[SVE][Inline-Asm] Add constraints for SVE ACLE types

[kmclaughlin]

Summary

Adds the constraints described below to ensure that we
can tie variables of SVE ACLE types to operands in inline-asm:

• y: SVE registers Z0-Z7
• Upl: One of the low eight SVE predicate registers (P0-P7)
• Upa: Full range of SVE predicate registers (P0-P15)

[efriedma]

clang/lib/Basic/Targets/AArch64.h

95

Is "@3" some LLVM thing? I don't think I've seen it before.

clang/test/CodeGen/aarch64-sve-inline-asm-datatypes.c

2

REQUIRES: aarch64-registered-target is necessary for any test that isn't using -emit-llvm.

Generally, I'd prefer tests using -emit-llvm so we can independently verify that the IR is what we expect, vs. the backend processing the IR the way we expect.

[kmclaughlin]

• Added -emit-llvm to the RUN line of aarch64-sve-inline-asm-datatypes.c test. Added some more tests here for the Upl & y constraints and removed aarch64-sve-inline-asm-vec-low.c.
• Addressed formatting suggestions on previous patch.

[kmclaughlin]

Thanks for reviewing this, @efriedma!

clang/lib/Basic/Targets/AArch64.h

95

The "@" is used to indicate that this is a multi-letter constraint of a given length. This is parsed later by InlineAsm.cpp, see https://reviews.llvm.org/D66524

clang/test/CodeGen/aarch64-sve-inline-asm-datatypes.c

2

I've added -emit-llvm here and updated the CHECK lines accordingly

[efriedma]

LGTM

clang/lib/Basic/Targets/AArch64.h

public:
   CallingConvCheckResult checkCallingConvention(CallingConv CC) const override;
 
   bool isCLZForZeroUndef() const override;
 
   BuiltinVaListKind getBuiltinVaListKind() const override;
 
   ArrayRef<const char *> getGCCRegNames() const override;
   ArrayRef<TargetInfo::GCCRegAlias> getGCCRegAliases() const override;
+
+  std::string convertConstraint(const char *&Constraint) const override {
+    std::string R;
+    switch (*Constraint) {
+    case 'U':   // Three-character constraint; add "@3" hint for later parsing.
+      R = std::string("@3") + std::string(Constraint, 3);
+      Constraint += 2;
+      break;
+    default:
+      R = std::string(1, *Constraint);
+      break;
+    }
+    return R;
+  }
+
   bool validateAsmConstraint(const char *&Name,
                              TargetInfo::ConstraintInfo &Info) const override;
   bool
   validateConstraintModifier(StringRef Constraint, char Modifier, unsigned Size,
                              std::string &SuggestedModifier) const override;
   const char *getClobbers() const override;
 
   StringRef getConstraintRegister(StringRef Constraint,

clang/lib/Basic/Targets/AArch64.cpp

case 'Z': // Integer constant zero
     return true;
   case 'Q': // A memory reference with base register and no offset
     Info.setAllowsMemory();
     return true;
   case 'S': // A symbolic address
     Info.setAllowsRegister();
     return true;
   case 'U':
+    if (Name[1] == 'p' && (Name[2] == 'l' || Name[2] == 'a')) {
+        // SVE predicate registers ("Upa"=P0-15, "Upl"=P0-P7)
+        Info.setAllowsRegister();
+        Name += 2;
+        return true;
+      }
     // Ump: A memory address suitable for ldp/stp in SI, DI, SF and DF modes.
     // Utf: A memory address suitable for ldp/stp in TF mode.
     // Usa: An absolute symbolic address.
     // Ush: The high part (bits 32:12) of a pc-relative symbolic address.
-    llvm_unreachable("FIXME: Unimplemented support for U* constraints.");
+
+    // Better to return an error saying that it's an unrecognised constraint
+    // even if this is a valid constraint in gcc.
+    return false;
   case 'z': // Zero register, wzr or xzr
     Info.setAllowsRegister();
     return true;
   case 'x': // Floating point and SIMD registers (V0-V15)
     Info.setAllowsRegister();
     return true;
+  case 'y': // SVE registers (V0-V7)
+    Info.setAllowsRegister();
+    return true;
   }
   return false;
 }
 
 bool AArch64TargetInfo::validateConstraintModifier(
     StringRef Constraint, char Modifier, unsigned Size,
     std::string &SuggestedModifier) const {
   // Strip off constraint modifiers.

