[llvm] [LV] Vectorize conditional scalar assignments (PR #158088)
Graham Hunter via llvm-commits
llvm-commits at lists.llvm.org
Thu Nov 20 05:01:23 PST 2025
================
@@ -0,0 +1,405 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --check-globals none --version 6
+; RUN: opt -passes=loop-vectorize -S < %s 2>&1 | FileCheck %s --check-prefix=NEON
+; RUN: opt -passes=loop-vectorize -mattr=+sve -S < %s 2>&1 | FileCheck %s --check-prefix=SVE
+
+target triple = "aarch64-linux-gnu"
+
+define i32 @simple_csa_int_select(i64 %N, ptr %data, i32 %a) {
+; NEON-LABEL: define i32 @simple_csa_int_select(
+; NEON-SAME: i64 [[N:%.*]], ptr [[DATA:%.*]], i32 [[A:%.*]]) {
+; NEON-NEXT: [[ENTRY:.*]]:
+; NEON-NEXT: br label %[[LOOP:.*]]
+; NEON: [[LOOP]]:
+; NEON-NEXT: [[IV:%.*]] = phi i64 [ 0, %[[ENTRY]] ], [ [[IV_NEXT:%.*]], %[[LOOP]] ]
+; NEON-NEXT: [[DATA_PHI:%.*]] = phi i32 [ -1, %[[ENTRY]] ], [ [[SELECT_DATA:%.*]], %[[LOOP]] ]
+; NEON-NEXT: [[LD_ADDR:%.*]] = getelementptr inbounds i32, ptr [[DATA]], i64 [[IV]]
+; NEON-NEXT: [[LD:%.*]] = load i32, ptr [[LD_ADDR]], align 4
+; NEON-NEXT: [[SELECT_CMP:%.*]] = icmp slt i32 [[A]], [[LD]]
+; NEON-NEXT: [[SELECT_DATA]] = select i1 [[SELECT_CMP]], i32 [[LD]], i32 [[DATA_PHI]]
+; NEON-NEXT: [[IV_NEXT]] = add nuw nsw i64 [[IV]], 1
+; NEON-NEXT: [[EXIT_CMP:%.*]] = icmp eq i64 [[IV_NEXT]], [[N]]
+; NEON-NEXT: br i1 [[EXIT_CMP]], label %[[EXIT:.*]], label %[[LOOP]]
+; NEON: [[EXIT]]:
+; NEON-NEXT: [[SELECT_DATA_LCSSA:%.*]] = phi i32 [ [[SELECT_DATA]], %[[LOOP]] ]
+; NEON-NEXT: ret i32 [[SELECT_DATA_LCSSA]]
+;
+; SVE-LABEL: define i32 @simple_csa_int_select(
+; SVE-SAME: i64 [[N:%.*]], ptr [[DATA:%.*]], i32 [[A:%.*]]) #[[ATTR0:[0-9]+]] {
+; SVE-NEXT: [[ENTRY:.*]]:
+; SVE-NEXT: [[TMP0:%.*]] = call i64 @llvm.vscale.i64()
+; SVE-NEXT: [[TMP1:%.*]] = shl nuw i64 [[TMP0]], 2
+; SVE-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[N]], [[TMP1]]
+; SVE-NEXT: br i1 [[MIN_ITERS_CHECK]], label %[[SCALAR_PH:.*]], label %[[VECTOR_PH:.*]]
+; SVE: [[VECTOR_PH]]:
+; SVE-NEXT: [[TMP2:%.*]] = call i64 @llvm.vscale.i64()
+; SVE-NEXT: [[TMP3:%.*]] = mul nuw i64 [[TMP2]], 4
+; SVE-NEXT: [[N_MOD_VF:%.*]] = urem i64 [[N]], [[TMP3]]
+; SVE-NEXT: [[N_VEC:%.*]] = sub i64 [[N]], [[N_MOD_VF]]
