Optionally print `!prof` metadata inline #130303

mtrofin · 2025-03-07T16:19:49Z

Inspired by PR #127944, this patch adds an option to print profile metadata inline with respect to the instruction (or function) it annotates - this saves one time from having to search up and down large textual modules to find this info.

mtrofin · 2025-03-07T16:20:06Z

Optionally print !prof metadata inline #130303 👈 (View in Graphite)
main

This stack of pull requests is managed by Graphite. Learn more about stacking.

llvmbot · 2025-03-07T16:23:28Z

@llvm/pr-subscribers-llvm-ir

Author: Mircea Trofin (mtrofin)

Changes

Inspired by PR #127944, this patch adds an option to print profile metadata inline with respect to the instruction (or function) it annotates - this saves one time from having to search up and down large textual modules to find this info.

Full diff: https://github.com/llvm/llvm-project/pull/130303.diff

3 Files Affected:

(modified) llvm/lib/IR/AsmWriter.cpp (+17)
(added) llvm/test/Other/print-prof-data.ll (+24)
(added) llvm/test/Other/print-prof-data.s (+23)

diff --git a/llvm/lib/IR/AsmWriter.cpp b/llvm/lib/IR/AsmWriter.cpp
index 88bb2359cd055..8399fa8a98f80 100644
--- a/llvm/lib/IR/AsmWriter.cpp
+++ b/llvm/lib/IR/AsmWriter.cpp
@@ -96,6 +96,10 @@ static cl::opt<bool> PrintInstDebugLocs(
     "print-inst-debug-locs", cl::Hidden,
     cl::desc("Pretty print debug locations of instructions when dumping"));
 
+static cl::opt<bool> PrintPerfData(
+    "print-perf-data", cl::Hidden,
+    cl::desc("Pretty print perf data (branch weights, etc) when dumping"));
+
 // Make virtual table appear in this compilation unit.
 AssemblyAnnotationWriter::~AssemblyAnnotationWriter() = default;
 
@@ -4161,6 +4165,11 @@ void AssemblyWriter::printFunction(const Function *F) {
     writeOperand(F->getPersonalityFn(), /*PrintType=*/true);
   }
 
+  if (PrintPerfData && F->getEntryCount()) {
+    Out << " ";
+    F->getMetadata(LLVMContext::MD_prof)->print(Out, TheModule, true);
+  }
+
   if (F->isDeclaration()) {
     Out << '\n';
   } else {
@@ -4287,6 +4296,14 @@ void AssemblyWriter::printInfoComment(const Value &V) {
       }
     }
   }
+  if (PrintPerfData) {
+    if (auto *I = dyn_cast<Instruction>(&V)) {
+      if (auto *MD = I->getMetadata(LLVMContext::MD_prof)) {
+        Out << " ; ";
+        MD->print(Out, TheModule, true);
+      }
+    }
+  }
 
