[C API] Fix LLVMGetOrdering/LLVMIsAtomicSingleThread for fence/memory instrs #65228

Benjins · 2023-09-03T12:48:39Z

LLVMGetOrdering previously did not support Fence instructions, and calling it on a fence would lead to a bad cast as it
assumed a load/store, or an AtomicRMWInst. This would either read a garbage memory order, or assertion

LLVMIsAtomicSingleThread did not support either Fence instructions, loads, or stores, and would similarly lead to a bad cast.
It happened to work out since the relevant types all have their synch scope ID at the same offset, but it still should be fixed

These cases are now fixed for the C API, and tests for these instructions are added. The echo test utility now also supports cloning Fence instructions, which it did not previously

From what I can tell, there's no unified API to pull getOrdering/getSyncScopeID from, and instead requires casting to individual types: if there is a better way of handling this I can switch to that

Benjins · 2023-09-03T13:39:55Z

From the build logs:

mt.exe : general error c101008d: Failed to write the updated manifest to the resource of file "unittests\Testing\Support\TestingSupportTests.exe". Operation did not complete successfully because the file contains a virus or potentially unwanted software.

Possibly Windows Defender triggering on it?

… instrs LLVMGetOrdering previously did not support Fence instructions, and calling it on a fence would lead to a bad cast as it assumed a load/store, or an AtomicRMWInst. This would either read a garbage memory order, or assertion LLVMIsAtomicSingleThread did not support either Fence instructions, loads, or stores, and would similarly lead to a bad cast These cases are now fixed for the C API, and tests for these instructions are added. The echo test utility now also supports cloning Fence instructions, which it did not previously

cdisselkoen

This all looks good to me (and I'm happy to see this change in order to resolve cdisselkoen/llvm-ir#44 and cdisselkoen/llvm-ir#15)

Benjins · 2023-09-30T12:46:33Z

Bumping this in case it was missed: if anyone with merge access would be relevant to review or merge, I can ping them. I tried locally merging this into latest main and the tests still passed, so it should still be good in this state

nikic · 2023-09-30T12:55:48Z

llvm/lib/IR/Core.cpp

@@ -3769,6 +3771,8 @@ void LLVMSetOrdering(LLVMValueRef MemAccessInst, LLVMAtomicOrdering Ordering) {

  if (LoadInst *LI = dyn_cast<LoadInst>(P))
    return LI->setOrdering(O);
+  else if (FenceInst *FI = dyn_cast<FenceInst>(P))
+    return FI->setOrdering(O);


Missing atomicrmw handling?

I can add this as well. My understanding is that the bindings tests will not use that, since it uses LLVMBuildAtomicRMW to set the ordering directly. Is that okay, or is there a way to also add a test for it?

llvmbot · 2023-09-30T13:52:33Z

@llvm/pr-subscribers-llvm-ir

Changes

Fixes #65227

LLVMGetOrdering previously did not support Fence instructions, and calling it on a fence would lead to a bad cast as it
assumed a load/store, or an AtomicRMWInst. This would either read a garbage memory order, or assertion

LLVMIsAtomicSingleThread did not support either Fence instructions, loads, or stores, and would similarly lead to a bad cast.
It happened to work out since the relevant types all have their synch scope ID at the same offset, but it still should be fixed

These cases are now fixed for the C API, and tests for these instructions are added. The echo test utility now also supports cloning Fence instructions, which it did not previously

From what I can tell, there's no unified API to pull getOrdering/getSyncScopeID from, and instead requires casting to individual types: if there is a better way of handling this I can switch to that

Full diff: https://github.com/llvm/llvm-project/pull/65228.diff

3 Files Affected:

(modified) llvm/lib/IR/Core.cpp (+18)
(modified) llvm/test/Bindings/llvm-c/atomics.ll (+34)
(modified) llvm/tools/llvm-c-test/echo.cpp (+8)

diff --git a/llvm/lib/IR/Core.cpp b/llvm/lib/IR/Core.cpp
index 17093fa0ac4ee1e..04d044b8d51e11b 100644
--- a/llvm/lib/IR/Core.cpp
+++ b/llvm/lib/IR/Core.cpp
@@ -3758,6 +3758,8 @@ LLVMAtomicOrdering LLVMGetOrdering(LLVMValueRef MemAccessInst) {
     O = LI->getOrdering();
   else if (StoreInst *SI = dyn_cast<StoreInst>(P))
     O = SI->getOrdering();
+  else if (FenceInst *FI = dyn_cast<FenceInst>(P))
+    O = FI->getOrdering();
   else
     O = cast<AtomicRMWInst>(P)->getOrdering();
   return mapToLLVMOrdering(O);
@@ -3769,6 +3771,10 @@ void LLVMSetOrdering(LLVMValueRef MemAccessInst, LLVMAtomicOrdering Ordering) {
 
   if (LoadInst *LI = dyn_cast<LoadInst>(P))
     return LI->setOrdering(O);
+  else if (FenceInst *FI = dyn_cast<FenceInst>(P))
+    return FI->setOrdering(O);
+  else if (AtomicRMWInst *ARWI = dyn_cast<AtomicRMWInst>(P))
+    return ARWI->setOrdering(O);
   return cast<StoreInst>(P)->setOrdering(O);
 }
 
