[PowerPC][NFC] Add test case for memset tail store

[tingwang]

Summary

Add test case to show something can be improved.

[nemanjai]

I think that for completeness, we should add tests where the size is less than a full vector's width (i.e. 1, 4, 5, 8, 9, 12, 15 bytes).

[tingwang]

Add more test cases according to comment.

[tingwang]

Add memset2TailV1Bx() cases which does not fit in vspltisb and may require constant pool.

[shchenz]

LGTM. Go ahead to commit these cases first. If there are more missing tests found during the review of D138883, we can add them by then

llvm/test/CodeGen/PowerPC/memset-tail.ll

+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -verify-machineinstrs -mcpu=pwr8 \
+; RUN:   -mtriple=powerpc64-unknown-aix < %s | FileCheck %s --check-prefix=P8-BE
+; RUN: llc -verify-machineinstrs -mcpu=pwr9 \
+; RUN:   -mtriple=powerpc64-unknown-aix < %s | FileCheck %s --check-prefix=P9-BE
+; RUN: llc -verify-machineinstrs -mcpu=pwr10 \
+; RUN:   -mtriple=powerpc64-unknown-aix < %s | FileCheck %s --check-prefix=P10-BE
+; RUN: llc -verify-machineinstrs -mcpu=pwr8 \
+; RUN:   -mtriple=powerpc64le-unknown-unknown < %s | FileCheck %s --check-prefix=P8-LE
+; RUN: llc -verify-machineinstrs -mcpu=pwr9 \
+; RUN:   -mtriple=powerpc64le-unknown-unknown < %s | FileCheck %s --check-prefix=P9-LE
+; RUN: llc -verify-machineinstrs -mcpu=pwr10 \
+; RUN:   -mtriple=powerpc64le-unknown-unknown < %s | FileCheck %s --check-prefix=P10-LE
+
+define dso_local void @memsetTailV1B15(ptr nocapture noundef writeonly %p) local_unnamed_addr {
+; P8-BE-LABEL: memsetTailV1B15:
+; P8-BE:       # %bb.0: # %entry
+; P8-BE-NEXT:    vspltisb 2, 15
+; P8-BE-NEXT:    li 4, 15
+; P8-BE-NEXT:    stxvw4x 34, 3, 4
+; P8-BE-NEXT:    stxvw4x 34, 0, 3
+; P8-BE-NEXT:    blr
+;
+; P9-BE-LABEL: memsetTailV1B15:
+; P9-BE:       # %bb.0: # %entry
+; P9-BE-NEXT:    xxspltib 0, 15
+; P9-BE-NEXT:    li 4, 15
+; P9-BE-NEXT:    stxvx 0, 3, 4
+; P9-BE-NEXT:    stxv 0, 0(3)
+; P9-BE-NEXT:    blr
+;
+; P10-BE-LABEL: memsetTailV1B15:
+; P10-BE:       # %bb.0: # %entry
+; P10-BE-NEXT:    xxspltib 0, 15
+; P10-BE-NEXT:    pstxv 0, 15(3), 0
+; P10-BE-NEXT:    stxv 0, 0(3)
+; P10-BE-NEXT:    blr
+;
+; P8-LE-LABEL: memsetTailV1B15:
+; P8-LE:       # %bb.0: # %entry
+; P8-LE-NEXT:    vspltisb 2, 15
+; P8-LE-NEXT:    li 4, 15
+; P8-LE-NEXT:    stxvd2x 34, 3, 4
+; P8-LE-NEXT:    stxvd2x 34, 0, 3
