[flang-commits] [flang] [mlir] [Flang] [OpenMP] atomic compare (PR #184761)

[Tom Eccles via flang-commits]

================
@@ -2452,6 +2452,95 @@ llvm.func @omp_atomic_capture_misc(
 
 // -----
 
+// CHECK-LABEL: @omp_atomic_compare
+// CHECK-SAME: (ptr %[[X:.*]], i32 %[[E:.*]], i32 %[[D:.*]], ptr %[[XF:.*]], float %[[EF:.*]], float %[[DF:.*]], ptr %[[XC:.*]], { float, float } %[[EC:.*]], { float, float } %[[DC:.*]])
+llvm.func @omp_atomic_compare(
+  %x : !llvm.ptr, %e : i32, %d : i32,
+  %xf : !llvm.ptr, %ef : f32, %df : f32,
+  %xc : !llvm.ptr, %ec : !llvm.struct<(f32, f32)>, %dc : !llvm.struct<(f32, f32)>) {
+
+  // Integer equality  →  cmpxchg
+  // CHECK: cmpxchg ptr %[[X]], i32 %[[E]], i32 %[[D]] monotonic monotonic
+  omp.atomic.compare %x : !llvm.ptr {
+  ^bb0(%xval : i32):
+    %cmp0 = llvm.icmp "eq" %xval, %e : i32
+    %sel0 = llvm.select %cmp0, %d, %xval : i1, i32
+    omp.yield(%sel0 : i32)
+  }
+
+  // Float equality  →  bitcast + cmpxchg
+  // CHECK: %[[EBC:.*]] = bitcast float %[[EF]] to i32
+  // CHECK: %[[DBC:.*]] = bitcast float %[[DF]] to i32
+  // CHECK: cmpxchg ptr %[[XF]], i32 %[[EBC]], i32 %[[DBC]] monotonic monotonic
----------------
tblah wrote:

I can't think of a better alternative than a CAS loop for a correct result. However it looks like clang takes the same approach as you and does a bitwise comparison: [godbolt](https://godbolt.org/#g:!((g:!((g:!((h:codeEditor,i:(filename:'1',fontScale:14,fontUsePx:'0',j:1,lang:c%2B%2B,selection:(endColumn:1,endLineNumber:6,positionColumn:1,positionLineNumber:6,selectionStartColumn:1,selectionStartLineNumber:6,startColumn:1,startLineNumber:6),source:'void+f(float+*x,+float+e,+float+d)+%7B%0A%23pragma+omp+atomic+compare%0A++if+(*x+%3D%3D+e)%0A++++*x+%3D+d%3B%0A%7D%0A%0A'),l:'5',n:'0',o:'C%2B%2B+source+%231',t:'0')),k:51.367169376964796,l:'4',n:'0',o:'',s:0,t:'0'),(g:!((g:!((h:compiler,i:(compiler:armv8-clang-trunk,filters:(b:'0',binary:'1',binaryObject:'1',commentOnly:'0',debugCalls:'1',demangle:'1',directives:'0',execute:'1',intel:'1',libraryCode:'0',trim:'1',verboseDemangling:'0'),flagsViewOpen:'1',fontScale:14,fontUsePx:'0',j:2,lang:c%2B%2B,libs:!(),options:'-fopenmp+-emit-llvm+-S+-O3',overrides:!(),selection:(endColumn:1,endLineNumber:10,positionColumn:1,positionLineNumber:10,selectionStartColumn:1,selectionStartLineNumber:10,startColumn:1,startLineNumber:10),source:1),l:'5',n:'0',o:'+armv8-a+clang+(trunk)+(Editor+%231)',t:'0')),header:(),k:48.632830623035204,l:'4',m:80.9840425531915,n:'0',o:'',s:0,t:'0'),(g:!((h:output,i:(compilerName:'armv8-a+clang+(all+architectural+features,+trunk)',editorid:1,fontScale:14,fontUsePx:'0',j:2,wrap:'1'),l:'5',n:'0',o:'Output+of+armv8-a+clang+(trunk)+(Compiler+%232)',t:'0')),header:(),l:'4',m:19.015957446808507,n:'0',o:'',s:0,t:'0')),k:48.632830623035204,l:'3',n:'0',o:'',t:'0')),l:'2',n:'0',o:'',t:'0')),version:4)

@Meinersbur do you have an opinion here? Should we copy clang or preserve correctness? 

https://github.com/llvm/llvm-project/pull/184761