[SVE] Add fatal error for unnamed SVE variadic arguments

[david-arm]

Summary

We don't currently support passing unnamed variadic SVE arguments
so I've added a fatal error if we hit such cases to prevent any
silent ABI issues in future.

[sdesmalen]

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

4811

Does this also report a fatal error if ArgVT is a named argument?

i.e. declare i32 @sve_printf(i8*, <vscale x 4 x i32> %v, ...) and then calling with:

call i32 (i8*, <vscale x 4 x i32>, ...) @sve_printf(i8* %f, <vscale x 4 x i32> %x)

[efriedma]

We should have corresponding handling for va_arg. (clang doesn't produce va_arg instructions, but other frontends do.)

[david-arm]

• Added fatal errors on the callee side when one of the variadic arguments is a scalable vector.

[efriedma]

LGTM

clang/lib/CodeGen/TargetInfo.cpp

private:
   Address EmitDarwinVAArg(Address VAListAddr, QualType Ty,
                           CodeGenFunction &CGF) const;
 
   Address EmitAAPCSVAArg(Address VAListAddr, QualType Ty,
                          CodeGenFunction &CGF) const;
 
   Address EmitVAArg(CodeGenFunction &CGF, Address VAListAddr,
                     QualType Ty) const override {
+    llvm::Type *BaseTy = CGF.ConvertType(Ty);
+    if (isa<llvm::ScalableVectorType>(BaseTy))
+      llvm::report_fatal_error("Passing SVE types to variadic functions is "
+                               "currently not supported");
+
     return Kind == Win64 ? EmitMSVAArg(CGF, VAListAddr, Ty)
                          : isDarwinPCS() ? EmitDarwinVAArg(VAListAddr, Ty, CGF)
                                          : EmitAAPCSVAArg(VAListAddr, Ty, CGF);
   }
 
   Address EmitMSVAArg(CodeGenFunction &CGF, Address VAListAddr,
                       QualType Ty) const override;
 

clang/test/CodeGen/aarch64-varargs-sve.c

+// REQUIRES: aarch64-registered-target
+// RUN: not %clang_cc1 -triple aarch64-linux-gnu -target-feature +sve -fallow-half-arguments-and-returns -emit-llvm -o - %s 2>&1 | FileCheck %s
+// RUN: not %clang_cc1 -triple arm64-apple-ios7 -target-abi darwinpcs -target-feature +sve -fallow-half-arguments-and-returns -emit-llvm -o - %s 2>&1 | FileCheck %s
+
+// CHECK: Passing SVE types to variadic functions is currently not supported
+
+#include <arm_sve.h>
+#include <stdarg.h>
+
+double foo(char *str, ...) {
+  va_list ap;
+  svfloat64_t v;
+  double x;
+
+  va_start(ap, str);
+  v = va_arg(ap, svfloat64_t);
+  x = va_arg(ap, double);
+  va_end(ap);
+
+  return x + svaddv(svptrue_b8(), v);
+}

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

AArch64TargetLowering::LowerCall(CallLoweringInfo &CLI,
 
   if (IsVarArg) {
     // Handle fixed and variable vector arguments differently.
     // Variable vector arguments always go into memory.
     unsigned NumArgs = Outs.size();
 
     for (unsigned i = 0; i != NumArgs; ++i) {
       MVT ArgVT = Outs[i].VT;
+      if (!Outs[i].IsFixed && ArgVT.isScalableVector())
+        report_fatal_error("Passing SVE types to variadic functions is "
+                           "currently not supported");
+
       ISD::ArgFlagsTy ArgFlags = Outs[i].Flags;
       CCAssignFn *AssignFn = CCAssignFnForCall(CallConv,
                                                /*IsVarArg=*/ !Outs[i].IsFixed);
       bool Res = AssignFn(i, ArgVT, ArgVT, CCValAssign::Full, ArgFlags, CCInfo);
       assert(!Res && "Call operand has unhandled type");
       (void)Res;
     }
   } else {
SDValue AArch64TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
   unsigned MinSlotSize = Subtarget->isTargetILP32() ? 4 : 8;
   auto PtrVT = getPointerTy(DAG.getDataLayout());
   auto PtrMemVT = getPointerMemTy(DAG.getDataLayout());
   SDValue VAList =
       DAG.getLoad(PtrMemVT, DL, Chain, Addr, MachinePointerInfo(V));
   Chain = VAList.getValue(1);
   VAList = DAG.getZExtOrTrunc(VAList, DL, PtrVT);
 
