[Attributor][OpenMP] Add assumption for non-call assembly instructions

[jdoerfert]

Summary

Inline assembly is scary but we need to support it for the OpenMP GPU
device runtime. The new assumption expresses the fact that it may not
have call semantics, that is, it will not call another function but
simply perform an operation or side-effect. This is important for
reachability in the presence of inline assembly.

[jdoerfert]

Add new device runtime user

[jdoerfert]

llvm/test/Transforms/Attributor/reachability.ll

20

This is work in progress, we don't use reachability in-tree and I'll update the test as soon as we do.

[jdoerfert]

Update test case

[jhuber6]

LG

llvm/include/llvm/IR/Assumptions.h

 
 /// A set of known assumption strings that are accepted without warning and
 /// which can be recommended as typo correction.
 extern StringSet<> KnownAssumptionStrings;
 
 /// Helper that allows to insert a new assumption string in the known assumption
 /// set by creating a (static) object.
 struct KnownAssumptionString {
+  KnownAssumptionString(const char *AssumptionStr)
+      : AssumptionStr(AssumptionStr) {
+    KnownAssumptionStrings.insert(AssumptionStr);
+  }
   KnownAssumptionString(StringRef AssumptionStr)
       : AssumptionStr(AssumptionStr) {
     KnownAssumptionStrings.insert(AssumptionStr);
   }
   operator StringRef() const { return AssumptionStr; }
 
 private:
   StringRef AssumptionStr;

llvm/lib/IR/Assumptions.cpp

bool llvm::hasAssumption(CallBase &CB,
   return ::hasAssumption(A, AssumptionStr);
 }
 
 StringSet<> llvm::KnownAssumptionStrings({
     "omp_no_openmp",          // OpenMP 5.1
     "omp_no_openmp_routines", // OpenMP 5.1
     "omp_no_parallelism",     // OpenMP 5.1
     "ompx_spmd_amenable",     // OpenMPOpt extension
+    "ompx_no_call_asm",       // OpenMPOpt extension
 });

llvm/lib/Transforms/IPO/AttributorAttributes.cpp

 #include "llvm/Analysis/InstructionSimplify.h"
 #include "llvm/Analysis/LazyValueInfo.h"
 #include "llvm/Analysis/MemoryBuiltins.h"
 #include "llvm/Analysis/OptimizationRemarkEmitter.h"
 #include "llvm/Analysis/ScalarEvolution.h"
 #include "llvm/Analysis/TargetTransformInfo.h"
 #include "llvm/Analysis/ValueTracking.h"
 #include "llvm/IR/Argument.h"
+#include "llvm/IR/Assumptions.h"
 #include "llvm/IR/Constants.h"
 #include "llvm/IR/IRBuilder.h"
 #include "llvm/IR/Instruction.h"
 #include "llvm/IR/Instructions.h"
 #include "llvm/IR/IntrinsicInst.h"
 #include "llvm/IR/Intrinsics.h"
 #include "llvm/IR/NoFolder.h"
 #include "llvm/Support/Alignment.h"
auto ProcessCalledOperand = [&](Value *V) {
         // callees.
         setHasUnknownCallee(true, Change);
       }
     };
 
     CallBase *CB = static_cast<CallBase *>(getCtxI());
 
     if (CB->isInlineAsm()) {
+      if (!hasAssumption(*CB->getCaller(), "ompx_no_call_asm") &&
+          !hasAssumption(*CB, "ompx_no_call_asm"))
         setHasUnknownCallee(false, Change);
       return Change;
     }
 
     // Process callee metadata if available.
     if (auto *MD = getCtxI()->getMetadata(LLVMContext::MD_callees)) {
       for (auto &Op : MD->operands()) {
         Function *Callee = mdconst::extract_or_null<Function>(Op);
         if (Callee)

llvm/test/Transforms/Attributor/reachability.ll

+
+define dso_local void @non_recursive_asm_fn() #0 {
+entry:
+  call void asm sideeffect "barrier.sync $0;", "r,~{memory},~{dirflag},~{fpsr},~{flags}"(i32 1)
+  ret void
+}
+
+define dso_local void @non_recursive_asm_cs() {
+entry:
+  call void asm sideeffect "barrier.sync $0;", "r,~{memory},~{dirflag},~{fpsr},~{flags}"(i32 1) #0
+  ret void
+}
+
+define dso_local void @recursive_asm() {
+entry:
+  call void asm sideeffect "barrier.sync $0;", "r,~{memory},~{dirflag},~{fpsr},~{flags}"(i32 1)
+  ret void
+}
+
+attributes #0 = { "llvm.assume"="ompx_no_call_asm" }

openmp/libomptarget/DeviceRTL/include/Types.h

 //===---------- Types.h - OpenMP types ---------------------------- C++ -*-===//
 //
 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
 // See https://llvm.org/LICENSE.txt for license information.
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 //===----------------------------------------------------------------------===//
 //
 //
 //===----------------------------------------------------------------------===//
 
 #ifndef OMPTARGET_TYPES_H
 #define OMPTARGET_TYPES_H
 
+// Tell the compiler that we do not have any "call-like" inline assembly in the
+// device rutime. That means we cannot have inline assembly which will call
+// another function but only inline assembly that performs some operation or
+// side-effect and then continues execution with something on the existing call
+// stack.
+//
+// TODO: Find a good place for this
+#pragma omp assumes ext_no_call_asm
+
 /// Base type declarations for freestanding mode
 ///
 ///{
 using int8_t = char;
 using uint8_t = unsigned char;
 using int16_t = short;
 using uint16_t = unsigned short;
 using int32_t = int;