   if (PrintInstAddrs)
     Out << " ; " << &V;
diff --git a/llvm/test/Other/print-prof-data.ll b/llvm/test/Other/print-prof-data.ll
new file mode 100644
index 0000000000000..a5a2d444222b5
--- /dev/null
+++ b/llvm/test/Other/print-prof-data.ll
@@ -0,0 +1,24 @@
+; RUN: opt %s -print-prof-data -S | FileCheck %s
+
+define void @foo(ptr %p) !prof !0 {
+  %isnull = icmp eq ptr %p, null
+  br i1 %isnull, label %yes, label %no, !prof !1
+yes:
+  %something = select i1 %isnull, i32 1, i32 2, !prof !2
+  br label %exit
+no:
+  call void %p(), !prof !3
+  br label %exit
+exit:
+  ret void
+}
+
+!0 = !{!"function_entry_count", i64 42}
+!1 = !{!"branch_weights", i64 20, i64 101}
+!2 = !{!"branch_weights", i64 5, i64 70}
+!3 = !{!"VP", i32 0, i64 4, i64 4445083295448962937, i64 2, i64 -2718743882639408571, i64 2}
+
+; CHECK: define void @foo(ptr %p) !0 = !{!"function_entry_count", i64 42} !prof !0 {
+; CHECK: br i1 %isnull, label %yes, label %no, !prof !1 ; !1 = !{!"branch_weights", i64 20, i64 101}
+; CHECK: %something = select i1 %isnull, i32 1, i32 2, !prof !2 ; !2 = !{!"branch_weights", i64 5, i64 70}
+; CHECK: call void %p(), !prof !3 ; !3 = !{!"VP", i32 0, i64 4, i64 4445083295448962937, i64 2, i64 -2718743882639408571, i64 2}
\ No newline at end of file
diff --git a/llvm/test/Other/print-prof-data.s b/llvm/test/Other/print-prof-data.s
new file mode 100644
index 0000000000000..79595393eab2c
--- /dev/null
+++ b/llvm/test/Other/print-prof-data.s
@@ -0,0 +1,23 @@
+	.file	"print-prof-data.ll"
+	.text
+	.globl	foo                             # -- Begin function foo
+	.p2align	4
+	.type	foo,@function
+foo:                                    # @foo
+	.cfi_startproc
+# %bb.0:
+	testq	%rdi, %rdi
+	je	.LBB0_2
+# %bb.1:                                # %no
+	pushq	%rax
+	.cfi_def_cfa_offset 16
+	callq	*%rdi
+	addq	$8, %rsp
+	.cfi_def_cfa_offset 8
+.LBB0_2:                                # %exit
+	retq
+.Lfunc_end0:
+	.size	foo, .Lfunc_end0-foo
+	.cfi_endproc
+                                        # -- End function
+	.section	".note.GNU-stack","",@progbits

teresajohnson

lgtm with some nits

llvm/test/Other/print-prof-data.s

llvm/lib/IR/AsmWriter.cpp

mingmingl-llvm

thanks!

mtrofin · 2025-03-07T20:20:32Z

Merge activity

Mar 7, 3:20 PM EST: A user started a stack merge that includes this pull request via Graphite.
Mar 7, 3:22 PM EST: A user merged this pull request with Graphite.

pcc · 2025-03-07T20:53:14Z

It may be more generally useful to have a flag for printing instruction and GV metadata inline (one level of expansion) instead of having a special case for !prof. This could also be useful for CFI, e.g.

@foo = constant i32 0, type !12345
!12345 = !{i64 0, !"foo"}

becomes

@foo = constant i32 0, !type !{i64 0, !"foo"}

IIRC inline metadata is parsed as expected so this could be useful for developers who are manually editing metadata e.g. for test cases.

teresajohnson · 2025-03-07T21:04:12Z

It may be more generally useful to have a flag for printing instruction and GV metadata inline (one level of expansion) instead of having a special case for !prof. This could also be useful for CFI, e.g.
@foo = constant i32 0, type !12345
!12345 = !{i64 0, !"foo"}
becomes
@foo = constant i32 0, !type !{i64 0, !"foo"}
IIRC inline metadata is parsed as expected so this could be useful for developers who are manually editing metadata e.g. for test cases.

I was thinking about suggesting that too, but there may be complications for metadata that references other metadata. I'm not sure how that gets rendered automatically, and whether and how much to expand it inline, and what would happen when parsed again. But yes, a broader expansion to other metadata types could be very useful.

nikic · 2025-03-07T21:06:51Z

Agree with @pcc, an option to generally print instruction metadata inline would be quite useful. (Excluding distinct and self-referential metadata.)

llvm-ci · 2025-03-07T21:46:46Z

LLVM Buildbot has detected a new failure on builder lld-x86_64-ubuntu-fast running on as-builder-4 while building llvm at step 6 "test-build-unified-tree-check-all".

Full details are available at: https://lab.llvm.org/buildbot/#/builders/33/builds/12631

Here is the relevant piece of the build log for the reference

Step 6 (test-build-unified-tree-check-all) failure: test (failure)
******************** TEST 'LLVM-Unit :: Support/./SupportTests/48/88' FAILED ********************
Script(shard):
--
GTEST_OUTPUT=json:/home/buildbot/worker/as-builder-4/ramdisk/lld-x86_64/build/unittests/Support/./SupportTests-LLVM-Unit-2861403-48-88.json GTEST_SHUFFLE=0 GTEST_TOTAL_SHARDS=88 GTEST_SHARD_INDEX=48 /home/buildbot/worker/as-builder-4/ramdisk/lld-x86_64/build/unittests/Support/./SupportTests
--