+; SVE-NEXT: [[BROADCAST_SPLATINSERT:%.*]] = insertelement <vscale x 4 x i32> poison, i32 [[A]], i64 0
+; SVE-NEXT: [[BROADCAST_SPLAT:%.*]] = shufflevector <vscale x 4 x i32> [[BROADCAST_SPLATINSERT]], <vscale x 4 x i32> poison, <vscale x 4 x i32> zeroinitializer
+; SVE-NEXT: br label %[[VECTOR_BODY:.*]]
+; SVE: [[VECTOR_BODY]]:
+; SVE-NEXT: [[INDEX:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
+; SVE-NEXT: [[VEC_PHI:%.*]] = phi <vscale x 4 x i32> [ splat (i32 -1), %[[VECTOR_PH]] ], [ [[TMP10:%.*]], %[[VECTOR_BODY]] ]
+; SVE-NEXT: [[TMP4:%.*]] = phi <vscale x 4 x i1> [ zeroinitializer, %[[VECTOR_PH]] ], [ [[TMP9:%.*]], %[[VECTOR_BODY]] ]
+; SVE-NEXT: [[TMP6:%.*]] = getelementptr inbounds i32, ptr [[DATA]], i64 [[INDEX]]
+; SVE-NEXT: [[WIDE_LOAD:%.*]] = load <vscale x 4 x i32>, ptr [[TMP6]], align 4
+; SVE-NEXT: [[TMP13:%.*]] = icmp slt <vscale x 4 x i32> [[BROADCAST_SPLAT]], [[WIDE_LOAD]]
+; SVE-NEXT: [[TMP7:%.*]] = freeze <vscale x 4 x i1> [[TMP13]]
+; SVE-NEXT: [[TMP8:%.*]] = call i1 @llvm.vector.reduce.or.nxv4i1(<vscale x 4 x i1> [[TMP7]])
+; SVE-NEXT: [[TMP9]] = select i1 [[TMP8]], <vscale x 4 x i1> [[TMP13]], <vscale x 4 x i1> [[TMP4]]
+; SVE-NEXT: [[TMP10]] = select i1 [[TMP8]], <vscale x 4 x i32> [[WIDE_LOAD]], <vscale x 4 x i32> [[VEC_PHI]]
+; SVE-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], [[TMP3]]
+; SVE-NEXT: [[TMP11:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
+; SVE-NEXT: br i1 [[TMP11]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
+; SVE: [[MIDDLE_BLOCK]]:
+; SVE-NEXT: [[TMP12:%.*]] = call i32 @llvm.experimental.vector.extract.last.active.nxv4i32(<vscale x 4 x i32> [[TMP10]], <vscale x 4 x i1> [[TMP9]], i32 -1)
+; SVE-NEXT: [[CMP_N:%.*]] = icmp eq i64 [[N]], [[N_VEC]]
+; SVE-NEXT: br i1 [[CMP_N]], label %[[EXIT:.*]], label %[[SCALAR_PH]]
+; SVE: [[SCALAR_PH]]:
+; SVE-NEXT: [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], %[[MIDDLE_BLOCK]] ], [ 0, %[[ENTRY]] ]
+; SVE-NEXT: [[BC_MERGE_RDX:%.*]] = phi i32 [ [[TMP12]], %[[MIDDLE_BLOCK]] ], [ -1, %[[ENTRY]] ]
+; SVE-NEXT: br label %[[LOOP:.*]]
+; SVE: [[LOOP]]:
+; SVE-NEXT: [[IV:%.*]] = phi i64 [ [[BC_RESUME_VAL]], %[[SCALAR_PH]] ], [ [[IV_NEXT:%.*]], %[[LOOP]] ]
+; SVE-NEXT: [[DATA_PHI:%.*]] = phi i32 [ [[BC_MERGE_RDX]], %[[SCALAR_PH]] ], [ [[SELECT_DATA:%.*]], %[[LOOP]] ]