clang/lib/CodeGen/CGCall.cpp

for (unsigned i = 0; i < IRCallArgs.size(); ++i) {
     if (i < IRFuncTy->getNumParams())
       assert(IRCallArgs[i]->getType() == IRFuncTy->getParamType(i));
   }
 #endif
 
   // Update the largest vector width if any arguments have vector types.
   for (unsigned i = 0; i < IRCallArgs.size(); ++i) {
     if (auto *VT = dyn_cast<llvm::VectorType>(IRCallArgs[i]->getType()))
       LargestVectorWidth = std::max((uint64_t)LargestVectorWidth,
-                                   VT->getPrimitiveSizeInBits().getFixedSize());
+                                   VT->getPrimitiveSizeInBits().getKnownMinSize());
   }
 
   // Compute the calling convention and attributes.
   unsigned CallingConv;
   llvm::AttributeList Attrs;
   CGM.ConstructAttributeList(CalleePtr->getName(), CallInfo,
                              Callee.getAbstractInfo(), Attrs, CallingConv,
                              /*AttrOnCallSite=*/true);
#endif
 
   // Apply various metadata.
 
   if (!CI->getType()->isVoidTy())
     CI->setName("call");
 
   // Update largest vector width from the return type.
   if (auto *VT = dyn_cast<llvm::VectorType>(CI->getType()))
     LargestVectorWidth = std::max((uint64_t)LargestVectorWidth,
-                                  VT->getPrimitiveSizeInBits().getFixedSize());
+                                  VT->getPrimitiveSizeInBits().getKnownMinSize());
 
   // Insert instrumentation or attach profile metadata at indirect call sites.
   // For more details, see the comment before the definition of
   // IPVK_IndirectCallTarget in InstrProfData.inc.
   if (!CI->getCalledFunction())
     PGO.valueProfile(Builder, llvm::IPVK_IndirectCallTarget,
                      CI, CalleePtr);
 

clang/lib/CodeGen/CGStmt.cpp

if (!Info.allowsMemory() && (hasScalarEvaluationKind(OutExpr->getType()) ||
       else {
         CGM.getDiags().Report(S.getAsmLoc(),
                               diag::err_asm_invalid_type_in_input)
             << OutExpr->getType() << OutputConstraint;
       }
 
       // Update largest vector width for any vector types.
       if (auto *VT = dyn_cast<llvm::VectorType>(ResultRegTypes.back()))
         LargestVectorWidth = std::max((uint64_t)LargestVectorWidth,
-                                   VT->getPrimitiveSizeInBits().getFixedSize());
+                                   VT->getPrimitiveSizeInBits().getKnownMinSize());
     } else {
       ArgTypes.push_back(Dest.getAddress(*this).getType());
       Args.push_back(Dest.getPointer(*this));
       Constraints += "=*";
       Constraints += OutputConstraint;
       ReadOnly = ReadNone = false;
     }
 
     if (Info.isReadWrite()) {
       InOutConstraints += ',';
 
       const Expr *InputExpr = S.getOutputExpr(i);
       llvm::Value *Arg = EmitAsmInputLValue(Info, Dest, InputExpr->getType(),
                                             InOutConstraints,
                                             InputExpr->getExprLoc());
 
       if (llvm::Type* AdjTy =
           getTargetHooks().adjustInlineAsmType(*this, OutputConstraint,
                                                Arg->getType()))
         Arg = Builder.CreateBitCast(Arg, AdjTy);
 
       // Update largest vector width for any vector types.
       if (auto *VT = dyn_cast<llvm::VectorType>(Arg->getType()))
         LargestVectorWidth = std::max((uint64_t)LargestVectorWidth,
-                                   VT->getPrimitiveSizeInBits().getFixedSize());
+                                   VT->getPrimitiveSizeInBits().getKnownMinSize());
       if (Info.allowsRegister())
         InOutConstraints += llvm::utostr(i);
       else
         InOutConstraints += OutputConstraint;
 