Script:
--
/home/buildbot/worker/as-builder-4/ramdisk/lld-x86_64/build/unittests/Support/./SupportTests --gtest_filter=Caching.NoCommit
--
/home/buildbot/worker/as-builder-4/ramdisk/lld-x86_64/llvm-project/llvm/unittests/Support/Caching.cpp:142: Failure
Value of: AddStream
  Actual: false
Expected: true


/home/buildbot/worker/as-builder-4/ramdisk/lld-x86_64/llvm-project/llvm/unittests/Support/Caching.cpp:142
Value of: AddStream
  Actual: false
Expected: true



********************

mtrofin · 2025-03-08T02:16:54Z

Agree with @pcc, an option to generally print instruction metadata inline would be quite useful. (Excluding distinct and self-referential metadata.)

I agree it would be nice - let me look into the complications @teresajohnson was mentioning. I didn't realize that inlined metadata would be parsed, that's good to know!

Inspired by PR llvm#127944, this patch adds an option to print profile metadata inline with respect to the instruction (or function) it annotates - this saves one time from having to search up and down large textual modules to find this info.

mtrofin requested review from pcc, rnk and teresajohnson March 7, 2025 16:22

mtrofin marked this pull request as ready for review March 7, 2025 16:22

llvmbot added the llvm:ir label Mar 7, 2025

mtrofin requested a review from mingmingl-llvm March 7, 2025 16:24

teresajohnson approved these changes Mar 7, 2025

View reviewed changes

llvm/test/Other/print-prof-data.s Outdated Show resolved Hide resolved

llvm/lib/IR/AsmWriter.cpp Outdated Show resolved Hide resolved

llvm/lib/IR/AsmWriter.cpp Outdated Show resolved Hide resolved

mingmingl-llvm approved these changes Mar 7, 2025

View reviewed changes

mtrofin force-pushed the users/mtrofin/03-07-optionally_print_prof_metadata_inline branch 2 times, most recently from 2d5a881 to 33718e8 Compare March 7, 2025 17:01

Optionally print !prof metadata inline

fd5e082

mtrofin force-pushed the users/mtrofin/03-07-optionally_print_prof_metadata_inline branch from 33718e8 to fd5e082 Compare March 7, 2025 19:07

mtrofin merged commit 11b1f15 into main Mar 7, 2025
8 of 11 checks passed

mtrofin deleted the users/mtrofin/03-07-optionally_print_prof_metadata_inline branch March 7, 2025 20:22

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Optionally print `!prof` metadata inline #130303

Optionally print `!prof` metadata inline #130303

Uh oh!

mtrofin commented Mar 7, 2025 •

edited

Loading

Uh oh!

mtrofin commented Mar 7, 2025

Uh oh!

llvmbot commented Mar 7, 2025

Uh oh!

teresajohnson left a comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

mingmingl-llvm left a comment

Uh oh!

mtrofin commented Mar 7, 2025 •

edited

Loading

Uh oh!

Uh oh!

pcc commented Mar 7, 2025

Uh oh!

teresajohnson commented Mar 7, 2025

Uh oh!

nikic commented Mar 7, 2025

Uh oh!

llvm-ci commented Mar 7, 2025

Uh oh!

mtrofin commented Mar 8, 2025

Uh oh!

Uh oh!

Optionally print !prof metadata inline #130303

Optionally print !prof metadata inline #130303

Uh oh!

Conversation

mtrofin commented Mar 7, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

mtrofin commented Mar 7, 2025

Uh oh!

llvmbot commented Mar 7, 2025

Uh oh!

teresajohnson left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

mingmingl-llvm left a comment

Choose a reason for hiding this comment

Uh oh!

mtrofin commented Mar 7, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Merge activity

Uh oh!

Uh oh!

pcc commented Mar 7, 2025

Uh oh!

teresajohnson commented Mar 7, 2025

Uh oh!

nikic commented Mar 7, 2025

Uh oh!

llvm-ci commented Mar 7, 2025

Uh oh!

mtrofin commented Mar 8, 2025

Uh oh!

Uh oh!

Optionally print `!prof` metadata inline #130303

Optionally print `!prof` metadata inline #130303

mtrofin commented Mar 7, 2025 •

edited

Loading

mtrofin commented Mar 7, 2025 •

edited

Loading