[Mlir-commits] [mlir] [MLIR][NVVM][NVGPU] Combine prefetch and prefetch.tensormap (PR #153134)

[Giacomo Castiglioni]

================
@@ -582,10 +582,10 @@ func.func @elect_one_leader_sync() {
 
 // CHECK-LABEL: @init_mbarrier_arrive_expect_tx
 llvm.func @init_mbarrier_arrive_expect_tx(%desc : !llvm.ptr, %pred : i1) {
-  //CHECK: llvm.inline_asm has_side_effects asm_dialect = att "prefetch.tensormap [$0];", "l"
-  nvvm.prefetch.tensormap %desc : !llvm.ptr
+  //CHECK: nvvm.prefetch tensormap, %{{.*}}
----------------
castigli wrote:

I am trying to fix the nvdsl [example](https://github.com/castigli/llvm-project/commit/bc56b2ae9ea8dcc9931846d998de5815b7105955) which got a bit out of sync not being tested in the CI.
I think this PR introduced a bug for Ch4 and Ch5: `nvvm.prefetch tensormap` doesn't get lowered.
I am missing something or shouldn't this op lower to `llvm.inline_asm has_side_effects asm_dialect = att "prefetch.tensormap [$0];", "l"` as the old check shows?

https://github.com/llvm/llvm-project/pull/153134