+; SVE-NEXT: [[LD_ADDR:%.*]] = getelementptr inbounds i32, ptr [[DATA]], i64 [[IV]]
+; SVE-NEXT: [[LD:%.*]] = load i32, ptr [[LD_ADDR]], align 4
+; SVE-NEXT: [[SELECT_CMP:%.*]] = icmp slt i32 [[A]], [[LD]]
+; SVE-NEXT: [[SELECT_DATA]] = select i1 [[SELECT_CMP]], i32 [[LD]], i32 [[DATA_PHI]]
+; SVE-NEXT: [[IV_NEXT]] = add nuw nsw i64 [[IV]], 1
+; SVE-NEXT: [[EXIT_CMP:%.*]] = icmp eq i64 [[IV_NEXT]], [[N]]
+; SVE-NEXT: br i1 [[EXIT_CMP]], label %[[EXIT]], label %[[LOOP]], !llvm.loop [[LOOP3:![0-9]+]]
+; SVE: [[EXIT]]:
+; SVE-NEXT: [[SELECT_DATA_LCSSA:%.*]] = phi i32 [ [[SELECT_DATA]], %[[LOOP]] ], [ [[TMP12]], %[[MIDDLE_BLOCK]] ]
+; SVE-NEXT: ret i32 [[SELECT_DATA_LCSSA]]
+;
+entry:
+ br label %loop
+
+loop:
+ %iv = phi i64 [ 0, %entry ], [ %iv.next, %loop ]
+ %data.phi = phi i32 [ -1, %entry ], [ %select.data, %loop ]
+ %ld.addr = getelementptr inbounds i32, ptr %data, i64 %iv
+ %ld = load i32, ptr %ld.addr, align 4
+ %select.cmp = icmp slt i32 %a, %ld
+ %select.data = select i1 %select.cmp, i32 %ld, i32 %data.phi
+ %iv.next = add nuw nsw i64 %iv, 1
+ %exit.cmp = icmp eq i64 %iv.next, %N
+ br i1 %exit.cmp, label %exit, label %loop
+
+exit:
+ ret i32 %select.data
+}
+
+define ptr @simple_csa_ptr_select(i64 %N, ptr %data, i64 %a, ptr %init) {
+; NEON-LABEL: define ptr @simple_csa_ptr_select(
+; NEON-SAME: i64 [[N:%.*]], ptr [[DATA:%.*]], i64 [[A:%.*]], ptr [[INIT:%.*]]) {
+; NEON-NEXT: [[ENTRY:.*]]:
+; NEON-NEXT: br label %[[LOOP:.*]]
+; NEON: [[LOOP]]:
+; NEON-NEXT: [[IV:%.*]] = phi i64 [ 0, %[[ENTRY]] ], [ [[IV_NEXT:%.*]], %[[LOOP]] ]
+; NEON-NEXT: [[DATA_PHI:%.*]] = phi ptr [ [[INIT]], %[[ENTRY]] ], [ [[SELECT_DATA:%.*]], %[[LOOP]] ]
+; NEON-NEXT: [[LD_ADDR:%.*]] = getelementptr inbounds ptr, ptr [[DATA]], i64 [[IV]]
+; NEON-NEXT: [[LD:%.*]] = load ptr, ptr [[LD_ADDR]], align 4
+; NEON-NEXT: [[LD_I64:%.*]] = ptrtoint ptr [[LD]] to i64
+; NEON-NEXT: [[SELECT_CMP:%.*]] = icmp slt i64 [[A]], [[LD_I64]]
+; NEON-NEXT: [[SELECT_DATA]] = select i1 [[SELECT_CMP]], ptr [[LD]], ptr [[DATA_PHI]]
+; NEON-NEXT: [[IV_NEXT]] = add nuw nsw i64 [[IV]], 1
+; NEON-NEXT: [[EXIT_CMP:%.*]] = icmp eq i64 [[IV_NEXT]], [[N]]
+; NEON-NEXT: br i1 [[EXIT_CMP]], label %[[EXIT:.*]], label %[[LOOP]]
+; NEON: [[EXIT]]:
