From 1988cf4a18f9c0d2a3fd6ebced72190a600f247f Mon Sep 17 00:00:00 2001 From: bjorn3 <17426603+bjorn3@users.noreply.github.com> Date: Sat, 25 Nov 2023 10:05:52 +0000 Subject: [PATCH] Merge commit '710c67909d034e1c663174a016ca82b95c2d6c12' into sync_cg_clif-2023-11-25 --- Cargo.lock | 60 ++--- Cargo.toml | 12 +- build_system/tests.rs | 5 + config.txt | 1 + example/polymorphize_coroutine.rs | 16 ++ rust-toolchain | 2 +- scripts/test_rustc_tests.sh | 5 - src/intrinsics/llvm_x86.rs | 372 ++++++++++++++++++++++++++++-- src/pretty_clif.rs | 16 +- src/value_and_place.rs | 34 ++- 10 files changed, 452 insertions(+), 71 deletions(-) create mode 100644 example/polymorphize_coroutine.rs diff --git a/Cargo.lock b/Cargo.lock index dcb6cc57584..901d1dbea66 100644 --- a/Cargo.lock +++ b/Cargo.lock @@ -21,9 +21,9 @@ checksum = "a4668cab20f66d8d020e1fbc0ebe47217433c1b6c8f2040faf858554e394ace6" [[package]] name = "arbitrary" -version = "1.3.0" +version = "1.3.2" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "e2d098ff73c1ca148721f37baad5ea6a465a13f9573aba8641fbbbae8164a54e" +checksum = "7d5a26814d8dcb93b0e5a0ff3c6d80a8843bafb21b39e8e18a6f05471870e110" [[package]] name = "bitflags" @@ -45,18 +45,18 @@ checksum = "baf1de4339761588bc0619e3cbc0120ee582ebb74b53b4efbf79117bd2da40fd" [[package]] name = "cranelift-bforest" -version = "0.101.2" +version = "0.102.0" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "f773437307980ac0f424bf9b9a5d0cd21a0f17248c6860c9a65bec8b5975f3fe" +checksum = "76eb38f2af690b5a4411d9a8782b6d77dabff3ca939e0518453ab9f9a4392d41" dependencies = [ "cranelift-entity", ] [[package]] name = "cranelift-codegen" -version = "0.101.2" +version = "0.102.0" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "443c2ac50e97fb7de1a0f862753fce3f27215558811a6fcee508eb0c3747fa79" +checksum = "39526c036b92912417e8931f52c1e235796688068d3efdbbd8b164f299d19156" dependencies = [ "bumpalo", "cranelift-bforest", @@ -75,39 +75,39 @@ dependencies = [ [[package]] name = "cranelift-codegen-meta" -version = "0.101.2" +version = "0.102.0" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "c5b174c411480c79ce0793c55042fa51bec27e486381d103a53cab3b480cb2db" +checksum = "fdb0deedc9fccf2db53a5a3c9c9d0163e44143b0d004dca9bf6ab6a0024cd79a" dependencies = [ "cranelift-codegen-shared", ] [[package]] name = "cranelift-codegen-shared" -version = "0.101.2" +version = "0.102.0" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "73fa0151a528066a369de6debeea4d4b23a32aba68b5add8c46d3dc8091ff434" +checksum = "cea2d1b274e45aa8e61e9103efa1ba82d4b5a19d12bd1fd10744c3b7380ba3ff" [[package]] name = "cranelift-control" -version = "0.101.2" +version = "0.102.0" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "b8adf1e6398493c9bea1190e37d28a0eb0eca5fddbc80e01e506cda34db92b1f" +checksum = "6ea5977559a71e63db79a263f0e81a89b996e8a38212c4281e37dd1dbaa8b65c" dependencies = [ "arbitrary", ] [[package]] name = "cranelift-entity" -version = "0.101.2" +version = "0.102.0" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "4917e2ed3bb5fe87d0ed88395ca6d644018d119a034faedd1f3e1f2c33cd52b2" +checksum = "2f871ada808b58158d84dfc43a6a2e2d2756baaf4ed1c51fd969ca8330e6ca5c" [[package]] name = "cranelift-frontend" -version = "0.101.2" +version = "0.102.0" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "9aaadf1e7cf28886bbf046eaf7ef538997bc8a7e020e578ea4957b39da87d5a1" +checksum = "e8e6890f587ef59824b3debe577e68fdf9b307b3808c54b8d93a18fd0b70941b" dependencies = [ "cranelift-codegen", "log", @@ -117,15 +117,15 @@ dependencies = [ [[package]] name = "cranelift-isle" -version = "0.101.2" +version = "0.102.0" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "a67fda31b9d69eaa1c49a2081939454c45857596a9d45af6744680541c628b4c" +checksum = "a8d5fc6d5d3b52d1917002b17a8ecce448c2621b5bf394bb4e77e2f676893537" [[package]] name = "cranelift-jit" -version = "0.101.2" +version = "0.102.0" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "d6bf32710628e7ff298739f1ed80a0bfdafc0c6a3e284c4540b23f18e8889d4b" +checksum = "e8a2d7744f743f59d9646d7589ad22ea17ed0d71e04906eb77c31e99bc13bd8b" dependencies = [ "anyhow", "cranelift-codegen", @@ -143,9 +143,9 @@ dependencies = [ [[package]] name = "cranelift-module" -version = "0.101.2" +version = "0.102.0" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "4d693e93a0fbf56b4bc93cffe6b107c2e52f070e1111950505fc8c83ac440b9d" +checksum = "b96cb196334698e612c197d7d0ae59af5e07667306ec20d7be414717db400873" dependencies = [ "anyhow", "cranelift-codegen", @@ -154,9 +154,9 @@ dependencies = [ [[package]] name = "cranelift-native" -version = "0.101.2" +version = "0.102.0" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "76fb52ba71be98312f35e798d9e98e45ab2586f27584231bf7c644fa9501e8af" +checksum = "3e10c2e7faa65d4ae7de9a83b44f2c31aca7dc638e17d0a79572fdf8103d720b" dependencies = [ "cranelift-codegen", "libc", @@ -165,9 +165,9 @@ dependencies = [ [[package]] name = "cranelift-object" -version = "0.101.2" +version = "0.102.0" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "2551b2e185022b89e9efa5e04c0f17f679b86ef73d9f7feabc48b608ff23120d" +checksum = "83ce94e18756058af8a66e3c0ba1123ae15517c72162d8060d0cb0974642adf2" dependencies = [ "anyhow", "cranelift-codegen", @@ -295,9 +295,9 @@ checksum = "dd8b5dd2ae5ed71462c540258bedcb51965123ad7e7ccf4b9a8cafaa4a63576d" [[package]] name = "regalloc2" -version = "0.9.2" +version = "0.9.3" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "5b4dcbd3a2ae7fb94b5813fa0e957c6ab51bf5d0a8ee1b69e0c2d0f1e6eb8485" +checksum = "ad156d539c879b7a24a363a2016d77961786e71f48f2e2fc8302a92abd2429a6" dependencies = [ "hashbrown 0.13.2", "log", @@ -374,9 +374,9 @@ checksum = "49874b5167b65d7193b8aba1567f5c7d93d001cafc34600cee003eda787e483f" [[package]] name = "wasmtime-jit-icache-coherence" -version = "14.0.2" +version = "15.0.0" source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "0980a96b16abbdaf829858d2389697b1d6cfc6a903873fd74b7e47a6b1045584" +checksum = "b73ad1395eda136baec5ece7e079e0536a82ef73488e345456cc9b89858ad0ec" dependencies = [ "cfg-if", "libc", diff --git a/Cargo.toml b/Cargo.toml index 30db10f7457..20fcd222732 100644 --- a/Cargo.toml +++ b/Cargo.toml @@ -8,12 +8,12 @@ crate-type = ["dylib"] [dependencies] # These have to be in sync with each other -cranelift-codegen = { version = "0.101.2", default-features = false, features = ["std", "unwind", "all-arch"] } -cranelift-frontend = { version = "0.101.2" } -cranelift-module = { version = "0.101.2" } -cranelift-native = { version = "0.101.2" } -cranelift-jit = { version = "0.101.2", optional = true } -cranelift-object = { version = "0.101.2" } +cranelift-codegen = { version = "0.102", default-features = false, features = ["std", "unwind", "all-arch"] } +cranelift-frontend = { version = "0.102" } +cranelift-module = { version = "0.102" } +cranelift-native = { version = "0.102" } +cranelift-jit = { version = "0.102", optional = true } +cranelift-object = { version = "0.102" } target-lexicon = "0.12.0" gimli = { version = "0.28", default-features = false, features = ["write"]} object = { version = "0.32", default-features = false, features = ["std", "read_core", "write", "archive", "coff", "elf", "macho", "pe"] } diff --git a/build_system/tests.rs b/build_system/tests.rs index aa50dbfdf35..3309a0a6abd 100644 --- a/build_system/tests.rs +++ b/build_system/tests.rs @@ -99,6 +99,10 @@ const BASE_SYSROOT_SUITE: &[TestCase] = &[ TestCase::build_bin_and_run("aot.mod_bench", "example/mod_bench.rs", &[]), TestCase::build_bin_and_run("aot.issue-72793", "example/issue-72793.rs", &[]), TestCase::build_bin("aot.issue-59326", "example/issue-59326.rs"), + TestCase::custom("aot.polymorphize_coroutine", &|runner| { + runner.run_rustc(&["example/polymorphize_coroutine.rs", "-Zpolymorphize"]); + runner.run_out_command("polymorphize_coroutine", &[]); + }), TestCase::build_bin_and_run("aot.neon", "example/neon.rs", &[]), TestCase::custom("aot.gen_block_iterate", &|runner| { runner.run_rustc([ @@ -466,6 +470,7 @@ impl<'a> TestRunner<'a> { cmd.arg("--target"); cmd.arg(&self.target_compiler.triple); cmd.arg("-Cpanic=abort"); + cmd.arg("-Zunstable-options"); cmd.arg("--check-cfg=cfg(no_unstable_features)"); cmd.arg("--check-cfg=cfg(jit)"); cmd.args(args); diff --git a/config.txt b/config.txt index 3cf295c003e..0b7cac18837 100644 --- a/config.txt +++ b/config.txt @@ -42,6 +42,7 @@ aot.float-minmax-pass aot.mod_bench aot.issue-72793 aot.issue-59326 +aot.polymorphize_coroutine aot.neon aot.gen_block_iterate diff --git a/example/polymorphize_coroutine.rs b/example/polymorphize_coroutine.rs new file mode 100644 index 00000000000..c965b34e13b --- /dev/null +++ b/example/polymorphize_coroutine.rs @@ -0,0 +1,16 @@ +#![feature(coroutines, coroutine_trait)] + +use std::ops::Coroutine; +use std::pin::Pin; + +fn main() { + run_coroutine::(); +} + +fn run_coroutine() { + let mut coroutine = || { + yield; + return; + }; + Pin::new(&mut coroutine).resume(()); +} diff --git a/rust-toolchain b/rust-toolchain index 80ef1e49f23..2997816d96c 100644 --- a/rust-toolchain +++ b/rust-toolchain @@ -1,3 +1,3 @@ [toolchain] -channel = "nightly-2023-11-16" +channel = "nightly-2023-11-25" components = ["rust-src", "rustc-dev", "llvm-tools"] diff --git a/scripts/test_rustc_tests.sh b/scripts/test_rustc_tests.sh index cdc78adcf85..a299b6de6b1 100755 --- a/scripts/test_rustc_tests.sh +++ b/scripts/test_rustc_tests.sh @@ -146,11 +146,6 @@ rm tests/ui/process/nofile-limit.rs # TODO some AArch64 linking issue rm tests/ui/stdio-is-blocking.rs # really slow with unoptimized libstd -# rustc bugs -# ========== -# https://github.com/rust-lang/rust/pull/116447#issuecomment-1790451463 -rm tests/ui/coroutine/gen_block_*.rs - cp ../dist/bin/rustdoc-clif ../dist/bin/rustdoc # some tests expect bin/rustdoc to exist # prevent $(RUSTDOC) from picking up the sysroot built by x.py. It conflicts with the one used by diff --git a/src/intrinsics/llvm_x86.rs b/src/intrinsics/llvm_x86.rs index 8dd2b6ed014..07b95b7933d 100644 --- a/src/intrinsics/llvm_x86.rs +++ b/src/intrinsics/llvm_x86.rs @@ -22,6 +22,11 @@ pub(crate) fn codegen_x86_llvm_intrinsic_call<'tcx>( // Spin loop hint } + "llvm.x86.avx.vzeroupper" => { + // https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm256_zeroupper&ig_expand=7218 + // Do nothing. It is a perf hint anyway. + } + // Used by is_x86_feature_detected!(); "llvm.x86.xgetbv" => { intrinsic_args!(fx, args => (xcr_no); intrinsic); @@ -69,6 +74,103 @@ pub(crate) fn codegen_x86_llvm_intrinsic_call<'tcx>( ret.write_cvalue(fx, val); } + "llvm.x86.avx2.gather.d.d" + | "llvm.x86.avx2.gather.d.q" + | "llvm.x86.avx2.gather.d.ps" + | "llvm.x86.avx2.gather.d.pd" + | "llvm.x86.avx2.gather.d.d.256" + | "llvm.x86.avx2.gather.d.q.256" + | "llvm.x86.avx2.gather.d.ps.256" + | "llvm.x86.avx2.gather.d.pd.256" + | "llvm.x86.avx2.gather.q.d" + | "llvm.x86.avx2.gather.q.q" + | "llvm.x86.avx2.gather.q.ps" + | "llvm.x86.avx2.gather.q.pd" + | "llvm.x86.avx2.gather.q.d.256" + | "llvm.x86.avx2.gather.q.q.256" + | "llvm.x86.avx2.gather.q.ps.256" + | "llvm.x86.avx2.gather.q.pd.256" => { + // https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm_i64gather_pd&ig_expand=3818 + // https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm_mask_i64gather_pd&ig_expand=3819 + // https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm256_i64gather_pd&ig_expand=3821 + // https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm256_mask_i64gather_pd&ig_expand=3822 + // ... + + intrinsic_args!(fx, args => (src, ptr, index, mask, scale); intrinsic); + + let (src_lane_count, src_lane_ty) = src.layout().ty.simd_size_and_type(fx.tcx); + let (index_lane_count, index_lane_ty) = index.layout().ty.simd_size_and_type(fx.tcx); + let (mask_lane_count, mask_lane_ty) = mask.layout().ty.simd_size_and_type(fx.tcx); + let (ret_lane_count, ret_lane_ty) = ret.layout().ty.simd_size_and_type(fx.tcx); + assert_eq!(src_lane_ty, ret_lane_ty); + assert!(index_lane_ty.is_integral()); + assert_eq!