+; P8-LE-NEXT:    blr
+;
+; P9-LE-LABEL: memsetTailV1B15:
+; P9-LE:       # %bb.0: # %entry
+; P9-LE-NEXT:    xxspltib 0, 15
+; P9-LE-NEXT:    li 4, 15
+; P9-LE-NEXT:    stxvx 0, 3, 4
+; P9-LE-NEXT:    stxv 0, 0(3)
+; P9-LE-NEXT:    blr
+;
+; P10-LE-LABEL: memsetTailV1B15:
+; P10-LE:       # %bb.0: # %entry
+; P10-LE-NEXT:    xxspltib 0, 15
+; P10-LE-NEXT:    pstxv 0, 15(3), 0
+; P10-LE-NEXT:    stxv 0, 0(3)
+; P10-LE-NEXT:    blr
+entry:
+  tail call void @llvm.memset.p0.i64(ptr %p, i8 15, i64 31, i1 false)
+  ret void
+}
+
+define dso_local void @memsetTailV1B12(ptr nocapture noundef writeonly %p) local_unnamed_addr {
+; P8-BE-LABEL: memsetTailV1B12:
+; P8-BE:       # %bb.0: # %entry
+; P8-BE-NEXT:    vspltisb 2, 15
+; P8-BE-NEXT:    li 4, 12
+; P8-BE-NEXT:    stxvw4x 34, 3, 4
+; P8-BE-NEXT:    stxvw4x 34, 0, 3
+; P8-BE-NEXT:    blr
+;
+; P9-BE-LABEL: memsetTailV1B12:
+; P9-BE:       # %bb.0: # %entry
+; P9-BE-NEXT:    xxspltib 0, 15
+; P9-BE-NEXT:    li 4, 12
+; P9-BE-NEXT:    stxvx 0, 3, 4
+; P9-BE-NEXT:    stxv 0, 0(3)
+; P9-BE-NEXT:    blr
+;
+; P10-BE-LABEL: memsetTailV1B12:
+; P10-BE:       # %bb.0: # %entry
+; P10-BE-NEXT:    xxspltib 0, 15
+; P10-BE-NEXT:    pstxv 0, 12(3), 0
+; P10-BE-NEXT:    stxv 0, 0(3)
+; P10-BE-NEXT:    blr
+;
+; P8-LE-LABEL: memsetTailV1B12:
+; P8-LE:       # %bb.0: # %entry
+; P8-LE-NEXT:    vspltisb 2, 15
+; P8-LE-NEXT:    li 4, 12
+; P8-LE-NEXT:    stxvd2x 34, 3, 4
+; P8-LE-NEXT:    stxvd2x 34, 0, 3
+; P8-LE-NEXT:    blr
+;
+; P9-LE-LABEL: memsetTailV1B12:
+; P9-LE:       # %bb.0: # %entry
+; P9-LE-NEXT:    xxspltib 0, 15
+; P9-LE-NEXT:    li 4, 12
+; P9-LE-NEXT:    stxvx 0, 3, 4
+; P9-LE-NEXT:    stxv 0, 0(3)
+; P9-LE-NEXT:    blr
+;
+; P10-LE-LABEL: memsetTailV1B12:
+; P10-LE:       # %bb.0: # %entry
+; P10-LE-NEXT:    xxspltib 0, 15
+; P10-LE-NEXT:    pstxv 0, 12(3), 0
+; P10-LE-NEXT:    stxv 0, 0(3)
+; P10-LE-NEXT:    blr
+entry:
+  tail call void @llvm.memset.p0.i64(ptr %p, i8 15, i64 28, i1 false)
+  ret void
+}
+
+define dso_local void @memsetTailV1B9(ptr nocapture noundef writeonly %p) local_unnamed_addr {
+; P8-BE-LABEL: memsetTailV1B9:
+; P8-BE:       # %bb.0: # %entry
+; P8-BE-NEXT:    vspltisb 2, 15
+; P8-BE-NEXT:    li 4, 9
+; P8-BE-NEXT:    stxvw4x 34, 3, 4
+; P8-BE-NEXT:    stxvw4x 34, 0, 3
+; P8-BE-NEXT:    blr
+;
+; P9-BE-LABEL: memsetTailV1B9:
+; P9-BE:       # %bb.0: # %entry
+; P9-BE-NEXT:    xxspltib 0, 15
