[llvm] [AArch64][ARM] Optimize more `tbl`/`tbx` calls into `shufflevector` (PR #169748)
David Green via llvm-commits
llvm-commits at lists.llvm.org
Thu Dec 4 10:11:24 PST 2025
================
@@ -0,0 +1,269 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt < %s -passes=instcombine -S | FileCheck %s
+
+target datalayout = "e-m:e-i8:8:32-i16:16:32-i64:64-i128:128-n32:64-S128"
+target triple = "aarch64"
+
+; We can turn a tbl/tbx intrinsic into a shufflevector instruction if the mask
+; is constant and references 2 or fewer operands.
+
+; Basic tbl1 with all in-bounds indices should optimize to shufflevector.
+define <16 x i8> @tbl1_basic(<16 x i8> %a) {
+; CHECK-LABEL: @tbl1_basic(
+; CHECK-NEXT: [[TMP1:%.*]] = shufflevector <16 x i8> [[A:%.*]], <16 x i8> poison, <16 x i32> <i32 15, i32 14, i32 13, i32 12, i32 11, i32 10, i32 9, i32 8, i32 7, i32 6, i32 5, i32 4, i32 3, i32 2, i32 1, i32 0>
+; CHECK-NEXT: ret <16 x i8> [[TMP1]]
+;
+ %tbl = call <16 x i8> @llvm.aarch64.neon.tbl1.v16i8(<16 x i8> %a, <16 x i8> <i8 15, i8 14, i8 13, i8 12, i8 11, i8 10, i8 9, i8 8, i8 7, i8 6, i8 5, i8 4, i8 3, i8 2, i8 1, i8 0>)
+ ret <16 x i8> %tbl
+}
+
+; tbl2 with both operands the same should optimize (1 unique source).
+define <16 x i8> @tbl2_duplicate_operands(<16 x i8> %a) {
+; CHECK-LABEL: @tbl2_duplicate_operands(
+; CHECK-NEXT: [[TMP1:%.*]] = shufflevector <16 x i8> [[A:%.*]], <16 x i8> poison, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
+; CHECK-NEXT: ret <16 x i8> [[TMP1]]
+;
+ %tbl = call <16 x i8> @llvm.aarch64.neon.tbl2.v16i8(<16 x i8> %a, <16 x i8> %a, <16 x i8> <i8 0, i8 1, i8 2, i8 3, i8 4, i8 5, i8 6, i8 7, i8 16, i8 17, i8 18, i8 19, i8 20, i8 21, i8 22, i8 23>)
+ ret <16 x i8> %tbl
+}
+
+; tbl4 with alternating duplicate operands should optimize (2 unique sources).
+define <16 x i8> @tbl4_duplicate_operands(<16 x i8> %a, <16 x i8> %b) {
+; CHECK-LABEL: @tbl4_duplicate_operands(
+; CHECK-NEXT: [[TMP1:%.*]] = shufflevector <16 x i8> [[A:%.*]], <16 x i8> [[B:%.*]], <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 16, i32 17, i32 18, i32 19, i32 0, i32 1, i32 2, i32 3, i32 16, i32 17, i32 18, i32 19>
+; CHECK-NEXT: ret <16 x i8> [[TMP1]]
+;
+ %tbl = call <16 x i8> @llvm.aarch64.neon.tbl4.v16i8(<16 x i8> %a, <16 x i8> %b, <16 x i8> %a, <16 x i8> %b, <16 x i8> <i8 0, i8 1, i8 2, i8 3, i8 16, i8 17, i8 18, i8 19, i8 32, i8 33, i8 34, i8 35, i8 48, i8 49, i8 50, i8 51>)
+ ret <16 x i8> %tbl
+}
+
+; tbl4 where mask only references first two operands should optimize.
