features"
+            shift
+            ;;
+        --features)
+            shift
+            flags="$flags --features $1"
+            shift
+            ;;
         *)
             echo "Unknown option $1"
             exit 1
@@ -33,21 +44,13 @@ fi
 export LD_LIBRARY_PATH="$GCC_PATH"
 export LIBRARY_PATH="$GCC_PATH"
 
-features=
-
-if [[ "$1" == "--features" ]]; then
-    shift
-    features="--features $1"
-    shift
-fi
-
 if [[ "$codegen_channel" == "release" ]]; then
     export CHANNEL='release'
-    CARGO_INCREMENTAL=1 cargo rustc --release $features
+    CARGO_INCREMENTAL=1 cargo rustc --release $flags
 else
     echo $LD_LIBRARY_PATH
     export CHANNEL='debug'
-    cargo rustc $features
+    cargo rustc $flags
 fi
 
 source config.sh

build_sysroot/build_sysroot.sh

@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash
 
 # Requires the CHANNEL env var to be set to `debug` or `release.`
 

build_sysroot/prepare_sysroot_src.sh

@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash
 set -e
 cd $(dirname "$0")
 

cargo.sh

@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash
 
 if [ -z $CHANNEL ]; then
 export CHANNEL='debug'
@@ -20,4 +20,4 @@ fi
 cmd=$1
 shift
 
-RUSTDOCFLAGS="$RUSTFLAGS" cargo +${TOOLCHAIN} $cmd --target $TARGET_TRIPLE $@
+RUSTDOCFLAGS="$RUSTFLAGS" cargo +${TOOLCHAIN} $cmd $@

clean_all.sh

@@ -1,5 +1,6 @@
-#!/bin/bash --verbose
+#!/usr/bin/env bash
 set -e
+set -v
 
 rm -rf target/ build_sysroot/{sysroot/,sysroot_src/,target/,Cargo.lock} perf.data{,.old}
 rm -rf regex/ simple-raytracer/

config.sh

@@ -2,7 +2,7 @@ set -e
 
 export CARGO_INCREMENTAL=0
 
-if [ -f ./gcc_path ]; then 
+if [ -f ./gcc_path ]; then
     export GCC_PATH=$(cat gcc_path)
 else
     echo 'Please put the path to your custom build of libgccjit in the file `gcc_path`, see Readme.md for details'
@@ -38,7 +38,7 @@ if [[ "$HOST_TRIPLE" != "$TARGET_TRIPLE" ]]; then
    fi
 fi
 
-export RUSTFLAGS="$linker -Cpanic=abort -Csymbol-mangling-version=v0 -Cdebuginfo=2 -Clto=off -Zpanic-abort-tests -Zcodegen-backend=$(pwd)/target/${CHANNEL:-debug}/librustc_codegen_gcc.$dylib_ext --sysroot $(pwd)/build_sysroot/sysroot"
+export RUSTFLAGS="$CG_RUSTFLAGS $linker -Cpanic=abort -Csymbol-mangling-version=v0 -Cdebuginfo=2 -Clto=off -Zpanic-abort-tests -Zcodegen-backend=$(pwd)/target/${CHANNEL:-debug}/librustc_codegen_gcc.$dylib_ext --sysroot $(pwd)/build_sysroot/sysroot"
 
 # FIXME(antoyo): remove once the atomic shim is gone
 if [[ `uname` == 'Darwin' ]]; then

crate_patches/0002-rand-Disable-failing-test.patch

@@ -0,0 +1,32 @@
+From a8fb97120d71252538b6b026695df40d02696bdb Mon Sep 17 00:00:00 2001
+From: bjorn3 <bjorn3@users.noreply.github.com>
+Date: Sat, 15 Aug 2020 20:04:38 +0200
+Subject: [PATCH] [rand] Disable failing test
+
+---
+ src/distributions/uniform.rs | 3 ++-
+ 1 file changed, 2 insertions(+), 1 deletion(-)
+
+diff --git a/src/distributions/uniform.rs b/src/distributions/uniform.rs
+index 480b859..c80bb6f 100644
+--- a/src/distributions/uniform.rs
++++ b/src/distributions/uniform.rs
+@@ -1085,7 +1085,7 @@ mod tests {
+             _ => panic!("`UniformDurationMode` was not serialized/deserialized correctly")
+         }
+     }
+-    
++
+     #[test]
+     #[cfg(feature = "serde1")]
+     fn test_uniform_serialization() {
+@@ -1314,6 +1314,7 @@ mod tests {
+         not(target_arch = "wasm32"),
+         not(target_arch = "asmjs")
+     ))]
++    #[ignore] // FIXME
+     fn test_float_assertions() {
+         use super::SampleUniform;
+         use std::panic::catch_unwind;
+-- 
+2.20.1

example/std_example.rs

@@ -93,9 +93,10 @@ fn main() {
 
     println!("{:?}", std::intrinsics::caller_location());
 
-    /*unsafe {
+    #[cfg(feature="master")]
+    unsafe {
         test_simd();
-    }*/
+    }
 
     Box::pin(move |mut _task_context| {
         yield ();
@@ -104,7 +105,8 @@ fn main() {
     println!("End");
 }
 
-/*#[target_feature(enable = "sse2")]
+#[cfg(feature="master")]
+#[target_feature(enable = "sse2")]
 unsafe fn test_simd() {
     let x = _mm_setzero_si128();
     let y = _mm_set1_epi16(7);
@@ -112,7 +114,7 @@ unsafe fn test_simd() {
     let cmp_eq = _mm_cmpeq_epi8(y, y);
     let cmp_lt = _mm_cmplt_epi8(y, y);
 
-    /*assert_eq!(std::mem::transmute::<_, [u16; 8]>(or), [7, 7, 7, 7, 7, 7, 7, 7]);
+    assert_eq!(std::mem::transmute::<_, [u16; 8]>(or), [7, 7, 7, 7, 7, 7, 7, 7]);
     assert_eq!(std::mem::transmute::<_, [u16; 8]>(cmp_eq), [0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff]);
     assert_eq!(std::mem::transmute::<_, [u16; 8]>(cmp_lt), [0, 0, 0, 0, 0, 0, 0, 0]);
 
@@ -124,14 +126,15 @@ unsafe fn test_simd() {
     test_mm_cvtepi8_epi16();
     test_mm_cvtsi128_si64();
 
-    // FIXME(#666) implement `#[rustc_arg_required_const(..)]` support
+    test_mm_extract_epi8();
-    //test_mm_extract_epi8();
+    test_mm_insert_epi16();
 
     let mask1 = _mm_movemask_epi8(dbg!(_mm_setr_epi8(255u8 as i8, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0)));
-    assert_eq!(mask1, 1);*/
+    assert_eq!(mask1, 1);
-}*/
+}
 
-/*#[target_feature(enable = "sse2")]
+#[cfg(feature="master")]
+#[target_feature(enable = "sse2")]
 unsafe fn test_mm_slli_si128() {
     #[rustfmt::skip]
     let a = _mm_setr_epi8(
@@ -155,22 +158,10 @@ unsafe fn test_mm_slli_si128() {
     );
     let r = _mm_slli_si128(a, 16);
     assert_eq_m128i(r, _mm_set1_epi8(0));
-
-    #[rustfmt::skip]
-    let a = _mm_setr_epi8(
-        1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16,
-    );
-    let r = _mm_slli_si128(a, -1);
-    assert_eq_m128i(_mm_set1_epi8(0), r);
-
-    #[rustfmt::skip]
-    let a = _mm_setr_epi8(
-        1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16,
-    );
-    let r = _mm_slli_si128(a, -0x80000000);
-    assert_eq_m128i(r, _mm_set1_epi8(0));
 }
 
+
+#[cfg(feature="master")]
 #[target_feature(enable = "sse2")]
 unsafe fn test_mm_movemask_epi8() {
     #[rustfmt::skip]
@@ -184,6 +175,7 @@ unsafe fn test_mm_movemask_epi8() {
     assert_eq!(r, 0b10100100_00100101);
 }
 
+#[cfg(feature="master")]
 #[target_feature(enable = "avx2")]
 unsafe fn test_mm256_movemask_epi8() {
     let a = _mm256_set1_epi8(-1);
@@ -192,6 +184,7 @@ unsafe fn test_mm256_movemask_epi8() {
     assert_eq!(r, e);
 }
 
+#[cfg(feature="master")]
 #[target_feature(enable = "sse2")]
 unsafe fn test_mm_add_epi8() {
     let a = _mm_setr_epi8(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15);
@@ -207,6 +200,7 @@ unsafe fn test_mm_add_epi8() {
     assert_eq_m128i(r, e);
 }
 
+#[cfg(feature="master")]
 #[target_feature(enable = "sse2")]
 unsafe fn test_mm_add_pd() {
     let a = _mm_setr_pd(1.0, 2.0);
@@ -215,12 +209,14 @@ unsafe fn test_mm_add_pd() {
     assert_eq_m128d(r, _mm_setr_pd(6.0, 12.0));
 }
 
+#[cfg(feature="master")]
 fn assert_eq_m128i(x: std::arch::x86_64::__m128i, y: std::arch::x86_64::__m128i) {
     unsafe {
         assert_eq!(std::mem::transmute::<_, [u8; 16]>(x), std::mem::transmute::<_, [u8; 16]>(y));
     }
 }
 
+#[cfg(feature="master")]
 #[target_feature(enable = "sse2")]
 pub unsafe fn assert_eq_m128d(a: __m128d, b: __m128d) {
     if _mm_movemask_pd(_mm_cmpeq_pd(a, b)) != 0b11 {
@@ -228,12 +224,14 @@ pub unsafe fn assert_eq_m128d(a: __m128d, b: __m128d) {
     }
 }
 
+#[cfg(feature="master")]
 #[target_feature(enable = "sse2")]
 unsafe fn test_mm_cvtsi128_si64() {
     let r = _mm_cvtsi128_si64(std::mem::transmute::<[i64; 2], _>([5, 0]));
     assert_eq!(r, 5);
 }
 
+#[cfg(feature="master")]
 #[target_feature(enable = "sse4.1")]
 unsafe fn test_mm_cvtepi8_epi16() {
     let a = _mm_set1_epi8(10);
@@ -246,6 +244,7 @@ unsafe fn test_mm_cvtepi8_epi16() {
     assert_eq_m128i(r, e);
 }
 
+#[cfg(feature="master")]
 #[target_feature(enable = "sse4.1")]
 unsafe fn test_mm_extract_epi8() {
     #[rustfmt::skip]
@@ -254,10 +253,19 @@ unsafe fn test_mm_extract_epi8() {
         8, 9, 10, 11, 12, 13, 14, 15
     );
     let r1 = _mm_extract_epi8(a, 0);
-    let r2 = _mm_extract_epi8(a, 19);
+    let r2 = _mm_extract_epi8(a, 3);
     assert_eq!(r1, 0xFF);
     assert_eq!(r2, 3);
-}*/
+}
+
+#[cfg(all(feature="master", target_arch = "x86_64"))]
+#[target_feature(enable = "sse2")]
+unsafe fn test_mm_insert_epi16() {
+    let a = _mm_setr_epi16(0, 1, 2, 3, 4, 5, 6, 7);
+    let r = _mm_insert_epi16(a, 9, 0);
+    let e = _mm_setr_epi16(9, 1, 2, 3, 4, 5, 6, 7);
+    assert_eq_m128i(r, e);
+}
 
 #[derive(PartialEq)]
 enum LoopState {

patches/0024-core-Disable-portable-simd-test.patch

@@ -7,167 +7,6 @@ Subject: [PATCH] [core] Disable portable-simd test
  library/core/tests/lib.rs | 1 -
  1 file changed, 1 deletion(-)
 
-diff --git a/library/core/src/lib.rs b/library/core/src/lib.rs
-index aa1ad93..95fbf55 100644
---- a/library/core/src/lib.rs
-+++ b/library/core/src/lib.rs
-@@ -398,23 +398,4 @@ pub mod arch {
-     }
- }
- 
--// Pull in the `core_simd` crate directly into libcore. The contents of
--// `core_simd` are in a different repository: rust-lang/portable-simd.
--//
--// `core_simd` depends on libcore, but the contents of this module are
--// set up in such a way that directly pulling it here works such that the
--// crate uses this crate as its libcore.
--#[path = "../../portable-simd/crates/core_simd/src/mod.rs"]
--#[allow(missing_debug_implementations, dead_code, unsafe_op_in_unsafe_fn, unused_unsafe)]
--#[allow(rustdoc::bare_urls)]
--#[unstable(feature = "portable_simd", issue = "86656")]
--mod core_simd;
--
--#[doc = include_str!("../../portable-simd/crates/core_simd/src/core_simd_docs.md")]
--#[unstable(feature = "portable_simd", issue = "86656")]
--pub mod simd {
--    #[unstable(feature = "portable_simd", issue = "86656")]
--    pub use crate::core_simd::simd::*;
--}
--
- include!("primitive_docs.rs");
-diff --git a/library/core/src/slice/mod.rs b/library/core/src/slice/mod.rs
-index cd38c3a..ad632dc 100644
---- a/library/core/src/slice/mod.rs
-+++ b/library/core/src/slice/mod.rs
-@@ -17,6 +17,5 @@ use crate::ptr;
- use crate::result::Result;
- use crate::result::Result::{Err, Ok};
--use crate::simd::{self, Simd};
- use crate::slice;
- 
- #[unstable(
-@@ -3475,121 +3474,6 @@ impl<T> [T] {
-         }
-     }
- 
--    /// Split a slice into a prefix, a middle of aligned SIMD types, and a suffix.
--    ///
--    /// This is a safe wrapper around [`slice::align_to`], so has the same weak
--    /// postconditions as that method.  You're only assured that
--    /// `self.len() == prefix.len() + middle.len() * LANES + suffix.len()`.
--    ///
--    /// Notably, all of the following are possible:
--    /// - `prefix.len() >= LANES`.
--    /// - `middle.is_empty()` despite `self.len() >= 3 * LANES`.
--    /// - `suffix.len() >= LANES`.
--    ///
--    /// That said, this is a safe method, so if you're only writing safe code,
--    /// then this can at most cause incorrect logic, not unsoundness.
--    ///
--    /// # Panics
--    ///
--    /// This will panic if the size of the SIMD type is different from
--    /// `LANES` times that of the scalar.
--    ///
--    /// At the time of writing, the trait restrictions on `Simd<T, LANES>` keeps
--    /// that from ever happening, as only power-of-two numbers of lanes are
--    /// supported.  It's possible that, in the future, those restrictions might
--    /// be lifted in a way that would make it possible to see panics from this
--    /// method for something like `LANES == 3`.
--    ///
--    /// # Examples
--    ///
--    /// ```
--    /// #![feature(portable_simd)]
--    ///
--    /// let short = &[1, 2, 3];
--    /// let (prefix, middle, suffix) = short.as_simd::<4>();
--    /// assert_eq!(middle, []); // Not enough elements for anything in the middle
--    ///
--    /// // They might be split in any possible way between prefix and suffix
--    /// let it = prefix.iter().chain(suffix).copied();
--    /// assert_eq!(it.collect::<Vec<_>>(), vec![1, 2, 3]);
--    ///
--    /// fn basic_simd_sum(x: &[f32]) -> f32 {
--    ///     use std::ops::Add;
--    ///     use std::simd::f32x4;
--    ///     let (prefix, middle, suffix) = x.as_simd();
--    ///     let sums = f32x4::from_array([
--    ///         prefix.iter().copied().sum(),
--    ///         0.0,
--    ///         0.0,
--    ///         suffix.iter().copied().sum(),
--    ///     ]);
--    ///     let sums = middle.iter().copied().fold(sums, f32x4::add);
--    ///     sums.reduce_sum()
--    /// }
--    ///
--    /// let numbers: Vec<f32> = (1..101).map(|x| x as _).collect();
--    /// assert_eq!(basic_simd_sum(&numbers[1..99]), 4949.0);
--    /// ```
--    #[unstable(feature = "portable_simd", issue = "86656")]
--    pub fn as_simd<const LANES: usize>(&self) -> (&[T], &[Simd<T, LANES>], &[T])
--    where
--        Simd<T, LANES>: AsRef<[T; LANES]>,
--        T: simd::SimdElement,
--        simd::LaneCount<LANES>: simd::SupportedLaneCount,
--    {
--        // These are expected to always match, as vector types are laid out like
--        // arrays per <https://llvm.org/docs/LangRef.html#vector-type>, but we
--        // might as well double-check since it'll optimize away anyhow.
--        assert_eq!(mem::size_of::<Simd<T, LANES>>(), mem::size_of::<[T; LANES]>());
--
--        // SAFETY: The simd types have the same layout as arrays, just with
--        // potentially-higher alignment, so the de-facto transmutes are sound.
--        unsafe { self.align_to() }
--    }
--
--    /// Split a slice into a prefix, a middle of aligned SIMD types, and a suffix.
--    ///
--    /// This is a safe wrapper around [`slice::align_to_mut`], so has the same weak
--    /// postconditions as that method.  You're only assured that
--    /// `self.len() == prefix.len() + middle.len() * LANES + suffix.len()`.
--    ///
--    /// Notably, all of the following are possible:
--    /// - `prefix.len() >= LANES`.
--    /// - `middle.is_empty()` despite `self.len() >= 3 * LANES`.
--    /// - `suffix.len() >= LANES`.
--    ///
--    /// That said, this is a safe method, so if you're only writing safe code,
--    /// then this can at most cause incorrect logic, not unsoundness.
--    ///
--    /// This is the mutable version of [`slice::as_simd`]; see that for examples.
--    ///
--    /// # Panics
--    ///
--    /// This will panic if the size of the SIMD type is different from
--    /// `LANES` times that of the scalar.
--    ///
--    /// At the time of writing, the trait restrictions on `Simd<T, LANES>` keeps
--    /// that from ever happening, as only power-of-two numbers of lanes are
--    /// supported.  It's possible that, in the future, those restrictions might
--    /// be lifted in a way that would make it possible to see panics from this
--    /// method for something like `LANES == 3`.
--    #[unstable(feature = "portable_simd", issue = "86656")]
--    pub fn as_simd_mut<const LANES: usize>(&mut self) -> (&mut [T], &mut [Simd<T, LANES>], &mut [T])
--    where
--        Simd<T, LANES>: AsMut<[T; LANES]>,
--        T: simd::SimdElement,
--        simd::LaneCount<LANES>: simd::SupportedLaneCount,
--    {
--        // These are expected to always match, as vector types are laid out like
--        // arrays per <https://llvm.org/docs/LangRef.html#vector-type>, but we
--        // might as well double-check since it'll optimize away anyhow.
--        assert_eq!(mem::size_of::<Simd<T, LANES>>(), mem::size_of::<[T; LANES]>());
--
--        // SAFETY: The simd types have the same layout as arrays, just with
--        // potentially-higher alignment, so the de-facto transmutes are sound.
--        unsafe { self.align_to_mut() }
--    }
--
-     /// Checks if the elements of this slice are sorted.
-     ///
-     /// That is, for each element `a` and its following element `b`, `a <= b` must hold. If the
 diff --git a/library/core/tests/lib.rs b/library/core/tests/lib.rs
 index 06c7be0..359e2e7 100644
 --- a/library/core/tests/lib.rs
@@ -188,41 +27,3 @@ index 06c7be0..359e2e7 100644
  mod slice;
  mod str;
  mod str_lossy;
-diff --git a/library/std/src/lib.rs b/library/std/src/lib.rs
-index 5dc586d..b6fc48f 100644
---- a/library/std/src/lib.rs
-+++ b/library/std/src/lib.rs
-@@ -312,6 +312,5 @@
- #![feature(panic_can_unwind)]
- #![feature(panic_unwind)]
- #![feature(platform_intrinsics)]
--#![feature(portable_simd)]
- #![feature(prelude_import)]
- #![feature(ptr_as_uninit)]
-@@ -508,23 +508,6 @@ pub mod time;
- #[unstable(feature = "once_cell", issue = "74465")]
- pub mod lazy;
- 
--// Pull in `std_float` crate  into libstd. The contents of
--// `std_float` are in a different repository: rust-lang/portable-simd.
--#[path = "../../portable-simd/crates/std_float/src/lib.rs"]
--#[allow(missing_debug_implementations, dead_code, unsafe_op_in_unsafe_fn, unused_unsafe)]
--#[allow(rustdoc::bare_urls)]
--#[unstable(feature = "portable_simd", issue = "86656")]
--mod std_float;
--
--#[doc = include_str!("../../portable-simd/crates/core_simd/src/core_simd_docs.md")]
--#[unstable(feature = "portable_simd", issue = "86656")]
--pub mod simd {
--    #[doc(inline)]
--    pub use crate::std_float::StdFloat;
--    #[doc(inline)]
--    pub use core::simd::*;
--}
--
- #[stable(feature = "futures_api", since = "1.36.0")]
- pub mod task {
-     //! Types and Traits for working with asynchronous tasks.
---
-2.26.2.7.g19db9cfb68
-

prepare.sh

@@ -1,10 +1,18 @@
-#!/bin/bash --verbose
+#!/usr/bin/env bash
 set -e
+set -v
 
 source prepare_build.sh
 
 cargo install hyperfine || echo "Skipping hyperfine install"
 
