Ignore device-side asm constraint errors while compiling CUDA code for host and vice versa.

[tra]

Summary

Ignore device-side asm constraint errors while compiling CUDA code for host and vice versa.

[eliben]

test/SemaCUDA/asm-constraints.cu

7 ↗

(On Diff #22118)

Are there checks anywhere that asm diagnostics are validated within device?

Perhaps a positive test would be helpful (make sure that an invalid constraint produces a diagnostic within a kernel)

[tra]

Good point. Will do.

[echristo]

As Eli said a positive test would be nice, otherwise LGTM.

-eric

[tra]

Added a test to verify that we do validate asm constraints in CUDA device compilation mode.
Addressed echristo@'s offline style comment.

[echristo]

LGTM.

Thanks!

-eric

[eliben]

LGTM

cfe/trunk/lib/Sema/SemaStmtAsm.cpp

StmtResult Sema::ActOnGCCAsmStmt(SourceLocation AsmLoc, bool IsSimple,
   StringLiteral *AsmString = cast<StringLiteral>(asmString);
   StringLiteral **Clobbers = reinterpret_cast<StringLiteral**>(clobbers.data());
 
   SmallVector<TargetInfo::ConstraintInfo, 4> OutputConstraintInfos;
 
   // The parser verifies that there is a string literal here.
   assert(AsmString->isAscii());
 
+  bool ValidateConstraints = true;
+  if (getLangOpts().CUDA) {
+    // In CUDA mode don't verify asm constraints in device functions during host
+    // compilation and vice versa.
+    bool InDeviceMode = getLangOpts().CUDAIsDevice;
+    FunctionDecl *FD = getCurFunctionDecl();
+    bool IsDeviceFunction =
+        FD && (FD->hasAttr<CUDADeviceAttr>() || FD->hasAttr<CUDAGlobalAttr>());
+    ValidateConstraints = IsDeviceFunction == InDeviceMode;
+  }
+
   for (unsigned i = 0; i != NumOutputs; i++) {
     StringLiteral *Literal = Constraints[i];
     assert(Literal->isAscii());
 
     StringRef OutputName;
     if (Names[i])
       OutputName = Names[i]->getName();
 
     TargetInfo::ConstraintInfo Info(Literal->getString(), OutputName);
-    if (!Context.getTargetInfo().validateOutputConstraint(Info))
+    if (ValidateConstraints &&
+        !Context.getTargetInfo().validateOutputConstraint(Info))
       return StmtError(Diag(Literal->getLocStart(),
                             diag::err_asm_invalid_output_constraint)
                        << Info.getConstraintStr());
 
     ExprResult ER = CheckPlaceholderExpr(Exprs[i]);
     if (ER.isInvalid())
       return StmtError();
     Exprs[i] = ER.get();
for (unsigned i = NumOutputs, e = NumOutputs + NumInputs; i != e; i++) {
     StringLiteral *Literal = Constraints[i];
     assert(Literal->isAscii());
 
     StringRef InputName;
     if (Names[i])
       InputName = Names[i]->getName();
 
     TargetInfo::ConstraintInfo Info(Literal->getString(), InputName);
-    if (!Context.getTargetInfo().validateInputConstraint(OutputConstraintInfos.data(),
-                                                NumOutputs, Info)) {
+    if (ValidateConstraints &&
+        !Context.getTargetInfo().validateInputConstraint(
+            OutputConstraintInfos.data(), NumOutputs, Info)) {
       return StmtError(Diag(Literal->getLocStart(),
                             diag::err_asm_invalid_input_constraint)
                        << Info.getConstraintStr());
     }
 
     ExprResult ER = CheckPlaceholderExpr(Exprs[i]);
     if (ER.isInvalid())
       return StmtError();

cfe/trunk/test/SemaCUDA/asm-constraints-device.cu

+// Verify that we do check for constraints in device-side inline
+// assembly. Passing an illegal input/output constraint and look 
+// for corresponding error
+// RUN: %clang_cc1 -triple nvptx-unknown-cuda -fsyntax-only -fcuda-is-device -verify %s
+
+__attribute__((device)) void df() {
+  short h;
+  int a;
+  // asm with PTX constraints. Some of them are PTX-specific.
+  __asm__("output constraints"
+          : "=h"(h), // .u16 reg, OK
+            "=a"(a)  // expected-error {{invalid output constraint '=a' in asm}}
+          :          // None
+          );
+  __asm__("input constraints"
+          :           // None
+          : "f"(0.0), // .f32 reg, OK
+            "d"(0.0), // .f64 reg, OK
+            "h"(0),   // .u16 reg, OK
+            "r"(0),   // .u32 reg, OK
+            "l"(0),   // .u64 reg, OK
+            "a"(0)    // expected-error {{invalid input constraint 'a' in asm}}
+          );
+}

cfe/trunk/test/SemaCUDA/asm-constraints-mixed.cu

+// RUN: %clang_cc1 -triple nvptx-unknown-cuda -fsyntax-only -fcuda-is-device -verify %s
+// RUN: %clang_cc1 -triple x86_64-unknown-linux-gnu -fsyntax-only -verify %s
+// expected-no-diagnostics
+
+__attribute__((device)) void df() {
+  short h;
+  // asm with PTX constraints. Some of them are PTX-specific.
+  __asm__("dont care" : "=h"(h): "f"(0.0), "d"(0.0), "h"(0), "r"(0), "l"(0));
+}
+
+void hf() {
+  int a;
+  // Asm with x86 constraints that are not supported by PTX.
+  __asm__("dont care" : "=a"(a): "a"(0), "b"(0), "c"(0));
+}