       InOutArgTypes.push_back(Arg->getType());
       InOutArgs.push_back(Arg);
     }
if (llvm::Type* AdjTy =
                                                    Arg->getType()))
       Arg = Builder.CreateBitCast(Arg, AdjTy);
     else
       CGM.getDiags().Report(S.getAsmLoc(), diag::err_asm_invalid_type_in_input)
           << InputExpr->getType() << InputConstraint;
 
     // Update largest vector width for any vector types.
     if (auto *VT = dyn_cast<llvm::VectorType>(Arg->getType()))
       LargestVectorWidth = std::max((uint64_t)LargestVectorWidth,
-                                   VT->getPrimitiveSizeInBits().getFixedSize());
+                                   VT->getPrimitiveSizeInBits().getKnownMinSize());
 
     ArgTypes.push_back(Arg->getType());
     Args.push_back(Arg);
     Constraints += InputConstraint;
   }
 
   // Labels
   SmallVector<llvm::BasicBlock *, 16> Transfer;

clang/lib/CodeGen/CodeGenFunction.cpp

if (NormalCleanupDest.isValid() && isCoroutine()) {
     llvm::PromoteMemToReg(
         cast<llvm::AllocaInst>(NormalCleanupDest.getPointer()), DT);
     NormalCleanupDest = Address::invalid();
   }
 
   // Scan function arguments for vector width.
   for (llvm::Argument &A : CurFn->args())
     if (auto *VT = dyn_cast<llvm::VectorType>(A.getType()))
       LargestVectorWidth = std::max((uint64_t)LargestVectorWidth,
-                                   VT->getPrimitiveSizeInBits().getFixedSize());
+                                   VT->getPrimitiveSizeInBits().getKnownMinSize());
 
   // Update vector width based on return type.
   if (auto *VT = dyn_cast<llvm::VectorType>(CurFn->getReturnType()))
     LargestVectorWidth = std::max((uint64_t)LargestVectorWidth,
-                                  VT->getPrimitiveSizeInBits().getFixedSize());
+                                  VT->getPrimitiveSizeInBits().getKnownMinSize());
 
   // Add the required-vector-width attribute. This contains the max width from:
   // 1. min-vector-width attribute used in the source program.
   // 2. Any builtins used that have a vector width specified.
   // 3. Values passed in and out of inline assembly.
   // 4. Width of vector arguments and return types for this function.
   // 5. Width of vector aguments and return types for functions called by this
   //    function.

clang/test/CodeGen/aarch64-sve-inline-asm-crash.c

+// RUN: not %clang_cc1 -triple aarch64-none-linux-gnu -target-feature +sve -fallow-half-arguments-and-returns \
+// RUN:   -target-feature +neon -S -O1 -o - %s 2>&1 | FileCheck %s
+
+// Set a vector constraint for an sve predicate register
+// As the wrong constraint is used for an SVBool,
+// the compiler will try to extend the nxv16i1 to an nxv16i8
+// TODO: We don't have patterns for this yet but once they are added this test
+// should be updated to check for an assembler error
+__SVBool_t funcB1(__SVBool_t in)
+{
+  __SVBool_t ret ;
+  asm volatile (
+    "mov %[ret].b, %[in].b \n"
+    : [ret] "=w" (ret)
+    : [in] "w" (in)
+    :);
+
+  return ret ;
+}
+
+// CHECK: funcB1
+// CHECK-ERROR: fatal error: error in backend: Cannot select