(src_lane_count, mask_lane_count); + assert_eq!(src_lane_count, ret_lane_count); + + let lane_clif_ty = fx.clif_type(ret_lane_ty).unwrap(); + let index_lane_clif_ty = fx.clif_type(index_lane_ty).unwrap(); + let mask_lane_clif_ty = fx.clif_type(mask_lane_ty).unwrap(); + let ret_lane_layout = fx.layout_of(ret_lane_ty); + + let ptr = ptr.load_scalar(fx); + let scale = scale.load_scalar(fx); + let scale = fx.bcx.ins().uextend(types::I64, scale); + for lane_idx in 0..std::cmp::min(src_lane_count, index_lane_count) { + let src_lane = src.value_lane(fx, lane_idx).load_scalar(fx); + let index_lane = index.value_lane(fx, lane_idx).load_scalar(fx); + let mask_lane = mask.value_lane(fx, lane_idx).load_scalar(fx); + let mask_lane = + fx.bcx.ins().bitcast(mask_lane_clif_ty.as_int(), MemFlags::new(), mask_lane); + + let if_enabled = fx.bcx.create_block(); + let if_disabled = fx.bcx.create_block(); + let next = fx.bcx.create_block(); + let res_lane = fx.bcx.append_block_param(next, lane_clif_ty); + + let mask_lane = match mask_lane_clif_ty { + types::I32 | types::F32 => { + fx.bcx.ins().band_imm(mask_lane, 0x8000_0000u64 as i64) + } + types::I64 | types::F64 => { + fx.bcx.ins().band_imm(mask_lane, 0x8000_0000_0000_0000u64 as i64) + } + _ => unreachable!(), + }; + fx.bcx.ins().brif(mask_lane, if_enabled, &[], if_disabled, &[]); + fx.bcx.seal_block(if_enabled); + fx.bcx.seal_block(if_disabled); + + fx.bcx.switch_to_block(if_enabled); + let index_lane = if index_lane_clif_ty != types::I64 { + fx.bcx.ins().sextend(types::I64, index_lane) + } else { + index_lane + }; + let offset = fx.bcx.ins().imul(index_lane, scale); + let lane_ptr = fx.bcx.ins().iadd(ptr, offset); + let res = fx.bcx.ins().load(lane_clif_ty, MemFlags::trusted(), lane_ptr, 0); + fx.bcx.ins().jump(next, &[res]); + + fx.bcx.switch_to_block(if_disabled); + fx.bcx.ins().jump(next, &[src_lane]); + + fx.bcx.seal_block(next); + fx.bcx.switch_to_block(next); + + fx.bcx.ins().nop(); + + ret.place_lane(fx, lane_idx) + .write_cvalue(fx, CValue::by_val(res_lane, ret_lane_layout)); + } + + for lane_idx in std::cmp::min(src_lane_count, index_lane_count)..ret_lane_count { + let zero_lane = fx.bcx.ins().iconst(mask_lane_clif_ty.as_int(), 0); + let zero_lane = fx.bcx.ins().bitcast(mask_lane_clif_ty, MemFlags::new(), zero_lane); + ret.place_lane(fx, lane_idx) + .write_cvalue(fx, CValue::by_val(zero_lane, ret_lane_layout)); + } + } + "llvm.x86.sse.cmp.ps" | "llvm.x86.sse2.cmp.pd" => { let (x, y, kind) = match args { [x, y, kind] => (x, y, kind), @@ -273,16 +375,31 @@ pub(crate) fn codegen_x86_llvm_intrinsic_call<'tcx>( ); } "llvm.x86.ssse3.pabs.b.128" | "llvm.x86.ssse3.pabs.w.128" | "llvm.x86.ssse3.pabs.d.128" => { - let a = match args { - [a] => a, - _ => bug!("wrong number of args for intrinsic {intrinsic}"), - }; - let a = codegen_operand(fx, a); + intrinsic_args!(fx, args => (a); intrinsic); simd_for_each_lane(fx, a, ret, &|fx, _lane_ty, _res_lane_ty, lane| { fx.bcx.ins().iabs(lane) }); } + "llvm.x86.sse2.cvttps2dq" => { + // https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm_cvttps_epi32&ig_expand=2429 + intrinsic_args!(fx, args => (a); intrinsic); + let a = a.load_scalar(fx); + + // Using inline asm instead of fcvt_to_sint_sat as unrepresentable values are turned + // into 0x80000000 for which Cranelift doesn't have a native instruction. + codegen_inline_asm_inner( + fx, + &[InlineAsmTemplatePiece::String(format!("cvttps2dq xmm0, xmm0"))], + &[CInlineAsmOperand::InOut { + reg: InlineAsmRegOrRegClass::Reg(InlineAsmReg::X86(X86InlineAsmReg::xmm0)), + _late: true, + in_value: a, + out_place: Some(ret), + }], + InlineAsmOptions::NOSTACK | InlineAsmOptions::PURE | InlineAsmOptions::NOMEM, + ); + } "llvm.x86.addcarry.32" | "llvm.x86.addcarry.64" => { intrinsic_args!