+git clone https://github.com/rust-random/rand.git || echo "rust-random/rand has already been cloned"
+pushd rand
+git checkout -- .
+git checkout 0f933f9c7176e53b2a3c7952ded484e1783f0bf1
+git am ../crate_patches/*-rand-*.patch
+popd
+
 git clone https://github.com/rust-lang/regex.git || echo "rust-lang/regex has already been cloned"
 pushd regex
 git checkout -- .

prepare_build.sh

@@ -1,4 +1,5 @@
-#!/bin/bash --verbose
+#!/usr/bin/env bash
 set -e
+set -v
 
 ./build_sysroot/prepare_sysroot_src.sh

rustc_patches/compile_test.patch

@@ -0,0 +1,14 @@
+diff --git a/src/tools/compiletest/src/header.rs b/src/tools/compiletest/src/header.rs
+index 887d27fd6dca4..2c2239f2b83d1 100644
+--- a/src/tools/compiletest/src/header.rs
++++ b/src/tools/compiletest/src/header.rs
+@@ -806,8 +806,8 @@ pub fn make_test_description<R: Read>(
+     cfg: Option<&str>,
+ ) -> test::TestDesc {
+     let mut ignore = false;
+     #[cfg(not(bootstrap))]
+-    let ignore_message: Option<String> = None;
++    let ignore_message: Option<&str> = None;
+     let mut should_fail = false;
+
+     let rustc_has_profiler_support = env::var_os("RUSTC_PROFILER_SUPPORT").is_some();

rustup.sh

@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash
 
 set -e
 

src/asm.rs

@@ -13,6 +13,7 @@ use std::borrow::Cow;
 use crate::builder::Builder;
 use crate::context::CodegenCx;
 use crate::type_of::LayoutGccExt;
+use crate::callee::get_fn;
 
 
 // Rust asm! and GCC Extended Asm semantics differ substantially.
@@ -116,7 +117,6 @@ impl<'a, 'gcc, 'tcx> AsmBuilderMethods<'tcx> for Builder<'a, 'gcc, 'tcx> {
         let asm_arch = self.tcx.sess.asm_arch.unwrap();
         let is_x86 = matches!(asm_arch, InlineAsmArch::X86 | InlineAsmArch::X86_64);
         let att_dialect = is_x86 && options.contains(InlineAsmOptions::ATT_SYNTAX);
-        let intel_dialect = is_x86 && !options.contains(InlineAsmOptions::ATT_SYNTAX);
 
         // GCC index of an output operand equals its position in the array
         let mut outputs = vec![];
@@ -348,9 +348,24 @@ impl<'a, 'gcc, 'tcx> AsmBuilderMethods<'tcx> for Builder<'a, 'gcc, 'tcx> {
                     // processed in the previous pass
                 }
 
-                InlineAsmOperandRef::Const { .. }
+                InlineAsmOperandRef::SymFn { instance } => {
-                | InlineAsmOperandRef::SymFn { .. }
+                    inputs.push(AsmInOperand {
-                | InlineAsmOperandRef::SymStatic { .. } => {
+                        constraint: "X".into(),
+                        rust_idx,
+                        val: self.cx.rvalue_as_function(get_fn(self.cx, instance))
+                            .get_address(None),
+                    });
+                }
+
+                InlineAsmOperandRef::SymStatic { def_id } => {
+                    inputs.push(AsmInOperand {
+                        constraint: "X".into(),
+                        rust_idx,
+                        val: self.cx.get_static(def_id).get_address(None),
+                    });
+                }
+
+                InlineAsmOperandRef::Const { .. } => {
                     // processed in the previous pass
                 }
             }
@@ -359,7 +374,7 @@ impl<'a, 'gcc, 'tcx> AsmBuilderMethods<'tcx> for Builder<'a, 'gcc, 'tcx> {
         // 3. Build the template string
 
         let mut template_str = String::with_capacity(estimate_template_length(template, constants_len, att_dialect));
-        if !intel_dialect {
+        if att_dialect {
             template_str.push_str(ATT_SYNTAX_INS);
         }
 
@@ -444,7 +459,7 @@ impl<'a, 'gcc, 'tcx> AsmBuilderMethods<'tcx> for Builder<'a, 'gcc, 'tcx> {
             }
         }
 
-        if !intel_dialect {
+        if att_dialect {
             template_str.push_str(INTEL_SYNTAX_INS);
         }
 
@@ -588,7 +603,7 @@ fn reg_to_gcc(reg: InlineAsmRegOrRegClass) -> ConstraintOrRegister {
             InlineAsmRegClass::X86(X86InlineAsmRegClass::xmm_reg)
             | InlineAsmRegClass::X86(X86InlineAsmRegClass::ymm_reg) => "x",
             InlineAsmRegClass::X86(X86InlineAsmRegClass::zmm_reg) => "v",
-            InlineAsmRegClass::X86(X86InlineAsmRegClass::kreg) => unimplemented!(),
+            InlineAsmRegClass::X86(X86InlineAsmRegClass::kreg) => "Yk",
             InlineAsmRegClass::X86(X86InlineAsmRegClass::kreg0) => unimplemented!(),
             InlineAsmRegClass::Wasm(WasmInlineAsmRegClass::local) => unimplemented!(),
             InlineAsmRegClass::X86(
@@ -672,8 +687,8 @@ impl<'gcc, 'tcx> AsmMethods<'tcx> for CodegenCx<'gcc, 'tcx> {
         let asm_arch = self.tcx.sess.asm_arch.unwrap();
 
         // Default to Intel syntax on x86
-        let intel_syntax = matches!(asm_arch, InlineAsmArch::X86 | InlineAsmArch::X86_64)
+        let att_dialect = matches!(asm_arch, InlineAsmArch::X86 | InlineAsmArch::X86_64)
-            && !options.contains(InlineAsmOptions::ATT_SYNTAX);
+            && options.contains(InlineAsmOptions::ATT_SYNTAX);
 
         // Build the template string
         let mut template_str = String::new();
@@ -723,11 +738,11 @@ impl<'gcc, 'tcx> AsmMethods<'tcx> for CodegenCx<'gcc, 'tcx> {
         }
 
         let template_str =
-            if intel_syntax {
+            if att_dialect {
-                format!("{}\n\t.intel_syntax noprefix", template_str)
+                format!(".att_syntax\n\t{}\n\t.intel_syntax noprefix", template_str)
             }
             else {
-                format!(".att_syntax\n\t{}\n\t.intel_syntax noprefix", template_str)
+                template_str
             };
         // NOTE: seems like gcc will put the asm in the wrong section, so set it to .text manually.
         let template_str = format!(".pushsection .text\n{}\n.popsection", template_str);

src/base.rs

@@ -78,9 +78,19 @@ pub fn compile_codegen_unit<'tcx>(tcx: TyCtxt<'tcx>, cgu_name: Symbol, supports_
         let context = Context::default();
         // TODO(antoyo): only set on x86 platforms.
         context.add_command_line_option("-masm=intel");
+        // TODO(antoyo): only add the following cli argument if the feature is supported.
+        context.add_command_line_option("-msse2");
+        context.add_command_line_option("-mavx2");
+        context.add_command_line_option("-msha");
+        context.add_command_line_option("-mpclmul");
+        // FIXME(antoyo): the following causes an illegal instruction on vmovdqu64 in std_example on my CPU.
+        // Only add if the CPU supports it.
+        //context.add_command_line_option("-mavx512f");
         for arg in &tcx.sess.opts.cg.llvm_args {
             context.add_command_line_option(arg);
         }
+        // NOTE: This is needed to compile the file src/intrinsic/archs.rs during a bootstrap of rustc.
+        context.add_command_line_option("-fno-var-tracking-assignments");
         // NOTE: an optimization (https://github.com/rust-lang/rustc_codegen_gcc/issues/53).
         context.add_command_line_option("-fno-semantic-interposition");
         // NOTE: Rust relies on LLVM not doing TBAA (https://github.com/rust-lang/unsafe-code-guidelines/issues/292).

src/builder.rs

@@ -3,11 +3,11 @@ use std::cell::Cell;
 use std::convert::TryFrom;
 use std::ops::Deref;
 
-use gccjit::FunctionType;
 use gccjit::{
     BinaryOp,
     Block,
     ComparisonOp,
+    Context,
     Function,
     LValue,
     RValue,
@@ -48,6 +48,7 @@ use rustc_target::spec::{HasTargetSpec, Target};
 
 use crate::common::{SignType, TypeReflection, type_is_pointer};
 use crate::context::CodegenCx;
+use crate::intrinsic::llvm;
 use crate::type_of::LayoutGccExt;
 
 // TODO(antoyo)
@@ -199,17 +200,28 @@ impl<'a, 'gcc, 'tcx> Builder<'a, 'gcc, 'tcx> {
             return Cow::Borrowed(args);
         }
 
+        let func_name = format!("{:?}", func_ptr);
+
         let casted_args: Vec<_> = param_types
             .into_iter()
             .zip(args.iter())
             .enumerate()
             .map(|(index, (expected_ty, &actual_val))| {
+                if llvm::ignore_arg_cast(&func_name, index, args.len()) {
+                    return actual_val;
+                }
+
                 let actual_ty = actual_val.get_type();
                 if expected_ty != actual_ty {
-                    if on_stack_param_indices.contains(&index) {
+                    if !actual_ty.is_vector() && !expected_ty.is_vector() && actual_ty.is_integral() && expected_ty.is_integral() && actual_ty.get_size() != expected_ty.get_size() {
+                        self.context.new_cast(None, actual_val, expected_ty)
+                    }
+                    else if on_stack_param_indices.contains(&index) {
                         actual_val.dereference(None).to_rvalue()
                     }
                     else {
+                        assert!(!((actual_ty.is_vector() && !expected_ty.is_vector()) || (!actual_ty.is_vector() && expected_ty.is_vector())), "{:?} ({}) -> {:?} ({}), index: {:?}[{}]", actual_ty, actual_ty.is_vector(), expected_ty, expected_ty.is_vector(), func_ptr, index);
+                        // TODO(antoyo): perhaps use __builtin_convertvector for vector casting.
                         self.bitcast(actual_val, expected_ty)
                     }
                 }
@@ -268,22 +280,20 @@ impl<'a, 'gcc, 'tcx> Builder<'a, 'gcc, 'tcx> {
         // gccjit requires to use the result of functions, even when it's not used.
         // That's why we assign the result to a local or call add_eval().
         let gcc_func = func_ptr.get_type().dyncast_function_ptr_type().expect("function ptr");
-        let mut return_type = gcc_func.get_return_type();
+        let return_type = gcc_func.get_return_type();
         let void_type = self.context.new_type::<()>();
         let current_func = self.block.get_function();
 
-        // FIXME(antoyo): As a temporary workaround for unsupported LLVM intrinsics.
-        if gcc_func.get_param_count() == 0 && format!("{:?}", func_ptr) == "__builtin_ia32_pmovmskb128" {
-            return_type = self.int_type;
-        }
-
         if return_type != void_type {
             unsafe { RETURN_VALUE_COUNT += 1 };
             let result = current_func.new_local(None, return_type, &format!("ptrReturnValue{}", unsafe { RETURN_VALUE_COUNT }));
+            let func_name = format!("{:?}", func_ptr);
+            let args = llvm::adjust_intrinsic_arguments(&self, gcc_func, args, &func_name);
             self.block.add_assignment(None, result, self.cx.context.new_call_through_ptr(None, func_ptr, &args));
             result.to_rvalue()
         }
         else {
+            #[cfg(not(feature="master"))]
             if gcc_func.get_param_count() == 0 {
                 // FIXME(antoyo): As a temporary workaround for unsupported LLVM intrinsics.
                 self.block.add_eval(None, self.cx.context.new_call_through_ptr(None, func_ptr, &[]));
@@ -291,6 +301,8 @@ impl<'a, 'gcc, 'tcx> Builder<'a, 'gcc, 'tcx> {
             else {
                 self.block.add_eval(None, self.cx.context.new_call_through_ptr(None, func_ptr, &args));
             }
+            #[cfg(feature="master")]
+            self.block.add_eval(None, self.cx.context.new_call_through_ptr(None, func_ptr, &args));
             // Return dummy value when not having return value.
             let result = current_func.new_local(None, self.isize_type, "dummyValueThatShouldNeverBeUsed");
             self.block.add_assignment(None, result, self.context.new_rvalue_from_long(self.isize_type, 0));
@@ -511,12 +523,12 @@ impl<'a, 'gcc, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'gcc, 'tcx> {
     }
 
     fn frem(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> {
-        if a.get_type() == self.cx.float_type {
+        if a.get_type().is_compatible_with(self.cx.float_type) {
             let fmodf = self.context.get_builtin_function("fmodf");
             // FIXME(antoyo): this seems to produce the wrong result.
             return self.context.new_call(None, fmodf, &[a, b]);
         }
-        assert_eq!(a.get_type(), self.cx.double_type);
+        assert_eq!(a.get_type().unqualified(), self.cx.double_type);
 
         let fmod = self.context.get_builtin_function("fmod");
         return self.context.new_call(None, fmod, &[a, b]);
@@ -632,18 +644,17 @@ impl<'a, 'gcc, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'gcc, 'tcx> {
         unimplemented!();
     }
 
-    fn load(&mut self, _ty: Type<'gcc>, ptr: RValue<'gcc>, _align: Align) -> RValue<'gcc> {
+    fn load(&mut self, pointee_ty: Type<'gcc>, ptr: RValue<'gcc>, _align: Align) -> RValue<'gcc> {
-        // TODO(antoyo): use ty.
         let block = self.llbb();
         let function = block.get_function();
         // NOTE: instead of returning the dereference here, we have to assign it to a variable in
         // the current basic block. Otherwise, it could be used in another basic block, causing a
         // dereference after a drop, for instance.
-        // TODO(antoyo): handle align.
+        // TODO(antoyo): handle align of the load instruction.
+        let ptr = self.context.new_cast(None, ptr, pointee_ty.make_pointer());
         let deref = ptr.dereference(None).to_rvalue();
-        let value_type = deref.get_type();
         unsafe { RETURN_VALUE_COUNT += 1 };
-        let loaded_value = function.new_local(None, value_type, &format!("loadedValue{}", unsafe { RETURN_VALUE_COUNT }));
+        let loaded_value = function.new_local(None, pointee_ty, &format!("loadedValue{}", unsafe { RETURN_VALUE_COUNT }));
         block.add_assignment(None, loaded_value, deref);
         loaded_value.to_rvalue()
     }
@@ -695,7 +706,11 @@ impl<'a, 'gcc, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'gcc, 'tcx> {
                 OperandValue::Ref(place.llval, Some(llextra), place.align)
             }
             else if place.layout.is_gcc_immediate() {
-                let load = self.load(place.llval.get_type(), place.llval, place.align);
+                let load = self.load(
+                    place.layout.gcc_type(self, false),
+                    place.llval,
+                    place.align,
+                );
                 if let abi::Abi::Scalar(ref scalar) = place.layout.abi {
                     scalar_load_metadata(self, load, scalar);
                 }
@@ -707,7 +722,8 @@ impl<'a, 'gcc, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'gcc, 'tcx> {
 
                 let mut load = |i, scalar: &abi::Scalar, align| {
                     let llptr = self.struct_gep(pair_type, place.llval, i as u64);
-                    let load = self.load(llptr.get_type(), llptr, align);
+                    let llty = place.layout.scalar_pair_element_gcc_type(self, i, false);
+                    let load = self.load(llty, llptr, align);
                     scalar_load_metadata(self, load, scalar);
                     if scalar.is_bool() { self.trunc(load, self.type_i1()) } else { load }
                 };
@@ -779,9 +795,16 @@ impl<'a, 'gcc, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'gcc, 'tcx> {
         self.store_with_flags(val, ptr, align, MemFlags::empty())
     }
 
-    fn store_with_flags(&mut self, val: RValue<'gcc>, ptr: RValue<'gcc>, _align: Align, _flags: MemFlags) -> RValue<'gcc> {
+    fn store_with_flags(&mut self, val: RValue<'gcc>, ptr: RValue<'gcc>, align: Align, _flags: MemFlags) -> RValue<'gcc> {
         let ptr = self.check_store(val, ptr);
-        self.llbb().add_assignment(None, ptr.dereference(None), val);
+        let destination = ptr.dereference(None);
+        // NOTE: libgccjit does not support specifying the alignment on the assignment, so we cast
+        // to type so it gets the proper alignment.
+        let destination_type = destination.to_rvalue().get_type().unqualified();
+        let aligned_type = destination_type.get_aligned(align.bytes()).make_pointer();
+        let aligned_destination = self.cx.context.new_bitcast(None, ptr, aligned_type);
+        let aligned_destination = aligned_destination.dereference(None);
+        self.llbb().add_assignment(None, aligned_destination, val);
         // TODO(antoyo): handle align and flags.
         // NOTE: dummy value here since it's never used. FIXME(antoyo): API should not return a value here?
         self.cx.context.new_rvalue_zero(self.type_i32())
@@ -953,7 +976,7 @@ impl<'a, 'gcc, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'gcc, 'tcx> {
     fn memmove(&mut self, dst: RValue<'gcc>, dst_align: Align, src: RValue<'gcc>, src_align: Align, size: RValue<'gcc>, flags: MemFlags) {
         if flags.contains(MemFlags::NONTEMPORAL) {
             // HACK(nox): This is inefficient but there is no nontemporal memmove.
-            let val = self.load(src.get_type(), src, src_align);
+            let val = self.load(src.get_type().get_pointee().expect("get_pointee"), src, src_align);
             let ptr = self.pointercast(dst, self.type_ptr_to(self.val_ty(val)));
             self.store_with_flags(val, ptr, dst_align, flags);
             return;
@@ -1269,16 +1292,183 @@ impl<'a, 'gcc, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'gcc, 'tcx> {
 }
 