clang/test/CodeGen/aarch64-sve-inline-asm-datatypes.c

+// RUN: %clang_cc1 -triple aarch64-none-linux-gnu -target-feature +sve -fallow-half-arguments-and-returns \
+// RUN:   -target-feature +neon -S -O1 -o - %s | FileCheck %s
+
+// Tests to check that all sve datatypes can be passed in as input operands
+// and passed out as output operands.
+
+#define SVINT_TEST(DT, KIND)\
+DT func_int_##DT##KIND(DT in)\
+{\
+  DT out;\
+  asm volatile (\
+    "ptrue p0.b\n"\
+    "mov %[out]." #KIND ", p0/m, %[in]." #KIND "\n"\
+    : [out] "=w" (out)\
+    : [in] "w" (in)\
+    : "p0"\
+    );\
+  return out;\
+}
+
+SVINT_TEST(__SVUint8_t,b);
+// CHECK: mov {{z[0-9]+}}.b, p0/m, {{z[0-9]+}}.b
+SVINT_TEST(__SVUint8_t,h);
+// CHECK: mov {{z[0-9]+}}.h, p0/m, {{z[0-9]+}}.h
+SVINT_TEST(__SVUint8_t,s);
+// CHECK: mov {{z[0-9]+}}.s, p0/m, {{z[0-9]+}}.s
+SVINT_TEST(__SVUint8_t,d);
+// CHECK: mov {{z[0-9]+}}.d, p0/m, {{z[0-9]+}}.d
+
+SVINT_TEST(__SVUint16_t,b);
+// CHECK: mov {{z[0-9]+}}.b, p0/m, {{z[0-9]+}}.b
+SVINT_TEST(__SVUint16_t,h);
+// CHECK: mov {{z[0-9]+}}.h, p0/m, {{z[0-9]+}}.h
+SVINT_TEST(__SVUint16_t,s);
+// CHECK: mov {{z[0-9]+}}.s, p0/m, {{z[0-9]+}}.s
+SVINT_TEST(__SVUint16_t,d);
+// CHECK: mov {{z[0-9]+}}.d, p0/m, {{z[0-9]+}}.d
+
+SVINT_TEST(__SVUint32_t,b);
+// CHECK: mov {{z[0-9]+}}.b, p0/m, {{z[0-9]+}}.b
+SVINT_TEST(__SVUint32_t,h);
+// CHECK: mov {{z[0-9]+}}.h, p0/m, {{z[0-9]+}}.h
+SVINT_TEST(__SVUint32_t,s);
+// CHECK: mov {{z[0-9]+}}.s, p0/m, {{z[0-9]+}}.s
+SVINT_TEST(__SVUint32_t,d);
+// CHECK: mov {{z[0-9]+}}.d, p0/m, {{z[0-9]+}}.d
+
+SVINT_TEST(__SVUint64_t,b);
+// CHECK: mov {{z[0-9]+}}.b, p0/m, {{z[0-9]+}}.b
+SVINT_TEST(__SVUint64_t,h);
+// CHECK: mov {{z[0-9]+}}.h, p0/m, {{z[0-9]+}}.h
+SVINT_TEST(__SVUint64_t,s);
+// CHECK: mov {{z[0-9]+}}.s, p0/m, {{z[0-9]+}}.s
+SVINT_TEST(__SVUint64_t,d);
+// CHECK: mov {{z[0-9]+}}.d, p0/m, {{z[0-9]+}}.d
+
+SVINT_TEST(__SVInt8_t,b);
+// CHECK: mov {{z[0-9]+}}.b, p0/m, {{z[0-9]+}}.b
+SVINT_TEST(__SVInt8_t,h);
+// CHECK: mov {{z[0-9]+}}.h, p0/m, {{z[0-9]+}}.h
+SVINT_TEST(__SVInt8_t,s);
+// CHECK: mov {{z[0-9]+}}.s, p0/m, {{z[0-9]+}}.s
+SVINT_TEST(__SVInt8_t,d);
+// CHECK: mov {{z[0-9]+}}.d, p0/m, {{z[0-9]+}}.d
+
+SVINT_TEST(__SVInt16_t,b);
+// CHECK: mov {{z[0-9]+}}.b, p0/m, {{z[0-9]+}}.b
+SVINT_TEST(__SVInt16_t,h);
+// CHECK: mov {{z[0-9]+}}.h, p0/m, {{z[0-9]+}}.h
