Auto merge of #55073 - alexcrichton:demote-simd, r=<try>

rustc: Fix (again) simd vectors by-val in ABI

The issue of passing around SIMD types as values between functions has
seen [quite a lot] of [discussion], and although we thought [we fixed
it][quite a lot] it [wasn't]! This PR is a change to rustc to, again,
try to fix this issue.

The fundamental problem here remains the same, if a SIMD vector argument
is passed by-value in LLVM's function type, then if the caller and
callee disagree on target features a miscompile happens. We solve this
by never passing SIMD vectors by-value, but LLVM will still thwart us
with its argument promotion pass to promote by-ref SIMD arguments to
by-val SIMD arguments.

This commit is an attempt to thwart LLVM thwarting us. We, just before
codegen, will take yet another look at the LLVM module and demote any
by-value SIMD arguments we see. This is a very manual attempt by us to
ensure the codegen for a module keeps working, and it unfortunately is
likely producing suboptimal code, even in release mode. The saving grace
for this, in theory, is that if SIMD types are passed by-value across
a boundary in release mode it's pretty unlikely to be performance
sensitive (as it's already doing a load/store, and otherwise
perf-sensitive bits should be inlined).

The implementation here is basically a big wad of C++. It was largely
copied from LLVM's own argument promotion pass, only doing the reverse.
In local testing this...

Closes #50154
Closes #52636
Closes #54583
Closes #55059

[quite a lot]: #47743
[discussion]: #44367
[wasn't]: #50154

src/librustc_codegen_llvm/back/lto.rs

@@ -80,9 +80,7 @@ impl LtoModuleCodegen {
                 let module = module.take().unwrap();
                 {
                     let config = cgcx.config(module.kind);
-                    let llmod = module.module_llvm.llmod();
-                    let tm = &*module.module_llvm.tm;
-                    run_pass_manager(cgcx, tm, llmod, config, false);
+                    run_pass_manager(cgcx, &module, config, false);
                     timeline.record("fat-done");
                 }
                 Ok(module)
@@ -557,8 +555,7 @@ fn thin_lto(cgcx: &CodegenContext,
 }
 
 fn run_pass_manager(cgcx: &CodegenContext,
-                    tm: &llvm::TargetMachine,
-                    llmod: &llvm::Module,
+                    module: &ModuleCodegen,
                     config: &ModuleConfig,
                     thin: bool) {
     // Now we have one massive module inside of llmod. Time to run the
@@ -569,7 +566,8 @@ fn run_pass_manager(cgcx: &CodegenContext,
     debug!("running the pass manager");
     unsafe {
         let pm = llvm::LLVMCreatePassManager();
-        llvm::LLVMRustAddAnalysisPasses(tm, pm, llmod);
+        let llmod = module.module_llvm.llmod();
+        llvm::LLVMRustAddAnalysisPasses(module.module_llvm.tm, pm, llmod);
 
         if config.verify_llvm_ir {
             let pass = llvm::LLVMRustFindAndCreatePass("verify\0".as_ptr() as *const _);
@@ -864,7 +862,7 @@ impl ThinModule {
             // little differently.
             info!("running thin lto passes over {}", module.name);
             let config = cgcx.config(module.kind);
-            run_pass_manager(cgcx, module.module_llvm.tm, llmod, config, true);
+            run_pass_manager(cgcx, &module, config, true);
             cgcx.save_temp_bitcode(&module, "thin-lto-after-pm");
             timeline.record("thin-done");
         }

src/librustc_codegen_llvm/back/write.rs

@@ -633,7 +633,7 @@ unsafe fn optimize(cgcx: &CodegenContext,
                  None,
                  &format!("llvm module passes [{}]", module_name.unwrap()),
                  || {
-            llvm::LLVMRunPassManager(mpm, llmod)
+            llvm::LLVMRunPassManager(mpm, llmod);
         });
 
         // Deallocate managers that we're now done with
@@ -691,6 +691,38 @@ unsafe fn codegen(cgcx: &CodegenContext,
             create_msvc_imps(cgcx, llcx, llmod);
         }
 
