[llvm] [AArch64][Machine-Combiner] Split loads into lanes of neon vectors into multiple vectors when possible (PR #142941)

[Jonathan Cohen via llvm-commits]

================
@@ -0,0 +1,268 @@
+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py UTC_ARGS: --version 5
+# RUN: llc -run-pass=machine-combiner -mcpu=neoverse-n2 -mtriple=aarch64-none-linux-gnu -verify-machineinstrs %s -o - | FileCheck %s
+
+---
+name:            split_loads_to_fpr128
+body:             |
+  bb.0.entry:
+    liveins: $x0, $x1, $x2, $x3, $x4
+
+    ; CHECK-LABEL: name: split_loads_to_fpr128
+    ; CHECK: [[COPY:%[0-9]+]]:gpr64common = COPY $x0
+    ; CHECK-NEXT: [[COPY1:%[0-9]+]]:gpr64common = COPY $x1
+    ; CHECK-NEXT: [[COPY2:%[0-9]+]]:gpr64common = COPY $x2
+    ; CHECK-NEXT: [[COPY3:%[0-9]+]]:gpr64common = COPY $x3
+    ; CHECK-NEXT: [[COPY4:%[0-9]+]]:gpr64common = COPY $x4
+    ; CHECK-NEXT: [[LD_i32:%[0-9]+]]:fpr32 = LDRSroX [[COPY]], killed [[COPY1]], 0, 1
+    ; CHECK-NEXT: [[FIRST_REG:%[0-9]+]]:fpr128 = SUBREG_TO_REG 0, killed [[LD_i32]], %subreg.ssub
+    ; CHECK-NEXT: [[LD0_1:%[0-9]+]]:fpr128 = LD1i32 [[FIRST_REG]], 1, killed [[COPY2]] 
+    ; CHECK-NEXT: [[LD1_0:%[0-9]+]]:fpr32 = LDRSui [[COPY3]], 0
+    ; CHECK-NEXT: [[SECOND_REG:%[0-9]+]]:fpr128 = SUBREG_TO_REG 0, killed [[LD1_0]], %subreg.ssub
+    ; CHECK-NEXT: [[LD1_1:%[0-9]+]]:fpr128 = LD1i32 [[SECOND_REG]], 1, killed [[COPY4]]
+    ; CHECK-NEXT: [[ZIP:%[0-9]+]]:fpr128 = ZIP1v2i64 [[LD0_1]], [[LD1_1]]
+    ; CHECK-NEXT: $q0 = COPY [[ZIP]]
+    ; CHECK-NEXT: RET_ReallyLR implicit $q0
+    %0:gpr64common = COPY $x0
+    %1:gpr64common = COPY $x1
+    %2:gpr64common = COPY $x2
+    %3:gpr64common = COPY $x3
+    %4:gpr64common = COPY $x4
+    %5:fpr32 = LDRSroX %0, killed %1, 0, 1
+    %6:fpr128 = SUBREG_TO_REG 0, killed %5, %subreg.ssub
----------------
jcohen-apple wrote:

There was a negative test at the bottom of the file (for not loading all lanes)
Added the other tests you suggested.
For the out of order lanes - I added that as a positive test case, and fixed the pattern.
Added an extra negative test case where the load to lanes have complex addressing mode (which I don't cover)

https://github.com/llvm/llvm-project/pull/142941