+; NEON-NEXT: [[SELECT_DATA_LCSSA:%.*]] = phi ptr [ [[SELECT_DATA]], %[[LOOP]] ]
+; NEON-NEXT: ret ptr [[SELECT_DATA_LCSSA]]
+;
+; SVE-LABEL: define ptr @simple_csa_ptr_select(
+; SVE-SAME: i64 [[N:%.*]], ptr [[DATA:%.*]], i64 [[A:%.*]], ptr [[INIT:%.*]]) #[[ATTR0]] {
+; SVE-NEXT: [[ENTRY:.*]]:
+; SVE-NEXT: br label %[[LOOP:.*]]
+; SVE: [[LOOP]]:
+; SVE-NEXT: [[IV:%.*]] = phi i64 [ 0, %[[ENTRY]] ], [ [[IV_NEXT:%.*]], %[[LOOP]] ]
+; SVE-NEXT: [[DATA_PHI:%.*]] = phi ptr [ [[INIT]], %[[ENTRY]] ], [ [[SELECT_DATA:%.*]], %[[LOOP]] ]
+; SVE-NEXT: [[LD_ADDR:%.*]] = getelementptr inbounds ptr, ptr [[DATA]], i64 [[IV]]
+; SVE-NEXT: [[LD:%.*]] = load ptr, ptr [[LD_ADDR]], align 4
+; SVE-NEXT: [[LD_I64:%.*]] = ptrtoint ptr [[LD]] to i64
+; SVE-NEXT: [[SELECT_CMP:%.*]] = icmp slt i64 [[A]], [[LD_I64]]
+; SVE-NEXT: [[SELECT_DATA]] = select i1 [[SELECT_CMP]], ptr [[LD]], ptr [[DATA_PHI]]
+; SVE-NEXT: [[IV_NEXT]] = add nuw nsw i64 [[IV]], 1
+; SVE-NEXT: [[EXIT_CMP:%.*]] = icmp eq i64 [[IV_NEXT]], [[N]]
+; SVE-NEXT: br i1 [[EXIT_CMP]], label %[[EXIT:.*]], label %[[LOOP]]
+; SVE: [[EXIT]]:
+; SVE-NEXT: [[SELECT_DATA_LCSSA:%.*]] = phi ptr [ [[SELECT_DATA]], %[[LOOP]] ]
+; SVE-NEXT: ret ptr [[SELECT_DATA_LCSSA]]
+;
+entry:
+ br label %loop
+
+loop:
+ %iv = phi i64 [ 0, %entry ], [ %iv.next, %loop ]
+ %data.phi = phi ptr [ %init, %entry ], [ %select.data, %loop ]
+ %ld.addr = getelementptr inbounds ptr, ptr %data, i64 %iv
+ %ld = load ptr, ptr %ld.addr, align 4
+ %ld.i64 = ptrtoint ptr %ld to i64
+ %select.cmp = icmp slt i64 %a, %ld.i64
+ %select.data = select i1 %select.cmp, ptr %ld, ptr %data.phi
+ %iv.next = add nuw nsw i64 %iv, 1
+ %exit.cmp = icmp eq i64 %iv.next, %N
+ br i1 %exit.cmp, label %exit, label %loop
+
+exit:
+ ret ptr %select.data
+}
+
+define float @simple_csa_float_select(i64 %N, ptr %data, float %a) {
+; NEON-LABEL: define float @simple_csa_float_select(
+; NEON-SAME: i64 [[N:%.*]], ptr [[DATA:%.*]], float [[A:%.*]]) {
+; NEON-NEXT: [[ENTRY:.*]]:
+; NEON-NEXT: br label %[[LOOP:.*]]
+; NEON: [[LOOP]]:
+; NEON-NEXT: [[IV:%.*]] = phi i64 [ 0, %[[ENTRY]] ], [ [[IV_NEXT:%.*]], %[[LOOP]] ]
+; NEON-NEXT: [[DATA_PHI:%.*]] = phi float [ -1.000000e+00, %[[ENTRY]] ], [ [[SELECT_DATA:%.*]], %[[LOOP]] ]
