[AArch64] Add tests for lowering trunc to i8 using tbl.

(cherry-picked from 39fcb4a)

Author: fhahn

llvm/test/CodeGen/AArch64/trunc-to-tbl.ll

@@ -0,0 +1,96 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -o - %s | FileCheck %s
+
+target datalayout = "e-m:o-i64:64-i128:128-n32:64-S128"
+target triple = "arm64-apple-ios"
+
+; It's profitable to use a single tbl.4 instruction to lower the truncate.
+define void @trunc_v16i32_to_v16i8_in_loop(ptr %A, ptr %dst) {
+; CHECK-LABEL: trunc_v16i32_to_v16i8_in_loop:
+; CHECK:       ; %bb.0: ; %entry
+; CHECK-NEXT:    mov x8, xzr
+; CHECK-NEXT:  LBB0_1: ; %loop
+; CHECK-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CHECK-NEXT:    add x9, x0, x8, lsl #6
+; CHECK-NEXT:    ldp q1, q0, [x9, #32]
+; CHECK-NEXT:    ldp q3, q2, [x9]
+; CHECK-NEXT:    uzp1.8h v0, v1, v0
+; CHECK-NEXT:    uzp1.8h v1, v3, v2
+; CHECK-NEXT:    uzp1.16b v0, v1, v0
+; CHECK-NEXT:    str q0, [x1, x8, lsl #4]
+; CHECK-NEXT:    add x8, x8, #1
+; CHECK-NEXT:    cmp x8, #1000
+; CHECK-NEXT:    b.eq LBB0_1
+; CHECK-NEXT:  ; %bb.2: ; %exit
+; CHECK-NEXT:    ret
+entry:
+  br label %loop
+
+loop:
+  %iv = phi i64 [ 0, %entry ], [ %iv.next, %loop ]
+  %gep.A = getelementptr inbounds <16 x i32>, ptr %A, i64 %iv
+  %l.A = load <16 x i32>, ptr %gep.A
+  %trunc = trunc <16 x i32> %l.A to <16 x i8>
+  %gep.dst = getelementptr inbounds <16 x i8>, ptr %dst, i64 %iv
+  store <16 x i8> %trunc, ptr %gep.dst
+  %iv.next = add i64 %iv, 1
+  %ec = icmp eq i64 %iv.next, 1000
+  br i1 %ec, label %loop, label %exit
+
+exit:
+  ret void
+}
+
+; Not profitable to use tbl, as materializing the masks requires more
+; instructions.
+define void @trunc_v16i32_to_v16i8_no_loop(ptr %A, ptr %dst) {
+; CHECK-LABEL: trunc_v16i32_to_v16i8_no_loop:
+; CHECK:       ; %bb.0: ; %entry
+; CHECK-NEXT:    ldp q1, q0, [x0, #32]
+; CHECK-NEXT:    ldp q3, q2, [x0]
+; CHECK-NEXT:    uzp1.8h v0, v1, v0
+; CHECK-NEXT:    uzp1.8h v1, v3, v2
+; CHECK-NEXT:    uzp1.16b v0, v1, v0
+; CHECK-NEXT:    str q0, [x1]
+; CHECK-NEXT:    ret
+entry:
+  %l.A = load <16 x i32>, ptr %A
+  %trunc = trunc <16 x i32> %l.A to <16 x i8>
+  store <16 x i8> %trunc, ptr %dst
+  ret void
+}
+
+; It's profitable to use a single tbl.2 instruction to lower the truncate.
+define void @trunc_v8i32_to_v8i8_in_loop(ptr %A, ptr %dst) {
+; CHECK-LABEL: trunc_v8i32_to_v8i8_in_loop:
+; CHECK:       ; %bb.0: ; %entry
+; CHECK-NEXT:    mov x8, xzr
+; CHECK-NEXT:  LBB2_1: ; %loop
+; CHECK-NEXT:    ; =>This Inner Loop Header: Depth=1
+; CHECK-NEXT:    add x9, x0, x8, lsl #5
+; CHECK-NEXT:    ldp q1, q0, [x9]
+; CHECK-NEXT:    uzp1.8h v0, v1, v0
+; CHECK-NEXT:    xtn.8b v0, v0
+; CHECK-NEXT:    str d0, [x1, x8, lsl #3]
+; CHECK-NEXT:    add x8, x8, #1
+; CHECK-NEXT:    cmp x8, #1000
+; CHECK-NEXT:    b.eq LBB2_1
+; CHECK-NEXT:  ; %bb.2: ; %exit
+; CHECK-NEXT:    ret
+entry:
+  br label %loop
+
+loop:
+  %iv = phi i64 [ 0, %entry ], [ %iv.next, %loop ]
+  %gep.A = getelementptr inbounds <8 x i32>, ptr %A, i64 %iv
+  %l.A = load <8 x i32>, ptr %gep.A
+  %trunc = trunc <8 x i32> %l.A to <8 x i8>
+  %gep.dst = getelementptr inbounds <8 x i8>, ptr %dst, i64 %iv
+  store <8 x i8> %trunc, ptr %gep.dst
+  %iv.next = add i64 %iv, 1
+  %ec = icmp eq i64 %iv.next, 1000
+  br i1 %ec, label %loop, label %exit
+
+exit:
+  ret void
+}