+        // Ok now this one's a super interesting invocations. SIMD in rustc is
+        // difficult where we want some parts of the program to be able to use
+        // some SIMD features while other parts of the program don't. The real
+        // tough part is that we want this to actually work correctly!
+        //
+        // We go to great lengths to make sure this works, and one crucial
+        // aspect is that vector arguments (simd types) are never passed by
+        // value in the ABI of functions. It turns out, however, that LLVM will
+        // undo our "clever work" of passing vector types by reference. Its
+        // argument promotion pass will promote these by-ref arguments to
+        // by-val. That, however, introduces codegen errors!
+        //
+        // The upstream LLVM bug [1] has unfortunatey not really seen a lot of
+        // activity. The Rust bug [2], however, has seen quite a lot of reports
+        // of this in the wild. As a result, this is worked around locally here.
+        // We have a custom transformation, `LLVMRustDemoteSimdArguments`, which
+        // does the opposite of argument promotion by demoting any by-value SIMD
+        // arguments in function signatures to pointers intead of being
+        // by-value.
+        //
+        // This operates at the LLVM IR layer because LLVM is thwarting our
+        // codegen and this is the only chance we get to make sure it's correct
+        // before we hit codegen.
+        //
+        // Hopefully one day the upstream LLVM bug will be fixed and we'll no
+        // longer need this!
+        //
+        // [1]: https://bugs.llvm.org/show_bug.cgi?id=37358
+        // [2]: https://github.com/rust-lang/rust/issues/50154
+        llvm::LLVMRustDemoteSimdArguments(llmod);
+        cgcx.save_temp_bitcode(&module, "simd-demoted");
+
         // A codegen-specific pass manager is used to generate object
         // files for an LLVM module.
         //

src/librustc_codegen_llvm/llvm/ffi.rs

@@ -1136,6 +1136,8 @@ extern "C" {
     /// Runs a pass manager on a module.
     pub fn LLVMRunPassManager(PM: &PassManager<'a>, M: &'a Module) -> Bool;
 
+    pub fn LLVMRustDemoteSimdArguments(M: &'a Module);
+
     pub fn LLVMInitializePasses();
 
     pub fn LLVMPassManagerBuilderCreate() -> &'static mut PassManagerBuilder;

src/librustc_llvm/build.rs

@@ -162,7 +162,9 @@ fn main() {
     }
 
     build_helper::rerun_if_changed_anything_in_dir(Path::new("../rustllvm"));
-    cfg.file("../rustllvm/PassWrapper.cpp")
+    cfg
+       .file("../rustllvm/DemoteSimd.cpp")
+       .file("../rustllvm/PassWrapper.cpp")
        .file("../rustllvm/RustWrapper.cpp")
        .file("../rustllvm/ArchiveWrapper.cpp")
        .file("../rustllvm/Linker.cpp")