+; P9-BE-NEXT:    li 4, 9
+; P9-BE-NEXT:    stxvx 0, 3, 4
+; P9-BE-NEXT:    stxv 0, 0(3)
+; P9-BE-NEXT:    blr
+;
+; P10-BE-LABEL: memsetTailV1B9:
+; P10-BE:       # %bb.0: # %entry
+; P10-BE-NEXT:    xxspltib 0, 15
+; P10-BE-NEXT:    pstxv 0, 9(3), 0
+; P10-BE-NEXT:    stxv 0, 0(3)
+; P10-BE-NEXT:    blr
+;
+; P8-LE-LABEL: memsetTailV1B9:
+; P8-LE:       # %bb.0: # %entry
+; P8-LE-NEXT:    vspltisb 2, 15
+; P8-LE-NEXT:    li 4, 9
+; P8-LE-NEXT:    stxvd2x 34, 3, 4
+; P8-LE-NEXT:    stxvd2x 34, 0, 3
+; P8-LE-NEXT:    blr
+;
+; P9-LE-LABEL: memsetTailV1B9:
+; P9-LE:       # %bb.0: # %entry
+; P9-LE-NEXT:    xxspltib 0, 15
+; P9-LE-NEXT:    li 4, 9
+; P9-LE-NEXT:    stxvx 0, 3, 4
+; P9-LE-NEXT:    stxv 0, 0(3)
+; P9-LE-NEXT:    blr
+;
+; P10-LE-LABEL: memsetTailV1B9:
+; P10-LE:       # %bb.0: # %entry
+; P10-LE-NEXT:    xxspltib 0, 15
+; P10-LE-NEXT:    pstxv 0, 9(3), 0
+; P10-LE-NEXT:    stxv 0, 0(3)
+; P10-LE-NEXT:    blr
+entry:
+  tail call void @llvm.memset.p0.i64(ptr %p, i8 15, i64 25, i1 false)
+  ret void
+}
+
+define dso_local void @memsetTailV1B8(ptr nocapture noundef writeonly %p) local_unnamed_addr {
+; P8-BE-LABEL: memsetTailV1B8:
+; P8-BE:       # %bb.0: # %entry
+; P8-BE-NEXT:    vspltisb 2, 15
+; P8-BE-NEXT:    lis 4, 3855
+; P8-BE-NEXT:    ori 4, 4, 3855
+; P8-BE-NEXT:    rldimi 4, 4, 32, 0
+; P8-BE-NEXT:    stxvw4x 34, 0, 3
+; P8-BE-NEXT:    std 4, 16(3)
+; P8-BE-NEXT:    blr
+;
+; P9-BE-LABEL: memsetTailV1B8:
+; P9-BE:       # %bb.0: # %entry
+; P9-BE-NEXT:    lis 4, 3855
+; P9-BE-NEXT:    xxspltib 0, 15
+; P9-BE-NEXT:    ori 4, 4, 3855
+; P9-BE-NEXT:    stxv 0, 0(3)
+; P9-BE-NEXT:    rldimi 4, 4, 32, 0
+; P9-BE-NEXT:    std 4, 16(3)
+; P9-BE-NEXT:    blr
+;
+; P10-BE-LABEL: memsetTailV1B8:
+; P10-BE:       # %bb.0: # %entry
+; P10-BE-NEXT:    pli 4, 252645135
+; P10-BE-NEXT:    rldimi 4, 4, 32, 0
+; P10-BE-NEXT:    std 4, 16(3)
+; P10-BE-NEXT:    xxspltib 0, 15
+; P10-BE-NEXT:    stxv 0, 0(3)
+; P10-BE-NEXT:    blr
+;
+; P8-LE-LABEL: memsetTailV1B8:
+; P8-LE:       # %bb.0: # %entry
+; P8-LE-NEXT:    lis 4, 3855
+; P8-LE-NEXT:    vspltisb 2, 15
+; P8-LE-NEXT:    ori 4, 4, 3855
+; P8-LE-NEXT:    rldimi 4, 4, 32, 0
+; P8-LE-NEXT:    std 4, 16(3)
+; P8-LE-NEXT:    stxvd2x 34, 0, 3
+; P8-LE-NEXT:    blr
+;
+; P9-LE-LABEL: memsetTailV1B8:
+; P9-LE:       # %bb.0: # %entry
+; P9-LE-NEXT:    lis 4, 3855
+; P9-LE-NEXT:    xxspltib 0, 15