@@ -4039,6 +4045,12 @@ LLVMBool LLVMIsAtomicSingleThread(LLVMValueRef AtomicInst) {
 
   if (AtomicRMWInst *I = dyn_cast<AtomicRMWInst>(P))
     return I->getSyncScopeID() == SyncScope::SingleThread;
+  else if (FenceInst *FI = dyn_cast<FenceInst>(P))
+    return FI->getSyncScopeID() == SyncScope::SingleThread;
+  else if (StoreInst *SI = dyn_cast<StoreInst>(P))
+    return SI->getSyncScopeID() == SyncScope::SingleThread;
+  else if (LoadInst *LI = dyn_cast<LoadInst>(P))
+    return LI->getSyncScopeID() == SyncScope::SingleThread;
   return cast<AtomicCmpXchgInst>(P)->getSyncScopeID() ==
              SyncScope::SingleThread;
 }
@@ -4049,6 +4061,12 @@ void LLVMSetAtomicSingleThread(LLVMValueRef AtomicInst, LLVMBool NewValue) {
 
   if (AtomicRMWInst *I = dyn_cast<AtomicRMWInst>(P))
     return I->setSyncScopeID(SSID);
+  else if (FenceInst *FI = dyn_cast<FenceInst>(P))
+    return FI->setSyncScopeID(SSID);
+  else if (StoreInst *SI = dyn_cast<StoreInst>(P))
+    return SI->setSyncScopeID(SSID);
+  else if (LoadInst *LI = dyn_cast<LoadInst>(P))
+    return LI->setSyncScopeID(SSID);
   return cast<AtomicCmpXchgInst>(P)->setSyncScopeID(SSID);
 }
 
diff --git a/llvm/test/Bindings/llvm-c/atomics.ll b/llvm/test/Bindings/llvm-c/atomics.ll
index ba29e5e1e170862..e64a29944ef9df9 100644
--- a/llvm/test/Bindings/llvm-c/atomics.ll
+++ b/llvm/test/Bindings/llvm-c/atomics.ll
@@ -2,6 +2,40 @@
 ; RUN: llvm-as < %s | llvm-c-test --echo > %t.echo
 ; RUN: diff -w %t.orig %t.echo
 
+
+define void @fence_instrs() {
+  fence acquire
+  fence release
+  fence acq_rel
+  fence seq_cst
+
+  fence syncscope("singlethread") acquire
+  fence syncscope("singlethread") release
+  fence syncscope("singlethread") acq_rel
+  fence syncscope("singlethread") seq_cst
+
+  ret void
+}
+
+define void @atomic_load_store(ptr %word) {
+  ; Test different atomic loads
+  %ld.1 = load atomic i32, ptr %word monotonic, align 4
+  %ld.2 = load atomic volatile i32, ptr %word acquire, align 4
+  %ld.3 = load atomic volatile i32, ptr %word seq_cst, align 4
+  %ld.4 = load atomic volatile i32, ptr %word syncscope("singlethread") acquire, align 4
+  %ld.5 = load atomic volatile i32, ptr %word syncscope("singlethread") seq_cst, align 4
+  %ld.6 = load atomic i32, ptr %word syncscope("singlethread") seq_cst, align 4
+
+  ; Test different atomic stores
+  store atomic i32 1, ptr %word monotonic, align 4
+  store atomic volatile i32 2, ptr %word release, align 4
+  store atomic volatile i32 3, ptr %word seq_cst, align 4
+  store atomic volatile i32 4, ptr %word syncscope("singlethread") release, align 4
+  store atomic volatile i32 5, ptr %word syncscope("singlethread") seq_cst, align 4
+  store atomic i32 6, ptr %word syncscope("singlethread") seq_cst, align 4
+  ret void
+}
+
 define i32 @main() {
   %1 = alloca i32, align 4
   %2 = cmpxchg ptr %1, i32 2, i32 3 seq_cst acquire
diff --git a/llvm/tools/llvm-c-test/echo.cpp b/llvm/tools/llvm-c-test/echo.cpp
index faf9838a0069afc..06966ce528eae4d 100644
--- a/llvm/tools/llvm-c-test/echo.cpp
+++ b/llvm/tools/llvm-c-test/echo.cpp
@@ -677,6 +677,7 @@ struct FunCloner {
         LLVMSetAlignment(Dst, LLVMGetAlignment(Src));
         LLVMSetOrdering(Dst, LLVMGetOrdering(Src));
         LLVMSetVolatile(Dst, LLVMGetVolatile(Src));
+        LLVMSetAtomicSingleThread(Dst, LLVMIsAtomicSingleThread(Src));
         break;
       }
       case LLVMStore: {
@@ -686,6 +687,7 @@ struct FunCloner {
         LLVMSetAlignment(Dst, LLVMGetAlignment(Src));
         LLVMSetOrdering(Dst, LLVMGetOrdering(Src));
         LLVMSetVolatile(Dst, LLVMGetVolatile(Src));
+        LLVMSetAtomicSingleThread(Dst, LLVMIsAtomicSingleThread(Src));
         break;
       }
       case LLVMGetElementPtr: {
@@ -891,6 +893,12 @@ struct FunCloner {
         Dst = LLVMBuildFreeze(Builder, Arg, Name);
         break;
       }
+      case LLVMFence: {
+        LLVMAtomicOrdering Ordering = LLVMGetOrdering(Src);
+        LLVMBool IsSingleThreaded = LLVMIsAtomicSingleThread(Src);
+        Dst = LLVMBuildFence(Builder, Ordering, IsSingleThreaded, Name);
+        break;
+      }
       default:
         break;
     }

