[AsmParser][MCA] Fix handling of multi-character comments

[alexander-shaposhnikov]

This PR adjusts the lexer (following the surrounding style) to make sure the comments are consumed
exactly after CommentString.
This enables to feed the output of clang on aarch64 into llvm-mca (previously broken - the comments
//LLVM-MCA-BEGIN and //LLVM-MCA-END were actually ignored (see e.g. https://godbolt.org/z/PEzW7Tvra)).

Test plan: ninja check-all

[llvmbot]

@llvm/pr-subscribers-mc

Author: Alexander Shaposhnikov (alexander-shaposhnikov)

Changes

This PR adjusts the lexer (following the surrounding style) to make sure the comments are consumed
exactly after CommentString.
This enables to feed the output of clang on aarch64 into llvm-mca (previously broken - the comments
//LLVM-MCA-BEGIN and //LLVM-MCA-END were actually ignored (see e.g. https://godbolt.org/z/PEzW7Tvra)).

Test plan: ninja check-all

---

Full diff: https://github.com/llvm/llvm-project/pull/147228.diff

3 Files Affected:

• (modified) llvm/lib/MC/MCParser/AsmLexer.cpp (+3-1)
• (added) llvm/test/MC/AsmParser/preserve-comments-aarch64-linux.s (+10)
• (added) llvm/test/tools/llvm-mca/AArch64/Neoverse/llvm-mca-markers.s (+36)

diff --git a/llvm/lib/MC/MCParser/AsmLexer.cpp b/llvm/lib/MC/MCParser/AsmLexer.cpp
index 3db9ed3199dd8..968ccf776440b 100644
--- a/llvm/lib/MC/MCParser/AsmLexer.cpp
+++ b/llvm/lib/MC/MCParser/AsmLexer.cpp
@@ -834,8 +834,10 @@ AsmToken AsmLexer::LexToken() {
       return LexLineComment();
   }
 
-  if (isAtStartOfComment(TokStart))
+  if (isAtStartOfComment(TokStart)) {
+    CurPtr += MAI.getCommentString().size() - 1;
     return LexLineComment();
+  }
 
   if (isAtStatementSeparator(TokStart)) {
     CurPtr += strlen(MAI.getSeparatorString()) - 1;
diff --git a/llvm/test/MC/AsmParser/preserve-comments-aarch64-linux.s b/llvm/test/MC/AsmParser/preserve-comments-aarch64-linux.s
new file mode 100644
index 0000000000000..acaa90b012008
--- /dev/null
+++ b/llvm/test/MC/AsmParser/preserve-comments-aarch64-linux.s
@@ -0,0 +1,10 @@
+	// REQUIRES: aarch64-registered-target
+	// RUN: llvm-mc -preserve-comments -n -triple aarch64-unknown-linux-gnu < %s > %t
+	// RUN: diff -b %s %t
+
+	.text
+
+foo:
+	// comment here
+	nop
+	// comment here too
diff --git a/llvm/test/tools/llvm-mca/AArch64/Neoverse/llvm-mca-markers.s b/llvm/test/tools/llvm-mca/AArch64/Neoverse/llvm-mca-markers.s
new file mode 100644
index 0000000000000..d5a291b1d7102
--- /dev/null
+++ b/llvm/test/tools/llvm-mca/AArch64/Neoverse/llvm-mca-markers.s
@@ -0,0 +1,36 @@
+# NOTE: Assertions have been autogenerated by utils/update_mca_test_checks.py
+// RUN: llvm-mca -mtriple=aarch64-unknown-linux-gnu -mcpu=neoverse-v2 -iterations=1 -resource-pressure=false < %s | FileCheck %s
+
+.text
+// LLVM-MCA-BEGIN Empty
+// Empty sequence
+// LLVM-MCA-END
+
+  mul x1, x1, x1
+// LLVM-MCA-BEGIN NotEmpty
+  add x0, x0, x1
+// LLVM-MCA-END
+  mul x2, x2, x2
+
+# CHECK:      [0] Code Region - NotEmpty
+
+# CHECK:      Iterations:        1
+# CHECK-NEXT: Instructions:      1
+# CHECK-NEXT: Total Cycles:      4
+# CHECK-NEXT: Total uOps:        1
+
+# CHECK:      Dispatch Width:    6
+# CHECK-NEXT: uOps Per Cycle:    0.25
+# CHECK-NEXT: IPC:               0.25
+# CHECK-NEXT: Block RThroughput: 0.2
+
+# CHECK:      Instruction Info:
+# CHECK-NEXT: [1]: #uOps
+# CHECK-NEXT: [2]: Latency
+# CHECK-NEXT: [3]: RThroughput
+# CHECK-NEXT: [4]: MayLoad
+# CHECK-NEXT: [5]: MayStore
+# CHECK-NEXT: [6]: HasSideEffects (U)
+
+# CHECK:      [1]    [2]    [3]    [4]    [5]    [6]    Instructions:
+# CHECK-NEXT:  1      1     0.17                        add	x0, x0, x1

[alexander-shaposhnikov]

no, llvm-mc still prints tabs (with -o), i think that's the reason why some other tests contain them as well