+define <16 x i8> @tbl4_unused_operands(<16 x i8> %a, <16 x i8> %b, <16 x i8> %c, <16 x i8> %d) {
+; CHECK-LABEL: @tbl4_unused_operands(
+; CHECK-NEXT: [[TMP1:%.*]] = shufflevector <16 x i8> [[A:%.*]], <16 x i8> [[B:%.*]], <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23>
+; CHECK-NEXT: ret <16 x i8> [[TMP1]]
+;
+ %tbl = call <16 x i8> @llvm.aarch64.neon.tbl4.v16i8(<16 x i8> %a, <16 x i8> %b, <16 x i8> %c, <16 x i8> %d, <16 x i8> <i8 0, i8 1, i8 2, i8 3, i8 4, i8 5, i8 6, i8 7, i8 16, i8 17, i8 18, i8 19, i8 20, i8 21, i8 22, i8 23>)
+ ret <16 x i8> %tbl
+}
+
+; tbl4 where mask only references one operand should optimize.
+define <16 x i8> @tbl4_single_operand_used(<16 x i8> %a, <16 x i8> %b, <16 x i8> %c, <16 x i8> %d) {
+; CHECK-LABEL: @tbl4_single_operand_used(
+; CHECK-NEXT: [[TMP1:%.*]] = shufflevector <16 x i8> [[A:%.*]], <16 x i8> poison, <16 x i32> <i32 15, i32 14, i32 13, i32 12, i32 11, i32 10, i32 9, i32 8, i32 7, i32 6, i32 5, i32 4, i32 3, i32 2, i32 1, i32 0>
+; CHECK-NEXT: ret <16 x i8> [[TMP1]]
+;
+ %tbl = call <16 x i8> @llvm.aarch64.neon.tbl4.v16i8(<16 x i8> %a, <16 x i8> %b, <16 x i8> %c, <16 x i8> %d, <16 x i8> <i8 15, i8 14, i8 13, i8 12, i8 11, i8 10, i8 9, i8 8, i8 7, i8 6, i8 5, i8 4, i8 3, i8 2, i8 1, i8 0>)
+ ret <16 x i8> %tbl
+}
+
+; tbl1 with some OOB indices should optimize (1 source + zero vector = 2 sources).
+define <16 x i8> @tbl1_with_oob(<16 x i8> %a) {
+; CHECK-LABEL: @tbl1_with_oob(
+; CHECK-NEXT: [[TMP1:%.*]] = shufflevector <16 x i8> [[A:%.*]], <16 x i8> <i8 0, i8 poison, i8 poison, i8 poison, i8 poison, i8 poison, i8 poison, i8 poison, i8 poison, i8 poison, i8 poison, i8 poison, i8 poison, i8 poison, i8 poison, i8 poison>, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 16, i32 16, i32 16, i32 16, i32 16, i32 16, i32 16, i32 16>
+; CHECK-NEXT: ret <16 x i8> [[TMP1]]
+;
+ %tbl = call <16 x i8> @llvm.aarch64.neon.tbl1.v16i8(<16 x i8> %a, <16 x i8> <i8 0, i8 1, i8 2, i8 3, i8 4, i8 5, i8 6, i8 7, i8 16, i8 17, i8 18, i8 19, i8 20, i8 21, i8 22, i8 23>)
+ ret <16 x i8> %tbl
+}
+
+; tbl2 with duplicate operands and OOB should optimize (1 unique source + zero vector = 2 sources).