+  if (VT.isScalableVector())
+    report_fatal_error("Passing SVE types to variadic functions is "
+                       "currently not supported");
+
   if (Align && *Align > MinSlotSize) {
     VAList = DAG.getNode(ISD::ADD, DL, PtrVT, VAList,
                          DAG.getConstant(Align->value() - 1, DL, PtrVT));
     VAList = DAG.getNode(ISD::AND, DL, PtrVT, VAList,
                          DAG.getConstant(-(int64_t)Align->value(), DL, PtrVT));
   }
 
   Type *ArgTy = VT.getTypeForEVT(*DAG.getContext());

llvm/test/CodeGen/AArch64/sve-varargs-callee-broken.ll

+; RUN: not --crash llc -mtriple arm64-apple-ios7 -mattr=+sve < %s 2>&1 | FileCheck %s
+
+; CHECK: Passing SVE types to variadic functions is currently not supported
+
+@.str = private unnamed_addr constant [4 x i8] c"fmt\00", align 1
+define void @foo(i8* %fmt, ...) nounwind {
+entry:
+  %fmt.addr = alloca i8*, align 8
+  %args = alloca i8*, align 8
+  %vc = alloca i32, align 4
+  %vv = alloca <vscale x 4 x i32>, align 16
+  store i8* %fmt, i8** %fmt.addr, align 8
+  %args1 = bitcast i8** %args to i8*
+  call void @llvm.va_start(i8* %args1)
+  %0 = va_arg i8** %args, i32
+  store i32 %0, i32* %vc, align 4
+  %1 = va_arg i8** %args, <vscale x 4 x i32>
+  store <vscale x 4 x i32> %1, <vscale x 4 x i32>* %vv, align 16
+  ret void
+}
+
+declare void @llvm.va_start(i8*) nounwind

llvm/test/CodeGen/AArch64/sve-varargs-caller-broken.ll

+; RUN: not --crash llc -mtriple aarch64-linux-gnu -mattr=+sve <%s 2>&1 | FileCheck %s
+
+declare i32 @sve_printf(i8*, <vscale x 4 x i32>, ...)
+
+@.str_1 = internal constant [6 x i8] c"boo!\0A\00"
+
+; CHECK: Passing SVE types to variadic functions is currently not supported
+define void @foo(<vscale x 4 x i32> %x) {
+  %f = getelementptr [6 x i8], [6 x i8]* @.str_1, i64 0, i64 0
+  call i32 (i8*, <vscale x 4 x i32>, ...) @sve_printf(i8* %f, <vscale x 4 x i32> %x, <vscale x 4 x i32> %x)
+  ret void
+}

llvm/test/CodeGen/AArch64/sve-varargs.ll

+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -mtriple=aarch64--linux-gnu -mattr=+sve < %s 2>%t | FileCheck %s
+; RUN: FileCheck --check-prefix=WARN --allow-empty %s <%t
+
+; If this check fails please read test/CodeGen/AArch64/README for instructions on how to resolve it.
+; WARN-NOT: warning
+
+declare i32 @sve_printf(i8*, <vscale x 4 x i32>, ...)
+
+@.str_1 = internal constant [6 x i8] c"boo!\0A\00"
+
+define void @foo(<vscale x 4 x i32> %x) {
+; CHECK-LABEL: foo:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    str x30, [sp, #-16]! // 8-byte Folded Spill
+; CHECK-NEXT:    .cfi_def_cfa_offset 16
+; CHECK-NEXT:    .cfi_offset w30, -16
+; CHECK-NEXT:    adrp x0, .str_1
+; CHECK-NEXT:    add x0, x0, :lo12:.str_1
+; CHECK-NEXT:    bl sve_printf
+; CHECK-NEXT:    ldr x30, [sp], #16 // 8-byte Folded Reload
+; CHECK-NEXT:    ret
+  %f = getelementptr [6 x i8], [6 x i8]* @.str_1, i64 0, i64 0
+  call i32 (i8*, <vscale x 4 x i32>, ...) @sve_printf(i8* %f, <vscale x 4 x i32> %x)
+  ret void
+}