src/rustllvm/DemoteSimd.cpp

@@ -0,0 +1,186 @@
+// Copyright 2018 The Rust Project Developers. See the COPYRIGHT
+// file at the top-level directory of this distribution and at
+// http://rust-lang.org/COPYRIGHT.
+//
+// Licensed under the Apache License, Version 2.0 <LICENSE-APACHE or
+// http://www.apache.org/licenses/LICENSE-2.0> or the MIT license
+// <LICENSE-MIT or http://opensource.org/licenses/MIT>, at your
+// option. This file may not be copied, modified, or distributed
+// except according to those terms.
+
+#include <vector>
+#include <set>
+
+#include "rustllvm.h"
+
+#include "llvm/IR/CallSite.h"
+#include "llvm/IR/Module.h"
+
+using namespace llvm;
+
+static std::vector<Function*>
+GetFunctionsWithSimdArgs(Module *M) {
+  std::vector<Function*> Ret;
+
+  for (auto &F : M->functions()) {
+    // Skip all intrinsic calls as these are always tightly controlled to "work
+    // correctly", so no need to fixup any of these.
+    if (F.isIntrinsic())
+      continue;
+
+    // We know that we started out with a `Module` that has what we want, so
+    // we're just trying to undo specifically the work of the
+    // `ArgumentPromotion` pass. That only runs in a select few circumstances,
+    // so make sure that we don't get anything surprising. For example, make
+    // sure we don't actually return a vector type because rustc shouldn't ever
+    // generate this and nor should passes make this happen.
+    assert(!F->getReturnType()->isVectorTy());
+
+    // If any argument to this function is a by-value vector type, then that's
+    // bad! The compiler didn't generate any functions that looked like this,
+    // and we try to rely on LLVM to not do this! Argument promotion may,
+    // however, promote arguments from behind references. In any case, figure
+    // out if we're interested in demoting this argument.
+    bool anyVector = false;
+    for (auto &Arg : F.args())
+      anyVector = anyVector || Arg.getType()->isVectorTy();
+    if (anyVector)
+      Ret.push_back(&F);
+  }
+
+  return Ret;
+}
+
+extern "C" void
+LLVMRustDemoteSimdArguments(LLVMModuleRef Mod) {
+  Module *M = unwrap(Mod);
+
+  auto Functions = GetFunctionsWithSimdArgs(M);
+
+  for (auto F : Functions) {
+    // The argument promotion pass in LLVM should only run on functions that
+    // have local linkage. We're modifying function signatures here, so make
+    // sure such a desctructive change doesn't affect the public ABI.
+    assert(F->hasLocalLinkage());
+
+    // Build up our list of new parameters and new argument attributes.
+    // We're only changing those arguments which are vector types.
+    SmallVector<Type*, 8> Params;
+    SmallVector<AttributeSet, 8> ArgAttrVec;
+    auto PAL = F->getAttributes();
+    for (auto &Arg : F->args()) {
+      auto *Ty = Arg.getType();
+      if (Ty->isVectorTy()) {
+        Params.push_back(PointerType::get(Ty, 0));
+        ArgAttrVec.push_back(AttributeSet());
+      } else {
+        Params.push_back(Ty);
+        ArgAttrVec.push_back(PAL.getParamAttributes(Arg.getArgNo()));
+      }
+    }
+
+    // Replace `F` with a new function with our new signature. I'm... not really
+    // sure how this works, but this is all the steps `ArgumentPromotion` does
+    // to replace a signature as well.
+    assert(!F->isVarArg()); // ArgumentPromotion should skip these fns
+    FunctionType *NFTy = FunctionType::get(F->getReturnType(), Params, false);
+    Function *NF = Function::Create(NFTy, F->getLinkage(), F->getName());
+    NF->copyAttributesFrom(F);
+    NF->setSubprogram(F->getSubprogram());
+    F->setSubprogram(nullptr);
+    NF->setAttributes(AttributeList::get(F->getContext(),
+                                         PAL.getFnAttributes(),
+                                         PAL.getRetAttributes(),
+                                         ArgAttrVec));
+    ArgAttrVec.clear();
+    F->getParent()->getFunctionList().insert(F->getIterator(), NF);
+    NF->takeName(F);
+
+    // Iterate over all invocations of `F`, updating all `call` instructions to
+    // store immediate vector types in a local `alloc` instead of a by-value
+    // vector.
+    //
+    // Like before, much of this is copied from the `ArgumentPromotion` pass in
+    // LLVM.
+    SmallVector<Value*, 16> Args;
+    while (!F->use_empty()) {
+      CallSite CS(F->user_back());
+      assert(CS.getCalledFunction() == F);
+      Instruction *Call = CS.getInstruction();
+      const AttributeList &CallPAL = CS.getAttributes();
+
+      // Loop over the operands, inserting an `alloca` and a store for any
+      // argument we're demoting to be by reference
+      //
+      // FIXME: we probably want to figure out an LLVM pass to run and clean up
+      // this function and instructions we're generating, we should in theory
+      // only generate a maximum number of `alloca` instructions rather than
+      // one-per-variable unconditionally.
+      CallSite::arg_iterator AI = CS.arg_begin();
+      size_t ArgNo = 0;
+      for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E;
+           ++I, ++AI, ++ArgNo) {
+        if (I->getType()->isVectorTy()) {
+          AllocaInst *AllocA = new AllocaInst(I->getType(), 0, nullptr, "", Call);
+          new StoreInst(*AI, AllocA, Call);
+          Args.push_back(AllocA);
+          ArgAttrVec.push_back(AttributeSet());
+        } else {
+          Args.push_back(*AI);
+          ArgAttrVec.push_back(CallPAL.getParamAttributes(ArgNo));
+        }
+      }
+      assert(AI == CS.arg_end());
+
+      // Create a new call instructions which we'll use to replace the old call
+      // instruction, copying over as many attributes and such as possible.
+      SmallVector<OperandBundleDef, 1> OpBundles;
+      CS.getOperandBundlesAsDefs(OpBundles);
+
+      CallSite NewCS;
+      if (InvokeInst *II = dyn_cast<InvokeInst>(Call)) {
+        InvokeInst::Create(NF, II->getNormalDest(), II->getUnwindDest(),
+                           Args, OpBundles, "", Call);
+      } else {
+        auto *NewCall = CallInst::Create(NF, Args, OpBundles, "", Call);
+        NewCall->setTailCallKind(cast<CallInst>(Call)->getTailCallKind());
+        NewCS = NewCall;
+      }
+      NewCS.setCallingConv(CS.getCallingConv());
+      NewCS.setAttributes(
+          AttributeList::get(F->getContext(), CallPAL.getFnAttributes(),
+                             CallPAL.getRetAttributes(), ArgAttrVec));
+      NewCS->setDebugLoc(Call->getDebugLoc());
+      Args.clear();
+      ArgAttrVec.clear();
+      Call->replaceAllUsesWith(NewCS.getInstruction());
+      NewCS->takeName(Call);
+      Call->eraseFromParent();
+    }
+
+    // Splice the body of the old function right into the new function.
+    NF->getBasicBlockList().splice(NF->begin(), F->getBasicBlockList());
+
+    // Update our new function to replace all uses of the by-value argument with
+    // loads of the pointer argument we've generated.
+    //
+    // FIXME: we probably want to only generate one load instruction per
+    // function? Or maybe run an LLVM pass to clean up this function?
+    for (Function::arg_iterator I = F->arg_begin(),
+                                E = F->arg_end(),
+                                I2 = NF->arg_begin();
+         I != E;
+         ++I, ++I2) {
+      if (I->getType()->isVectorTy()) {
+        I->replaceAllUsesWith(new LoadInst(&*I2, "", &NF->begin()->front()));
+      } else {
+        I->replaceAllUsesWith(&*I2);
+      }
+      I2->takeName(&*I);
+    }
+
+    // Delete all references to the old function, it should be entirely dead
+    // now.
+    M->getFunctionList().remove(F);
+  }
+}