+; P9-LE-NEXT:    ori 4, 4, 3855
+; P9-LE-NEXT:    stxv 0, 0(3)
+; P9-LE-NEXT:    rldimi 4, 4, 32, 0
+; P9-LE-NEXT:    std 4, 16(3)
+; P9-LE-NEXT:    blr
+;
+; P10-LE-LABEL: memsetTailV1B8:
+; P10-LE:       # %bb.0: # %entry
+; P10-LE-NEXT:    pli 4, 252645135
+; P10-LE-NEXT:    rldimi 4, 4, 32, 0
+; P10-LE-NEXT:    std 4, 16(3)
+; P10-LE-NEXT:    xxspltib 0, 15
+; P10-LE-NEXT:    stxv 0, 0(3)
+; P10-LE-NEXT:    blr
+entry:
+  tail call void @llvm.memset.p0.i64(ptr %p, i8 15, i64 24, i1 false)
+  ret void
+}
+
+define dso_local void @memsetTailV1B7(ptr nocapture noundef writeonly %p) local_unnamed_addr {
+; P8-BE-LABEL: memsetTailV1B7:
+; P8-BE:       # %bb.0: # %entry
+; P8-BE-NEXT:    lis 4, 3855
+; P8-BE-NEXT:    vspltisb 2, 15
+; P8-BE-NEXT:    li 5, 15
+; P8-BE-NEXT:    ori 4, 4, 3855
+; P8-BE-NEXT:    rldimi 4, 4, 32, 0
+; P8-BE-NEXT:    stdx 4, 3, 5
+; P8-BE-NEXT:    stxvw4x 34, 0, 3
+; P8-BE-NEXT:    blr
+;
+; P9-BE-LABEL: memsetTailV1B7:
+; P9-BE:       # %bb.0: # %entry
+; P9-BE-NEXT:    lis 4, 3855
+; P9-BE-NEXT:    li 5, 15
+; P9-BE-NEXT:    ori 4, 4, 3855
+; P9-BE-NEXT:    rldimi 4, 4, 32, 0
+; P9-BE-NEXT:    stdx 4, 3, 5
+; P9-BE-NEXT:    xxspltib 0, 15
+; P9-BE-NEXT:    stxv 0, 0(3)
+; P9-BE-NEXT:    blr
+;
+; P10-BE-LABEL: memsetTailV1B7:
+; P10-BE:       # %bb.0: # %entry
+; P10-BE-NEXT:    pli 4, 252645135
+; P10-BE-NEXT:    rldimi 4, 4, 32, 0
+; P10-BE-NEXT:    pstd 4, 15(3), 0
+; P10-BE-NEXT:    xxspltib 0, 15
+; P10-BE-NEXT:    stxv 0, 0(3)
+; P10-BE-NEXT:    blr
+;
+; P8-LE-LABEL: memsetTailV1B7:
+; P8-LE:       # %bb.0: # %entry
+; P8-LE-NEXT:    lis 4, 3855
+; P8-LE-NEXT:    vspltisb 2, 15
+; P8-LE-NEXT:    li 5, 15
+; P8-LE-NEXT:    ori 4, 4, 3855
+; P8-LE-NEXT:    rldimi 4, 4, 32, 0
+; P8-LE-NEXT:    stdx 4, 3, 5
+; P8-LE-NEXT:    stxvd2x 34, 0, 3
+; P8-LE-NEXT:    blr
+;
+; P9-LE-LABEL: memsetTailV1B7:
+; P9-LE:       # %bb.0: # %entry
+; P9-LE-NEXT:    lis 4, 3855
+; P9-LE-NEXT:    li 5, 15
+; P9-LE-NEXT:    ori 4, 4, 3855
+; P9-LE-NEXT:    rldimi 4, 4, 32, 0
+; P9-LE-NEXT:    stdx 4, 3, 5
+; P9-LE-NEXT:    xxspltib 0, 15
+; P9-LE-NEXT:    stxv 0, 0(3)
+; P9-LE-NEXT:    blr
+;
+; P10-LE-LABEL: memsetTailV1B7:
+; P10-LE:       # %bb.0: # %entry
+; P10-LE-NEXT:    pli 4, 252645135
+; P10-LE-NEXT:    rldimi 4, 4, 32, 0
+; P10-LE-NEXT:    pstd 4, 15(3), 0
+; P10-LE-NEXT:    xxspltib 0, 15
+; P10-LE-NEXT:    stxv 0, 0(3)