+; NEON-NEXT: [[LD_ADDR:%.*]] = getelementptr inbounds float, ptr [[DATA]], i64 [[IV]]
+; NEON-NEXT: [[LD:%.*]] = load float, ptr [[LD_ADDR]], align 4
+; NEON-NEXT: [[SELECT_CMP:%.*]] = fcmp olt float [[A]], [[LD]]
+; NEON-NEXT: [[SELECT_DATA]] = select i1 [[SELECT_CMP]], float [[LD]], float [[DATA_PHI]]
+; NEON-NEXT: [[IV_NEXT]] = add nuw nsw i64 [[IV]], 1
+; NEON-NEXT: [[EXIT_CMP:%.*]] = icmp eq i64 [[IV_NEXT]], [[N]]
+; NEON-NEXT: br i1 [[EXIT_CMP]], label %[[EXIT:.*]], label %[[LOOP]]
+; NEON: [[EXIT]]:
+; NEON-NEXT: [[SELECT_DATA_LCSSA:%.*]] = phi float [ [[SELECT_DATA]], %[[LOOP]] ]
+; NEON-NEXT: ret float [[SELECT_DATA_LCSSA]]
+;
+; SVE-LABEL: define float @simple_csa_float_select(
+; SVE-SAME: i64 [[N:%.*]], ptr [[DATA:%.*]], float [[A:%.*]]) #[[ATTR0]] {
+; SVE-NEXT: [[ENTRY:.*]]:
+; SVE-NEXT: br label %[[LOOP:.*]]
+; SVE: [[LOOP]]:
+; SVE-NEXT: [[IV:%.*]] = phi i64 [ 0, %[[ENTRY]] ], [ [[IV_NEXT:%.*]], %[[LOOP]] ]
+; SVE-NEXT: [[DATA_PHI:%.*]] = phi float [ -1.000000e+00, %[[ENTRY]] ], [ [[SELECT_DATA:%.*]], %[[LOOP]] ]
+; SVE-NEXT: [[LD_ADDR:%.*]] = getelementptr inbounds float, ptr [[DATA]], i64 [[IV]]
+; SVE-NEXT: [[LD:%.*]] = load float, ptr [[LD_ADDR]], align 4
+; SVE-NEXT: [[SELECT_CMP:%.*]] = fcmp olt float [[A]], [[LD]]
+; SVE-NEXT: [[SELECT_DATA]] = select i1 [[SELECT_CMP]], float [[LD]], float [[DATA_PHI]]
+; SVE-NEXT: [[IV_NEXT]] = add nuw nsw i64 [[IV]], 1
+; SVE-NEXT: [[EXIT_CMP:%.*]] = icmp eq i64 [[IV_NEXT]], [[N]]
+; SVE-NEXT: br i1 [[EXIT_CMP]], label %[[EXIT:.*]], label %[[LOOP]]
+; SVE: [[EXIT]]:
+; SVE-NEXT: [[SELECT_DATA_LCSSA:%.*]] = phi float [ [[SELECT_DATA]], %[[LOOP]] ]
+; SVE-NEXT: ret float [[SELECT_DATA_LCSSA]]
+;
+entry:
+ br label %loop
+
+loop:
+ %iv = phi i64 [ 0, %entry ], [ %iv.next, %loop ]
+ %data.phi = phi float [ -1.0, %entry ], [ %select.data, %loop ]
+ %ld.addr = getelementptr inbounds float, ptr %data, i64 %iv
+ %ld = load float, ptr %ld.addr, align 4
+ %select.cmp = fcmp olt float %a, %ld
+ %select.data = select i1 %select.cmp, float %ld, float %data.phi
+ %iv.next = add nuw nsw i64 %iv, 1
+ %exit.cmp = icmp eq i64 %iv.next, %N
+ br i1 %exit.cmp, label %exit, label %loop
+
+exit:
+ ret float %select.data
+}
+
+define i32 @multi_user_csa_int_select(i64 %N, ptr %data, ptr %results, i32 %a) {