 impl<'a, 'gcc, 'tcx> Builder<'a, 'gcc, 'tcx> {
+    #[cfg(feature="master")]
     pub fn shuffle_vector(&mut self, v1: RValue<'gcc>, v2: RValue<'gcc>, mask: RValue<'gcc>) -> RValue<'gcc> {
-        let return_type = v1.get_type();
+        let struct_type = mask.get_type().is_struct().expect("mask of struct type");
-        let params = [
+
-            self.context.new_parameter(None, return_type, "v1"),
+        // TODO(antoyo): use a recursive unqualified() here.
-            self.context.new_parameter(None, return_type, "v2"),
+        let vector_type = v1.get_type().unqualified().dyncast_vector().expect("vector type");
-            self.context.new_parameter(None, mask.get_type(), "mask"),
+        let element_type = vector_type.get_element_type();
-        ];
+        let vec_num_units = vector_type.get_num_units();
-        let shuffle = self.context.new_function(None, FunctionType::Extern, return_type, &params, "_mm_shuffle_epi8", false);
+
-        self.context.new_call(None, shuffle, &[v1, v2, mask])
+        let mask_num_units = struct_type.get_field_count();
+        let mut vector_elements = vec![];
+        let mask_element_type =
+            if element_type.is_integral() {
+                element_type
+            }
+            else {
+                #[cfg(feature="master")]
+                {
+                    self.cx.type_ix(element_type.get_size() as u64 * 8)
+                }
+                #[cfg(not(feature="master"))]
+                self.int_type
+            };
+        for i in 0..mask_num_units {
+            let field = struct_type.get_field(i as i32);
+            vector_elements.push(self.context.new_cast(None, mask.access_field(None, field).to_rvalue(), mask_element_type));
+        }
+
+        // NOTE: the mask needs to be the same length as the input vectors, so add the missing
+        // elements in the mask if needed.
+        for _ in mask_num_units..vec_num_units {
+            vector_elements.push(self.context.new_rvalue_zero(mask_element_type));
+        }
+
+        let array_type = self.context.new_array_type(None, element_type, vec_num_units as i32);
+        let result_type = self.context.new_vector_type(element_type, mask_num_units as u64);
+        let (v1, v2) =
+            if vec_num_units < mask_num_units {
+                // NOTE: the mask needs to be the same length as the input vectors, so join the 2
+                // vectors and create a dummy second vector.
+                // TODO(antoyo): switch to using new_vector_access.
+                let array = self.context.new_bitcast(None, v1, array_type);
+                let mut elements = vec![];
+                for i in 0..vec_num_units {
+                    elements.push(self.context.new_array_access(None, array, self.context.new_rvalue_from_int(self.int_type, i as i32)).to_rvalue());
+                }
+                // TODO(antoyo): switch to using new_vector_access.
+                let array = self.context.new_bitcast(None, v2, array_type);
+                for i in 0..(mask_num_units - vec_num_units) {
+                    elements.push(self.context.new_array_access(None, array, self.context.new_rvalue_from_int(self.int_type, i as i32)).to_rvalue());
+                }
+                let v1 = self.context.new_rvalue_from_vector(None, result_type, &elements);
+                let zero = self.context.new_rvalue_zero(element_type);
+                let v2 = self.context.new_rvalue_from_vector(None, result_type, &vec![zero; mask_num_units]);
+                (v1, v2)
+            }
+            else {
+                (v1, v2)
+            };
+
+        let new_mask_num_units = std::cmp::max(mask_num_units, vec_num_units);
+        let mask_type = self.context.new_vector_type(mask_element_type, new_mask_num_units as u64);
+        let mask = self.context.new_rvalue_from_vector(None, mask_type, &vector_elements);
+        let result = self.context.new_rvalue_vector_perm(None, v1, v2, mask);
+
+        if vec_num_units != mask_num_units {
+            // NOTE: if padding was added, only select the number of elements of the masks to
+            // remove that padding in the result.
+            let mut elements = vec![];
+            // TODO(antoyo): switch to using new_vector_access.
+            let array = self.context.new_bitcast(None, result, array_type);
+            for i in 0..mask_num_units {
+                elements.push(self.context.new_array_access(None, array, self.context.new_rvalue_from_int(self.int_type, i as i32)).to_rvalue());
+            }
+            self.context.new_rvalue_from_vector(None, result_type, &elements)
+        }
+        else {
+            result
+        }
     }
+
+    #[cfg(not(feature="master"))]
+    pub fn shuffle_vector(&mut self, _v1: RValue<'gcc>, _v2: RValue<'gcc>, _mask: RValue<'gcc>) -> RValue<'gcc> {
+        unimplemented!();
+    }
+
+    #[cfg(feature="master")]
+    pub fn vector_reduce<F>(&mut self, src: RValue<'gcc>, op: F) -> RValue<'gcc>
+    where F: Fn(RValue<'gcc>, RValue<'gcc>, &'gcc Context<'gcc>) -> RValue<'gcc>
+    {
+        let vector_type = src.get_type().unqualified().dyncast_vector().expect("vector type");
+        let element_count = vector_type.get_num_units();
+        let mut vector_elements = vec![];
+        for i in 0..element_count {
+            vector_elements.push(i);
+        }
+        let mask_type = self.context.new_vector_type(self.int_type, element_count as u64);
+        let mut shift = 1;
+        let mut res = src;
+        while shift < element_count {
+            let vector_elements: Vec<_> =
+                vector_elements.iter()
+                    .map(|i| self.context.new_rvalue_from_int(self.int_type, ((i + shift) % element_count) as i32))
+                    .collect();
+            let mask = self.context.new_rvalue_from_vector(None, mask_type, &vector_elements);
+            let shifted = self.context.new_rvalue_vector_perm(None, res, res, mask);
+            shift *= 2;
+            res = op(res, shifted, &self.context);
+        }
+        self.context.new_vector_access(None, res, self.context.new_rvalue_zero(self.int_type))
+            .to_rvalue()
+    }
+
+    #[cfg(not(feature="master"))]
+    pub fn vector_reduce<F>(&mut self, src: RValue<'gcc>, op: F) -> RValue<'gcc>
+    where F: Fn(RValue<'gcc>, RValue<'gcc>, &'gcc Context<'gcc>) -> RValue<'gcc>
+    {
+        unimplemented!();
+    }
+
+    pub fn vector_reduce_op(&mut self, src: RValue<'gcc>, op: BinaryOp) -> RValue<'gcc> {
+        self.vector_reduce(src, |a, b, context| context.new_binary_op(None, op, a.get_type(), a, b))
+    }
+
+    pub fn vector_reduce_fadd_fast(&mut self, _acc: RValue<'gcc>, _src: RValue<'gcc>) -> RValue<'gcc> {
+        unimplemented!();
+    }
+
+    pub fn vector_reduce_fmul_fast(&mut self, _acc: RValue<'gcc>, _src: RValue<'gcc>) -> RValue<'gcc> {
+        unimplemented!();
+    }
+
+    // Inspired by Hacker's Delight min implementation.
+    pub fn vector_reduce_min(&mut self, src: RValue<'gcc>) -> RValue<'gcc> {
+        self.vector_reduce(src, |a, b, context| {
+            let differences_or_zeros = difference_or_zero(a, b, context);
+            context.new_binary_op(None, BinaryOp::Minus, a.get_type(), a, differences_or_zeros)
+        })
+    }
+
+    // Inspired by Hacker's Delight max implementation.
+    pub fn vector_reduce_max(&mut self, src: RValue<'gcc>) -> RValue<'gcc> {
+        self.vector_reduce(src, |a, b, context| {
+            let differences_or_zeros = difference_or_zero(a, b, context);
+            context.new_binary_op(None, BinaryOp::Plus, b.get_type(), b, differences_or_zeros)
+        })
+    }
+
+    pub fn vector_select(&mut self, cond: RValue<'gcc>, then_val: RValue<'gcc>, else_val: RValue<'gcc>) -> RValue<'gcc> {
+        // cond is a vector of integers, not of bools.
+        let cond_type = cond.get_type();
+        let vector_type = cond_type.unqualified().dyncast_vector().expect("vector type");
+        let num_units = vector_type.get_num_units();
+        let element_type = vector_type.get_element_type();
+        let zeros = vec![self.context.new_rvalue_zero(element_type); num_units];
+        let zeros = self.context.new_rvalue_from_vector(None, cond_type, &zeros);
+
+        let masks = self.context.new_comparison(None, ComparisonOp::NotEquals, cond, zeros);
+        let then_vals = masks & then_val;
+
+        let ones = vec![self.context.new_rvalue_one(element_type); num_units];
+        let ones = self.context.new_rvalue_from_vector(None, cond_type, &ones);
+        let inverted_masks = masks + ones;
+        // NOTE: sometimes, the type of else_val can be different than the type of then_val in
+        // libgccjit (vector of int vs vector of int32_t), but they should be the same for the AND
+        // operation to work.
+        let else_val = self.context.new_bitcast(None, else_val, then_val.get_type());
+        let else_vals = inverted_masks & else_val;
+
+        then_vals | else_vals
+    }
+}
+
+fn difference_or_zero<'gcc>(a: RValue<'gcc>, b: RValue<'gcc>, context: &'gcc Context<'gcc>) -> RValue<'gcc> {
+    let difference = a - b;
+    let masks = context.new_comparison(None, ComparisonOp::GreaterThanEquals, b, a);
+    difference & masks
 }
 
 impl<'a, 'gcc, 'tcx> StaticBuilderMethods for Builder<'a, 'gcc, 'tcx> {

src/common.rs

@@ -121,8 +121,8 @@ impl<'gcc, 'tcx> ConstMethods<'tcx> for CodegenCx<'gcc, 'tcx> {
         unimplemented!();
     }
 
-    fn const_real(&self, _t: Type<'gcc>, _val: f64) -> RValue<'gcc> {
+    fn const_real(&self, typ: Type<'gcc>, val: f64) -> RValue<'gcc> {
-        unimplemented!();
+        self.context.new_rvalue_from_double(typ, val)
     }
 
     fn const_str(&self, s: Symbol) -> (RValue<'gcc>, RValue<'gcc>) {
@@ -326,6 +326,8 @@ pub trait TypeReflection<'gcc, 'tcx>  {
 
     fn is_f32(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool;
     fn is_f64(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool;
+
+    fn is_vector(&self) -> bool;
 }
 
 impl<'gcc, 'tcx> TypeReflection<'gcc, 'tcx> for Type<'gcc> {
@@ -396,4 +398,21 @@ impl<'gcc, 'tcx> TypeReflection<'gcc, 'tcx> for Type<'gcc> {
     fn is_f64(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool {
         self.unqualified() == cx.context.new_type::<f64>()
     }
+
+    fn is_vector(&self) -> bool {
+        let mut typ = self.clone();
+        loop {
+            if typ.dyncast_vector().is_some() {
+                return true;
+            }
+
+            let old_type = typ;
+            typ = typ.unqualified();
+            if old_type == typ {
+                break;
+            }
+        }
+
+        false
+    }
 }

src/consts.rs

@@ -25,7 +25,9 @@ impl<'gcc, 'tcx> CodegenCx<'gcc, 'tcx> {
                 }
             }
         }
-        self.context.new_bitcast(None, value, typ)
+        // NOTE: since bitcast makes a value non-constant, don't bitcast if not necessary as some
+        // SIMD builtins require a constant value.
+        self.bitcast_if_needed(value, typ)
     }
 }
 
@@ -45,7 +47,10 @@ impl<'gcc, 'tcx> StaticMethods for CodegenCx<'gcc, 'tcx> {
             }
         }
         let global_value = self.static_addr_of_mut(cv, align, kind);
-        // TODO(antoyo): set global constant.
+        #[cfg(feature = "master")]
+        self.global_lvalues.borrow().get(&global_value)
+            .expect("`static_addr_of_mut` did not add the global to `self.global_lvalues`")
+            .global_set_readonly();
         self.const_globals.borrow_mut().insert(cv, global_value);
         global_value
     }
@@ -79,20 +84,15 @@ impl<'gcc, 'tcx> StaticMethods for CodegenCx<'gcc, 'tcx> {
 
         // TODO(antoyo): set alignment.
 
-        let value =
+        let value = self.bitcast_if_needed(value, gcc_type);
-            if value.get_type() != gcc_type {
-                self.context.new_bitcast(None, value, gcc_type)
-            }
-            else {
-                value
-            };
         global.global_set_initializer_rvalue(value);
 
         // As an optimization, all shared statics which do not have interior
         // mutability are placed into read-only memory.
         if !is_mutable {
             if self.type_is_freeze(ty) {
-                // TODO(antoyo): set global constant.
+                #[cfg(feature = "master")]
+                global.global_set_readonly();
             }
         }
 
@@ -171,8 +171,9 @@ impl<'gcc, 'tcx> CodegenCx<'gcc, 'tcx> {
                 Some(kind) if !self.tcx.sess.fewer_names() => {
                     let name = self.generate_local_symbol_name(kind);
                     // TODO(antoyo): check if it's okay that no link_section is set.
-                    // TODO(antoyo): set alignment here as well.
+
-                    let global = self.declare_private_global(&name[..], self.val_ty(cv));
+                    let typ = self.val_ty(cv).get_aligned(align.bytes());
+                    let global = self.declare_private_global(&name[..], typ);
                     global
                 }
                 _ => {

src/context.rs

@@ -35,6 +35,7 @@ pub struct CodegenCx<'gcc, 'tcx> {
     pub normal_function_addresses: RefCell<FxHashSet<RValue<'gcc>>>,
 
     pub functions: RefCell<FxHashMap<String, Function<'gcc>>>,
+    pub intrinsics: RefCell<FxHashMap<String, Function<'gcc>>>,
 
     pub tls_model: gccjit::TlsModel,
 
@@ -184,6 +185,7 @@ impl<'gcc, 'tcx> CodegenCx<'gcc, 'tcx> {
             current_func: RefCell::new(None),
             normal_function_addresses: Default::default(),
             functions: RefCell::new(functions),
+            intrinsics: RefCell::new(FxHashMap::default()),
 
             tls_model,
 
@@ -269,16 +271,25 @@ impl<'gcc, 'tcx> CodegenCx<'gcc, 'tcx> {
     }
 
     pub fn is_native_int_type_or_bool(&self, typ: Type<'gcc>) -> bool {
-        self.is_native_int_type(typ) || typ == self.bool_type
+        self.is_native_int_type(typ) || typ.is_compatible_with(self.bool_type)
     }
 
     pub fn is_int_type_or_bool(&self, typ: Type<'gcc>) -> bool {
-        self.is_native_int_type(typ) || self.is_non_native_int_type(typ) || typ == self.bool_type
+        self.is_native_int_type(typ) || self.is_non_native_int_type(typ) || typ.is_compatible_with(self.bool_type)
     }
 
     pub fn sess(&self) -> &Session {
         &self.tcx.sess
     }
+
+    pub fn bitcast_if_needed(&self, value: RValue<'gcc>, expected_type: Type<'gcc>) -> RValue<'gcc> {
+        if value.get_type() != expected_type {
+            self.context.new_bitcast(None, value, expected_type)
+        }
+        else {
+            value
+        }
+    }
 }
 
 impl<'gcc, 'tcx> BackendTypes for CodegenCx<'gcc, 'tcx> {
@@ -306,8 +317,16 @@ impl<'gcc, 'tcx> MiscMethods<'tcx> for CodegenCx<'gcc, 'tcx> {
     }
 
     fn get_fn_addr(&self, instance: Instance<'tcx>) -> RValue<'gcc> {
-        let func = get_fn(self, instance);
+        let func_name = self.tcx.symbol_name(instance).name;
-        let func = self.rvalue_as_function(func);
+
+        let func =
+            if self.intrinsics.borrow().contains_key(func_name) {
+                self.intrinsics.borrow()[func_name].clone()
+            }
+            else {
+                let func = get_fn(self, instance);
+                self.rvalue_as_function(func)
+            };
         let ptr = func.get_address(None);
 
         // TODO(antoyo): don't do this twice: i.e. in declare_fn and here.

src/declare.rs

@@ -11,7 +11,7 @@ use crate::intrinsic::llvm;
 impl<'gcc, 'tcx> CodegenCx<'gcc, 'tcx> {
     pub fn get_or_insert_global(&self, name: &str, ty: Type<'gcc>, is_tls: bool, link_section: Option<Symbol>) -> LValue<'gcc> {
         if self.globals.borrow().contains_key(name) {
-            let typ = self.globals.borrow().get(name).expect("global").get_type();
+            let typ = self.globals.borrow()[name].get_type();
             let global = self.context.new_global(None, GlobalKind::Imported, typ, name);
             if is_tls {
                 global.set_tls_model(self.tls_model);
@@ -103,11 +103,13 @@ impl<'gcc, 'tcx> CodegenCx<'gcc, 'tcx> {
 /// update the declaration and return existing Value instead.
 fn declare_raw_fn<'gcc>(cx: &CodegenCx<'gcc, '_>, name: &str, _callconv: () /*llvm::CallConv*/, return_type: Type<'gcc>, param_types: &[Type<'gcc>], variadic: bool) -> Function<'gcc> {
     if name.starts_with("llvm.") {
-        return llvm::intrinsic(name, cx);
+        let intrinsic = llvm::intrinsic(name, cx);
+        cx.intrinsics.borrow_mut().insert(name.to_string(), intrinsic);
+        return intrinsic;
     }
     let func =
         if cx.functions.borrow().contains_key(name) {
-            *cx.functions.borrow().get(name).expect("function")
+            cx.functions.borrow()[name]
         }
         else {
             let params: Vec<_> = param_types.into_iter().enumerate()

src/int.rs

@@ -153,8 +153,15 @@ impl<'a, 'gcc, 'tcx> Builder<'a, 'gcc, 'tcx> {
         let a_type = a.get_type();
         let b_type = b.get_type();
         if self.is_native_int_type_or_bool(a_type) && self.is_native_int_type_or_bool(b_type) {
-            if a.get_type() != b.get_type() {
+            if a_type != b_type {
-                b = self.context.new_cast(None, b, a.get_type());
+                if a_type.is_vector() {
+                    // Vector types need to be bitcast.
+                    // TODO(antoyo): perhaps use __builtin_convertvector for vector casting.
+                    b = self.context.new_bitcast(None, b, a.get_type());
+                }
+                else {
+                    b = self.context.new_cast(None, b, a.get_type());
+                }
             }
             self.context.new_binary_op(None, operation, a_type, a, b)
         }
@@ -593,7 +600,10 @@ impl<'gcc, 'tcx> CodegenCx<'gcc, 'tcx> {
         let b_type = b.get_type();
         let a_native = self.is_native_int_type_or_bool(a_type);
         let b_native = self.is_native_int_type_or_bool(b_type);
-        if a_native && b_native {
+        if a_type.is_vector() && b_type.is_vector() {
+            self.context.new_binary_op(None, operation, a_type, a, b)
+        }
+        else if a_native && b_native {
             if a_type != b_type {
                 b = self.context.new_cast(None, b, a_type);
             }
@@ -639,6 +649,8 @@ impl<'gcc, 'tcx> CodegenCx<'gcc, 'tcx> {
         else {
             // Since u128 and i128 are the only types that can be unsupported, we know the type of
             // value and the destination type have the same size, so a bitcast is fine.
+
+            // TODO(antoyo): perhaps use __builtin_convertvector for vector casting.
             self.context.new_bitcast(None, value, dest_typ)
         }
     }

src/intrinsic/llvm.rs

@@ -1,22 +1,250 @@
-use gccjit::Function;
+use std::borrow::Cow;
 
-use crate::context::CodegenCx;
+use gccjit::{Function, FunctionPtrType, RValue, ToRValue};
 
-pub fn intrinsic<'gcc, 'tcx>(name: &str, cx: &CodegenCx<'gcc, 'tcx>) -> Function<'gcc> {
+use crate::{context::CodegenCx, builder::Builder};
-    let _gcc_name =
-        match name {
-            "llvm.x86.xgetbv" => {
-                let gcc_name = "__builtin_trap";
-                let func = cx.context.get_builtin_function(gcc_name);
-                cx.functions.borrow_mut().insert(gcc_name.to_string(), func);
-                return func;
-            },
-            // NOTE: this doc specifies the equivalent GCC builtins: http://huonw.github.io/llvmint/llvmint/x86/index.html
-            "llvm.x86.sse2.cmp.pd" => "__builtin_ia32_cmppd",
-            "llvm.x86.sse2.movmsk.pd" => "__builtin_ia32_movmskpd",
-            "llvm.x86.sse2.pmovmskb.128" => "__builtin_ia32_pmovmskb128",
-            _ => unimplemented!("unsupported LLVM intrinsic {}", name)
-        };
 