+define <16 x i8> @tbl2_duplicate_with_oob(<16 x i8> %a) {
+; CHECK-LABEL: @tbl2_duplicate_with_oob(
+; CHECK-NEXT: [[TMP1:%.*]] = shufflevector <16 x i8> [[A:%.*]], <16 x i8> <i8 0, i8 poison, i8 poison, i8 poison, i8 poison, i8 poison, i8 poison, i8 poison, i8 poison, i8 poison, i8 poison, i8 poison, i8 poison, i8 poison, i8 poison, i8 poison>, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 16, i32 16, i32 16, i32 16, i32 16, i32 16, i32 16, i32 16>
+; CHECK-NEXT: ret <16 x i8> [[TMP1]]
+;
+ %tbl = call <16 x i8> @llvm.aarch64.neon.tbl2.v16i8(<16 x i8> %a, <16 x i8> %a, <16 x i8> <i8 0, i8 1, i8 2, i8 3, i8 16, i8 17, i8 18, i8 19, i8 99, i8 99, i8 99, i8 99, i8 99, i8 99, i8 99, i8 99>)
+ ret <16 x i8> %tbl
+}
+
+; tbl2 with OOB indices should NOT optimize (2 sources + zero vector = 3 sources).
+define <16 x i8> @tbl2_with_oob_bail(<16 x i8> %a, <16 x i8> %b) {
+; CHECK-LABEL: @tbl2_with_oob_bail(
+; CHECK-NEXT: [[TBL:%.*]] = call <16 x i8> @llvm.aarch64.neon.tbl2.v16i8(<16 x i8> [[A:%.*]], <16 x i8> [[B:%.*]], <16 x i8> <i8 0, i8 1, i8 2, i8 3, i8 16, i8 17, i8 18, i8 19, i8 99, i8 99, i8 99, i8 99, i8 99, i8 99, i8 99, i8 99>)
+; CHECK-NEXT: ret <16 x i8> [[TBL]]
+;
+ %tbl = call <16 x i8> @llvm.aarch64.neon.tbl2.v16i8(<16 x i8> %a, <16 x i8> %b, <16 x i8> <i8 0, i8 1, i8 2, i8 3, i8 16, i8 17, i8 18, i8 19, i8 99, i8 99, i8 99, i8 99, i8 99, i8 99, i8 99, i8 99>)
+ ret <16 x i8> %tbl
+}
+
+; tbl1 with all OOB indices should optimize to zero vector.
+define <16 x i8> @tbl1_all_oob(<16 x i8> %a) {
+; CHECK-LABEL: @tbl1_all_oob(
+; CHECK-NEXT: ret <16 x i8> zeroinitializer
+;
+ %tbl = call <16 x i8> @llvm.aarch64.neon.tbl1.v16i8(<16 x i8> %a, <16 x i8> <i8 99, i8 99, i8 99, i8 99, i8 99, i8 99, i8 99, i8 99, i8 99, i8 99, i8 99, i8 99, i8 99, i8 99, i8 99, i8 99>)
+ ret <16 x i8> %tbl
+}
+
+; tbl3 referencing all 3 operands should NOT optimize.
+define <16 x i8> @tbl3_three_sources_bail(<16 x i8> %a, <16 x i8> %b, <16 x i8> %c) {
+; CHECK-LABEL: @tbl3_three_sources_bail(
+; CHECK-NEXT: [[TBL:%.*]] = call <16 x i8> @llvm.aarch64.neon.tbl3.v16i8(<16 x i8> [[A:%.*]], <16 x i8> [[B:%.*]], <16 x i8> [[C:%.*]], <16 x i8> <i8 0, i8 1, i8 2, i8 3, i8 16, i8 17, i8 18, i8 19, i8 32, i8 33, i8 34, i8 35, i8 0, i8 0, i8 0, i8 0>)
+; CHECK-NEXT: ret <16 x i8> [[TBL]]
+;
+ %tbl = call <16 x i8> @llvm.aarch64.neon.tbl3.v16i8(<16 x i8> %a, <16 x i8> %b, <16 x i8> %c, <16 x i8> <i8 0, i8 1, i8 2, i8 3, i8 16, i8 17, i8 18, i8 19, i8 32, i8 33, i8 34, i8 35, i8 0, i8 0, i8 0, i8 0>)
----------------
davemgreen wrote:
Add a tbl3 that does optimize?
https://github.com/llvm/llvm-project/pull/169748
More information about the llvm-commits
mailing list