+; P10-LE-NEXT:    blr
+entry:
+  tail call void @llvm.memset.p0.i64(ptr %p, i8 15, i64 23, i1 false)
+  ret void
+}
+
+define dso_local void @memsetTailV1B4(ptr nocapture noundef writeonly %p) local_unnamed_addr {
+; P8-BE-LABEL: memsetTailV1B4:
+; P8-BE:       # %bb.0: # %entry
+; P8-BE-NEXT:    vspltisb 2, 15
+; P8-BE-NEXT:    lis 4, 3855
+; P8-BE-NEXT:    ori 4, 4, 3855
+; P8-BE-NEXT:    stw 4, 16(3)
+; P8-BE-NEXT:    stxvw4x 34, 0, 3
+; P8-BE-NEXT:    blr
+;
+; P9-BE-LABEL: memsetTailV1B4:
+; P9-BE:       # %bb.0: # %entry
+; P9-BE-NEXT:    lis 4, 3855
+; P9-BE-NEXT:    ori 4, 4, 3855
+; P9-BE-NEXT:    stw 4, 16(3)
+; P9-BE-NEXT:    xxspltib 0, 15
+; P9-BE-NEXT:    stxv 0, 0(3)
+; P9-BE-NEXT:    blr
+;
+; P10-BE-LABEL: memsetTailV1B4:
+; P10-BE:       # %bb.0: # %entry
+; P10-BE-NEXT:    pli 4, 252645135
+; P10-BE-NEXT:    stw 4, 16(3)
+; P10-BE-NEXT:    xxspltib 0, 15
+; P10-BE-NEXT:    stxv 0, 0(3)
+; P10-BE-NEXT:    blr
+;
+; P8-LE-LABEL: memsetTailV1B4:
+; P8-LE:       # %bb.0: # %entry
+; P8-LE-NEXT:    vspltisb 2, 15
+; P8-LE-NEXT:    lis 4, 3855
+; P8-LE-NEXT:    ori 4, 4, 3855
+; P8-LE-NEXT:    stw 4, 16(3)
+; P8-LE-NEXT:    stxvd2x 34, 0, 3
+; P8-LE-NEXT:    blr
+;
+; P9-LE-LABEL: memsetTailV1B4:
+; P9-LE:       # %bb.0: # %entry
+; P9-LE-NEXT:    lis 4, 3855
+; P9-LE-NEXT:    ori 4, 4, 3855
+; P9-LE-NEXT:    stw 4, 16(3)
+; P9-LE-NEXT:    xxspltib 0, 15
+; P9-LE-NEXT:    stxv 0, 0(3)
+; P9-LE-NEXT:    blr
+;
+; P10-LE-LABEL: memsetTailV1B4:
+; P10-LE:       # %bb.0: # %entry
+; P10-LE-NEXT:    pli 4, 252645135
+; P10-LE-NEXT:    stw 4, 16(3)
+; P10-LE-NEXT:    xxspltib 0, 15
+; P10-LE-NEXT:    stxv 0, 0(3)
+; P10-LE-NEXT:    blr
+entry:
+  tail call void @llvm.memset.p0.i32(ptr %p, i8 15, i32 20, i1 false)
+  ret void
+}
+
+define dso_local void @memsetTailV1B3(ptr nocapture noundef writeonly %p) local_unnamed_addr {
+; P8-BE-LABEL: memsetTailV1B3:
+; P8-BE:       # %bb.0: # %entry
+; P8-BE-NEXT:    vspltisb 2, 15
+; P8-BE-NEXT:    lis 4, 3855
+; P8-BE-NEXT:    ori 4, 4, 3855
+; P8-BE-NEXT:    stxvw4x 34, 0, 3
+; P8-BE-NEXT:    stw 4, 15(3)
+; P8-BE-NEXT:    blr
+;
+; P9-BE-LABEL: memsetTailV1B3:
+; P9-BE:       # %bb.0: # %entry
+; P9-BE-NEXT:    lis 4, 3855
+; P9-BE-NEXT:    ori 4, 4, 3855
+; P9-BE-NEXT:    stw 4, 15(3)
+; P9-BE-NEXT:    xxspltib 0, 15
+; P9-BE-NEXT:    stxv 0, 0(3)
+; P9-BE-NEXT:    blr
+;
+; P10-BE-LABEL: memsetTailV1B3:
