[AsmParser] Implicitly declare intrinsics (#78251)

We currently require that all referenced globals have an explicit
declaration or definition in the IR. For intrinsics, this requirement is
redundant, because they cannot be called indirectly (including "direct"
calls with mismatched function type). The function type used in the call
directly determines the function type of the intrinsic declaration.

Relax this requirement, and implicitly declare any intrinsics that do
not have an explicit declaration. This will remove a common annoyance
when writing tests and alive2 proofs.

(I also plan to introduce a mode where declarations for all missing
symbols will be automatically added, to make working with incomplete IR
easier -- but that will be behind a default-disabled flag.)

Author: nikic

llvm/lib/AsmParser/LLParser.cpp

@@ -246,6 +246,34 @@ bool LLParser::validateEndOfModule(bool UpgradeDebugInfo) {
                  "use of undefined comdat '$" +
                      ForwardRefComdats.begin()->first + "'");
 
+  // Automatically create declarations for intrinsics. Intrinsics can only be
+  // called directly, so the call function type directly determines the
+  // declaration function type.
+  for (const auto &[Name, Info] : make_early_inc_range(ForwardRefVals)) {
+    if (!StringRef(Name).starts_with("llvm."))
+      continue;
+
+    // Don't do anything if the intrinsic is called with different function
+    // types. This would result in a verifier error anyway.
+    auto GetCommonFunctionType = [](Value *V) -> FunctionType * {
+      FunctionType *FTy = nullptr;
+      for (User *U : V->users()) {
+        auto *CB = dyn_cast<CallBase>(U);
+        if (!CB || (FTy && FTy != CB->getFunctionType()))
+          return nullptr;
+        FTy = CB->getFunctionType();
+      }
+      return FTy;
+    };
+    if (FunctionType *FTy = GetCommonFunctionType(Info.first)) {
+      Function *Fn =
+          Function::Create(FTy, GlobalValue::ExternalLinkage, Name, M);
+      Info.first->replaceAllUsesWith(Fn);
+      Info.first->eraseFromParent();
+      ForwardRefVals.erase(Name);
+    }
+  }
+
   if (!ForwardRefVals.empty())
     return error(ForwardRefVals.begin()->second.second,
                  "use of undefined value '@" + ForwardRefVals.begin()->first +

llvm/test/Assembler/implicit-intrinsic-declaration-invalid.ll

@@ -0,0 +1,11 @@
+; RUN: not llvm-as < %s 2>&1 | FileCheck %s
+
+; Check that intrinsics do not get automatically declared if they are used
+; with different function types.
+
+; CHECK: error: use of undefined value '@llvm.umax'
+define void @test() {
+  call i8 @llvm.umax(i8 0, i8 1)
+  call i16 @llvm.umax(i16 0, i16 1)
+  ret void
+}

llvm/test/Assembler/implicit-intrinsic-declaration.ll

@@ -0,0 +1,33 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 4
+; RUN: opt -S < %s | FileCheck %s
+; RUN: opt -S -passes=instcombine < %s | FileCheck %s --check-prefix=INSTCOMBINE
+
+; llvm.umax is intentionally missing the mangling suffix here, to show that
+; this works fine with auto-upgrade.
+define i16 @test(i8 %x, i8 %y) {
+; CHECK-LABEL: define i16 @test(
+; CHECK-SAME: i8 [[X:%.*]], i8 [[Y:%.*]]) {
+; CHECK-NEXT:    [[CMP:%.*]] = icmp sgt i8 [[X]], -1
+; CHECK-NEXT:    call void @llvm.assume(i1 [[CMP]])
+; CHECK-NEXT:    [[MAX1:%.*]] = call i8 @llvm.umax.i8(i8 [[X]], i8 [[Y]])
+; CHECK-NEXT:    [[X_EXT:%.*]] = zext i8 [[X]] to i16
+; CHECK-NEXT:    [[Y_EXT:%.*]] = zext i8 [[Y]] to i16
+; CHECK-NEXT:    [[MAX2:%.*]] = call i16 @llvm.umax.i16(i16 [[X_EXT]], i16 [[Y_EXT]])
+; CHECK-NEXT:    ret i16 [[MAX2]]
+;
+; INSTCOMBINE-LABEL: define i16 @test(
+; INSTCOMBINE-SAME: i8 [[X:%.*]], i8 [[Y:%.*]]) {
+; INSTCOMBINE-NEXT:    [[CMP:%.*]] = icmp sgt i8 [[X]], -1
+; INSTCOMBINE-NEXT:    call void @llvm.assume(i1 [[CMP]])
+; INSTCOMBINE-NEXT:    [[TMP1:%.*]] = call i8 @llvm.umax.i8(i8 [[X]], i8 [[Y]])
+; INSTCOMBINE-NEXT:    [[MAX2:%.*]] = zext i8 [[TMP1]] to i16
+; INSTCOMBINE-NEXT:    ret i16 [[MAX2]]
+;
+  %cmp = icmp sgt i8 %x, -1
+  call void @llvm.assume(i1 %cmp)
+  %max1 = call i8 @llvm.umax(i8 %x, i8 %y)
+  %x.ext = zext i8 %x to i16
+  %y.ext = zext i8 %y to i16
+  %max2 = call i16 @llvm.umax.i16(i16 %x.ext, i16 %y.ext)
+  ret i16 %max2
+}