Benjins · 2023-09-30T14:47:10Z

Should I merge again from main, or is this good to land?

nikic · 2023-09-30T14:48:38Z

Should be good to land.

Benjins · 2023-09-30T14:50:19Z

Is that something you can do, or should I reach out to someone else? I don't have access rights to merge it

nikic · 2023-09-30T14:53:04Z

Oh, I thought you have commit access because you are marked as "Member". Apparently that also includes people in bz-contributors that do not have commit access as well.

Local branch amd-gfx 89049b0 Merged main:dc1dc60ea878 into amd-gfx:b14bde5d0245 Remote branch main d222c5e [C API] Fix LLVMGetOrdering/LLVMIsAtomicSingleThread for fence/memory instrs (llvm#65228)

Benjins requested a review from a team as a code owner September 3, 2023 12:48

Benjins mentioned this pull request Sep 3, 2023

[C API] GetOrdering/IsAtomicSingleThread do not work for fence/load/store #65227

Closed

Benjins force-pushed the bns-store-atomics-memory-order-c-api branch from df4d141 to b20e1a3 Compare September 3, 2023 14:13

cdisselkoen approved these changes Sep 6, 2023

View reviewed changes

nikic reviewed Sep 30, 2023

View reviewed changes

Add case for AtomicRMWInst in LLVMSetOrdering

1d3c77a

llvmbot added the llvm:ir label Sep 30, 2023

nikic approved these changes Sep 30, 2023

View reviewed changes

nikic merged commit d222c5e into llvm:main Sep 30, 2023
2 of 4 checks passed

Benjins deleted the bns-store-atomics-memory-order-c-api branch September 30, 2023 14:53

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[C API] Fix LLVMGetOrdering/LLVMIsAtomicSingleThread for fence/memory instrs #65228

[C API] Fix LLVMGetOrdering/LLVMIsAtomicSingleThread for fence/memory instrs #65228

Benjins commented Sep 3, 2023 •

edited

Loading

Benjins commented Sep 3, 2023

cdisselkoen left a comment

Benjins commented Sep 30, 2023

nikic Sep 30, 2023

Benjins Sep 30, 2023 •

edited

Loading

llvmbot commented Sep 30, 2023

Benjins commented Sep 30, 2023

nikic commented Sep 30, 2023

Benjins commented Sep 30, 2023

nikic commented Sep 30, 2023

[C API] Fix LLVMGetOrdering/LLVMIsAtomicSingleThread for fence/memory instrs #65228

[C API] Fix LLVMGetOrdering/LLVMIsAtomicSingleThread for fence/memory instrs #65228

Conversation

Benjins commented Sep 3, 2023 • edited Loading

Benjins commented Sep 3, 2023

cdisselkoen left a comment

Choose a reason for hiding this comment

Benjins commented Sep 30, 2023

nikic Sep 30, 2023

Choose a reason for hiding this comment

Benjins Sep 30, 2023 • edited Loading

Choose a reason for hiding this comment

llvmbot commented Sep 30, 2023

Benjins commented Sep 30, 2023

nikic commented Sep 30, 2023

Benjins commented Sep 30, 2023

nikic commented Sep 30, 2023

Benjins commented Sep 3, 2023 •

edited

Loading

Benjins Sep 30, 2023 •

edited

Loading