[ARM] Allow automatically deducing the thumb instruction size for .inst

[mstorsjo]

Summary

This matches GAS, that allows unsuffixed .inst for thumb.

[peter.smith]

That looks like the right behaviour to me. For reference the Arm ARM definition of Thumb Instruction width from encodeing is in https://developer.arm.com/products/architecture/cpu-architecture/a-profile/docs/ddi0406/latest/arm-architecture-reference-manual-armv7-a-and-armv7-r-edition 6.1 Thumb Instruction set encoding.

test/MC/ARM/inst-thumb-suffixes-auto.s

2

Would be good to add a test with triple armebv7-linux-gnueabi, my understanding is that the .inst and CHECK should remain the same.

[mstorsjo]

test/MC/ARM/inst-thumb-suffixes-auto.s

2

Yes, these ones remain the same here since it just tests the output assembly, not the generated binary. I'll amend the test.

[mstorsjo]

Added a test RUN line with a big-endian arch.

lib/Target/ARM/AsmParser/ARMAsmParser.cpp

   if (isThumb()) {
     switch (Suffix) {
     case 'n':
       Width = 2;
       break;
     case 'w':
       break;
     default:
-      return Error(Loc, "cannot determine Thumb instruction size, "
-                        "use inst.n/inst.w instead");
+      Width = 0;
+      break;
     }
   } else {
     if (Suffix)
       return Error(Loc, "width suffixes are invalid in ARM mode");
   }
 
   auto parseOne = [&]() -> bool {
     const MCExpr *Expr;
     if (getParser().parseExpression(Expr))
       return true;
     const MCConstantExpr *Value = dyn_cast_or_null<MCConstantExpr>(Expr);
     if (!Value) {
       return Error(Loc, "expected constant expression");
     }
 
+    char CurSuffix = Suffix;
     switch (Width) {
     case 2:
       if (Value->getValue() > 0xffff)
         return Error(Loc, "inst.n operand is too big, use inst.w instead");
       break;
     case 4:
       if (Value->getValue() > 0xffffffff)
         return Error(Loc, StringRef(Suffix ? "inst.w" : "inst") +
                               " operand is too big");
       break;
+    case 0:
+      // Thumb mode, no width indicated. Guess from the opcode, if possible.
+      if (Value->getValue() < 0xe800)
+        CurSuffix = 'n';
+      else if (Value->getValue() >= 0xe8000000)
+        CurSuffix = 'w';
+      else
+        return Error(Loc, "cannot determine Thumb instruction size, "
+                          "use inst.n/inst.w instead");
+      break;
     default:
       llvm_unreachable("only supported widths are 2 and 4");
     }
 
-    getTargetStreamer().emitInst(Value->getValue(), Suffix);
+    getTargetStreamer().emitInst(Value->getValue(), CurSuffix);
     return false;
   };
 
   if (parseOptionalToken(AsmToken::EndOfStatement))
     return Error(Loc, "expected expression following directive");
   if (parseMany(parseOne))
     return true;
   return false;

test/MC/ARM/inst-thumb-suffixes-auto.s

+@ RUN: llvm-mc %s -triple armv7-linux-gnueabi -filetype asm -o - \
+@ RUN:   | FileCheck %s
+@ RUN: llvm-mc %s -triple armebv7-linux-gnueabi -filetype asm -o - \
+@ RUN:   | FileCheck %s
+
+  .syntax unified
+  .thumb
+
+  .align 2
+  .global inst_n
+  .type inst_n,%function
+inst_n:
+  @ bx lr, mov.w r0, #42
+  .inst 0x4770, 0xf04f002a
+@ CHECK: .inst.n 0x4770
+@ CHECK: .inst.w 0xf04f002a

test/MC/ARM/inst-thumb-suffixes.s

 @ RUN: not llvm-mc %s -triple armv7-linux-gnueabi -filetype asm -o - 2>&1 \
 @ RUN:   | FileCheck -check-prefix CHECK-ERROR %s
 
   .syntax unified
   .thumb
 
   .align 2
   .global suffixes_required_in_thumb
   .type suffixes_required_in_thumb,%function
 suffixes_required_in_thumb:
-  .inst 0x0000
+  .inst 0xff00
 @ CHECK-ERROR: cannot determine Thumb instruction size, use inst.n/inst.w instead