src/test/codegen/repr-transparent.rs

@@ -108,7 +108,7 @@ struct f32x4(f32, f32, f32, f32);
 #[repr(transparent)]
 pub struct Vector(f32x4);
 
-// CHECK: define <4 x float> @test_Vector(<4 x float> %arg0)
+// CHECK: define <4 x float> @test_Vector(<4 x float>* %arg0)
 #[no_mangle]
 pub extern fn test_Vector(_: Vector) -> Vector { loop {} }
 

src/test/run-make/simd-argument-promotion-thwarted/Makefile

@@ -0,0 +1,13 @@
+-include ../../run-make-fulldeps/tools.mk
+
+ifeq ($(TARGET),x86_64-unknown-linux-gnu)
+all:
+	$(RUSTC) t1.rs -C opt-level=3
+	$(TMPDIR)/t1
+	$(RUSTC) t2.rs -C opt-level=3
+	$(TMPDIR)/t2
+	$(RUSTC) t3.rs -C opt-level=3
+	$(TMPDIR)/t3
+else
+all:
+endif

src/test/run-make/simd-argument-promotion-thwarted/t1.rs

@@ -0,0 +1,21 @@
+use std::arch::x86_64;
+
+fn main() {
+    if !is_x86_feature_detected!("avx2") {
+        return println!("AVX2 is not supported on this machine/build.");
+    }
+    let load_bytes: [u8; 32] = [0x0f; 32];
+    let lb_ptr = load_bytes.as_ptr();
+    let reg_load = unsafe {
+        x86_64::_mm256_loadu_si256(
+            lb_ptr as *const x86_64::__m256i
+        )
+    };
+    println!("{:?}", reg_load);
+    let mut store_bytes: [u8; 32] = [0; 32];
+    let sb_ptr = store_bytes.as_mut_ptr();
+    unsafe {
+        x86_64::_mm256_storeu_si256(sb_ptr as *mut x86_64::__m256i, reg_load);
+    }
+    assert_eq!(load_bytes, store_bytes);
+}

src/test/run-make/simd-argument-promotion-thwarted/t2.rs

@@ -0,0 +1,14 @@
+use std::arch::x86_64::*;
+
+fn main() {
+    if !is_x86_feature_detected!("avx") {
+        return println!("AVX is not supported on this machine/build.");
+    }
+    unsafe {
+        let f = _mm256_set_pd(2.0, 2.0, 2.0, 2.0);
+        let r = _mm256_mul_pd(f, f);
+
+        union A { a: __m256d, b: [f64; 4] }
+        assert_eq!(A { a: r }.b, [4.0, 4.0, 4.0, 4.0]);
+    }
+}