-    unimplemented!();
+pub fn adjust_intrinsic_arguments<'a, 'b, 'gcc, 'tcx>(builder: &Builder<'a, 'gcc, 'tcx>, gcc_func: FunctionPtrType<'gcc>, mut args: Cow<'b, [RValue<'gcc>]>, func_name: &str) -> Cow<'b, [RValue<'gcc>]> {
+    // Some LLVM intrinsics do not map 1-to-1 to GCC intrinsics, so we add the missing
+    // arguments here.
+    if gcc_func.get_param_count() != args.len() {
+        match &*func_name {
+            "__builtin_ia32_pmuldq512_mask" | "__builtin_ia32_pmuludq512_mask"
+                // FIXME(antoyo): the following intrinsics has 4 (or 5) arguments according to the doc, but is defined with 2 (or 3) arguments in library/stdarch/crates/core_arch/src/x86/avx512f.rs.
+                | "__builtin_ia32_pmaxsd512_mask" | "__builtin_ia32_pmaxsq512_mask" | "__builtin_ia32_pmaxsq256_mask"
+                | "__builtin_ia32_pmaxsq128_mask" | "__builtin_ia32_maxps512_mask" | "__builtin_ia32_maxpd512_mask"
+                | "__builtin_ia32_pmaxud512_mask" | "__builtin_ia32_pmaxuq512_mask" | "__builtin_ia32_pmaxuq256_mask"
+                | "__builtin_ia32_pmaxuq128_mask"
+                | "__builtin_ia32_pminsd512_mask" | "__builtin_ia32_pminsq512_mask" | "__builtin_ia32_pminsq256_mask"
+                | "__builtin_ia32_pminsq128_mask" | "__builtin_ia32_minps512_mask" | "__builtin_ia32_minpd512_mask"
+                | "__builtin_ia32_pminud512_mask" | "__builtin_ia32_pminuq512_mask" | "__builtin_ia32_pminuq256_mask"
+                | "__builtin_ia32_pminuq128_mask" | "__builtin_ia32_sqrtps512_mask" | "__builtin_ia32_sqrtpd512_mask"
+                => {
+                    // TODO: refactor by separating those intrinsics outside of this branch.
+                    let add_before_last_arg =
+                        match &*func_name {
+                            "__builtin_ia32_maxps512_mask" | "__builtin_ia32_maxpd512_mask"
+                                | "__builtin_ia32_minps512_mask" | "__builtin_ia32_minpd512_mask"
+                                | "__builtin_ia32_sqrtps512_mask" | "__builtin_ia32_sqrtpd512_mask" => true,
+                            _ => false,
+                        };
+                    let new_first_arg_is_zero =
+                        match &*func_name {
+                            "__builtin_ia32_pmaxuq256_mask" | "__builtin_ia32_pmaxuq128_mask"
+                                | "__builtin_ia32_pminuq256_mask" | "__builtin_ia32_pminuq128_mask" => true,
+                            _ => false
+                        };
+                    let arg3_index =
+                        match &*func_name {
+                            "__builtin_ia32_sqrtps512_mask" | "__builtin_ia32_sqrtpd512_mask" => 1,
+                            _ => 2,
+                        };
+                    let mut new_args = args.to_vec();
+                    let arg3_type = gcc_func.get_param_type(arg3_index);
+                    let first_arg =
+                        if new_first_arg_is_zero {
+                            let vector_type = arg3_type.dyncast_vector().expect("vector type");
+                            let zero = builder.context.new_rvalue_zero(vector_type.get_element_type());
+                            let num_units = vector_type.get_num_units();
+                            builder.context.new_rvalue_from_vector(None, arg3_type, &vec![zero; num_units])
+                        }
+                        else {
+                            builder.current_func().new_local(None, arg3_type, "undefined_for_intrinsic").to_rvalue()
+                        };
+                    if add_before_last_arg {
+                        new_args.insert(new_args.len() - 1, first_arg);
+                    }
+                    else {
+                        new_args.push(first_arg);
+                    }
+                    let arg4_index =
+                        match &*func_name {
+                            "__builtin_ia32_sqrtps512_mask" | "__builtin_ia32_sqrtpd512_mask" => 2,
+                            _ => 3,
+                        };
+                    let arg4_type = gcc_func.get_param_type(arg4_index);
+                    let minus_one = builder.context.new_rvalue_from_int(arg4_type, -1);
+                    if add_before_last_arg {
+                        new_args.insert(new_args.len() - 1, minus_one);
+                    }
+                    else {
+                        new_args.push(minus_one);
+                    }
+                    args = new_args.into();
+                },
+                "__builtin_ia32_pternlogd512_mask" | "__builtin_ia32_pternlogd256_mask"
+                    | "__builtin_ia32_pternlogd128_mask" | "__builtin_ia32_pternlogq512_mask"
+                    | "__builtin_ia32_pternlogq256_mask" | "__builtin_ia32_pternlogq128_mask" => {
+                        let mut new_args = args.to_vec();
+                        let arg5_type = gcc_func.get_param_type(4);
+                        let minus_one = builder.context.new_rvalue_from_int(arg5_type, -1);
+                        new_args.push(minus_one);
+                        args = new_args.into();
+                    },
+                    "__builtin_ia32_vfmaddps512_mask" | "__builtin_ia32_vfmaddpd512_mask" => {
+                        let mut new_args = args.to_vec();
+
+                        let mut last_arg = None;
+                        if args.len() == 4 {
+                            last_arg = new_args.pop();
+                        }
+
+                        let arg4_type = gcc_func.get_param_type(3);
+                        let minus_one = builder.context.new_rvalue_from_int(arg4_type, -1);
+                        new_args.push(minus_one);
+
+                        if args.len() == 3 {
+                            // Both llvm.fma.v16f32 and llvm.x86.avx512.vfmadd.ps.512 maps to
+                            // the same GCC intrinsic, but the former has 3 parameters and the
+                            // latter has 4 so it doesn't require this additional argument.
+                            let arg5_type = gcc_func.get_param_type(4);
+                            new_args.push(builder.context.new_rvalue_from_int(arg5_type, 4));
+                        }
+
+                        if let Some(last_arg) = last_arg {
+                            new_args.push(last_arg);
+                        }
+
+                        args = new_args.into();
+                    },
+                    "__builtin_ia32_addps512_mask" | "__builtin_ia32_addpd512_mask"
+                        | "__builtin_ia32_subps512_mask" | "__builtin_ia32_subpd512_mask"
+                        | "__builtin_ia32_mulps512_mask" | "__builtin_ia32_mulpd512_mask"
+                        | "__builtin_ia32_divps512_mask" | "__builtin_ia32_divpd512_mask" => {
+                        let mut new_args = args.to_vec();
+                        let last_arg = new_args.pop().expect("last arg");
+                        let arg3_type = gcc_func.get_param_type(2);
+                        let undefined = builder.current_func().new_local(None, arg3_type, "undefined_for_intrinsic").to_rvalue();
+                        new_args.push(undefined);
+                        let arg4_type = gcc_func.get_param_type(3);
+                        let minus_one = builder.context.new_rvalue_from_int(arg4_type, -1);
+                        new_args.push(minus_one);
+                        new_args.push(last_arg);
+                        args = new_args.into();
+                    },
+                    "__builtin_ia32_vfmaddsubps512_mask" | "__builtin_ia32_vfmaddsubpd512_mask" => {
+                        let mut new_args = args.to_vec();
+                        let last_arg = new_args.pop().expect("last arg");
+                        let arg4_type = gcc_func.get_param_type(3);
+                        let minus_one = builder.context.new_rvalue_from_int(arg4_type, -1);
+                        new_args.push(minus_one);
+                        new_args.push(last_arg);
+                        args = new_args.into();
+                    },
+                    _ => (),
+        }
+    }
+
+    args
+}
+
+pub fn ignore_arg_cast(func_name: &str, index: usize, args_len: usize) -> bool {
+    // NOTE: these intrinsics have missing parameters before the last one, so ignore the
+    // last argument type check.
+    // FIXME(antoyo): find a way to refactor in order to avoid this hack.
+    match func_name {
+        "__builtin_ia32_maxps512_mask" | "__builtin_ia32_maxpd512_mask"
+            | "__builtin_ia32_minps512_mask" | "__builtin_ia32_minpd512_mask" | "__builtin_ia32_sqrtps512_mask"
+            | "__builtin_ia32_sqrtpd512_mask" | "__builtin_ia32_addps512_mask" | "__builtin_ia32_addpd512_mask"
+            | "__builtin_ia32_subps512_mask" | "__builtin_ia32_subpd512_mask"
+            | "__builtin_ia32_mulps512_mask" | "__builtin_ia32_mulpd512_mask"
+            | "__builtin_ia32_divps512_mask" | "__builtin_ia32_divpd512_mask"
+            | "__builtin_ia32_vfmaddsubps512_mask" | "__builtin_ia32_vfmaddsubpd512_mask" => {
+                if index == args_len - 1 {
+                    return true;
+                }
+            },
+        "__builtin_ia32_vfmaddps512_mask" | "__builtin_ia32_vfmaddpd512_mask" => {
+            // Since there are two LLVM intrinsics that map to each of these GCC builtins and only
+            // one of them has a missing parameter before the last one, we check the number of
+            // arguments to distinguish those cases.
+            if args_len == 4 && index == args_len - 1 {
+                return true;
+            }
+        },
+        _ => (),
+    }
+
+    false
+}
+
+#[cfg(not(feature="master"))]
+pub fn intrinsic<'gcc, 'tcx>(name: &str, cx: &CodegenCx<'gcc, 'tcx>) -> Function<'gcc> {
+    match name {
+        "llvm.x86.xgetbv" => {
+            let gcc_name = "__builtin_trap";
+            let func = cx.context.get_builtin_function(gcc_name);
+            cx.functions.borrow_mut().insert(gcc_name.to_string(), func);
+            return func;
+        },
+        _ => unimplemented!("unsupported LLVM intrinsic {}", name),
+    }
+}
+
+#[cfg(feature="master")]
+pub fn intrinsic<'gcc, 'tcx>(name: &str, cx: &CodegenCx<'gcc, 'tcx>) -> Function<'gcc> {
+    let gcc_name = match name {
+        "llvm.x86.xgetbv" => "__builtin_ia32_xgetbv",
+        // NOTE: this doc specifies the equivalent GCC builtins: http://huonw.github.io/llvmint/llvmint/x86/index.html
+        "llvm.sqrt.v2f64" => "__builtin_ia32_sqrtpd",
+        "llvm.x86.avx512.pmul.dq.512" => "__builtin_ia32_pmuldq512_mask",
+        "llvm.x86.avx512.pmulu.dq.512" => "__builtin_ia32_pmuludq512_mask",
+        "llvm.x86.avx512.mask.pmaxs.q.256" => "__builtin_ia32_pmaxsq256_mask",
+        "llvm.x86.avx512.mask.pmaxs.q.128" => "__builtin_ia32_pmaxsq128_mask",
+        "llvm.x86.avx512.max.ps.512" => "__builtin_ia32_maxps512_mask",
+        "llvm.x86.avx512.max.pd.512" => "__builtin_ia32_maxpd512_mask",
+        "llvm.x86.avx512.mask.pmaxu.q.256" => "__builtin_ia32_pmaxuq256_mask",
+        "llvm.x86.avx512.mask.pmaxu.q.128" => "__builtin_ia32_pmaxuq128_mask",
+        "llvm.x86.avx512.mask.pmins.q.256" => "__builtin_ia32_pminsq256_mask",
+        "llvm.x86.avx512.mask.pmins.q.128" => "__builtin_ia32_pminsq128_mask",
+        "llvm.x86.avx512.min.ps.512" => "__builtin_ia32_minps512_mask",
+        "llvm.x86.avx512.min.pd.512" => "__builtin_ia32_minpd512_mask",
+        "llvm.x86.avx512.mask.pminu.q.256" => "__builtin_ia32_pminuq256_mask",
+        "llvm.x86.avx512.mask.pminu.q.128" => "__builtin_ia32_pminuq128_mask",
+        "llvm.fma.v16f32" => "__builtin_ia32_vfmaddps512_mask",
+        "llvm.fma.v8f64" => "__builtin_ia32_vfmaddpd512_mask",
+        "llvm.x86.avx512.vfmaddsub.ps.512" => "__builtin_ia32_vfmaddsubps512_mask",
+        "llvm.x86.avx512.vfmaddsub.pd.512" => "__builtin_ia32_vfmaddsubpd512_mask",
+        "llvm.x86.avx512.pternlog.d.512" => "__builtin_ia32_pternlogd512_mask",
+        "llvm.x86.avx512.pternlog.d.256" => "__builtin_ia32_pternlogd256_mask",
+        "llvm.x86.avx512.pternlog.d.128" => "__builtin_ia32_pternlogd128_mask",
+        "llvm.x86.avx512.pternlog.q.512" => "__builtin_ia32_pternlogq512_mask",
+        "llvm.x86.avx512.pternlog.q.256" => "__builtin_ia32_pternlogq256_mask",
+        "llvm.x86.avx512.pternlog.q.128" => "__builtin_ia32_pternlogq128_mask",
+        "llvm.x86.avx512.add.ps.512" => "__builtin_ia32_addps512_mask",
+        "llvm.x86.avx512.add.pd.512" => "__builtin_ia32_addpd512_mask",
+        "llvm.x86.avx512.sub.ps.512" => "__builtin_ia32_subps512_mask",
+        "llvm.x86.avx512.sub.pd.512" => "__builtin_ia32_subpd512_mask",
+        "llvm.x86.avx512.mul.ps.512" => "__builtin_ia32_mulps512_mask",
+        "llvm.x86.avx512.mul.pd.512" => "__builtin_ia32_mulpd512_mask",
+        "llvm.x86.avx512.div.ps.512" => "__builtin_ia32_divps512_mask",
+        "llvm.x86.avx512.div.pd.512" => "__builtin_ia32_divpd512_mask",
+        "llvm.x86.avx512.vfmadd.ps.512" => "__builtin_ia32_vfmaddps512_mask",
+        "llvm.x86.avx512.vfmadd.pd.512" => "__builtin_ia32_vfmaddpd512_mask",
+
+        // The above doc points to unknown builtins for the following, so override them:
+        "llvm.x86.avx2.gather.d.d" => "__builtin_ia32_gathersiv4si",
+        "llvm.x86.avx2.gather.d.d.256" => "__builtin_ia32_gathersiv8si",
+        "llvm.x86.avx2.gather.d.ps" => "__builtin_ia32_gathersiv4sf",
+        "llvm.x86.avx2.gather.d.ps.256" => "__builtin_ia32_gathersiv8sf",
+        "llvm.x86.avx2.gather.d.q" => "__builtin_ia32_gathersiv2di",
+        "llvm.x86.avx2.gather.d.q.256" => "__builtin_ia32_gathersiv4di",
+        "llvm.x86.avx2.gather.d.pd" => "__builtin_ia32_gathersiv2df",
+        "llvm.x86.avx2.gather.d.pd.256" => "__builtin_ia32_gathersiv4df",
+        "llvm.x86.avx2.gather.q.d" => "__builtin_ia32_gatherdiv4si",
+        "llvm.x86.avx2.gather.q.d.256" => "__builtin_ia32_gatherdiv4si256",
+        "llvm.x86.avx2.gather.q.ps" => "__builtin_ia32_gatherdiv4sf",
+        "llvm.x86.avx2.gather.q.ps.256" => "__builtin_ia32_gatherdiv4sf256",
+        "llvm.x86.avx2.gather.q.q" => "__builtin_ia32_gatherdiv2di",
+        "llvm.x86.avx2.gather.q.q.256" => "__builtin_ia32_gatherdiv4di",
+        "llvm.x86.avx2.gather.q.pd" => "__builtin_ia32_gatherdiv2df",
+        "llvm.x86.avx2.gather.q.pd.256" => "__builtin_ia32_gatherdiv4df",
+        "" => "",
+        // NOTE: this file is generated by https://github.com/GuillaumeGomez/llvmint/blob/master/generate_list.py
+        _ => include!("archs.rs"),
+    };
+
+    let func = cx.context.get_target_builtin_function(gcc_name);
+    cx.functions.borrow_mut().insert(gcc_name.to_string(), func);
+    func
 }

src/intrinsic/mod.rs

@@ -967,34 +967,55 @@ impl<'a, 'gcc, 'tcx> Builder<'a, 'gcc, 'tcx> {
     }
 
     fn saturating_add(&mut self, lhs: RValue<'gcc>, rhs: RValue<'gcc>, signed: bool, width: u64) -> RValue<'gcc> {
-        let func = self.current_func.borrow().expect("func");
+        let result_type = lhs.get_type();
-
         if signed {
-            // Algorithm from: https://stackoverflow.com/a/56531252/389119
+            // Based on algorithm from: https://stackoverflow.com/a/56531252/389119
-            let after_block = func.new_block("after");
+            let func = self.current_func.borrow().expect("func");
-            let func_name =
-                match width {
-                    8 => "__builtin_add_overflow",
-                    16 => "__builtin_add_overflow",
-                    32 => "__builtin_sadd_overflow",
-                    64 => "__builtin_saddll_overflow",
-                    128 => "__builtin_add_overflow",
-                    _ => unreachable!(),
-                };
-            let overflow_func = self.context.get_builtin_function(func_name);
-            let result_type = lhs.get_type();
             let res = func.new_local(None, result_type, "saturating_sum");
-            let overflow = self.overflow_call(overflow_func, &[lhs, rhs, res.get_address(None)], None);
+            let supports_native_type = self.is_native_int_type(result_type);
+            let overflow =
+                if supports_native_type {
+                    let func_name =
+                        match width {
+                            8 => "__builtin_add_overflow",
+                            16 => "__builtin_add_overflow",
+                            32 => "__builtin_sadd_overflow",
+                            64 => "__builtin_saddll_overflow",
+                            128 => "__builtin_add_overflow",
+                            _ => unreachable!(),
+                        };
+                    let overflow_func = self.context.get_builtin_function(func_name);
+                    self.overflow_call(overflow_func, &[lhs, rhs, res.get_address(None)], None)
+                }
+                else {
+                    let func_name =
+                        match width {
+                            128 => "__rust_i128_addo",
+                            _ => unreachable!(),
+                        };
+                    let param_a = self.context.new_parameter(None, result_type, "a");
+                    let param_b = self.context.new_parameter(None, result_type, "b");
+                    let result_field = self.context.new_field(None, result_type, "result");
+                    let overflow_field = self.context.new_field(None, self.bool_type, "overflow");
+                    let return_type = self.context.new_struct_type(None, "result_overflow", &[result_field, overflow_field]);
+                    let func = self.context.new_function(None, FunctionType::Extern, return_type.as_type(), &[param_a, param_b], func_name, false);
+                    let result = self.context.new_call(None, func, &[lhs, rhs]);
+                    let overflow = result.access_field(None, overflow_field);
+                    let int_result = result.access_field(None, result_field);
+                    self.llbb().add_assignment(None, res, int_result);
+                    overflow
+                };
 
             let then_block = func.new_block("then");
+            let after_block = func.new_block("after");
 
-            let unsigned_type = self.context.new_int_type(width as i32 / 8, false);
+            // Return `result_type`'s maximum or minimum value on overflow
-            let shifted = self.context.new_cast(None, lhs, unsigned_type) >> self.context.new_rvalue_from_int(unsigned_type, width as i32 - 1);
+            // NOTE: convert the type to unsigned to have an unsigned shift.
-            let uint_max = self.context.new_unary_op(None, UnaryOp::BitwiseNegate, unsigned_type,
+            let unsigned_type = result_type.to_unsigned(&self.cx);
-                self.context.new_rvalue_from_int(unsigned_type, 0)
+            let shifted = self.gcc_lshr(self.gcc_int_cast(lhs, unsigned_type), self.gcc_int(unsigned_type, width as i64 - 1));
-            );
+            let uint_max = self.gcc_not(self.gcc_int(unsigned_type, 0));
-            let int_max = uint_max >> self.context.new_rvalue_one(unsigned_type);
+            let int_max = self.gcc_lshr(uint_max, self.gcc_int(unsigned_type, 1));
-            then_block.add_assignment(None, res, self.context.new_cast(None, shifted + int_max, result_type));
+            then_block.add_assignment(None, res, self.gcc_int_cast(self.gcc_add(shifted, int_max), result_type));
             then_block.end_with_jump(None, after_block);
 
             self.llbb().end_with_conditional(None, overflow, then_block, after_block);
@@ -1007,19 +1028,18 @@ impl<'a, 'gcc, 'tcx> Builder<'a, 'gcc, 'tcx> {
         }
         else {
             // Algorithm from: http://locklessinc.com/articles/sat_arithmetic/
-            let res = lhs + rhs;
+            let res = self.gcc_add(lhs, rhs);
-            let res_type = res.get_type();
+            let cond = self.gcc_icmp(IntPredicate::IntULT, res, lhs);
-            let cond = self.context.new_comparison(None, ComparisonOp::LessThan, res, lhs);
+            let value = self.gcc_neg(self.gcc_int_cast(cond, result_type));
-            let value = self.context.new_unary_op(None, UnaryOp::Minus, res_type, self.context.new_cast(None, cond, res_type));
+            self.gcc_or(res, value)
-            res | value
         }
     }
 