(fx, args => (c_in, a, b); intrinsic); let c_in = c_in.load_scalar(fx); @@ -364,9 +481,11 @@ pub(crate) fn codegen_x86_llvm_intrinsic_call<'tcx>( for out_lane_idx in 0..lane_count / 8 { let mut lane_diff_acc = fx.bcx.ins().iconst(types::I64, 0); - for lane_idx in out_lane_idx * 8..out_lane_idx * 8 + 1 { + for lane_idx in out_lane_idx * 8..out_lane_idx * 8 + 8 { let a_lane = a.value_lane(fx, lane_idx).load_scalar(fx); + let a_lane = fx.bcx.ins().uextend(types::I16, a_lane); let b_lane = b.value_lane(fx, lane_idx).load_scalar(fx); + let b_lane = fx.bcx.ins().uextend(types::I16, b_lane); let lane_diff = fx.bcx.ins().isub(a_lane, b_lane); let abs_lane_diff = fx.bcx.ins().iabs(lane_diff); @@ -437,12 +556,12 @@ pub(crate) fn codegen_x86_llvm_intrinsic_call<'tcx>( let ret_lane_layout = fx.layout_of(fx.tcx.types.i32); for out_lane_idx in 0..lane_count / 2 { let a_lane0 = a.value_lane(fx, out_lane_idx * 2).load_scalar(fx); - let a_lane0 = fx.bcx.ins().uextend(types::I32, a_lane0); + let a_lane0 = fx.bcx.ins().sextend(types::I32, a_lane0); let b_lane0 = b.value_lane(fx, out_lane_idx * 2).load_scalar(fx); let b_lane0 = fx.bcx.ins().sextend(types::I32, b_lane0); let a_lane1 = a.value_lane(fx, out_lane_idx * 2 + 1).load_scalar(fx); - let a_lane1 = fx.bcx.ins().uextend(types::I32, a_lane1); + let a_lane1 = fx.bcx.ins().sextend(types::I32, a_lane1); let b_lane1 = b.value_lane(fx, out_lane_idx * 2 + 1).load_scalar(fx); let b_lane1 = fx.bcx.ins().sextend(types::I32, b_lane1); @@ -597,14 +716,14 @@ pub(crate) fn codegen_x86_llvm_intrinsic_call<'tcx>( assert_eq!(ret_lane_ty, fx.tcx.types.i16); assert_eq!(lane_count * 2, ret_lane_count); - let min_i16 = fx.bcx.ins().iconst(types::I32, i64::from(i16::MIN as u16)); - let max_i16 = fx.bcx.ins().iconst(types::I32, i64::from(i16::MAX as u16)); + let min_i16 = fx.bcx.ins().iconst(types::I32, i32::from(i16::MIN) as u32 as i64); + let max_i16 = fx.bcx.ins().iconst(types::I32, i32::from(i16::MAX) as u32 as i64); let ret_lane_layout = fx.layout_of(fx.tcx.types.i16); for idx in 0..lane_count { let lane = a.value_lane(fx, idx).load_scalar(fx); let sat = fx.bcx.ins().smax(lane, min_i16); - let sat = fx.bcx.ins().umin(sat, max_i16); + let sat = fx.bcx.ins().smin(sat, max_i16); let res = fx.bcx.ins().ireduce(types::I16, sat); let res_lane = CValue::by_val(res, ret_lane_layout); @@ -614,7 +733,7 @@ pub(crate) fn codegen_x86_llvm_intrinsic_call<'tcx>( for idx in 0..lane_count { let lane = b.value_lane(fx, idx).load_scalar(fx); let sat = fx.bcx.ins().smax(lane, min_i16); - let sat = fx.bcx.ins().umin(sat, max_i16); + let sat = fx.bcx.ins().smin(sat, max_i16); let res = fx.bcx.ins().ireduce(types::I16, sat); let res_lane = CValue::by_val(res, ret_lane_layout); @@ -641,8 +760,8 @@ pub(crate) fn codegen_x86_llvm_intrinsic_call<'tcx>( for idx in 0..lane_count { let lane = a.value_lane(fx, idx).load_scalar(fx); - let sat = fx.bcx.ins().umax(lane, min_u16); - let sat = fx.bcx.ins().umin(sat, max_u16); + let sat = fx.bcx.ins().smax(lane, min_u16); + let sat = fx.bcx.ins().smin(sat, max_u16); let res = fx.bcx.ins().ireduce(types::I16, sat); let res_lane = CValue::by_val(res, ret_lane_layout); @@ -651,8 +770,8 @@ pub(crate) fn codegen_x86_llvm_intrinsic_call<'tcx>( for idx in 0..lane_count { let lane = b.value_lane(fx, idx).load_scalar(fx); - let sat = fx.bcx.ins().umax(lane, min_u16); - let sat = fx.bcx.ins().umin(sat, max_u16); + let sat = fx.bcx.ins().smax(lane, min_u16); + let sat = fx.bcx.ins().smin(sat, max_u16); let res = fx.bcx.ins().ireduce(types::I16, sat); let res_lane = CValue::by_val(res, ret_lane_layout); @@ -673,14 +792,14 @@ pub(crate) fn codegen_x86_llvm_intrinsic_call<'tcx>( assert_eq!(ret_lane_ty, fx.tcx.types.i16); assert_eq!