+SVINT_TEST(__SVInt16_t,s);
+// CHECK: mov {{z[0-9]+}}.s, p0/m, {{z[0-9]+}}.s
+SVINT_TEST(__SVInt16_t,d);
+// CHECK: mov {{z[0-9]+}}.d, p0/m, {{z[0-9]+}}.d
+
+SVINT_TEST(__SVInt32_t,b);
+// CHECK: mov {{z[0-9]+}}.b, p0/m, {{z[0-9]+}}.b
+SVINT_TEST(__SVInt32_t,h);
+// CHECK: mov {{z[0-9]+}}.h, p0/m, {{z[0-9]+}}.h
+SVINT_TEST(__SVInt32_t,s);
+// CHECK: mov {{z[0-9]+}}.s, p0/m, {{z[0-9]+}}.s
+SVINT_TEST(__SVInt32_t,d);
+// CHECK: mov {{z[0-9]+}}.d, p0/m, {{z[0-9]+}}.d
+
+SVINT_TEST(__SVInt64_t,b);
+// CHECK: mov {{z[0-9]+}}.b, p0/m, {{z[0-9]+}}.b
+SVINT_TEST(__SVInt64_t,h);
+// CHECK: mov {{z[0-9]+}}.h, p0/m, {{z[0-9]+}}.h
+SVINT_TEST(__SVInt64_t,s);
+// CHECK: mov {{z[0-9]+}}.s, p0/m, {{z[0-9]+}}.s
+SVINT_TEST(__SVInt64_t,d);
+// CHECK: mov {{z[0-9]+}}.d, p0/m, {{z[0-9]+}}.d
+
+
+//Test that floats can also be used as datatypes for integer instructions
+//and check all the variants which would not be possible with a float
+//instruction
+SVINT_TEST(__SVFloat16_t,b);
+// CHECK: mov {{z[0-9]+}}.b, p0/m, {{z[0-9]+}}.b
+SVINT_TEST(__SVFloat16_t,h);
+// CHECK: mov {{z[0-9]+}}.h, p0/m, {{z[0-9]+}}.h
+SVINT_TEST(__SVFloat16_t,s);
+// CHECK: mov {{z[0-9]+}}.s, p0/m, {{z[0-9]+}}.s
+SVINT_TEST(__SVFloat16_t,d);
+// CHECK: mov {{z[0-9]+}}.d, p0/m, {{z[0-9]+}}.d
+
+SVINT_TEST(__SVFloat32_t,b);
+// CHECK: mov {{z[0-9]+}}.b, p0/m, {{z[0-9]+}}.b
+SVINT_TEST(__SVFloat32_t,h);
+// CHECK: mov {{z[0-9]+}}.h, p0/m, {{z[0-9]+}}.h
+SVINT_TEST(__SVFloat32_t,s);
+// CHECK: mov {{z[0-9]+}}.s, p0/m, {{z[0-9]+}}.s
+SVINT_TEST(__SVFloat32_t,d);
+// CHECK: mov {{z[0-9]+}}.d, p0/m, {{z[0-9]+}}.d
+
+SVINT_TEST(__SVFloat64_t,b);
+// CHECK: mov {{z[0-9]+}}.b, p0/m, {{z[0-9]+}}.b
+SVINT_TEST(__SVFloat64_t,h);
+// CHECK: mov {{z[0-9]+}}.h, p0/m, {{z[0-9]+}}.h
+SVINT_TEST(__SVFloat64_t,s);
+// CHECK: mov {{z[0-9]+}}.s, p0/m, {{z[0-9]+}}.s
+SVINT_TEST(__SVFloat64_t,d);
+// CHECK: mov {{z[0-9]+}}.d, p0/m, {{z[0-9]+}}.d
+
+
+#define SVBOOL_TEST(KIND)\
+__SVBool_t func_bool_##KIND(__SVBool_t in1, __SVBool_t in2)\
+{\
+  __SVBool_t out;\
+  asm volatile (\
+    "zip1 %[out]." #KIND ", %[in1]." #KIND ", %[in2]." #KIND "\n"\
+    : [out] "=Upa" (out)\
+    :  [in1] "Upa" (in1),\
+      [in2] "Upl" (in2)\
+    :);\
+  return out;\
+}
+
+SVBOOL_TEST(b) ;
+// CHECK: zip1 {{p[0-9]+}}.b, {{p[0-9]+}}.b, {{p[0-9]+}}.b
+SVBOOL_TEST(h) ;
+// CHECK: zip1 {{p[0-9]+}}.h, {{p[0-9]+}}.h, {{p[0-9]+}}.h