+; P10-BE:       # %bb.0: # %entry
+; P10-BE-NEXT:    pli 4, 252645135
+; P10-BE-NEXT:    stw 4, 15(3)
+; P10-BE-NEXT:    xxspltib 0, 15
+; P10-BE-NEXT:    stxv 0, 0(3)
+; P10-BE-NEXT:    blr
+;
+; P8-LE-LABEL: memsetTailV1B3:
+; P8-LE:       # %bb.0: # %entry
+; P8-LE-NEXT:    vspltisb 2, 15
+; P8-LE-NEXT:    lis 4, 3855
+; P8-LE-NEXT:    ori 4, 4, 3855
+; P8-LE-NEXT:    stw 4, 15(3)
+; P8-LE-NEXT:    stxvd2x 34, 0, 3
+; P8-LE-NEXT:    blr
+;
+; P9-LE-LABEL: memsetTailV1B3:
+; P9-LE:       # %bb.0: # %entry
+; P9-LE-NEXT:    lis 4, 3855
+; P9-LE-NEXT:    ori 4, 4, 3855
+; P9-LE-NEXT:    stw 4, 15(3)
+; P9-LE-NEXT:    xxspltib 0, 15
+; P9-LE-NEXT:    stxv 0, 0(3)
+; P9-LE-NEXT:    blr
+;
+; P10-LE-LABEL: memsetTailV1B3:
+; P10-LE:       # %bb.0: # %entry
+; P10-LE-NEXT:    pli 4, 252645135
+; P10-LE-NEXT:    stw 4, 15(3)
+; P10-LE-NEXT:    xxspltib 0, 15
+; P10-LE-NEXT:    stxv 0, 0(3)
+; P10-LE-NEXT:    blr
+entry:
+  tail call void @llvm.memset.p0.i64(ptr %p, i8 15, i64 19, i1 false)
+  ret void
+}
+
+define dso_local void @memsetTailV1B2(ptr nocapture noundef writeonly %p) local_unnamed_addr {
+; P8-BE-LABEL: memsetTailV1B2:
+; P8-BE:       # %bb.0: # %entry
+; P8-BE-NEXT:    vspltisb 2, 15
+; P8-BE-NEXT:    li 4, 3855
+; P8-BE-NEXT:    sth 4, 16(3)
+; P8-BE-NEXT:    stxvw4x 34, 0, 3
+; P8-BE-NEXT:    blr
+;
+; P9-BE-LABEL: memsetTailV1B2:
+; P9-BE:       # %bb.0: # %entry
+; P9-BE-NEXT:    li 4, 3855
+; P9-BE-NEXT:    sth 4, 16(3)
+; P9-BE-NEXT:    xxspltib 0, 15
+; P9-BE-NEXT:    stxv 0, 0(3)
+; P9-BE-NEXT:    blr
+;
+; P10-BE-LABEL: memsetTailV1B2:
+; P10-BE:       # %bb.0: # %entry
+; P10-BE-NEXT:    li 4, 3855
+; P10-BE-NEXT:    sth 4, 16(3)
+; P10-BE-NEXT:    xxspltib 0, 15
+; P10-BE-NEXT:    stxv 0, 0(3)
+; P10-BE-NEXT:    blr
+;
+; P8-LE-LABEL: memsetTailV1B2:
+; P8-LE:       # %bb.0: # %entry
+; P8-LE-NEXT:    vspltisb 2, 15
+; P8-LE-NEXT:    li 4, 3855
+; P8-LE-NEXT:    sth 4, 16(3)
+; P8-LE-NEXT:    stxvd2x 34, 0, 3
+; P8-LE-NEXT:    blr
+;
+; P9-LE-LABEL: memsetTailV1B2:
+; P9-LE:       # %bb.0: # %entry
+; P9-LE-NEXT:    li 4, 3855
+; P9-LE-NEXT:    sth 4, 16(3)
+; P9-LE-NEXT:    xxspltib 0, 15
+; P9-LE-NEXT:    stxv 0, 0(3)
+; P9-LE-NEXT:    blr
+;
+; P10-LE-LABEL: memsetTailV1B2:
+; P10-LE:       # %bb.0: # %entry
+; P10-LE-NEXT:    li 4, 3855
+; P10-LE-NEXT:    sth 4, 16(3)
+; P10-LE-NEXT:    xxspltib 0, 15
+; P10-LE-NEXT:    stxv 0, 0(3)
