[BACKEND] Make ExternElementwise op implement ConditionallySpeculatable (#5079)

ExternElementwise ops have a `pure` attribute that marks the op as pure.
If an op is pure, it should also be speculatable. In the reduction/scan
ttgir->llvm passes, checks for speculatability are failing for
ExternElementwise ops, causing additional conditional handling to be
added.

This PR makes ExternElementwise ops implement ConditionallySpeculatable,
and mark the op as speculatable if the op is marked as pure. This
removes the conditional branches from the generated scan/reduction code.

Author: davidberard98

include/triton/Dialect/Triton/IR/TritonOps.td

@@ -778,7 +778,8 @@ def TT_ScanReturnOp: TT_Op<"scan.return",
 def TT_ExternElementwiseOp : TT_Op<"extern_elementwise", [Elementwise,
                                                             SameOperandsAndResultEncoding,
                                                             SameVariadicOperandSize,
-                                                            DeclareOpInterfaceMethods<MemoryEffectsOpInterface>]> {
+                                                            DeclareOpInterfaceMethods<MemoryEffectsOpInterface>,
+                                                            ConditionallySpeculatable]> {
 
     let description = [{
         call an external function $symbol implemented in $libpath/$libname with $args
@@ -790,6 +791,12 @@ def TT_ExternElementwiseOp : TT_Op<"extern_elementwise", [Elementwise,
     let results = (outs TT_Type:$result);
 
     let assemblyFormat = "operands attr-dict `:` functional-type(operands, $result)";
+
+    let extraClassDeclaration = [{
+      // Interface method for ConditionallySpeculatable.
+      Speculation::Speculatability getSpeculatability();
+    }];
+
 }
 
 //

lib/Dialect/Triton/IR/Ops.cpp

@@ -1039,6 +1039,12 @@ void ExternElementwiseOp::getEffects(
                        SideEffects::DefaultResource::get());
 }
 
+Speculation::Speculatability ExternElementwiseOp::getSpeculatability() {
+  if (getPure())
+    return Speculation::Speculatable;
+  return Speculation::NotSpeculatable;
+}
+
 // -- ExperimentalTensormapCreateOp --
 LogicalResult ExperimentalTensormapCreateOp::verify() {
   auto rank = getBoxDim().size();

test/Conversion/tritongpu_to_llvm.mlir

@@ -1855,3 +1855,21 @@ module attributes {"triton_gpu.num-ctas" = 1 : i32, "triton_gpu.num-warps" = 4 :
 #loc3 = loc("inner_call":29:28)
 #loc4 = loc(callsite(#loc3 at #loc1))
 #loc5 = loc(callsite(#loc4 at #loc2))
+
+// -----
+
+#blocked = #triton_gpu.blocked<{sizePerThread = [1, 1], threadsPerWarp = [32, 1], warpsPerCTA = [1, 4], order = [0, 1]}>
+module attributes {"triton_gpu.num-ctas" = 1 : i32, "triton_gpu.num-warps" = 4 : i32, triton_gpu.target = "cuda:90", "triton_gpu.threads-per-warp" = 32 : i32} {
+  tt.func public @log1pf_scan(%39: tensor<32x16xf32, #blocked>) attributes {noinline = false} {
+    // CHECK: log1pf_scan
+    // non-speculatable ops will introduce a cond_br; extern_elementwise with pure = true should be considered speculatable.
+    // CHECK-NOT: llvm.cond_br
+    %40 = "tt.scan"(%39) <{axis = 1 : i32, reverse = false}> ({
+    ^bb0(%arg5: f32, %arg6: f32):
+      %43 = tt.extern_elementwise %arg5 {libname = "", libpath = "", pure = true, symbol = "__nv_log1pf"} : (f32) -> f32
+      %44 = arith.addf %43, %43 : f32
+      tt.scan.return %44 : f32
+    }) : (tensor<32x16xf32, #blocked>) -> tensor<32x16xf32, #blocked>
+    tt.return
+  }
+}