+; NEON-LABEL: define i32 @multi_user_csa_int_select(
+; NEON-SAME: i64 [[N:%.*]], ptr [[DATA:%.*]], ptr [[RESULTS:%.*]], i32 [[A:%.*]]) {
+; NEON-NEXT: [[ENTRY:.*]]:
+; NEON-NEXT: br label %[[LOOP:.*]]
+; NEON: [[LOOP]]:
+; NEON-NEXT: [[IV:%.*]] = phi i64 [ 0, %[[ENTRY]] ], [ [[IV_NEXT:%.*]], %[[LOOP]] ]
+; NEON-NEXT: [[DATA_PHI:%.*]] = phi i32 [ -1, %[[ENTRY]] ], [ [[SELECT_DATA:%.*]], %[[LOOP]] ]
+; NEON-NEXT: [[LD_ADDR:%.*]] = getelementptr inbounds i32, ptr [[DATA]], i64 [[IV]]
+; NEON-NEXT: [[LD:%.*]] = load i32, ptr [[LD_ADDR]], align 4
+; NEON-NEXT: [[SELECT_CMP:%.*]] = icmp slt i32 [[A]], [[LD]]
+; NEON-NEXT: [[SELECT_DATA]] = select i1 [[SELECT_CMP]], i32 [[LD]], i32 [[DATA_PHI]]
+; NEON-NEXT: [[RES_ADDR:%.*]] = getelementptr inbounds i32, ptr [[RESULTS]], i64 [[IV]]
+; NEON-NEXT: store i32 [[SELECT_DATA]], ptr [[RES_ADDR]], align 4
+; NEON-NEXT: [[IV_NEXT]] = add nuw nsw i64 [[IV]], 1
+; NEON-NEXT: [[EXIT_CMP:%.*]] = icmp eq i64 [[IV_NEXT]], [[N]]
+; NEON-NEXT: br i1 [[EXIT_CMP]], label %[[EXIT:.*]], label %[[LOOP]]
+; NEON: [[EXIT]]:
+; NEON-NEXT: [[SELECT_DATA_LCSSA:%.*]] = phi i32 [ [[SELECT_DATA]], %[[LOOP]] ]
+; NEON-NEXT: ret i32 [[SELECT_DATA_LCSSA]]
+;
+; SVE-LABEL: define i32 @multi_user_csa_int_select(
+; SVE-SAME: i64 [[N:%.*]], ptr [[DATA:%.*]], ptr [[RESULTS:%.*]], i32 [[A:%.*]]) #[[ATTR0]] {
+; SVE-NEXT: [[ENTRY:.*]]:
+; SVE-NEXT: br label %[[LOOP:.*]]
+; SVE: [[LOOP]]:
+; SVE-NEXT: [[IV:%.*]] = phi i64 [ 0, %[[ENTRY]] ], [ [[IV_NEXT:%.*]], %[[LOOP]] ]
+; SVE-NEXT: [[DATA_PHI:%.*]] = phi i32 [ -1, %[[ENTRY]] ], [ [[SELECT_DATA:%.*]], %[[LOOP]] ]
+; SVE-NEXT: [[LD_ADDR:%.*]] = getelementptr inbounds i32, ptr [[DATA]], i64 [[IV]]
+; SVE-NEXT: [[LD:%.*]] = load i32, ptr [[LD_ADDR]], align 4
+; SVE-NEXT: [[SELECT_CMP:%.*]] = icmp slt i32 [[A]], [[LD]]
+; SVE-NEXT: [[SELECT_DATA]] = select i1 [[SELECT_CMP]], i32 [[LD]], i32 [[DATA_PHI]]
+; SVE-NEXT: [[RES_ADDR:%.*]] = getelementptr inbounds i32, ptr [[RESULTS]], i64 [[IV]]
+; SVE-NEXT: store i32 [[SELECT_DATA]], ptr [[RES_ADDR]], align 4
+; SVE-NEXT: [[IV_NEXT]] = add nuw nsw i64 [[IV]], 1
+; SVE-NEXT: [[EXIT_CMP:%.*]] = icmp eq i64 [[IV_NEXT]], [[N]]
+; SVE-NEXT: br i1 [[EXIT_CMP]], label %[[EXIT:.*]], label %[[LOOP]]
+; SVE: [[EXIT]]:
+; SVE-NEXT: [[SELECT_DATA_LCSSA:%.*]] = phi i32 [ [[SELECT_DATA]], %[[LOOP]] ]
+; SVE-NEXT: ret i32 [[SELECT_DATA_LCSSA]]
+;
+entry:
+ br label %loop
+
+loop:
+ %iv = phi i64 [ 0, %entry ], [ %iv.next, %loop ]
+ %data.phi = phi i32 [ -1, %entry ], [ %select.data, %loop ]
+ %ld.addr = getelementptr inbounds i32, ptr %data, i64 %iv
+ %ld = load i32, ptr %ld.addr, align 4
+ %select.cmp = icmp slt i32 %a, %ld
+ %select.data = select i1 %select.cmp, i32 %ld, i32 %data.phi
+ %res.addr = getelementptr inbounds i32, ptr %results, i64 %iv
+ store i32 %select.data, ptr %res.addr, align 4
+ %iv.next = add nuw nsw i64 %iv, 1
+ %exit.cmp = icmp eq i64 %iv.next, %N
+ br i1 %exit.cmp, label %exit, label %loop
+
+exit:
+ ret i32 %select.data
+}
+
+
+define i32 @multi_use_cmp_for_csa_int_select(i64 %N, ptr %data, i32 %a) {
+; NEON-LABEL: define i32 @multi_use_cmp_for_csa_int_select(
+; NEON-SAME: i64 [[N:%.*]], ptr [[DATA:%.*]], i32 [[A:%.*]]) {
+; NEON-NEXT: [[ENTRY:.*]]:
+; NEON-NEXT: br label %[[LOOP:.*]]
+; NEON: [[LOOP]]:
+; NEON-NEXT: [[IV:%.*]] = phi i64 [ 0, %[[ENTRY]] ], [ [[IV_NEXT:%.*]], %[[LOOP]] ]
+; NEON-NEXT: [[DATA_PHI:%.*]] = phi i32 [ -1, %[[ENTRY]] ], [ [[SELECT_DATA:%.*]], %[[LOOP]] ]
+; NEON-NEXT: [[IDX_PHI:%.*]] = phi i64 [ -1, %[[ENTRY]] ], [ [[SELECT_IDX:%.*]], %[[LOOP]] ]
+; NEON-NEXT: [[LD_ADDR:%.*]] = getelementptr inbounds i32, ptr [[DATA]], i64 [[IV]]
+; NEON-NEXT: [[LD:%.*]] = load i32, ptr [[LD_ADDR]], align 4
+; NEON-NEXT: [[SELECT_CMP:%.*]] = icmp slt i32 [[A]], [[LD]]
+; NEON-NEXT: [[SELECT_DATA]] = select i1 [[SELECT_CMP]], i32 [[LD]], i32 [[DATA_PHI]]
+; NEON-NEXT: [[SELECT_IDX]] = select i1 [[SELECT_CMP]], i64 [[IV]], i64 [[IDX_PHI]]
+; NEON-NEXT: [[IV_NEXT]] = add nuw nsw i64 [[IV]], 1
+; NEON-NEXT: [[EXIT_CMP:%.*]] = icmp eq i64 [[IV_NEXT]], [[N]]
+; NEON-NEXT: br i1 [[EXIT_CMP]], label %[[EXIT:.*]], label %[[LOOP]]
+; NEON: [[EXIT]]:
+; NEON-NEXT: [[SELECT_DATA_LCSSA:%.*]] = phi i32 [ [[SELECT_DATA]], %[[LOOP]] ]
+; NEON-NEXT: [[SELECT_IDX_LCSSA:%.*]] = phi i64 [ [[SELECT_IDX]], %[[LOOP]] ]
+; NEON-NEXT: [[IDX:%.*]] = trunc i64 [[SELECT_IDX_LCSSA]] to i32
+; NEON-NEXT: [[RES:%.*]] = add i32 [[IDX]], [[SELECT_DATA_LCSSA]]
+; NEON-NEXT: ret i32 [[RES]]
+;
+; SVE-LABEL: define i32 @multi_use_cmp_for_csa_int_select(