+SVBOOL_TEST(s) ;
+// CHECK: zip1 {{p[0-9]+}}.s, {{p[0-9]+}}.s, {{p[0-9]+}}.s
+SVBOOL_TEST(d) ;
+// CHECK: zip1 {{p[0-9]+}}.d, {{p[0-9]+}}.d, {{p[0-9]+}}.d
+
+
+#define SVFLOAT_TEST(DT,KIND)\
+DT func_float_##DT##KIND(DT inout1, DT in2)\
+{\
+  asm volatile (\
+    "ptrue p0." #KIND ", #1 \n"\
+    "fsub %[inout1]." #KIND ", p0/m, %[inout1]." #KIND ", %[in2]." #KIND "\n"\
+    : [inout1] "=w" (inout1)\
+    : "[inout1]" (inout1),\
+      [in2] "w" (in2)\
+    : "p0");\
+  return inout1 ;\
+}\
+
+SVFLOAT_TEST(__SVFloat16_t,s);
+// CHECK: fsub {{z[0-9]+}}.s, p0/m, {{z[0-9]+}}.s, {{z[0-9]+}}.s
+SVFLOAT_TEST(__SVFloat16_t,d);
+// CHECK: fsub {{z[0-9]+}}.d, p0/m, {{z[0-9]+}}.d, {{z[0-9]+}}.d
+
+SVFLOAT_TEST(__SVFloat32_t,s);
+// CHECK: fsub {{z[0-9]+}}.s, p0/m, {{z[0-9]+}}.s, {{z[0-9]+}}.s
+SVFLOAT_TEST(__SVFloat32_t,d);
+// CHECK: fsub {{z[0-9]+}}.d, p0/m, {{z[0-9]+}}.d, {{z[0-9]+}}.d
+
+SVFLOAT_TEST(__SVFloat64_t,s);
+// CHECK: fsub {{z[0-9]+}}.s, p0/m, {{z[0-9]+}}.s, {{z[0-9]+}}.s
+SVFLOAT_TEST(__SVFloat64_t,d);
+// CHECK: fsub {{z[0-9]+}}.d, p0/m, {{z[0-9]+}}.d, {{z[0-9]+}}.d
+
+
+// Another test for floats to include h suffix
+
+#define SVFLOAT_CVT_TEST(DT1,KIND1,DT2,KIND2)\
+DT1 func_float_cvt_##DT1##KIND1##DT2##KIND2(DT2 in1)\
+{\
+  DT1 out1 ;\
+  asm volatile (\
+    "ptrue p0." #KIND2 ", #1 \n"\
+    "fcvt %[out1]." #KIND1 ", p0/m, %[in1]." #KIND2 "\n"\
+    : [out1] "=w" (out1)\
+    : [in1] "w" (in1)\
+    : "p0");\
+  return out1 ;\
+}\
+
+SVFLOAT_CVT_TEST(__SVFloat64_t,d,__SVFloat32_t,s);
+// CHECK: fcvt {{z[0-9]+}}.d, p0/m, {{z[0-9]+}}.s
+SVFLOAT_CVT_TEST(__SVFloat64_t,d,__SVFloat16_t,h);
+// CHECK: fcvt {{z[0-9]+}}.d, p0/m, {{z[0-9]+}}.h
+SVFLOAT_CVT_TEST(__SVFloat32_t,s,__SVFloat16_t,h);
+// CHECK: fcvt {{z[0-9]+}}.s, p0/m, {{z[0-9]+}}.h
+SVFLOAT_CVT_TEST(__SVFloat32_t,s,__SVFloat64_t,d);
+// CHECK: fcvt {{z[0-9]+}}.s, p0/m, {{z[0-9]+}}.d
+SVFLOAT_CVT_TEST(__SVFloat16_t,h,__SVFloat64_t,d);
+// CHECK: fcvt {{z[0-9]+}}.h, p0/m, {{z[0-9]+}}.d
+SVFLOAT_CVT_TEST(__SVFloat16_t,h,__SVFloat32_t,s);
+// CHECK: fcvt {{z[0-9]+}}.h, p0/m, {{z[0-9]+}}.s
+
+//Test a mix of float and ints
+SVFLOAT_CVT_TEST(__SVInt16_t,h,__SVFloat32_t,s);
+// CHECK: fcvt {{z[0-9]+}}.h, p0/m, {{z[0-9]+}}.s
+SVFLOAT_CVT_TEST(__SVFloat16_t,s,__SVUint32_t,d);
+// CHECK: fcvt {{z[0-9]+}}.s, p0/m, {{z[0-9]+}}.d