     // Algorithm from: https://locklessinc.com/articles/sat_arithmetic/
     fn saturating_sub(&mut self, lhs: RValue<'gcc>, rhs: RValue<'gcc>, signed: bool, width: u64) -> RValue<'gcc> {
+        let result_type = lhs.get_type();
         if signed {
-            // Also based on algorithm from: https://stackoverflow.com/a/56531252/389119
+            // Based on algorithm from: https://stackoverflow.com/a/56531252/389119
-            let result_type = lhs.get_type();
             let func = self.current_func.borrow().expect("func");
             let res = func.new_local(None, result_type, "saturating_diff");
             let supports_native_type = self.is_native_int_type(result_type);
@@ -1059,6 +1079,7 @@ impl<'a, 'gcc, 'tcx> Builder<'a, 'gcc, 'tcx> {
             let then_block = func.new_block("then");
             let after_block = func.new_block("after");
 
+            // Return `result_type`'s maximum or minimum value on overflow
             // NOTE: convert the type to unsigned to have an unsigned shift.
             let unsigned_type = result_type.to_unsigned(&self.cx);
             let shifted = self.gcc_lshr(self.gcc_int_cast(lhs, unsigned_type), self.gcc_int(unsigned_type, width as i64 - 1));
@@ -1076,11 +1097,10 @@ impl<'a, 'gcc, 'tcx> Builder<'a, 'gcc, 'tcx> {
             res.to_rvalue()
         }
         else {
-            let res = lhs - rhs;
+            let res = self.gcc_sub(lhs, rhs);
-            let comparison = self.context.new_comparison(None, ComparisonOp::LessThanEquals, res, lhs);
+            let comparison = self.gcc_icmp(IntPredicate::IntULE, res, lhs);
-            let comparison = self.context.new_cast(None, comparison, lhs.get_type());
+            let value = self.gcc_neg(self.gcc_int_cast(comparison, result_type));
-            let unary_op = self.context.new_unary_op(None, UnaryOp::Minus, comparison.get_type(), comparison);
+            self.gcc_and(res, value)
-            self.and(res, unary_op)
         }
     }
 }

src/intrinsic/simd.rs

@@ -1,15 +1,20 @@
-use gccjit::{RValue, Type};
+use std::cmp::Ordering;
+
+use gccjit::{BinaryOp, RValue, Type, ToRValue};
 use rustc_codegen_ssa::base::compare_simd_types;
 use rustc_codegen_ssa::common::{TypeKind, span_invalid_monomorphization_error};
 use rustc_codegen_ssa::mir::operand::OperandRef;
+use rustc_codegen_ssa::mir::place::PlaceRef;
 use rustc_codegen_ssa::traits::{BaseTypeMethods, BuilderMethods};
 use rustc_hir as hir;
 use rustc_middle::span_bug;
 use rustc_middle::ty::layout::HasTyCtxt;
 use rustc_middle::ty::{self, Ty};
 use rustc_span::{Span, Symbol, sym};
+use rustc_target::abi::Align;
 
 use crate::builder::Builder;
+use crate::intrinsic;
 
 pub fn generic_simd_intrinsic<'a, 'gcc, 'tcx>(bx: &mut Builder<'a, 'gcc, 'tcx>, name: Symbol, callee_ty: Ty<'tcx>, args: &[OperandRef<'tcx, RValue<'gcc>>], ret_ty: Ty<'tcx>, llret_ty: Type<'gcc>, span: Span) -> Result<RValue<'gcc>, ()> {
     // macros for error handling:
@@ -53,7 +58,53 @@ pub fn generic_simd_intrinsic<'a, 'gcc, 'tcx>(bx: &mut Builder<'a, 'gcc, 'tcx>,
     let sig =
         tcx.normalize_erasing_late_bound_regions(ty::ParamEnv::reveal_all(), callee_ty.fn_sig(tcx));
     let arg_tys = sig.inputs();
-    let name_str = name.as_str();
+
+    if name == sym::simd_select_bitmask {
+        require_simd!(arg_tys[1], "argument");
+        let (len, _) = arg_tys[1].simd_size_and_type(bx.tcx());
+
+        let expected_int_bits = (len.max(8) - 1).next_power_of_two();
+        let expected_bytes = len / 8 + ((len % 8 > 0) as u64);
+
+        let mask_ty = arg_tys[0];
+        let mut mask = match mask_ty.kind() {
+            ty::Int(i) if i.bit_width() == Some(expected_int_bits) => args[0].immediate(),
+            ty::Uint(i) if i.bit_width() == Some(expected_int_bits) => args[0].immediate(),
+            ty::Array(elem, len)
+                if matches!(elem.kind(), ty::Uint(ty::UintTy::U8))
+                    && len.try_eval_usize(bx.tcx, ty::ParamEnv::reveal_all())
+                        == Some(expected_bytes) =>
+            {
+                let place = PlaceRef::alloca(bx, args[0].layout);
+                args[0].val.store(bx, place);
+                let int_ty = bx.type_ix(expected_bytes * 8);
+                let ptr = bx.pointercast(place.llval, bx.cx.type_ptr_to(int_ty));
+                bx.load(int_ty, ptr, Align::ONE)
+            }
+            _ => return_error!(
+                "invalid bitmask `{}`, expected `u{}` or `[u8; {}]`",
+                mask_ty,
+                expected_int_bits,
+                expected_bytes
+            ),
+        };
+
+        let arg1 = args[1].immediate();
+        let arg1_type = arg1.get_type();
+        let arg1_vector_type = arg1_type.unqualified().dyncast_vector().expect("vector type");
+        let arg1_element_type = arg1_vector_type.get_element_type();
+
+        let mut elements = vec![];
+        let one = bx.context.new_rvalue_one(mask.get_type());
+        for _ in 0..len {
+            let element = bx.context.new_cast(None, mask & one, arg1_element_type);
+            elements.push(element);
+            mask = mask >> one;
+        }
+        let vector_mask = bx.context.new_rvalue_from_vector(None, arg1_type, &elements);
+
+        return Ok(bx.vector_select(vector_mask, arg1, args[2].immediate()));
+    }
 
     // every intrinsic below takes a SIMD vector as its first argument
     require_simd!(arg_tys[0], "input");
@@ -100,10 +151,28 @@ pub fn generic_simd_intrinsic<'a, 'gcc, 'tcx>(bx: &mut Builder<'a, 'gcc, 'tcx>,
         ));
     }
 
-    if let Some(stripped) = name_str.strip_prefix("simd_shuffle") {
+    if let Some(stripped) = name.as_str().strip_prefix("simd_shuffle") {
-        let n: u64 = stripped.parse().unwrap_or_else(|_| {
+        let n: u64 =
-            span_bug!(span, "bad `simd_shuffle` instruction only caught in codegen?")
+            if stripped.is_empty() {
-        });
+                // Make sure this is actually an array, since typeck only checks the length-suffixed
+                // version of this intrinsic.
+                match args[2].layout.ty.kind() {
+                    ty::Array(ty, len) if matches!(ty.kind(), ty::Uint(ty::UintTy::U32)) => {
+                        len.try_eval_usize(bx.cx.tcx, ty::ParamEnv::reveal_all()).unwrap_or_else(|| {
+                            span_bug!(span, "could not evaluate shuffle index array length")
+                        })
+                    }
+                    _ => return_error!(
+                        "simd_shuffle index must be an array of `u32`, got `{}`",
+                        args[2].layout.ty
+                    ),
+                }
+            }
+            else {
+                stripped.parse().unwrap_or_else(|_| {
+                    span_bug!(span, "bad `simd_shuffle` instruction only caught in codegen?")
+                })
+            };
 
         require_simd!(ret_ty, "return");
 
@@ -134,6 +203,225 @@ pub fn generic_simd_intrinsic<'a, 'gcc, 'tcx>(bx: &mut Builder<'a, 'gcc, 'tcx>,
         ));
     }
 
+    #[cfg(feature="master")]
+    if name == sym::simd_insert {
+        require!(
+            in_elem == arg_tys[2],
+            "expected inserted type `{}` (element of input `{}`), found `{}`",
+            in_elem,
+            in_ty,
+            arg_tys[2]
+        );
+        let vector = args[0].immediate();
+        let index = args[1].immediate();
+        let value = args[2].immediate();
+        // TODO(antoyo): use a recursive unqualified() here.
+        let vector_type = vector.get_type().unqualified().dyncast_vector().expect("vector type");
+        let element_type = vector_type.get_element_type();
+        // NOTE: we cannot cast to an array and assign to its element here because the value might
+        // not be an l-value. So, call a builtin to set the element.
+        // TODO(antoyo): perhaps we could create a new vector or maybe there's a GIMPLE instruction for that?
+        // TODO(antoyo): don't use target specific builtins here.
+        let func_name =
+            match in_len {
+                2 => {
+                    if element_type == bx.i64_type {
+                        "__builtin_ia32_vec_set_v2di"
+                    }
+                    else {
+                        unimplemented!();
+                    }
+                },
+                4 => {
+                    if element_type == bx.i32_type {
+                        "__builtin_ia32_vec_set_v4si"
+                    }
+                    else {
+                        unimplemented!();
+                    }
+                },
+                8 => {
+                    if element_type == bx.i16_type {
+                        "__builtin_ia32_vec_set_v8hi"
+                    }
+                    else {
+                        unimplemented!();
+                    }
+                },
+                _ => unimplemented!("Len: {}", in_len),
+            };
+        let builtin = bx.context.get_target_builtin_function(func_name);
+        let param1_type = builtin.get_param(0).to_rvalue().get_type();
+        // TODO(antoyo): perhaps use __builtin_convertvector for vector casting.
+        let vector = bx.cx.bitcast_if_needed(vector, param1_type);
+        let result = bx.context.new_call(None, builtin, &[vector, value, bx.context.new_cast(None, index, bx.int_type)]);
+        // TODO(antoyo): perhaps use __builtin_convertvector for vector casting.
+        return Ok(bx.context.new_bitcast(None, result, vector.get_type()));
+    }
+
+    #[cfg(feature="master")]
+    if name == sym::simd_extract {
+        require!(
+            ret_ty == in_elem,
+            "expected return type `{}` (element of input `{}`), found `{}`",
+            in_elem,
+            in_ty,
+            ret_ty
+        );
+        let vector = args[0].immediate();
+        return Ok(bx.context.new_vector_access(None, vector, args[1].immediate()).to_rvalue());
+    }
+
+    if name == sym::simd_select {
+        let m_elem_ty = in_elem;
+        let m_len = in_len;
+        require_simd!(arg_tys[1], "argument");
+        let (v_len, _) = arg_tys[1].simd_size_and_type(bx.tcx());
+        require!(
+            m_len == v_len,
+            "mismatched lengths: mask length `{}` != other vector length `{}`",
+            m_len,
+            v_len
+        );
+        match m_elem_ty.kind() {
+            ty::Int(_) => {}
+            _ => return_error!("mask element type is `{}`, expected `i_`", m_elem_ty),
+        }
+        return Ok(bx.vector_select(args[0].immediate(), args[1].immediate(), args[2].immediate()));
+    }
+
+    if name == sym::simd_cast {
+        require_simd!(ret_ty, "return");
+        let (out_len, out_elem) = ret_ty.simd_size_and_type(bx.tcx());
+        require!(
+            in_len == out_len,
+            "expected return type with length {} (same as input type `{}`), \
+                  found `{}` with length {}",
+            in_len,
+            in_ty,
+            ret_ty,
+            out_len
+        );
+        // casting cares about nominal type, not just structural type
+        if in_elem == out_elem {
+            return Ok(args[0].immediate());
+        }
+
+        enum Style {
+            Float,
+            Int(/* is signed? */ bool),
+            Unsupported,
+        }
+
+        let (in_style, in_width) = match in_elem.kind() {
+            // vectors of pointer-sized integers should've been
+            // disallowed before here, so this unwrap is safe.
+            ty::Int(i) => (
+                Style::Int(true),
+                i.normalize(bx.tcx().sess.target.pointer_width).bit_width().unwrap(),
+            ),
+            ty::Uint(u) => (
+                Style::Int(false),
+                u.normalize(bx.tcx().sess.target.pointer_width).bit_width().unwrap(),
+            ),
+            ty::Float(f) => (Style::Float, f.bit_width()),
+            _ => (Style::Unsupported, 0),
+        };
+        let (out_style, out_width) = match out_elem.kind() {
+            ty::Int(i) => (
+                Style::Int(true),
+                i.normalize(bx.tcx().sess.target.pointer_width).bit_width().unwrap(),
+            ),
+            ty::Uint(u) => (
+                Style::Int(false),
+                u.normalize(bx.tcx().sess.target.pointer_width).bit_width().unwrap(),
+            ),
+            ty::Float(f) => (Style::Float, f.bit_width()),
+            _ => (Style::Unsupported, 0),
+        };
+
+        let extend = |in_type, out_type| {
+            let vector_type = bx.context.new_vector_type(out_type, 8);
+            let vector = args[0].immediate();
+            let array_type = bx.context.new_array_type(None, in_type, 8);
+            // TODO(antoyo): switch to using new_vector_access or __builtin_convertvector for vector casting.
+            let array = bx.context.new_bitcast(None, vector, array_type);
+
+            let cast_vec_element = |index| {
+                let index = bx.context.new_rvalue_from_int(bx.int_type, index);
+                bx.context.new_cast(None, bx.context.new_array_access(None, array, index).to_rvalue(), out_type)
+            };
+
+            bx.context.new_rvalue_from_vector(None, vector_type, &[
+                cast_vec_element(0),
+                cast_vec_element(1),
+                cast_vec_element(2),
+                cast_vec_element(3),
+                cast_vec_element(4),
+                cast_vec_element(5),
+                cast_vec_element(6),
+                cast_vec_element(7),
+            ])
+        };
+
+        match (in_style, out_style) {
+            (Style::Int(in_is_signed), Style::Int(_)) => {
+                return Ok(match in_width.cmp(&out_width) {
+                    Ordering::Greater => bx.trunc(args[0].immediate(), llret_ty),
+                    Ordering::Equal => args[0].immediate(),
+                    Ordering::Less => {
+                        if in_is_signed {
+                            match (in_width, out_width) {
+                                // FIXME(antoyo): the function _mm_cvtepi8_epi16 should directly
+                                // call an intrinsic equivalent to __builtin_ia32_pmovsxbw128 so that
+                                // we can generate a call to it.
+                                (8, 16) => extend(bx.i8_type, bx.i16_type),
+                                (8, 32) => extend(bx.i8_type, bx.i32_type),
+                                (8, 64) => extend(bx.i8_type, bx.i64_type),
+                                (16, 32) => extend(bx.i16_type, bx.i32_type),
+                                (32, 64) => extend(bx.i32_type, bx.i64_type),
+                                (16, 64) => extend(bx.i16_type, bx.i64_type),
+                                _ => unimplemented!("in: {}, out: {}", in_width, out_width),
+                            }
+                        } else {
+                            match (in_width, out_width) {
+                                (8, 16) => extend(bx.u8_type, bx.u16_type),
+                                (8, 32) => extend(bx.u8_type, bx.u32_type),
+                                (8, 64) => extend(bx.u8_type, bx.u64_type),
+                                (16, 32) => extend(bx.u16_type, bx.u32_type),
+                                (16, 64) => extend(bx.u16_type, bx.u64_type),
+                                (32, 64) => extend(bx.u32_type, bx.u64_type),
+                                _ => unimplemented!("in: {}, out: {}", in_width, out_width),
+                            }
+                        }
+                    }
+                });
+            }
+            (Style::Int(_), Style::Float) => {
+                // TODO: add support for internal functions in libgccjit to get access to IFN_VEC_CONVERT which is
+                // doing like __builtin_convertvector?
+                // Or maybe provide convert_vector as an API since it might not easy to get the
+                // types of internal functions.
+                unimplemented!();
+            }
+            (Style::Float, Style::Int(_)) => {
+                unimplemented!();
+            }
+            (Style::Float, Style::Float) => {
+                unimplemented!();
+            }
+            _ => { /* Unsupported. Fallthrough. */ }
+        }
+        require!(
+            false,
+            "unsupported cast from `{}` with element `{}` to `{}` with element `{}`",
+            in_ty,
+            in_elem,
+            ret_ty,
+            out_elem
+        );
+    }
+
     macro_rules! arith_binary {
         ($($name: ident: $($($p: ident),* => $call: ident),*;)*) => {
             $(if name == sym::$name {
@@ -151,6 +439,105 @@ pub fn generic_simd_intrinsic<'a, 'gcc, 'tcx>(bx: &mut Builder<'a, 'gcc, 'tcx>,
         }
     }
 
+    fn simd_simple_float_intrinsic<'gcc, 'tcx>(
+        name: Symbol,
+        in_elem: Ty<'_>,
+        in_ty: Ty<'_>,
+        in_len: u64,
+        bx: &mut Builder<'_, 'gcc, 'tcx>,
+        span: Span,
+        args: &[OperandRef<'tcx, RValue<'gcc>>],
+    ) -> Result<RValue<'gcc>, ()> {
+        macro_rules! emit_error {
+            ($msg: tt) => {
+                emit_error!($msg, )
+            };
+            ($msg: tt, $($fmt: tt)*) => {
+                span_invalid_monomorphization_error(
+                    bx.sess(), span,
+                    &format!(concat!("invalid monomorphization of `{}` intrinsic: ", $msg),
+                             name, $($fmt)*));
+            }
+        }
+        macro_rules! return_error {
+            ($($fmt: tt)*) => {
+                {
+                    emit_error!($($fmt)*);
+                    return Err(());
+                }
+            }
+        }
+
+        let (elem_ty_str, elem_ty) =
+            if let ty::Float(f) = in_elem.kind() {
+                let elem_ty = bx.cx.type_float_from_ty(*f);
+                match f.bit_width() {
+                    32 => ("f32", elem_ty),
+                    64 => ("f64", elem_ty),
+                    _ => {
+                        return_error!(
+                            "unsupported element type `{}` of floating-point vector `{}`",
+                            f.name_str(),
+                            in_ty
+                        );
+                    }
+                }
+            }
+            else {
+                return_error!("`{}` is not a floating-point type", in_ty);
+            };
+
+        let vec_ty = bx.cx.type_vector(elem_ty, in_len);
+
+        let (intr_name, fn_ty) =
+            match name {
+                sym::simd_ceil => ("ceil", bx.type_func(&[vec_ty], vec_ty)),
+                sym::simd_fabs => ("fabs", bx.type_func(&[vec_ty], vec_ty)), // TODO(antoyo): pand with 170141183420855150465331762880109871103
+                sym::simd_fcos => ("cos", bx.type_func(&[vec_ty], vec_ty)),
+                sym::simd_fexp2 => ("exp2", bx.type_func(&[vec_ty], vec_ty)),
+                sym::simd_fexp => ("exp", bx.type_func(&[vec_ty], vec_ty)),
+                sym::simd_flog10 => ("log10", bx.type_func(&[vec_ty], vec_ty)),
+                sym::simd_flog2 => ("log2", bx.type_func(&[vec_ty], vec_ty)),
+                sym::simd_flog => ("log", bx.type_func(&[vec_ty], vec_ty)),
+                sym::simd_floor => ("floor", bx.type_func(&[vec_ty], vec_ty)),
+                sym::simd_fma => ("fma", bx.type_func(&[vec_ty, vec_ty, vec_ty], vec_ty)),
+                sym::simd_fpowi => ("powi", bx.type_func(&[vec_ty, bx.type_i32()], vec_ty)),
+                sym::simd_fpow => ("pow", bx.type_func(&[vec_ty, vec_ty], vec_ty)),
+                sym::simd_fsin => ("sin", bx.type_func(&[vec_ty], vec_ty)),
+                sym::simd_fsqrt => ("sqrt", bx.type_func(&[vec_ty], vec_ty)),
+                sym::simd_round => ("round", bx.type_func(&[vec_ty], vec_ty)),
+                sym::simd_trunc => ("trunc", bx.type_func(&[vec_ty], vec_ty)),
+                _ => return_error!("unrecognized intrinsic `{}`", name),
+            };
+        let llvm_name = &format!("llvm.{0}.v{1}{2}", intr_name, in_len, elem_ty_str);
+        let function = intrinsic::llvm::intrinsic(llvm_name, &bx.cx);
+        let function: RValue<'gcc> = unsafe { std::mem::transmute(function) };
+        let c = bx.call(fn_ty, function, &args.iter().map(|arg| arg.immediate()).collect::<Vec<_>>(), None);
+        Ok(c)
+    }
+
+    if std::matches!(
+        name,
+        sym::simd_ceil
+            | sym::simd_fabs
+            | sym::simd_fcos
+            | sym::simd_fexp2
+            | sym::simd_fexp
+            | sym::simd_flog10
+            | sym::simd_flog2
+            | sym::simd_flog
+            | sym::simd_floor
+            | sym::simd_fma
+            | sym::simd_fpow
+            | sym::simd_fpowi
+            | sym::simd_fsin
+            | sym::simd_fsqrt
+            | sym::simd_round
+            | sym::simd_trunc
+    ) {
+        return simd_simple_float_intrinsic(name, in_elem, in_ty, in_len, bx, span, args);
+    }
+
     arith_binary! {
         simd_add: Uint, Int => add, Float => fadd;
         simd_sub: Uint, Int => sub, Float => fsub;
@@ -185,5 +572,183 @@ pub fn generic_simd_intrinsic<'a, 'gcc, 'tcx>(bx: &mut Builder<'a, 'gcc, 'tcx>,
         simd_neg: Int => neg, Float => fneg;
     }
 