+; SVE-SAME: i64 [[N:%.*]], ptr [[DATA:%.*]], i32 [[A:%.*]]) #[[ATTR0]] {
+; SVE-NEXT: [[ENTRY:.*]]:
+; SVE-NEXT: br label %[[LOOP:.*]]
+; SVE: [[LOOP]]:
+; SVE-NEXT: [[IV:%.*]] = phi i64 [ 0, %[[ENTRY]] ], [ [[IV_NEXT:%.*]], %[[LOOP]] ]
+; SVE-NEXT: [[DATA_PHI:%.*]] = phi i32 [ -1, %[[ENTRY]] ], [ [[SELECT_DATA:%.*]], %[[LOOP]] ]
+; SVE-NEXT: [[IDX_PHI:%.*]] = phi i64 [ -1, %[[ENTRY]] ], [ [[SELECT_IDX:%.*]], %[[LOOP]] ]
+; SVE-NEXT: [[LD_ADDR:%.*]] = getelementptr inbounds i32, ptr [[DATA]], i64 [[IV]]
+; SVE-NEXT: [[LD:%.*]] = load i32, ptr [[LD_ADDR]], align 4
+; SVE-NEXT: [[SELECT_CMP:%.*]] = icmp slt i32 [[A]], [[LD]]
+; SVE-NEXT: [[SELECT_DATA]] = select i1 [[SELECT_CMP]], i32 [[LD]], i32 [[DATA_PHI]]
+; SVE-NEXT: [[SELECT_IDX]] = select i1 [[SELECT_CMP]], i64 [[IV]], i64 [[IDX_PHI]]
+; SVE-NEXT: [[IV_NEXT]] = add nuw nsw i64 [[IV]], 1
+; SVE-NEXT: [[EXIT_CMP:%.*]] = icmp eq i64 [[IV_NEXT]], [[N]]
+; SVE-NEXT: br i1 [[EXIT_CMP]], label %[[EXIT:.*]], label %[[LOOP]]
+; SVE: [[EXIT]]:
+; SVE-NEXT: [[SELECT_DATA_LCSSA:%.*]] = phi i32 [ [[SELECT_DATA]], %[[LOOP]] ]
+; SVE-NEXT: [[SELECT_IDX_LCSSA:%.*]] = phi i64 [ [[SELECT_IDX]], %[[LOOP]] ]
+; SVE-NEXT: [[IDX:%.*]] = trunc i64 [[SELECT_IDX_LCSSA]] to i32
+; SVE-NEXT: [[RES:%.*]] = add i32 [[IDX]], [[SELECT_DATA_LCSSA]]
+; SVE-NEXT: ret i32 [[RES]]
+;
+entry:
+ br label %loop
+
+loop:
+ %iv = phi i64 [ 0, %entry ], [ %iv.next, %loop ]
+ %data.phi = phi i32 [ -1, %entry ], [ %select.data, %loop ]
+ %idx.phi = phi i64 [ -1, %entry ], [ %select.idx, %loop ]
+ %ld.addr = getelementptr inbounds i32, ptr %data, i64 %iv
+ %ld = load i32, ptr %ld.addr, align 4
+ %select.cmp = icmp slt i32 %a, %ld
+ %select.data = select i1 %select.cmp, i32 %ld, i32 %data.phi
+ %select.idx = select i1 %select.cmp, i64 %iv, i64 %idx.phi
+ %iv.next = add nuw nsw i64 %iv, 1
+ %exit.cmp = icmp eq i64 %iv.next, %N
+ br i1 %exit.cmp, label %exit, label %loop
+
+exit:
+ %idx = trunc i64 %select.idx to i32
+ %res = add i32 %idx, %select.data
+ ret i32 %res
+}
+
+
+define i32 @chained_select_for_csa_int_select(i64 %N, ptr %data1, ptr %data2, i32 %a, i32 %b) {
----------------
huntergr-arm wrote:
done
https://github.com/llvm/llvm-project/pull/158088
More information about the llvm-commits
mailing list