[Mlir-commits] [mlir] [mlir][ArmSME] Add rudimentary support for tile spills to the stack (PR #76086)

[Benjamin Maxwell]

================
@@ -0,0 +1,77 @@
+// RUN: mlir-opt %s \
+// RUN:   -convert-vector-to-arm-sme -allocate-arm-sme-tiles  \
+// RUN:   -convert-vector-to-arm-sme -convert-arm-sme-to-scf \
+// RUN:   -enable-arm-streaming="streaming-mode=streaming-locally za-mode=new-za only-if-required-by-ops"  \
+// RUN:   -convert-vector-to-scf -cse -arm-sve-legalize-vector-storage \
+// RUN:   -convert-arm-sme-to-llvm -convert-vector-to-llvm=enable-arm-sve -cse \
+// RUN:   -canonicalize -test-lower-to-llvm -verify-diagnostics | \
+// RUN: %mcr_aarch64_cmd \
+// RUN:   -e=main -entry-point-result=void \
+// RUN:   -march=aarch64 -mattr="+sve,+sme" \
+// RUN:   -shared-libs=%mlir_runner_utils,%mlir_c_runner_utils,%arm_sme_abi_shlib | \
+// RUN: FileCheck %s
+
+/// This function uses too many tiles! There's only two i16 tiles (ZA0.H and
+/// ZA1.H), but this function uses five i16 tiles! Very expensive spills/reloads
+/// will be inserted to emulate the extra three tiles. Note: This is only done
+/// to avoid the compiler erroring out but is expected to have very poor
+/// performance (hence the warning).
+func.func @use_too_many_tiles(%a: memref<?x?xi16>, %b:  memref<?x?xi16>, %c: memref<?x?xi16>) {
+  %c0 = arith.constant 0 : index
+  %tile_a = arith.constant dense<0> : vector<[8]x[8]xi16>
+  %tile_b = arith.constant dense<1> : vector<[8]x[8]xi16>
+  // expected-warning @below {{failed to allocate SME virtual tile to operation, all tile operations will go through memory, expect degraded performance}}
+  %tile_c = arm_sme.tile_load %a[%c0, %c0] : memref<?x?xi16>, vector<[8]x[8]xi16>
+  // expected-warning @below {{failed to allocate SME virtual tile to operation, all tile operations will go through memory, expect degraded performance}}
+  %tile_d = arm_sme.tile_load %b[%c0, %c0] : memref<?x?xi16>, vector<[8]x[8]xi16>
+  // expected-warning @below {{failed to allocate SME virtual tile to operation, all tile operations will go through memory, expect degraded performance}}
+  %tile_e = arm_sme.tile_load %c[%c0, %c0] : memref<?x?xi16>, vector<[8]x[8]xi16>
+
+  // CHECK-LABEL: tile_a:
+  // CHECK-COUNT-8: ( 0, 0, 0, 0, 0, 0, 0, 0
+  vector.print str "tile_a:"
+  vector.print %tile_a : vector<[8]x[8]xi16>
+  // CHECK-LABEL: tile_b:
+  // CHECK-COUNT-8: ( 1, 1, 1, 1, 1, 1, 1, 1
+  vector.print str "tile_b:"
+  vector.print %tile_b : vector<[8]x[8]xi16>
+  // CHECK-LABEL: tile_c:
+  // CHECK-COUNT-8: ( 2, 2, 2, 2, 2, 2, 2, 2
+  vector.print str "tile_c:"
+  vector.print %tile_c : vector<[8]x[8]xi16>
+  // CHECK-LABEL: tile_d:
+  // CHECK-COUNT-8: ( 3, 3, 3, 3, 3, 3, 3, 3
+  vector.print str "tile_d:"
+  vector.print %tile_d : vector<[8]x[8]xi16>
+  // CHECK-LABEL: tile_e:
+  // CHECK-COUNT-8: ( 4, 4, 4, 4, 4, 4, 4, 4
+  vector.print str "tile_e:"
+  vector.print %tile_e : vector<[8]x[8]xi16>
+  return
+}
+
+func.func @get_svl() -> index  attributes { enable_arm_streaming_ignore, arm_locally_streaming }{
----------------
MacDue wrote:

I know, I'm just being explicit as I've manually set the streaming mode to ensure this returns SVL.

https://github.com/llvm/llvm-project/pull/76086