+    #[cfg(feature="master")]
+    if name == sym::simd_saturating_add || name == sym::simd_saturating_sub {
+        let lhs = args[0].immediate();
+        let rhs = args[1].immediate();
+        let is_add = name == sym::simd_saturating_add;
+        let ptr_bits = bx.tcx().data_layout.pointer_size.bits() as _;
+        let (signed, elem_width, elem_ty) = match *in_elem.kind() {
+            ty::Int(i) => (true, i.bit_width().unwrap_or(ptr_bits), bx.cx.type_int_from_ty(i)),
+            ty::Uint(i) => (false, i.bit_width().unwrap_or(ptr_bits), bx.cx.type_uint_from_ty(i)),
+            _ => {
+                return_error!(
+                    "expected element type `{}` of vector type `{}` \
+                     to be a signed or unsigned integer type",
+                    arg_tys[0].simd_size_and_type(bx.tcx()).1,
+                    arg_tys[0]
+                );
+            }
+        };
+        let builtin_name =
+            match (signed, is_add, in_len, elem_width) {
+                (true, true, 32, 8) => "__builtin_ia32_paddsb256", // TODO(antoyo): cast arguments to unsigned.
+                (false, true, 32, 8) => "__builtin_ia32_paddusb256",
+                (true, true, 16, 16) => "__builtin_ia32_paddsw256",
+                (false, true, 16, 16) => "__builtin_ia32_paddusw256",
+                (true, false, 16, 16) => "__builtin_ia32_psubsw256",
+                (false, false, 16, 16) => "__builtin_ia32_psubusw256",
+                (true, false, 32, 8) => "__builtin_ia32_psubsb256",
+                (false, false, 32, 8) => "__builtin_ia32_psubusb256",
+                _ => unimplemented!("signed: {}, is_add: {}, in_len: {}, elem_width: {}", signed, is_add, in_len, elem_width),
+            };
+        let vec_ty = bx.cx.type_vector(elem_ty, in_len as u64);
+
+        let func = bx.context.get_target_builtin_function(builtin_name);
+        let param1_type = func.get_param(0).to_rvalue().get_type();
+        let param2_type = func.get_param(1).to_rvalue().get_type();
+        let lhs = bx.cx.bitcast_if_needed(lhs, param1_type);
+        let rhs = bx.cx.bitcast_if_needed(rhs, param2_type);
+        let result = bx.context.new_call(None, func, &[lhs, rhs]);
+        // TODO(antoyo): perhaps use __builtin_convertvector for vector casting.
+        return Ok(bx.context.new_bitcast(None, result, vec_ty));
+    }
+
+    macro_rules! arith_red {
+        ($name:ident : $vec_op:expr, $float_reduce:ident, $ordered:expr, $op:ident,
+         $identity:expr) => {
+            if name == sym::$name {
+                require!(
+                    ret_ty == in_elem,
+                    "expected return type `{}` (element of input `{}`), found `{}`",
+                    in_elem,
+                    in_ty,
+                    ret_ty
+                );
+                return match in_elem.kind() {
+                    ty::Int(_) | ty::Uint(_) => {
+                        let r = bx.vector_reduce_op(args[0].immediate(), $vec_op);
+                        if $ordered {
+                            // if overflow occurs, the result is the
+                            // mathematical result modulo 2^n:
+                            Ok(bx.$op(args[1].immediate(), r))
+                        }
+                        else {
+                            Ok(bx.vector_reduce_op(args[0].immediate(), $vec_op))
+                        }
+                    }
+                    ty::Float(_) => {
+                        if $ordered {
+                            // ordered arithmetic reductions take an accumulator
+                            let acc = args[1].immediate();
+                            Ok(bx.$float_reduce(acc, args[0].immediate()))
+                        }
+                        else {
+                            Ok(bx.vector_reduce_op(args[0].immediate(), $vec_op))
+                        }
+                    }
+                    _ => return_error!(
+                        "unsupported {} from `{}` with element `{}` to `{}`",
+                        sym::$name,
+                        in_ty,
+                        in_elem,
+                        ret_ty
+                    ),
+                };
+            }
+        };
+    }
+
+    arith_red!(
+        simd_reduce_add_unordered: BinaryOp::Plus,
+        vector_reduce_fadd_fast,
+        false,
+        add,
+        0.0 // TODO: Use this argument.
+    );
+    arith_red!(
+        simd_reduce_mul_unordered: BinaryOp::Mult,
+        vector_reduce_fmul_fast,
+        false,
+        mul,
+        1.0
+    );
+
+    macro_rules! minmax_red {
+        ($name:ident: $reduction:ident) => {
+            if name == sym::$name {
+                require!(
+                    ret_ty == in_elem,
+                    "expected return type `{}` (element of input `{}`), found `{}`",
+                    in_elem,
+                    in_ty,
+                    ret_ty
+                );
+                return match in_elem.kind() {
+                    ty::Int(_) | ty::Uint(_) | ty::Float(_) => Ok(bx.$reduction(args[0].immediate())),
+                    _ => return_error!(
+                        "unsupported {} from `{}` with element `{}` to `{}`",
+                        sym::$name,
+                        in_ty,
+                        in_elem,
+                        ret_ty
+                    ),
+                };
+            }
+        };
+    }
+
+    minmax_red!(simd_reduce_min: vector_reduce_min);
+    minmax_red!(simd_reduce_max: vector_reduce_max);
+
+    macro_rules! bitwise_red {
+        ($name:ident : $op:expr, $boolean:expr) => {
+            if name == sym::$name {
+                let input = if !$boolean {
+                    require!(
+                        ret_ty == in_elem,
+                        "expected return type `{}` (element of input `{}`), found `{}`",
+                        in_elem,
+                        in_ty,
+                        ret_ty
+                    );
+                    args[0].immediate()
+                } else {
+                    match in_elem.kind() {
+                        ty::Int(_) | ty::Uint(_) => {}
+                        _ => return_error!(
+                            "unsupported {} from `{}` with element `{}` to `{}`",
+                            sym::$name,
+                            in_ty,
+                            in_elem,
+                            ret_ty
+                        ),
+                    }
+
+                    // boolean reductions operate on vectors of i1s:
+                    let i1 = bx.type_i1();
+                    let i1xn = bx.type_vector(i1, in_len as u64);
+                    bx.trunc(args[0].immediate(), i1xn)
+                };
+                return match in_elem.kind() {
+                    ty::Int(_) | ty::Uint(_) => {
+                        let r = bx.vector_reduce_op(input, $op);
+                        Ok(if !$boolean { r } else { bx.zext(r, bx.type_bool()) })
+                    }
+                    _ => return_error!(
+                        "unsupported {} from `{}` with element `{}` to `{}`",
+                        sym::$name,
+                        in_ty,
+                        in_elem,
+                        ret_ty
+                    ),
+                };
+            }
+        };
+    }
+
+    bitwise_red!(simd_reduce_and: BinaryOp::BitwiseAnd, false);
+    bitwise_red!(simd_reduce_or: BinaryOp::BitwiseOr, false);
+
     unimplemented!("simd {}", name);
 }

src/lib.rs

@@ -203,7 +203,7 @@ impl WriteBackendMethods for GccCodegenBackend {
     fn run_fat_lto(_cgcx: &CodegenContext<Self>, mut modules: Vec<FatLTOInput<Self>>, _cached_modules: Vec<(SerializedModule<Self::ModuleBuffer>, WorkProduct)>) -> Result<LtoModuleCodegen<Self>, FatalError> {
         // TODO(antoyo): implement LTO by sending -flto to libgccjit and adding the appropriate gcc linker plugins.
         // NOTE: implemented elsewhere.
-        // TODO: what is implemented elsewhere ^ ?
+        // TODO(antoyo): what is implemented elsewhere ^ ?
         let module =
             match modules.remove(0) {
                 FatLTOInput::InMemory(module) => module,
@@ -301,7 +301,22 @@ pub fn target_features(sess: &Session) -> Vec<Symbol> {
         )
         .filter(|_feature| {
             // TODO(antoyo): implement a way to get enabled feature in libgccjit.
-            false
+            // Probably using the equivalent of __builtin_cpu_supports.
+            #[cfg(feature="master")]
+            {
+                _feature.contains("sse") || _feature.contains("avx")
+            }
+            #[cfg(not(feature="master"))]
+            {
+                false
+            }
+            /*
+               adx, aes, avx, avx2, avx512bf16, avx512bitalg, avx512bw, avx512cd, avx512dq, avx512er, avx512f, avx512gfni,
+               avx512ifma, avx512pf, avx512vaes, avx512vbmi, avx512vbmi2, avx512vl, avx512vnni, avx512vp2intersect, avx512vpclmulqdq,
+               avx512vpopcntdq, bmi1, bmi2, cmpxchg16b, ermsb, f16c, fma, fxsr, lzcnt, movbe, pclmulqdq, popcnt, rdrand, rdseed, rtm,
+               sha, sse, sse2, sse3, sse4.1, sse4.2, sse4a, ssse3, tbm, xsave, xsavec, xsaveopt, xsaves
+             */
+            //false
         })
         .map(|feature| Symbol::intern(feature))
         .collect()

src/type_.rs

@@ -3,10 +3,11 @@ use std::convert::TryInto;
 use gccjit::{RValue, Struct, Type};
 use rustc_codegen_ssa::traits::{BaseTypeMethods, DerivedTypeMethods};
 use rustc_codegen_ssa::common::TypeKind;
-use rustc_middle::bug;
+use rustc_middle::{bug, ty};
 use rustc_middle::ty::layout::TyAndLayout;
 use rustc_target::abi::{AddressSpace, Align, Integer, Size};
 
+use crate::common::TypeReflection;
 use crate::context::CodegenCx;
 use crate::type_of::LayoutGccExt;
 
@@ -60,6 +61,17 @@ impl<'gcc, 'tcx> CodegenCx<'gcc, 'tcx> {
         let ity = Integer::approximate_align(self, align);
         self.type_from_integer(ity)
     }
+
+    pub fn type_vector(&self, ty: Type<'gcc>, len: u64) -> Type<'gcc> {
+        self.context.new_vector_type(ty, len)
+    }
+
+    pub fn type_float_from_ty(&self, t: ty::FloatTy) -> Type<'gcc> {
+        match t {
+            ty::FloatTy::F32 => self.type_f32(),
+            ty::FloatTy::F64 => self.type_f64(),
+        }
+    }
 }
 
 impl<'gcc, 'tcx> BaseTypeMethods<'tcx> for CodegenCx<'gcc, 'tcx> {
@@ -103,7 +115,7 @@ impl<'gcc, 'tcx> BaseTypeMethods<'tcx> for CodegenCx<'gcc, 'tcx> {
         self.context.new_function_pointer_type(None, return_type, params, false)
     }
 
-    fn type_struct(&self, fields: &[Type<'gcc>], _packed: bool) -> Type<'gcc> {
+    fn type_struct(&self, fields: &[Type<'gcc>], packed: bool) -> Type<'gcc> {
         let types = fields.to_vec();
         if let Some(typ) = self.struct_types.borrow().get(fields) {
             return typ.clone();
@@ -111,8 +123,11 @@ impl<'gcc, 'tcx> BaseTypeMethods<'tcx> for CodegenCx<'gcc, 'tcx> {
         let fields: Vec<_> = fields.iter().enumerate()
             .map(|(index, field)| self.context.new_field(None, *field, &format!("field{}_TODO", index)))
             .collect();
-        // TODO(antoyo): use packed.
         let typ = self.context.new_struct_type(None, "struct", &fields).as_type();
+        if packed {
+            #[cfg(feature="master")]
+            typ.set_packed();
+        }
         self.struct_types.borrow_mut().insert(types, typ);
         typ
     }
@@ -127,7 +142,7 @@ impl<'gcc, 'tcx> BaseTypeMethods<'tcx> for CodegenCx<'gcc, 'tcx> {
         else if typ.is_compatible_with(self.double_type) {
             TypeKind::Double
         }
-        else if typ.dyncast_vector().is_some() {
+        else if typ.is_vector() {
             TypeKind::Vector
         }
         else {
@@ -141,7 +156,7 @@ impl<'gcc, 'tcx> BaseTypeMethods<'tcx> for CodegenCx<'gcc, 'tcx> {
     }
 
     fn type_ptr_to_ext(&self, ty: Type<'gcc>, _address_space: AddressSpace) -> Type<'gcc> {
-        // TODO(antoyo): use address_space
+        // TODO(antoyo): use address_space, perhaps with TYPE_ADDR_SPACE?
         ty.make_pointer()
     }
 
@@ -167,10 +182,10 @@ impl<'gcc, 'tcx> BaseTypeMethods<'tcx> for CodegenCx<'gcc, 'tcx> {
     fn float_width(&self, typ: Type<'gcc>) -> usize {
         let f32 = self.context.new_type::<f32>();
         let f64 = self.context.new_type::<f64>();
-        if typ == f32 {
+        if typ.is_compatible_with(f32) {
             32
         }
-        else if typ == f64 {
+        else if typ.is_compatible_with(f64) {
             64
         }
         else {
@@ -197,12 +212,15 @@ impl<'gcc, 'tcx> CodegenCx<'gcc, 'tcx> {
         self.type_array(self.type_from_integer(unit), size / unit_size)
     }
 
-    pub fn set_struct_body(&self, typ: Struct<'gcc>, fields: &[Type<'gcc>], _packed: bool) {
+    pub fn set_struct_body(&self, typ: Struct<'gcc>, fields: &[Type<'gcc>], packed: bool) {
-        // TODO(antoyo): use packed.
         let fields: Vec<_> = fields.iter().enumerate()
             .map(|(index, field)| self.context.new_field(None, *field, &format!("field_{}", index)))
             .collect();
         typ.set_fields(None, &fields);
+        if packed {
+            #[cfg(feature="master")]
+            typ.as_type().set_packed();
+        }
     }
 
     pub fn type_named_struct(&self, name: &str) -> Struct<'gcc> {
@@ -229,6 +247,10 @@ impl<'gcc, 'tcx> CodegenCx<'gcc, 'tcx> {
 
         self.context.new_array_type(None, ty, len)
     }
+
+    pub fn type_bool(&self) -> Type<'gcc> {
+        self.context.new_type::<bool>()
+    }
 }
 
 pub fn struct_fields<'gcc, 'tcx>(cx: &CodegenCx<'gcc, 'tcx>, layout: TyAndLayout<'tcx>) -> (Vec<Type<'gcc>>, bool) {

src/type_of.rs

@@ -24,6 +24,30 @@ impl<'gcc, 'tcx> CodegenCx<'gcc, 'tcx> {
             I128 => self.type_u128(),
         }
     }
+
+    #[cfg(feature="master")]
+    pub fn type_int_from_ty(&self, t: ty::IntTy) -> Type<'gcc> {
+        match t {
+            ty::IntTy::Isize => self.type_isize(),
+            ty::IntTy::I8 => self.type_i8(),
+            ty::IntTy::I16 => self.type_i16(),
+            ty::IntTy::I32 => self.type_i32(),
+            ty::IntTy::I64 => self.type_i64(),
+            ty::IntTy::I128 => self.type_i128(),
+        }
+    }
+
+    #[cfg(feature="master")]
+    pub fn type_uint_from_ty(&self, t: ty::UintTy) -> Type<'gcc> {
+        match t {
+            ty::UintTy::Usize => self.type_isize(),
+            ty::UintTy::U8 => self.type_i8(),
+            ty::UintTy::U16 => self.type_i16(),
+            ty::UintTy::U32 => self.type_i32(),
+            ty::UintTy::U64 => self.type_i64(),
+            ty::UintTy::U128 => self.type_i128(),
+        }
+    }
 }
 
 pub fn uncached_gcc_type<'gcc, 'tcx>(cx: &CodegenCx<'gcc, 'tcx>, layout: TyAndLayout<'tcx>, defer: &mut Option<(Struct<'gcc>, TyAndLayout<'tcx>)>) -> Type<'gcc> {

test.sh

@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash
 
 # TODO(antoyo): rewrite to cargo-make (or just) or something like that to only rebuild the sysroot when needed?
 
@@ -14,25 +14,87 @@ fi
 export LD_LIBRARY_PATH="$GCC_PATH"
 export LIBRARY_PATH="$GCC_PATH"
 
-features=
+flags=
+gcc_master_branch=1
+channel="debug"
+func=all
+build_only=0
 
-if [[ "$1" == "--features" ]]; then
+while [[ $# -gt 0 ]]; do
-    shift
+    case $1 in
-    features="--features $1"
+        --release)
-    shift
+            codegen_channel=release
-fi
+            shift
+            ;;
+        --release-sysroot)
+            sysroot_channel=release
+            shift
+            ;;
+        --no-default-features)
+            gcc_master_branch=0
+            flags="$flags --no-default-features"
+            shift
+            ;;
+        --features)
+            shift
+            flags="$flags --features $1"
+            shift
+            ;;
+        --release)
+            channel="release"
+            shift
+            ;;
+        "--test-rustc")
+            func=test_rustc
+            shift
+            ;;
 
-if [[ "$1" == "--release" ]]; then
+        "--test-libcore")
+            func=test_libcore
+            shift
+            ;;
+
+        "--clean-ui-tests")
+            func=clean_ui_tests
+            shift
+            ;;
+
+        "--std-tests")
+            func=std_tests
+            shift
+            ;;
+
+        "--extended-tests")
+            func=extended_sysroot_tests
+            shift
+            ;;
+
+        "--build-sysroot")
+            func=build_sysroot
+            shift
+            ;;
+        "--build")
+            build_only=1
+            shift
+            ;;
+        *)
+            echo "Unknown option $1"
+            exit 1
+            ;;
+    esac
+done
+
+if [[ $channel == "release" ]]; then
     export CHANNEL='release'
-    CARGO_INCREMENTAL=1 cargo rustc --release $features
+    CARGO_INCREMENTAL=1 cargo rustc --release $flags
     shift
 else
     echo $LD_LIBRARY_PATH
     export CHANNEL='debug'
-    cargo rustc $features
+    cargo rustc $flags
 fi
 
-if [[ "$1" == "--build" ]]; then
+if (( $build_only == 1 )); then
     exit
 fi
 
@@ -78,7 +140,11 @@ function std_tests() {
     $RUN_WRAPPER ./target/out/dst_field_align || (echo $?; false)
 
     echo "[AOT] std_example"
-    $RUSTC example/std_example.rs --crate-type bin --target $TARGET_TRIPLE
+    std_flags="--cfg feature=\"master\""
+    if (( $gcc_master_branch == 0 )); then
+        std_flags=""
+    fi
+    $RUSTC example/std_example.rs --crate-type bin --target $TARGET_TRIPLE $std_flags
     $RUN_WRAPPER ./target/out/std_example --target $TARGET_TRIPLE
 
     echo "[AOT] subslice-patterns-const-eval"
@@ -97,25 +163,6 @@ function std_tests() {
 #echo "[BUILD] sysroot in release mode"
 #./build_sysroot/build_sysroot.sh --release
 