clang/test/CodeGen/aarch64-sve-inline-asm-negative-test.c

+// RUN: not %clang_cc1 -triple aarch64-none-linux-gnu -target-feature +sve -fallow-half-arguments-and-returns \
+// RUN:   -target-feature +neon -S -O1 -o - %s | FileCheck %s
+
+// Assembler error
+// Output constraint : Set a vector constraint on an integer
+__SVFloat32_t funcB2()
+{
+  __SVFloat32_t ret ;
+  asm volatile (
+    "fmov %[ret], wzr \n"
+    : [ret] "=w" (ret)
+    :
+    :);
+
+  return ret ;
+}
+
+// CHECK: funcB2
+// CHECK-ERROR: error: invalid operand for instruction

clang/test/CodeGen/aarch64-sve-inline-asm-vec-low.c

+// RUN: %clang_cc1 -triple aarch64-none-linux-gnu -target-feature +sve -fallow-half-arguments-and-returns -S -O1 -Werror -W -Wall -o - %s | FileCheck %s
+
+#define svfcmla_test_f16(Zda, Zn, Zm, rot, imm3) \
+        ({ \
+            __SVFloat16_t _Zn = Zn;\
+            __SVFloat16_t _Zm = Zm;\
+            __SVFloat16_t _res = Zda;\
+            __asm__("fcmla %[__res].h, %[__Zn].h, %[__Zm].h[" #imm3 "], %[__rot]"\
+                        : [__res] "+w" (_res) \
+                        : [__Zn] "w" (_Zn), [__Zm] "y" (_Zm), [__rot] "i" (rot) \
+                        :  \
+                        );\
+            _res; \
+        })
+
+
+// CHECK: fcmla {{z[0-9]+\.h}}, {{z[0-9]+\.h}}, {{z[0-7]\.h}}{{\[[0-9]+\]}}, #270
+__SVFloat16_t test_svfcmla_lane_f16(__SVFloat16_t aZda, __SVFloat16_t aZn, __SVFloat16_t aZm) {
+    return svfcmla_test_f16(aZda, aZn, aZm, 270, 0);
+}
+
+#define svfcmla_test_f32(Zda, Zn, Zm, rot, imm3) \
+        ({ \
+            __SVFloat32_t _Zn = Zn;\
+            __SVFloat32_t _Zm = Zm;\
+            __SVFloat32_t _res = Zda;\
+            __asm__("fcmla %[__res].s, %[__Zn].s, %[__Zm].s[" #imm3 "], %[__rot]"\
+                        : [__res] "+w" (_res) \
+                        : [__Zn] "w" (_Zn), [__Zm] "x" (_Zm), [__rot] "i" (rot) \
+                        :  \
+                        );\
+            _res; \
+        })
+
+
+// CHECK: fcmla {{z[0-9]+\.s}}, {{z[0-9]+\.s}}, {{z[0-9][0-5]?\.s}}{{\[[0-9]+\]}}, #270
+__SVFloat32_t test_svfcmla_lane_f(__SVFloat32_t aZda, __SVFloat32_t aZn, __SVFloat32_t aZm) {
+    return svfcmla_test_f32(aZda, aZn, aZm, 270, 0);
+}