[llvm] [AMDGPU][SDAG] Legalise v2i32 or/xor/and instructions to make use of 64-bit wide instructions (PR #140694)

[via llvm-commits]

================
@@ -63,9 +63,11 @@ define i64 @shl_metadata_cant_be_narrowed_to_i32(i64 %arg0, ptr %arg1.ptr) {
 ; CHECK-LABEL: shl_metadata_cant_be_narrowed_to_i32:
 ; CHECK:       ; %bb.0:
 ; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; CHECK-NEXT:    flat_load_dword v2, v[2:3]
+; CHECK-NEXT:    flat_load_dword v1, v[2:3]
 ; CHECK-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; CHECK-NEXT:    v_lshlrev_b64 v[0:1], v2, v[0:1]
+; CHECK-NEXT:    v_and_b32_e32 v1, 7, v1
+; CHECK-NEXT:    v_lshlrev_b32_e32 v1, v1, v0
+; CHECK-NEXT:    v_mov_b32_e32 v0, 0
----------------
LU-JOHN wrote:

The metadata range in this testcase has a Known.getMinValue().getZExtValue() of 0 and thus the reduction of the 64-bit shift to 32-bit should not be done.

The test in llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp line 4112 blocks this transformation in the old version of this code.


https://github.com/llvm/llvm-project/pull/140694