(lane_count * 2, ret_lane_count); - let min_i16 = fx.bcx.ins().iconst(types::I32, i64::from(i16::MIN as u16)); - let max_i16 = fx.bcx.ins().iconst(types::I32, i64::from(i16::MAX as u16)); + let min_i16 = fx.bcx.ins().iconst(types::I32, i32::from(i16::MIN) as u32 as i64); + let max_i16 = fx.bcx.ins().iconst(types::I32, i32::from(i16::MAX) as u32 as i64); let ret_lane_layout = fx.layout_of(fx.tcx.types.i16); for idx in 0..lane_count / 2 { let lane = a.value_lane(fx, idx).load_scalar(fx); let sat = fx.bcx.ins().smax(lane, min_i16); - let sat = fx.bcx.ins().umin(sat, max_i16); + let sat = fx.bcx.ins().smin(sat, max_i16); let res = fx.bcx.ins().ireduce(types::I16, sat); let res_lane = CValue::by_val(res, ret_lane_layout); @@ -690,7 +809,7 @@ pub(crate) fn codegen_x86_llvm_intrinsic_call<'tcx>( for idx in 0..lane_count / 2 { let lane = b.value_lane(fx, idx).load_scalar(fx); let sat = fx.bcx.ins().smax(lane, min_i16); - let sat = fx.bcx.ins().umin(sat, max_i16); + let sat = fx.bcx.ins().smin(sat, max_i16); let res = fx.bcx.ins().ireduce(types::I16, sat); let res_lane = CValue::by_val(res, ret_lane_layout); @@ -700,7 +819,7 @@ pub(crate) fn codegen_x86_llvm_intrinsic_call<'tcx>( for idx in 0..lane_count / 2 { let lane = a.value_lane(fx, idx).load_scalar(fx); let sat = fx.bcx.ins().smax(lane, min_i16); - let sat = fx.bcx.ins().umin(sat, max_i16); + let sat = fx.bcx.ins().smin(sat, max_i16); let res = fx.bcx.ins().ireduce(types::I16, sat); let res_lane = CValue::by_val(res, ret_lane_layout); @@ -710,7 +829,7 @@ pub(crate) fn codegen_x86_llvm_intrinsic_call<'tcx>( for idx in 0..lane_count / 2 { let lane = b.value_lane(fx, idx).load_scalar(fx); let sat = fx.bcx.ins().smax(lane, min_i16); - let sat = fx.bcx.ins().umin(sat, max_i16); + let sat = fx.bcx.ins().smin(sat, max_i16); let res = fx.bcx.ins().ireduce(types::I16, sat); let res_lane = CValue::by_val(res, ret_lane_layout); @@ -718,6 +837,215 @@ pub(crate) fn codegen_x86_llvm_intrinsic_call<'tcx>( } } + "llvm.x86.fma.vfmaddsub.ps" + | "llvm.x86.fma.vfmaddsub.pd" + | "llvm.x86.fma.vfmaddsub.ps.256" + | "llvm.x86.fma.vfmaddsub.pd.256" => { + // https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm_fmaddsub_ps&ig_expand=3205 + // https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm_fmaddsub_pd&ig_expand=3181 + // https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm256_fmaddsub_ps&ig_expand=3209 + // https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm256_fmaddsub_pd&ig_expand=3185 + intrinsic_args!(fx, args => (a, b, c); intrinsic); + + assert_eq!(a.layout(), b.layout()); + assert_eq!(a.layout(), c.layout()); + let layout = a.layout(); + + let (lane_count, lane_ty) = layout.ty.simd_size_and_type(fx.tcx); + let (ret_lane_count, ret_lane_ty) = ret.layout().ty.simd_size_and_type(fx.tcx); + assert!(lane_ty.is_floating_point()); + assert!(ret_lane_ty.is_floating_point()); + assert_eq!(lane_count, ret_lane_count); + let ret_lane_layout = fx.layout_of(ret_lane_ty); + + for idx in 0..lane_count { + let a_lane = a.value_lane(fx, idx).load_scalar(fx); + let b_lane = b.value_lane(fx, idx).load_scalar(fx); + let c_lane = c.value_lane(fx, idx).load_scalar(fx); + + let mul = fx.bcx.ins().fmul(a_lane, b_lane); + let res = if idx & 1 == 0 { + fx.bcx.ins().fsub(mul, c_lane) + } else { + fx.bcx.ins().fadd(mul, c_lane) + }; + + let res_lane = CValue::by_val(res, ret_lane_layout); + ret.place_lane(fx, idx).write_cvalue(fx, res_lane); + } + } + + "llvm.x86.fma.vfmsubadd.ps" + | "llvm.x86.fma.vfmsubadd.pd" + | "llvm.x86.fma.vfmsubadd.ps.256" + | "llvm.x86.fma.vfmsubadd.pd.256" => { + // https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm_fmsubadd_ps&ig_expand=3325 + // https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm_fmsubadd_pd&ig_expand=3301 + // https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm256_fmsubadd_ps&ig_expand=3329 + // https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm256_fmsubadd_pd&ig_expand=3305 + intrinsic_args!(fx, args => (a, b, c); intrinsic); + + assert_eq!(a.layout(), b.layout()); + assert_eq!(a.layout(), c.layout()); + let layout = a.layout(); + + let (lane_count, lane_ty) = layout.ty.simd_size_and_type(fx.tcx); + let (ret_lane_count, ret_lane_ty) = ret.layout().ty.simd_size_and_type(fx.tcx); + assert!(lane_ty.is_floating_point()); + assert!(ret_lane_ty.is_floating_point()); + assert_eq!