diff --git a/llvm/lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp b/llvm/lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp
--- a/llvm/lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp
+++ b/llvm/lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp
@@ -5020,9 +5020,11 @@
       while (!getLexer().is(AsmToken::EndOfStatement)) {
         Parser.Lex();
       }
+      Parser.Lex();
       return true;
     }
   }
+  Parser.Lex();
 
   return false;
 }
diff --git a/llvm/test/MC/AMDGPU/round-trip.s b/llvm/test/MC/AMDGPU/round-trip.s
new file mode 100644
--- /dev/null
+++ b/llvm/test/MC/AMDGPU/round-trip.s
@@ -0,0 +1,13 @@
+# RUN: llvm-mc -preserve-comments -triple amdgcn-amd-amdhsa %s >%t-1.s
+# RUN: llvm-mc -preserve-comments -triple amdgcn-amd-amdhsa %t-1.s >%t-2.s
+# RUN: diff %t-1.s %t-2.s
+
+# Test that AMDGPU assembly round-trips when run through MC; the first
+# transition from hand-written to "canonical" output may introduce some small
+# differences, so we don't include the initial input in the comparison.
+
+.text
+
+# The AMDGPU asm parser didn't consume the end of statement
+# consistently, which led to extra empty lines in the output.
+s_nop 0