-# TODO(antoyo): uncomment when it works.
-#pushd simple-raytracer
-#if [[ "$HOST_TRIPLE" = "$TARGET_TRIPLE" ]]; then
-    #echo "[BENCH COMPILE] ebobby/simple-raytracer"
-    #hyperfine --runs ${RUN_RUNS:-10} --warmup 1 --prepare "rm -r target/*/debug || true" \
-    #"RUSTFLAGS='' cargo build --target $TARGET_TRIPLE" \
-    #"../cargo.sh build"
-
-    #echo "[BENCH RUN] ebobby/simple-raytracer"
-    #cp ./target/*/debug/main ./raytracer_cg_gccjit
-    #hyperfine --runs ${RUN_RUNS:-10} ./raytracer_cg_llvm ./raytracer_cg_gccjit
-#else
-    #echo "[BENCH COMPILE] ebobby/simple-raytracer (skipped)"
-    #echo "[COMPILE] ebobby/simple-raytracer"
-    #../cargo.sh build
-    #echo "[BENCH RUN] ebobby/simple-raytracer (skipped)"
-#fi
-#popd
-
 function test_libcore() {
     pushd build_sysroot/sysroot_src/library/core/tests
     echo "[TEST] libcore"
@@ -124,19 +171,6 @@ function test_libcore() {
     popd
 }
 
-# TODO(antoyo): uncomment when it works.
-#pushd regex
-#echo "[TEST] rust-lang/regex example shootout-regex-dna"
-#../cargo.sh clean
-## Make sure `[codegen mono items] start` doesn't poison the diff
-#../cargo.sh build --example shootout-regex-dna
-#cat examples/regexdna-input.txt | ../cargo.sh run --example shootout-regex-dna | grep -v "Spawned thread" > res.txt
-#diff -u res.txt examples/regexdna-output.txt
-
-#echo "[TEST] rust-lang/regex tests"
-#../cargo.sh test --tests -- --exclude-should-panic --test-threads 1 -Zunstable-options
-#popd
-
 #echo
 #echo "[BENCH COMPILE] mod_bench"
 
@@ -153,6 +187,44 @@ function test_libcore() {
 #echo "[BENCH RUN] mod_bench"
 #hyperfine --runs ${RUN_RUNS:-10} ./target/out/mod_bench{,_inline} ./target/out/mod_bench_llvm_*
 
+function extended_sysroot_tests() {
+    if (( $gcc_master_branch == 0 )); then
+        return
+    fi
+
+    pushd rand
+    cargo clean
+    echo "[TEST] rust-random/rand"
+    ../cargo.sh test --workspace
+    popd
+
+    #pushd simple-raytracer
+    #echo "[BENCH COMPILE] ebobby/simple-raytracer"
+    #hyperfine --runs "${RUN_RUNS:-10}" --warmup 1 --prepare "cargo clean" \
+    #"RUSTC=rustc RUSTFLAGS='' cargo build" \
+    #"../cargo.sh build"
+
+    #echo "[BENCH RUN] ebobby/simple-raytracer"
+    #cp ./target/debug/main ./raytracer_cg_gcc
+    #hyperfine --runs "${RUN_RUNS:-10}" ./raytracer_cg_llvm ./raytracer_cg_gcc
+    #popd
+
+    pushd regex
+    echo "[TEST] rust-lang/regex example shootout-regex-dna"
+    cargo clean
+    export CG_RUSTFLAGS="--cap-lints warn" # newer aho_corasick versions throw a deprecation warning
+    # Make sure `[codegen mono items] start` doesn't poison the diff
+    ../cargo.sh build --example shootout-regex-dna
+    cat examples/regexdna-input.txt \
+        | ../cargo.sh run --example shootout-regex-dna \
+        | grep -v "Spawned thread" > res.txt
+    diff -u res.txt examples/regexdna-output.txt
+
+    echo "[TEST] rust-lang/regex tests"
+    ../cargo.sh test --tests -- --exclude-should-panic --test-threads 1 -Zunstable-options -q
+    popd
+}
+
 function test_rustc() {
     echo
     echo "[TEST] rust-lang/rust"
@@ -165,23 +237,7 @@ function test_rustc() {
     git checkout $(rustc -V | cut -d' ' -f3 | tr -d '(')
     export RUSTFLAGS=
 
-    git apply - <<EOF
+    git apply ../rustc_patches/compile_test.patch || true
-diff --git a/src/tools/compiletest/src/header.rs b/src/tools/compiletest/src/header.rs
-index 887d27fd6dca4..2c2239f2b83d1 100644
---- a/src/tools/compiletest/src/header.rs
-+++ b/src/tools/compiletest/src/header.rs
-@@ -806,8 +806,8 @@ pub fn make_test_description<R: Read>(
-     cfg: Option<&str>,
- ) -> test::TestDesc {
-     let mut ignore = false;
-     #[cfg(not(bootstrap))]
--    let ignore_message: Option<String> = None;
-+    let ignore_message: Option<&str> = None;
-     let mut should_fail = false;
-
-     let rustc_has_profiler_support = env::var_os("RUSTC_PROFILER_SUPPORT").is_some();
-
-EOF
 
     rm config.toml || true
 
@@ -205,7 +261,7 @@ EOF
 
     git checkout -- src/test/ui/issues/auxiliary/issue-3136-a.rs # contains //~ERROR, but shouldn't be removed
 
-    rm -r src/test/ui/{abi*,extern/,panic-runtime/,panics/,unsized-locals/,proc-macro/,threads-sendsync/,thinlto/,simd*,borrowck/,test*,*lto*.rs} || true
+    rm -r src/test/ui/{abi*,extern/,panic-runtime/,panics/,unsized-locals/,proc-macro/,threads-sendsync/,thinlto/,borrowck/,test*,*lto*.rs} || true
     for test in $(rg --files-with-matches "catch_unwind|should_panic|thread|lto" src/test/ui); do
       rm $test
     done
@@ -222,33 +278,14 @@ function clean_ui_tests() {
     find rust/build/x86_64-unknown-linux-gnu/test/ui/ -name stamp -exec rm -rf {} \;
 }
 
-case $1 in
+function all() {
-    "--test-rustc")
+    clean
-        test_rustc
+    mini_tests
-        ;;
+    build_sysroot
+    std_tests
+    test_libcore
+    extended_sysroot_tests
+    test_rustc
+}
 
-    "--test-libcore")
+$func
-        test_libcore
-        ;;
-
-    "--clean-ui-tests")
-        clean_ui_tests
-        ;;
-
-    "--std-tests")
-        std_tests
-        ;;
-
-    "--build-sysroot")
-        build_sysroot
-        ;;
-
-    *)
-        clean
-        mini_tests
-        build_sysroot
-        std_tests
-        test_libcore
-        test_rustc
-        ;;
-esac

tests/lang_tests_common.rs

@@ -1,3 +1,4 @@
+//! The common code for `tests/lang_tests_*.rs`
 use std::{
     env::{self, current_dir},
     path::PathBuf,
@@ -7,7 +8,15 @@ use std::{
 use lang_tester::LangTester;
 use tempfile::TempDir;
 
-fn main() {
+/// Controls the compile options (e.g., optimization level) used to compile
+/// test code.
+#[allow(dead_code)] // Each test crate picks one variant
+pub enum Profile {
+    Debug,
+    Release,
+}
+
+pub fn main_inner(profile: Profile) {
     let tempdir = TempDir::new().expect("temp dir");
     let current_dir = current_dir().expect("current dir");
     let current_dir = current_dir.to_str().expect("current dir").to_string();
@@ -42,6 +51,15 @@ fn main() {
                 "-o", exe.to_str().expect("to_str"),
                 path.to_str().expect("to_str"),
             ]);
+            match profile {
+                Profile::Debug => {}
+                Profile::Release => {
+                    compiler.args(&[
+                        "-C", "opt-level=3",
+                        "-C", "lto=no",
+                    ]);
+                }
+            }
             // Test command 2: run `tempdir/x`.
             let runtime = Command::new(exe);
             vec![("Compiler", compiler), ("Run-time", runtime)]

tests/lang_tests_debug.rs

@@ -0,0 +1,5 @@
+mod lang_tests_common;
+
+fn main() {
+    lang_tests_common::main_inner(lang_tests_common::Profile::Debug);
+}

tests/lang_tests_release.rs

@@ -0,0 +1,5 @@
+mod lang_tests_common;
+
+fn main() {
+    lang_tests_common::main_inner(lang_tests_common::Profile::Release);
+}

tests/run/int.rs

@@ -3,32 +3,13 @@
 // Run-time:
 //   status: 0
 
-#![feature(arbitrary_self_types, auto_traits, core_intrinsics, lang_items, start, intrinsics)]
+#![feature(bench_black_box, const_black_box, core_intrinsics, start)]
 
 #![no_std]
 
-mod intrinsics {
-    extern "rust-intrinsic" {
-        pub fn abort() -> !;
-    }
-}
-
-/*
- * Core
- */
-
-mod libc {
-    #[link(name = "c")]
-    extern "C" {
-        pub fn puts(s: *const u8) -> i32;
-    }
-}
-
 #[panic_handler]
 fn panic_handler(_: &core::panic::PanicInfo) -> ! {
-    unsafe {
+    core::intrinsics::abort();
-        core::intrinsics::abort();
-    }
 }
 
 /*
@@ -36,118 +17,324 @@ fn panic_handler(_: &core::panic::PanicInfo) -> ! {
  */
 
 #[start]
-fn main(argc: isize, _argv: *const *const u8) -> isize {
+fn main(_argc: isize, _argv: *const *const u8) -> isize {
-    let var = 134217856_u128;
+    use core::hint::black_box;
-    let var2 = 10475372733397991552_u128;
-    let var3 = 193236519889708027473620326106273939584_u128;
-    let var4 = 123236519889708027473620326106273939584_u128;
-    let var5 = 153236519889708027473620326106273939584_u128;
-    let var6 = 18446744073709551616_i128;
-    let var7 = 170141183460469231731687303715884105728_u128;
 
-    // Shifts.
+    macro_rules! check {
-    assert_eq!(var << (argc as u128 - 1), var);
+        ($ty:ty, $expr:expr) => {
-    assert_eq!(var << argc as u128, 268435712);
+            {
-    assert_eq!(var << (argc + 32) as u128, 1152922604118474752);
+                const EXPECTED: $ty = $expr;
-    assert_eq!(var << (argc + 48) as u128, 75557935783508361347072);
+                assert_eq!($expr, EXPECTED);
-    assert_eq!(var << (argc + 60) as u128, 309485304969250248077606912);
+            }
-    assert_eq!(var << (argc + 62) as u128, 1237941219877000992310427648);
+        };
-    assert_eq!(var << (argc + 63) as u128, 2475882439754001984620855296);
+    }
-    assert_eq!(var << (argc + 80) as u128, 324518863143436548128224745357312);
 
-    assert_eq!(var2 << argc as u128, 20950745466795983104);
+    check!(u32, (2220326408_u32 + black_box(1)) >> (32 - 6));
-    assert_eq!(var2 << (argc as u128 - 1), var2);
-    assert_eq!(var2 << (argc + 32) as u128, 89982766606709001335848566784);
-    assert_eq!(var2 << (argc + 48) as u128, 5897110592337281111546171672756224);
-    assert_eq!(var2 << (argc + 60) as u128, 24154564986213503432893119171609493504);
-    assert_eq!(var2 << (argc + 62) as u128, 96618259944854013731572476686437974016);
-    assert_eq!(var2 << (argc + 63) as u128, 193236519889708027463144953372875948032);
 
-    assert_eq!(var3 << argc as u128, 46190672858477591483866044780779667712);
+    /// Generate `check!` tests for integer types at least as wide as 128 bits.
-    assert_eq!(var3 << (argc as u128 - 1), var3);
+    macro_rules! check_ops128 {
-    assert_eq!(var3 << (argc + 32) as u128, 21267668304951024224840338247585366016);
+        () => {
-    assert_eq!(var3 << (argc + 48) as u128, 1335125106377253154015353231953100800);
+            check_ops64!();
-    assert_eq!(var3 << (argc + 60) as u128, 24154564986213503432893119171609493504);
-    assert_eq!(var3 << (argc + 62) as u128, 96618259944854013731572476686437974016);
-    assert_eq!(var3 << (argc + 63) as u128, 193236519889708027463144953372875948032);
 
-    assert_eq!((2220326408_u32 + argc as u32) >> (32 - 6), 33);
+            // Shifts.
+            check!(T, VAL1 << black_box(64));
+            check!(T, VAL1 << black_box(81));
+            check!(T, VAL3 << black_box(63));
+            check!(T, VAL3 << black_box(64));
 
-    assert_eq!(var >> (argc as u128 - 1), var);
+            check!(T, VAL1 >> black_box(64));
-    assert_eq!(var >> argc as u128, 67108928);
+            check!(T, VAL2 >> black_box(64));
-    assert_eq!(var >> (argc + 32) as u128, 0);
+            check!(T, VAL3 >> black_box(64));
-    assert_eq!(var >> (argc + 48) as u128, 0);
+            check!(T, VAL3 >> black_box(81));
-    assert_eq!(var >> (argc + 60) as u128, 0);
+        };
-    assert_eq!(var >> (argc + 62) as u128, 0);
+    }
-    assert_eq!(var >> (argc + 63) as u128, 0);
 
-    assert_eq!(var2 >> argc as u128, 5237686366698995776);
+    /// Generate `check!` tests for integer types at least as wide as 64 bits.
-    assert_eq!(var2 >> (argc as u128 - 1), var2);
+    macro_rules! check_ops64 {
-    assert_eq!(var2 >> (argc + 32) as u128, 1219493888);
+        () => {
-    assert_eq!(var2 >> (argc + 48) as u128, 18608);
+            check_ops32!();
-    assert_eq!(var2 >> (argc + 60) as u128, 4);
-    assert_eq!(var2 >> (argc + 62) as u128, 1);
-    assert_eq!(var2 >> (argc + 63) as u128, 0);
 
-    assert_eq!(var3 >> (argc as u128 - 1), var3);
+            // Shifts.
-    assert_eq!(var3 >> argc as u128, 96618259944854013736810163053136969792);
+            check!(T, VAL2 << black_box(33));
-    assert_eq!(var3 >> (argc + 32) as u128, 22495691651677250335181635584);
+            check!(T, VAL2 << black_box(49));
-    assert_eq!(var3 >> (argc + 48) as u128, 343257013727985387194544);
+            check!(T, VAL2 << black_box(61));
-    assert_eq!(var3 >> (argc + 60) as u128, 83802981867183932420);
+            check!(T, VAL2 << black_box(63));
-    assert_eq!(var3 >> (argc + 62) as u128, 20950745466795983105);
-    assert_eq!(var3 >> (argc + 63) as u128, 10475372733397991552);
-    assert_eq!(var3 >> (argc + 80) as u128, 79920751444992);
 
-    assert_eq!(var6 >> argc as u128, 9223372036854775808);
+            check!(T, VAL3 << black_box(33));
-    assert_eq!((var6 - 1) >> argc as u128, 9223372036854775807);
+            check!(T, VAL3 << black_box(49));
-    assert_eq!(var7 >> argc as u128, 85070591730234615865843651857942052864);
+            check!(T, VAL3 << black_box(61));
 
-    // Casts
+            check!(T, VAL1 >> black_box(33));
-    assert_eq!((var >> (argc + 32) as u128) as u64, 0);
+            check!(T, VAL1 >> black_box(49));
-    assert_eq!((var >> argc as u128) as u64, 67108928);
+            check!(T, VAL1 >> black_box(61));
+            check!(T, VAL1 >> black_box(63));
 
-    // Addition.
+            check!(T, VAL2 >> black_box(33));
-    assert_eq!(var + argc as u128, 134217857);
+            check!(T, VAL2 >> black_box(49));
+            check!(T, VAL2 >> black_box(61));
+            check!(T, VAL2 >> black_box(63));
 
-    assert_eq!(var2 + argc as u128, 10475372733397991553);
+            check!(T, VAL3 >> black_box(33));
-    assert_eq!(var2 + (var2 + argc as u128) as u128, 20950745466795983105);
+            check!(T, VAL3 >> black_box(49));
+            check!(T, VAL3 >> black_box(61));
+            check!(T, VAL3 >> black_box(63));
+        };
+    }
 
-    assert_eq!(var3 + argc as u128, 193236519889708027473620326106273939585);
+    /// Generate `check!` tests for integer types at least as wide as 32 bits.
+    macro_rules! check_ops32 {
+        () => {
+            // Shifts.
+            check!(T, VAL2 << black_box(1));
+            check!(T, VAL2 << black_box(0));
 
-    // Subtraction
+            check!(T, VAL3 << black_box(1));
-    assert_eq!(var - argc as u128, 134217855);
+            check!(T, VAL3 << black_box(0));
 
-    assert_eq!(var2 - argc as u128, 10475372733397991551);
+            check!(T, VAL1.wrapping_shl(black_box(0)));
+            check!(T, VAL1.wrapping_shl(black_box(1)));
+            check!(T, VAL1.wrapping_shl(black_box(33)));
+            check!(T, VAL1.wrapping_shl(black_box(49)));
+            check!(T, VAL1.wrapping_shl(black_box(61)));
+            check!(T, VAL1.wrapping_shl(black_box(63)));
+            check!(T, VAL1.wrapping_shl(black_box(64)));
+            check!(T, VAL1.wrapping_shl(black_box(81)));
 
-    assert_eq!(var3 - argc as u128, 193236519889708027473620326106273939583);
+            check!(Option<T>, VAL1.checked_shl(black_box(0)));
+            check!(Option<T>, VAL1.checked_shl(black_box(1)));
+            check!(Option<T>, VAL1.checked_shl(black_box(33)));
+            check!(Option<T>, VAL1.checked_shl(black_box(49)));
+            check!(Option<T>, VAL1.checked_shl(black_box(61)));
+            check!(Option<T>, VAL1.checked_shl(black_box(63)));
+            check!(Option<T>, VAL1.checked_shl(black_box(64)));
+            check!(Option<T>, VAL1.checked_shl(black_box(81)));
 
-    // Multiplication
+            check!(T, VAL1 >> black_box(0));
-    assert_eq!(var * (argc + 1) as u128, 268435712);
+            check!(T, VAL1 >> black_box(1));
-    assert_eq!(var * (argc as u128 + var2), 1405982069077538020949770368);
 
-    assert_eq!(var2 * (argc + 1) as u128, 20950745466795983104);
+            check!(T, VAL2 >> black_box(1));
-    assert_eq!(var2 * (argc as u128 + var2), 109733433903618109003204073240861360256);
+            check!(T, VAL2 >> black_box(0));
 
-    assert_eq!(var3 * argc as u128, 193236519889708027473620326106273939584);
+            check!(T, VAL3 >> black_box(0));
+            check!(T, VAL3 >> black_box(1));
 
-    assert_eq!(var4 * (argc + 1) as u128, 246473039779416054947240652212547879168);
+            check!(T, VAL1.wrapping_shr(black_box(0)));
+            check!(T, VAL1.wrapping_shr(black_box(1)));
+            check!(T, VAL1.wrapping_shr(black_box(33)));
+            check!(T, VAL1.wrapping_shr(black_box(49)));
+            check!(T, VAL1.wrapping_shr(black_box(61)));
+            check!(T, VAL1.wrapping_shr(black_box(63)));
+            check!(T, VAL1.wrapping_shr(black_box(64)));
+            check!(T, VAL1.wrapping_shr(black_box(81)));
 
-    assert_eq!(var5 * (argc + 1) as u128, 306473039779416054947240652212547879168);
+            check!(Option<T>, VAL1.checked_shr(black_box(0)));
+            check!(Option<T>, VAL1.checked_shr(black_box(1)));
+            check!(Option<T>, VAL1.checked_shr(black_box(33)));
+            check!(Option<T>, VAL1.checked_shr(black_box(49)));
+            check!(Option<T>, VAL1.checked_shr(black_box(61)));
+            check!(Option<T>, VAL1.checked_shr(black_box(63)));
+            check!(Option<T>, VAL1.checked_shr(black_box(64)));
+            check!(Option<T>, VAL1.checked_shr(black_box(81)));
 
-    // Division.
+            // Casts
-    assert_eq!(var / (argc + 1) as u128, 67108928);
+            check!(u64, (VAL1 >> black_box(1)) as u64);
-    assert_eq!(var / (argc + 2) as u128, 44739285);
 