(lane_count, ret_lane_count); + let ret_lane_layout = fx.layout_of(ret_lane_ty); + + for idx in 0..lane_count { + let a_lane = a.value_lane(fx, idx).load_scalar(fx); + let b_lane = b.value_lane(fx, idx).load_scalar(fx); + let c_lane = c.value_lane(fx, idx).load_scalar(fx); + + let mul = fx.bcx.ins().fmul(a_lane, b_lane); + let res = if idx & 1 == 0 { + fx.bcx.ins().fadd(mul, c_lane) + } else { + fx.bcx.ins().fsub(mul, c_lane) + }; + + let res_lane = CValue::by_val(res, ret_lane_layout); + ret.place_lane(fx, idx).write_cvalue(fx, res_lane); + } + } + + "llvm.x86.fma.vfnmadd.ps" + | "llvm.x86.fma.vfnmadd.pd" + | "llvm.x86.fma.vfnmadd.ps.256" + | "llvm.x86.fma.vfnmadd.pd.256" => { + // https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm_fnmadd_ps&ig_expand=3391 + // https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm_fnmadd_pd&ig_expand=3367 + // https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm256_fnmadd_ps&ig_expand=3395 + // https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm256_fnmadd_pd&ig_expand=3371 + intrinsic_args!(fx, args => (a, b, c); intrinsic); + + assert_eq!(a.layout(), b.layout()); + assert_eq!(a.layout(), c.layout()); + let layout = a.layout(); + + let (lane_count, lane_ty) = layout.ty.simd_size_and_type(fx.tcx); + let (ret_lane_count, ret_lane_ty) = ret.layout().ty.simd_size_and_type(fx.tcx); + assert!(lane_ty.is_floating_point()); + assert!(ret_lane_ty.is_floating_point()); + assert_eq!(lane_count, ret_lane_count); + let ret_lane_layout = fx.layout_of(ret_lane_ty); + + for idx in 0..lane_count { + let a_lane = a.value_lane(fx, idx).load_scalar(fx); + let b_lane = b.value_lane(fx, idx).load_scalar(fx); + let c_lane = c.value_lane(fx, idx).load_scalar(fx); + + let mul = fx.bcx.ins().fmul(a_lane, b_lane); + let neg_mul = fx.bcx.ins().fneg(mul); + let res = fx.bcx.ins().fadd(neg_mul, c_lane); + + let res_lane = CValue::by_val(res, ret_lane_layout); + ret.place_lane(fx, idx).write_cvalue(fx, res_lane); + } + } + + "llvm.x86.sse42.pcmpestri128" => { + // https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm_cmpestri&ig_expand=939 + intrinsic_args!(fx, args => (a, la, b, lb, _imm8); intrinsic); + + let a = a.load_scalar(fx); + let la = la.load_scalar(fx); + let b = b.load_scalar(fx); + let lb = lb.load_scalar(fx); + + let imm8 = if let Some(imm8) = crate::constant::mir_operand_get_const_val(fx, &args[4]) + { + imm8 + } else { + fx.tcx.sess.span_fatal(span, "Index argument for `_mm_cmpestri` is not a constant"); + }; + + let imm8 = imm8.try_to_u8().unwrap_or_else(|_| panic!("kind not scalar: {:?}", imm8)); + + codegen_inline_asm_inner( + fx, + &[InlineAsmTemplatePiece::String(format!("pcmpestri xmm0, xmm1, {imm8}"))], + &[ + CInlineAsmOperand::In { + reg: InlineAsmRegOrRegClass::Reg(InlineAsmReg::X86(X86InlineAsmReg::xmm0)), + value: a, + }, + CInlineAsmOperand::In { + reg: InlineAsmRegOrRegClass::Reg(InlineAsmReg::X86(X86InlineAsmReg::xmm1)), + value: b, + }, + // Implicit argument to the pcmpestri intrinsic + CInlineAsmOperand::In { + reg: InlineAsmRegOrRegClass::Reg(InlineAsmReg::X86(X86InlineAsmReg::ax)), + value: la, + }, + // Implicit argument to the pcmpestri intrinsic + CInlineAsmOperand::In { + reg: InlineAsmRegOrRegClass::Reg(InlineAsmReg::X86(X86InlineAsmReg::dx)), + value: lb, + }, + // Implicit result of the pcmpestri intrinsic + CInlineAsmOperand::Out { + reg: InlineAsmRegOrRegClass::Reg(InlineAsmReg::X86(X86InlineAsmReg::cx)), + late: true, + place: Some(ret), + }, + ], + InlineAsmOptions::NOSTACK | InlineAsmOptions::PURE | InlineAsmOptions::NOMEM, + ); + } + + "llvm.x86.sse42.pcmpestrm128" => { + // https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm_cmpestrm&ig_expand=940 + intrinsic_args!(fx, args => (a, la, b, lb, _imm8); intrinsic); + + let a = a.load_scalar(fx); + let la = la.load_scalar(fx); + let b = b.load_scalar(fx); + let lb = lb.load_scalar(fx); + + let imm8 = if let Some(imm8) = crate::constant::mir_operand_get_const_val(fx, &args[4]) + { + imm8 + } else { + fx.tcx.sess.span_fatal(span, "Index argument for `_mm_cmpestrm` is not a constant"); + }; + + let imm8 = imm8.try_to_u8().unwrap_or_else(|_| panic!