[CUDA] Don't pass top-level -march down to device cc1 or ptxas.

[jlebar]

Summary

Previously if you did e.g.

$ clang -march=haswell -x cuda foo.cu

we would pass "-march=haswell -march=sm_20" down to the ptxas tool.
This causes it to assert, and rightly so!

[tra]

test/Driver/cuda-march.cu

15–16 ↗

(On Diff #60927)

These look redundant -- we only care whether we eliminate -march on device side. It does not depend on the value of -march or on particular GPU arch. What do I miss?

[jlebar]

Remove redundant test.

[jlebar]

test/Driver/cuda-march.cu

15–16 ↗

(On Diff #60927)

Removed the last one. I don't think the first two are redundant, because I wanted to check that the *implicit* sm_20 still overrides the explicit -mhaswell.

[tra]

test/Driver/cuda-march.cu

22–28 ↗

(On Diff #60931)

You don't need SM30 checks these now.
Speaking of checks, you only need one check label now as all runs check for both SM20 and HASWELL.

[jlebar]

Fix tests for real this time.

[tra]

LGTM.

cfe/trunk/lib/Driver/ToolChains.cpp

if (A->getOption().matches(options::OPT_Xarch__)) {
       }
       XarchArg->setBaseArg(A);
       A = XarchArg.release();
       DAL->AddSynthesizedArg(A);
     }
     DAL->append(A);
   }
 
-  if (BoundArch)
+  if (BoundArch) {
+    DAL->eraseArg(options::OPT_march_EQ);
     DAL->AddJoinedArg(nullptr, Opts.getOption(options::OPT_march_EQ), BoundArch);
+  }
   return DAL;
 }
 
 Tool *CudaToolChain::buildAssembler() const {
   return new tools::NVPTX::Assembler(*this);
 }
 
 Tool *CudaToolChain::buildLinker() const {

cfe/trunk/test/Driver/cuda-march.cu

+// Checks that cuda compilation does the right thing when passed -march.
+// (Specifically, we want to pass it to host compilation, but not to device
+// compilation or ptxas!)
+//
+// REQUIRES: clang-driver
+// REQUIRES: x86-registered-target
+// REQUIRES: nvptx-registered-target
+
+// RUN: %clang -### -target x86_64-linux-gnu -c -march=haswell %s 2>&1 | FileCheck %s
+
+// RUN: %clang -### -target x86_64-linux-gnu -c -march=haswell --cuda-gpu-arch=sm_20 %s 2>&1 | \
+// RUN: FileCheck %s
+
+// CHECK:clang
+// CHECK: "-cc1"
+// CHECK-SAME: "-triple" "nvptx
+// CHECK-SAME: "-target-cpu" "sm_20"
+
+// CHECK: ptxas
+// CHECK-SAME: "--gpu-name" "sm_20"
+
+// CHECK:clang
+// CHECK-SAME: "-cc1"
+// CHECK-SAME: "-target-cpu" "haswell"