[AMD] Use more detailed convert_layout.mlir test (#7554)

Add more detailed checks in general swizzling part of convert_layout test;
Replace canonicalize with cse, it should be more robust, all this test
needs is to merge constants for easier pattern matching.

Co-authored-by: Alexander Efimov <[email protected]>

Author: alefimov-amd

test/Conversion/amd/convert_layout.mlir

@@ -1,14 +1,25 @@
-// RUN: triton-opt %s -split-input-file --allocate-shared-memory --convert-triton-amdgpu-to-llvm=arch=gfx942 --canonicalize| FileCheck %s
+// RUN: triton-opt %s -split-input-file --allocate-shared-memory --convert-triton-amdgpu-to-llvm=arch=gfx942 --cse| FileCheck %s
 
 #blocked0 = #ttg.blocked<{sizePerThread = [1, 4], threadsPerWarp = [16, 4], warpsPerCTA = [2, 2], order = [1, 0], CTAsPerCGA = [1, 1], CTASplitNum = [1, 1], CTAOrder = [1, 0]}>
 #blocked1 = #ttg.blocked<{sizePerThread = [4, 1], threadsPerWarp = [4, 16], warpsPerCTA = [2, 2], order = [0, 1], CTAsPerCGA = [1, 1], CTASplitNum = [1, 1], CTAOrder = [1, 0]}>
 module attributes {"ttg.num-ctas" = 1 : i32, "ttg.num-warps" = 4 : i32, ttg.target = "hip:gfx942", "ttg.threads-per-warp" = 64 : i32} {
   // CHECK: llvm.mlir.global external @global_smem
   tt.func @convert_layout_general_swizzling(%arg0: tensor<64x64xf32, #blocked0>, %arg1: tensor<64x64x!tt.ptr<f32>, #blocked1>) {
 
-    // verify that following convert layout uses general swizzling
+    // verify that following convert layout uses general swizzling path
 
-    // CHECK-NOT: llvm.lshr
+    // CHECK: [[CST_128:%.*]] = llvm.mlir.constant(128 : i32) : i32
+
+    // Part of offset computation generated by applyLinearLayout function
+    // CHECK: [[SEL:%.*]]= llvm.select {{.*}}, {{.*}}, [[CST_128]]
+    // CHECK: [[OFFSET_0:%.*]] = llvm.xor {{.*}}, [[SEL]]
+    // CHECK: [[OFFSET_1:%.*]] = llvm.xor {{.*}}, [[OFFSET_0]] : i32
+
+    // Part of offset computation generated by lowerLdSt function after applyLinearLayout
+    // CHECK: [[OFFSET_2:%.*]] = llvm.xor [[OFFSET_1]], {{.*}} : i32
+    // CHECK: [[OFFSET_3:%.*]] = llvm.xor [[OFFSET_2]], {{.*}} : i32
+    // CHECK: [[OFFSET_4:%.*]] = llvm.add [[OFFSET_3]], {{.*}} : i32
+    // CHECK: llvm.getelementptr inbounds {{.*}}{{\[}}[[OFFSET_4]]{{\]}}
 
     %0 = ttg.convert_layout %arg0 : tensor<64x64xf32, #blocked0> -> tensor<64x64xf32, #blocked1>
     tt.store %arg1, %0 : tensor<64x64x!tt.ptr<f32>, #blocked1>