("kind not scalar: {:?}", imm8)); + + codegen_inline_asm_inner( + fx, + &[InlineAsmTemplatePiece::String(format!("pcmpestrm xmm0, xmm1, {imm8}"))], + &[ + CInlineAsmOperand::InOut { + reg: InlineAsmRegOrRegClass::Reg(InlineAsmReg::X86(X86InlineAsmReg::xmm0)), + _late: true, + in_value: a, + out_place: Some(ret), + }, + CInlineAsmOperand::In { + reg: InlineAsmRegOrRegClass::Reg(InlineAsmReg::X86(X86InlineAsmReg::xmm1)), + value: b, + }, + // Implicit argument to the pcmpestri intrinsic + CInlineAsmOperand::In { + reg: InlineAsmRegOrRegClass::Reg(InlineAsmReg::X86(X86InlineAsmReg::ax)), + value: la, + }, + // Implicit argument to the pcmpestri intrinsic + CInlineAsmOperand::In { + reg: InlineAsmRegOrRegClass::Reg(InlineAsmReg::X86(X86InlineAsmReg::dx)), + value: lb, + }, + ], + InlineAsmOptions::NOSTACK | InlineAsmOptions::PURE | InlineAsmOptions::NOMEM, + ); + } + "llvm.x86.pclmulqdq" => { // https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm_clmulepi64_si128&ig_expand=772 intrinsic_args!(fx, args => (a, b, _imm8); intrinsic); diff --git a/src/pretty_clif.rs b/src/pretty_clif.rs index da84e54a916..02c0dcb8b1b 100644 --- a/src/pretty_clif.rs +++ b/src/pretty_clif.rs @@ -58,11 +58,10 @@ use std::fmt; use std::io::Write; -use cranelift_codegen::{ - entity::SecondaryMap, - ir::entities::AnyEntity, - write::{FuncWriter, PlainWriter}, -}; +use cranelift_codegen::entity::SecondaryMap; +use cranelift_codegen::ir::entities::AnyEntity; +use cranelift_codegen::ir::Fact; +use cranelift_codegen::write::{FuncWriter, PlainWriter}; use rustc_middle::ty::layout::FnAbiOf; use rustc_middle::ty::print::with_no_trimmed_paths; use rustc_session::config::{OutputFilenames, OutputType}; @@ -155,8 +154,13 @@ impl FuncWriter for &'_ CommentWriter { _func: &Function, entity: AnyEntity, value: &dyn fmt::Display, + maybe_fact: Option<&Fact>, ) -> fmt::Result { - write!(w, " {} = {}", entity, value)?; + if let Some(fact) = maybe_fact { + write!(w, " {} ! {} = {}", entity, fact, value)?; + } else { + write!(w, " {} = {}", entity, value)?; + } if let Some(comment) = self.entity_comments.get(&entity) { writeln!(w, " ; {}", comment.replace('\n', "\n; ")) diff --git a/src/value_and_place.rs b/src/value_and_place.rs index 21ad2a835fc..f52f59716a8 100644 --- a/src/value_and_place.rs +++ b/src/value_and_place.rs @@ -329,7 +329,13 @@ impl<'tcx> CValue<'tcx> { let msb = fx.bcx.ins().iconst(types::I64, (const_val >> 64) as u64 as i64); fx.bcx.ins().iconcat(lsb, msb) } - ty::Bool | ty::Char | ty::Uint(_) | ty::Int(_) | ty::Ref(..) | ty::RawPtr(..) => { + ty::Bool + | ty::Char + | ty::Uint(_) + | ty::Int(_) + | ty::Ref(..) + | ty::RawPtr(..) + | ty::FnPtr(..) => { let raw_val = const_val.size().truncate(const_val.to_bits(layout.size).unwrap()); fx.bcx.ins().iconst(clif_ty, raw_val as i64) } @@ -971,6 +977,32 @@ pub(crate) fn assert_assignable<'tcx>( } } } + (&ty::Coroutine(def_id_a, args_a, mov_a), &ty::Coroutine(def_id_b, args_b, mov_b)) + if def_id_a == def_id_b && mov_a == mov_b => + { + let mut types_a = args_a.types(); + let mut types_b = args_b.types(); + loop { + match (types_a.next(), types_b.next()) { + (Some(a), Some(b)) => assert_assignable(fx, a, b, limit - 1), + (None, None) => return, + (Some(_), None) | (None, Some(_)) => panic!("{:#?}/{:#?}", from_ty, to_ty), + } + } + } + (&ty::CoroutineWitness(def_id_a, args_a), &ty::CoroutineWitness(def_id_b, args_b)) + if def_id_a == def_id_b => + { + let mut types_a = args_a.types(); + let mut types_b = args_b.types(); + loop { + match (types_a.next(), types_b.next()) { + (Some(a), Some(b)) => assert_assignable(fx, a, b, limit - 1), + (None, None) => return, + (Some(_), None) | (None, Some(_)) => panic!("{:#?}/{:#?}", from_ty, to_ty), + } + } + } (ty::Param(_), _) | (_, ty::Param(_)) if fx.tcx.sess.opts.unstable_opts.polymorphize => { // No way to check if it is correct or not with polymorphization enabled }