-    assert_eq!(var2 / (argc + 1) as u128, 5237686366698995776);
+            // Addition.
-    assert_eq!(var2 / (argc + 2) as u128, 3491790911132663850);
+            check!(T, VAL1 + black_box(1));
+            check!(T, VAL2 + black_box(1));
+            check!(T, VAL2 + (VAL2 + black_box(1)));
+            check!(T, VAL3 + black_box(1));
 
-    assert_eq!(var3 / (argc + 1) as u128, 96618259944854013736810163053136969792);
+            check!(Option<T>, VAL1.checked_add(black_box(1)));
-    assert_eq!(var3 / (argc + 2) as u128, 64412173296569342491206775368757979861);
+            check!(Option<T>, VAL2.checked_add(black_box(1)));
-    assert_eq!(var3 / (argc as u128 + var4), 1);
+            check!(Option<T>, VAL2.checked_add(VAL2 + black_box(1)));
-    assert_eq!(var3 / (argc as u128 + var2), 18446744073709551615);
+            check!(Option<T>, VAL3.checked_add(T::MAX));
+            check!(Option<T>, VAL3.checked_add(T::MIN));
 
-    assert_eq!(var4 / (argc + 1) as u128, 61618259944854013736810163053136969792);
+            check!(T, VAL1.wrapping_add(black_box(1)));
-    assert_eq!(var4 / (argc + 2) as u128, 41078839963236009157873442035424646528);
+            check!(T, VAL2.wrapping_add(black_box(1)));
+            check!(T, VAL2.wrapping_add(VAL2 + black_box(1)));
+            check!(T, VAL3.wrapping_add(T::MAX));
+            check!(T, VAL3.wrapping_add(T::MIN));
+
+            check!((T, bool), VAL1.overflowing_add(black_box(1)));
+            check!((T, bool), VAL2.overflowing_add(black_box(1)));
+            check!((T, bool), VAL2.overflowing_add(VAL2 + black_box(1)));
+            check!((T, bool), VAL3.overflowing_add(T::MAX));
+            check!((T, bool), VAL3.overflowing_add(T::MIN));
+
+            check!(T, VAL1.saturating_add(black_box(1)));
+            check!(T, VAL2.saturating_add(black_box(1)));
+            check!(T, VAL2.saturating_add(VAL2 + black_box(1)));
+            check!(T, VAL3.saturating_add(T::MAX));
+            check!(T, VAL3.saturating_add(T::MIN));
+
+            // Subtraction
+            check!(T, VAL1 - black_box(1));
+            check!(T, VAL2 - black_box(1));
+            check!(T, VAL3 - black_box(1));
+
+            check!(Option<T>, VAL1.checked_sub(black_box(1)));
+            check!(Option<T>, VAL2.checked_sub(black_box(1)));
+            check!(Option<T>, VAL2.checked_sub(VAL2 + black_box(1)));
+            check!(Option<T>, VAL3.checked_sub(T::MAX));
+            check!(Option<T>, VAL3.checked_sub(T::MIN));
+
+            check!(T, VAL1.wrapping_sub(black_box(1)));
+            check!(T, VAL2.wrapping_sub(black_box(1)));
+            check!(T, VAL2.wrapping_sub(VAL2 + black_box(1)));
+            check!(T, VAL3.wrapping_sub(T::MAX));
+            check!(T, VAL3.wrapping_sub(T::MIN));
+
+            check!((T, bool), VAL1.overflowing_sub(black_box(1)));
+            check!((T, bool), VAL2.overflowing_sub(black_box(1)));
+            check!((T, bool), VAL2.overflowing_sub(VAL2 + black_box(1)));
+            check!((T, bool), VAL3.overflowing_sub(T::MAX));
+            check!((T, bool), VAL3.overflowing_sub(T::MIN));
+
+            check!(T, VAL1.saturating_sub(black_box(1)));
+            check!(T, VAL2.saturating_sub(black_box(1)));
+            check!(T, VAL2.saturating_sub(VAL2 + black_box(1)));
+            check!(T, VAL3.saturating_sub(T::MAX));
+            check!(T, VAL3.saturating_sub(T::MIN));
+
+            // Multiplication
+            check!(T, VAL1 * black_box(2));
+            check!(T, VAL1 * (black_box(1) + VAL2));
+            check!(T, VAL2 * black_box(2));
+            check!(T, VAL2 * (black_box(1) + VAL2));
+            check!(T, VAL3 * black_box(1));
+            check!(T, VAL4 * black_box(2));
+            check!(T, VAL5 * black_box(2));
+
+            check!(Option<T>, VAL1.checked_mul(black_box(2)));
+            check!(Option<T>, VAL1.checked_mul(black_box(1) + VAL2));
+            check!(Option<T>, VAL3.checked_mul(VAL3));
+            check!(Option<T>, VAL4.checked_mul(black_box(2)));
+            check!(Option<T>, VAL5.checked_mul(black_box(2)));
+
+            check!(T, VAL1.wrapping_mul(black_box(2)));
+            check!(T, VAL1.wrapping_mul((black_box(1) + VAL2)));
+            check!(T, VAL3.wrapping_mul(VAL3));
+            check!(T, VAL4.wrapping_mul(black_box(2)));
+            check!(T, VAL5.wrapping_mul(black_box(2)));
+
+            check!((T, bool), VAL1.overflowing_mul(black_box(2)));
+            check!((T, bool), VAL1.overflowing_mul(black_box(1) + VAL2));
+            check!((T, bool), VAL3.overflowing_mul(VAL3));
+            check!((T, bool), VAL4.overflowing_mul(black_box(2)));
+            check!((T, bool), VAL5.overflowing_mul(black_box(2)));
+
+            check!(T, VAL1.saturating_mul(black_box(2)));
+            check!(T, VAL1.saturating_mul(black_box(1) + VAL2));
+            check!(T, VAL3.saturating_mul(VAL3));
+            check!(T, VAL4.saturating_mul(black_box(2)));
+            check!(T, VAL5.saturating_mul(black_box(2)));
+
+            // Division.
+            check!(T, VAL1 / black_box(2));
+            check!(T, VAL1 / black_box(3));
+
+            check!(T, VAL2 / black_box(2));
+            check!(T, VAL2 / black_box(3));
+
+            check!(T, VAL3 / black_box(2));
+            check!(T, VAL3 / black_box(3));
+            check!(T, VAL3 / (black_box(1) + VAL4));
+            check!(T, VAL3 / (black_box(1) + VAL2));
+
+            check!(T, VAL4 / black_box(2));
+            check!(T, VAL4 / black_box(3));
+
+            check!(Option<T>, VAL1.checked_div(black_box(2)));
+            check!(Option<T>, VAL1.checked_div(black_box(1) + VAL2));
+            check!(Option<T>, VAL3.checked_div(VAL3));
+            check!(Option<T>, VAL4.checked_div(black_box(2)));
+            check!(Option<T>, VAL5.checked_div(black_box(2)));
+            check!(Option<T>, (T::MIN).checked_div(black_box(0 as T).wrapping_sub(1)));
+            check!(Option<T>, VAL5.checked_div(black_box(0))); // var5 / 0
+
+            check!(T, VAL1.wrapping_div(black_box(2)));
+            check!(T, VAL1.wrapping_div(black_box(1) + VAL2));
+            check!(T, VAL3.wrapping_div(VAL3));
+            check!(T, VAL4.wrapping_div(black_box(2)));
+            check!(T, VAL5.wrapping_div(black_box(2)));
+            check!(T, (T::MIN).wrapping_div(black_box(0 as T).wrapping_sub(1)));
+
+            check!((T, bool), VAL1.overflowing_div(black_box(2)));
+            check!((T, bool), VAL1.overflowing_div(black_box(1) + VAL2));
+            check!((T, bool), VAL3.overflowing_div(VAL3));
+            check!((T, bool), VAL4.overflowing_div(black_box(2)));
+            check!((T, bool), VAL5.overflowing_div(black_box(2)));
+            check!((T, bool), (T::MIN).overflowing_div(black_box(0 as T).wrapping_sub(1)));
+
+            check!(T, VAL1.saturating_div(black_box(2)));
+            check!(T, VAL1.saturating_div((black_box(1) + VAL2)));
+            check!(T, VAL3.saturating_div(VAL3));
+            check!(T, VAL4.saturating_div(black_box(2)));
+            check!(T, VAL5.saturating_div(black_box(2)));
+            check!(T, (T::MIN).saturating_div((0 as T).wrapping_sub(black_box(1))));
+        };
+    }
+
+    {
+        type T = u32;
+        const VAL1: T = 14162_u32;
+        const VAL2: T = 14556_u32;
+        const VAL3: T = 323656954_u32;
+        const VAL4: T = 2023651954_u32;
+        const VAL5: T = 1323651954_u32;
+        check_ops32!();
+    }
+
+    {
+        type T = i32;
+        const VAL1: T = 13456_i32;
+        const VAL2: T = 10475_i32;
+        const VAL3: T = 923653954_i32;
+        const VAL4: T = 993198738_i32;
+        const VAL5: T = 1023653954_i32;
+        check_ops32!();
+    }
+
+    {
+        type T = u64;
+        const VAL1: T = 134217856_u64;
+        const VAL2: T = 104753732_u64;
+        const VAL3: T = 12323651988970863954_u64;
+        const VAL4: T = 7323651988970863954_u64;
+        const VAL5: T = 8323651988970863954_u64;
+        check_ops64!();
+    }
+
+    {
+        type T = i64;
+        const VAL1: T = 134217856_i64;
+        const VAL2: T = 104753732_i64;
+        const VAL3: T = 6323651988970863954_i64;
+        const VAL4: T = 2323651988970863954_i64;
+        const VAL5: T = 3323651988970863954_i64;
+        check_ops64!();
+    }
+
+    {
+        type T = u128;
+        const VAL1: T = 134217856_u128;
+        const VAL2: T = 10475372733397991552_u128;
+        const VAL3: T = 193236519889708027473620326106273939584_u128;
+        const VAL4: T = 123236519889708027473620326106273939584_u128;
+        const VAL5: T = 153236519889708027473620326106273939584_u128;
+        check_ops128!();
+    }
+    {
+        type T = i128;
+        const VAL1: T = 134217856_i128;
+        const VAL2: T = 10475372733397991552_i128;
+        const VAL3: T = 83236519889708027473620326106273939584_i128;
+        const VAL4: T = 63236519889708027473620326106273939584_i128;
+        const VAL5: T = 73236519889708027473620326106273939584_i128;
+        check_ops128!();
+    }
 
     0
 }

tests/run/int_overflow.rs

@@ -1,7 +1,7 @@
 // Compiler:
 //
 // Run-time:
-//   stdout: Panicking
+//   stdout: Success
 //   status: signal
 
 #![allow(unused_attributes)]
@@ -64,7 +64,9 @@ mod intrinsics {
 #[no_mangle]
 pub fn panic(_msg: &str) -> ! {
     unsafe {
-        libc::puts("Panicking\0" as *const str as *const u8);
+        // Panicking is expected iff overflow checking is enabled.
+        #[cfg(debug_assertions)]
+        libc::puts("Success\0" as *const str as *const u8);
         libc::fflush(libc::stdout);
         intrinsics::abort();
     }
@@ -124,6 +126,15 @@ impl Add for isize {
 #[start]
 fn main(mut argc: isize, _argv: *const *const u8) -> isize {
     let int = 9223372036854775807isize;
-    let int = int + argc;
+    let int = int + argc;  // overflow
+
+    // If overflow checking is disabled, we should reach here.
+    #[cfg(not(debug_assertions))]
+    unsafe {
+        libc::puts("Success\0" as *const str as *const u8);
+        libc::fflush(libc::stdout);
+        intrinsics::abort();
+    }
+
     int
 }

tools/generate_intrinsics.py

@@ -0,0 +1,238 @@
+import json
+import os
+import re
+import sys
+import subprocess
+from os import walk
+
+
+def run_command(command, cwd=None):
+    p = subprocess.Popen(command, cwd=cwd)
+    if p.wait() != 0:
+        print("command `{}` failed...".format(" ".join(command)))
+        sys.exit(1)
+
+
+def clone_repository(repo_name, path, repo_url, sub_path=None):
+    if os.path.exists(path):
+        while True:
+            choice = input("There is already a `{}` folder, do you want to update it? [y/N]".format(path))
+            if choice == "" or choice.lower() == "n":
+                print("Skipping repository update.")
+                return
+            elif choice.lower() == "y":
+                print("Updating repository...")
+                run_command(["git", "pull", "origin"], cwd=path)
+                return
+            else:
+                print("Didn't understand answer...")
+    print("Cloning {} repository...".format(repo_name))
+    if sub_path is None:
+        run_command(["git", "clone", repo_url, "--depth", "1", path])
+    else:
+        run_command(["git", "clone", repo_url, "--filter=tree:0", "--no-checkout", path])
+        run_command(["git", "sparse-checkout", "init"], cwd=path)
+        run_command(["git", "sparse-checkout", "set", "add", sub_path], cwd=path)
+        run_command(["git", "checkout"], cwd=path)
+
+
+def append_intrinsic(array, intrinsic_name, translation):
+    array.append((intrinsic_name, translation))
+
+
+def extract_instrinsics(intrinsics, file):
+    print("Extracting intrinsics from `{}`...".format(file))
+    with open(file, "r", encoding="utf8") as f:
+        content = f.read()
+
+    lines = content.splitlines()
+    pos = 0
+    current_arch = None
+    while pos < len(lines):
+        line = lines[pos].strip()
+        if line.startswith("let TargetPrefix ="):
+            current_arch = line.split('"')[1].strip()
+            if len(current_arch) == 0:
+                current_arch = None
+        elif current_arch is None:
+            pass
+        elif line == "}":
+            current_arch = None
+        elif line.startswith("def "):
+            content = ""
+            while not content.endswith(";") and not content.endswith("}") and pos < len(lines):
+                line = lines[pos].split(" // ")[0].strip()
+                content += line
+                pos += 1
+            entries = re.findall('GCCBuiltin<"(\\w+)">', content)
+            if len(entries) > 0:
+                intrinsic = content.split("def ")[1].strip().split(":")[0].strip()
+                intrinsic = intrinsic.split("_")
+                if len(intrinsic) < 2 or intrinsic[0] != "int":
+                    continue
+                intrinsic[0] = "llvm"
+                intrinsic = ".".join(intrinsic)
+                if current_arch not in intrinsics:
+                    intrinsics[current_arch] = []
+                for entry in entries:
+                    append_intrinsic(intrinsics[current_arch], intrinsic, entry)
+            continue
+        pos += 1
+        continue
+    print("Done!")
+
+
+def extract_instrinsics_from_llvm(llvm_path, intrinsics):
+    files = []
+    intrinsics_path = os.path.join(llvm_path, "llvm/include/llvm/IR")
+    for (dirpath, dirnames, filenames) in walk(intrinsics_path):
+        files.extend([os.path.join(intrinsics_path, f) for f in filenames if f.endswith(".td")])
+
+    for file in files:
+        extract_instrinsics(intrinsics, file)
+
+
+def append_translation(json_data, p, array):
+    it = json_data["index"][p]
+    content = it["docs"].split('`')
+    if len(content) != 5:
+        return
+    append_intrinsic(array, content[1], content[3])
+
+
+def extract_instrinsics_from_llvmint(llvmint, intrinsics):
+    archs = [
+        "AMDGPU",
+        "aarch64",
+        "arm",
+        "cuda",
+        "hexagon",
+        "mips",
+        "nvvm",
+        "ppc",
+        "ptx",
+        "x86",
+        "xcore",
+    ]
+
+    json_file = os.path.join(llvmint, "target/doc/llvmint.json")
+    # We need to regenerate the documentation!
+    run_command(
+        ["cargo", "rustdoc", "--", "-Zunstable-options", "--output-format", "json"],
+        cwd=llvmint,
+    )
+    with open(json_file, "r", encoding="utf8") as f:
+        json_data = json.loads(f.read())
+    for p in json_data["paths"]:
+        it = json_data["paths"][p]
+        if it["crate_id"] != 0:
+            # This is from an external crate.
+            continue
+        if it["kind"] != "function":
+            # We're only looking for functions.
+            continue
+        # if len(it["path"]) == 2:
+        #   # This is a "general" intrinsic, not bound to a specific arch.
+        #   append_translation(json_data, p, general)
+        #   continue
+        if len(it["path"]) != 3 or it["path"][1] not in archs:
+            continue
+        arch = it["path"][1]
+        if arch not in intrinsics:
+            intrinsics[arch] = []
+        append_translation(json_data, p, intrinsics[arch])
+
+
+def fill_intrinsics(intrinsics, from_intrinsics, all_intrinsics):
+    for arch in from_intrinsics:
+        if arch not in intrinsics:
+            intrinsics[arch] = []
+        for entry in from_intrinsics[arch]:
+            if entry[0] in all_intrinsics:
+                if all_intrinsics[entry[0]] == entry[1]:
+                    # This is a "full" duplicate, both the LLVM instruction and the GCC
+                    # translation are the same.
+                    continue
+                intrinsics[arch].append((entry[0], entry[1], True))
+            else:
+                intrinsics[arch].append((entry[0], entry[1], False))
+                all_intrinsics[entry[0]] = entry[1]
+
+
+def update_intrinsics(llvm_path, llvmint, llvmint2):
+    intrinsics_llvm = {}
+    intrinsics_llvmint = {}
+    all_intrinsics = {}
+
+    extract_instrinsics_from_llvm(llvm_path, intrinsics_llvm)
+    extract_instrinsics_from_llvmint(llvmint, intrinsics_llvmint)
+    extract_instrinsics_from_llvmint(llvmint2, intrinsics_llvmint)
+
+    intrinsics = {}
+    # We give priority to translations from LLVM over the ones from llvmint.
+    fill_intrinsics(intrinsics, intrinsics_llvm, all_intrinsics)
+    fill_intrinsics(intrinsics, intrinsics_llvmint, all_intrinsics)
+
+    archs = [arch for arch in intrinsics]
+    archs.sort()
+
+    output_file = os.path.join(
+        os.path.dirname(os.path.abspath(__file__)),
+        "../src/intrinsic/archs.rs",
+    )
+    print("Updating content of `{}`...".format(output_file))
+    with open(output_file, "w", encoding="utf8") as out:
+        out.write("// File generated by `rustc_codegen_gcc/tools/generate_intrinsics.py`\n")
+        out.write("// DO NOT EDIT IT!\n")
+        out.write("match name {\n")
+        for arch in archs:
+            if len(intrinsics[arch]) == 0:
+                continue
+            intrinsics[arch].sort(key=lambda x: (x[0], x[2]))
+            out.write('    // {}\n'.format(arch))
+            for entry in intrinsics[arch]:
+                if entry[2] == True: # if it is a duplicate
+                    out.write('    // [DUPLICATE]: "{}" => "{}",\n'.format(entry[0], entry[1]))
+                else:
+                    out.write('    "{}" => "{}",\n'.format(entry[0], entry[1]))
+        out.write('    _ => unimplemented!("***** unsupported LLVM intrinsic {}", name),\n')
+        out.write("}\n")
+    print("Done!")
+
+
+def main():
+    llvm_path = os.path.join(
+        os.path.dirname(os.path.abspath(__file__)),
+        "llvm-project",
+    )
+    llvmint_path = os.path.join(
+        os.path.dirname(os.path.abspath(__file__)),
+        "llvmint",
+    )
+    llvmint2_path = os.path.join(
+        os.path.dirname(os.path.abspath(__file__)),
+        "llvmint-2",
+    )
+
+    # First, we clone the LLVM repository if it's not already here.
+    clone_repository(
+        "llvm-project",
+        llvm_path,
+        "https://github.com/llvm/llvm-project",
+        sub_path="llvm/include/llvm/IR",
+    )
+    clone_repository(
+        "llvmint",
+        llvmint_path,
+        "https://github.com/GuillaumeGomez/llvmint",
+    )
+    clone_repository(
+        "llvmint2",
+        llvmint2_path,
+        "https://github.com/antoyo/llvmint",
+    )
+    update_intrinsics(llvm_path, llvmint_path, llvmint2_path)
+
+
+if __name__ == "__main__":
+    sys.exit(main())