[ARM] Fix subtarget feature set truncation when using .cpu directive

[bsmith]

Summary

When using the .cpu directive with the integrated assembler on ARM, various subtarget features are getting incorrectly disabled. This is caused by storing the 64-bit feature bitset in a 32-bit variable, causing the top half of the available feature set to be discarded.

[jroelofs]

... other than that, LGTM.

lib/Target/ARM/AsmParser/ARMAsmParser.cpp

9186

I would change these two lines to:

setAvailableFeatures(ComputeAvailableFeatures(STI.getFeatureBits()));

[rengolin]

Ouch! LGTM with Jonathan's comment. Thanks!

[bsmith]

Thanks, committed as 228151 (with the tweak).

lib/Target/ARM/AsmParser/ARMAsmParser.cpp

if (!STI.isCPUStringValid(CPU)) {
     return false;
   }
 
   // FIXME: This switches the CPU features globally, therefore it might
   // happen that code you would not expect to assemble will. For details
   // see: http://llvm.org/bugs/show_bug.cgi?id=20757
   STI.InitMCProcessorInfo(CPU, "");
   STI.InitCPUSchedModel(CPU);
-  unsigned FB = ComputeAvailableFeatures(STI.getFeatureBits());
+  uint64_t FB = ComputeAvailableFeatures(STI.getFeatureBits());
   setAvailableFeatures(FB);
 
   return false;
 }
 
 // FIXME: This is duplicated in getARMFPUFeatures() in
 // tools/clang/lib/Driver/Tools.cpp
 static const struct {
   const unsigned ID;

test/MC/ARM/cpu-test.s

 // RUN: not llvm-mc -o - -triple arm-gnueabi-freebsd11.0 < %s > %t 2> %t2
 // RUN: FileCheck %s < %t
 // RUN: FileCheck %s --check-prefix=CHECK-ERROR < %t2
 
 // CHECK: .cpu cortex-a8
 .cpu cortex-a8
 // CHECK: dsb     sy
 dsb
 .cpu arm9       
 // CHECK-ERROR: error: instruction requires: data-barriers
 dsb
 // CHECK-ERROR: error: Unknown CPU name
 .cpu foobar
+// CHECK: .cpu cortex-m3
+.cpu cortex-m3
+// CHECK: sub sp, #16
+sub sp,#16