+; P10-LE-NEXT:    blr
+entry:
+  tail call void @llvm.memset.p0.i64(ptr %p, i8 15, i64 18, i1 false)
+  ret void
+}
+
+define dso_local void @memsetTailV1B1(ptr nocapture noundef writeonly %p) local_unnamed_addr {
+; P8-BE-LABEL: memsetTailV1B1:
+; P8-BE:       # %bb.0: # %entry
+; P8-BE-NEXT:    vspltisb 2, 15
+; P8-BE-NEXT:    li 4, 15
+; P8-BE-NEXT:    stb 4, 16(3)
+; P8-BE-NEXT:    stxvw4x 34, 0, 3
+; P8-BE-NEXT:    blr
+;
+; P9-BE-LABEL: memsetTailV1B1:
+; P9-BE:       # %bb.0: # %entry
+; P9-BE-NEXT:    li 4, 15
+; P9-BE-NEXT:    stb 4, 16(3)
+; P9-BE-NEXT:    xxspltib 0, 15
+; P9-BE-NEXT:    stxv 0, 0(3)
+; P9-BE-NEXT:    blr
+;
+; P10-BE-LABEL: memsetTailV1B1:
+; P10-BE:       # %bb.0: # %entry
+; P10-BE-NEXT:    li 4, 15
+; P10-BE-NEXT:    stb 4, 16(3)
+; P10-BE-NEXT:    xxspltib 0, 15
+; P10-BE-NEXT:    stxv 0, 0(3)
+; P10-BE-NEXT:    blr
+;
+; P8-LE-LABEL: memsetTailV1B1:
+; P8-LE:       # %bb.0: # %entry
+; P8-LE-NEXT:    vspltisb 2, 15
+; P8-LE-NEXT:    li 4, 15
+; P8-LE-NEXT:    stb 4, 16(3)
+; P8-LE-NEXT:    stxvd2x 34, 0, 3
+; P8-LE-NEXT:    blr
+;
+; P9-LE-LABEL: memsetTailV1B1:
+; P9-LE:       # %bb.0: # %entry
+; P9-LE-NEXT:    li 4, 15
+; P9-LE-NEXT:    stb 4, 16(3)
+; P9-LE-NEXT:    xxspltib 0, 15
+; P9-LE-NEXT:    stxv 0, 0(3)
+; P9-LE-NEXT:    blr
+;
+; P10-LE-LABEL: memsetTailV1B1:
+; P10-LE:       # %bb.0: # %entry
+; P10-LE-NEXT:    li 4, 15
+; P10-LE-NEXT:    stb 4, 16(3)
+; P10-LE-NEXT:    xxspltib 0, 15
+; P10-LE-NEXT:    stxv 0, 0(3)
+; P10-LE-NEXT:    blr
+entry:
+  tail call void @llvm.memset.p0.i64(ptr %p, i8 15, i64 17, i1 false)
+  ret void
+}
+
+define dso_local void @memset2TailV1B15(ptr nocapture noundef writeonly %p) local_unnamed_addr {
+; P8-BE-LABEL: memset2TailV1B15:
+; P8-BE:       # %bb.0: # %entry
+; P8-BE-NEXT:    ld 4, L..C0(2) # %const.0
+; P8-BE-NEXT:    lxvw4x 0, 0, 4
+; P8-BE-NEXT:    li 4, 15
+; P8-BE-NEXT:    stxvw4x 0, 3, 4
+; P8-BE-NEXT:    stxvw4x 0, 0, 3
+; P8-BE-NEXT:    blr
+;
+; P9-BE-LABEL: memset2TailV1B15:
+; P9-BE:       # %bb.0: # %entry
+; P9-BE-NEXT:    xxspltib 0, 165
+; P9-BE-NEXT:    li 4, 15
+; P9-BE-NEXT:    stxvx 0, 3, 4
+; P9-BE-NEXT:    stxv 0, 0(3)
+; P9-BE-NEXT:    blr
+;
+; P10-BE-LABEL: memset2TailV1B15:
+; P10-BE:       # %bb.0: # %entry
+; P10-BE-NEXT:    xxspltib 0, 165
+; P10-BE-NEXT:    pstxv 0, 15(3), 0
+; P10-BE-NEXT:    stxv 0, 0(3)
+; P10-BE-NEXT:    blr
+;
+; P8-LE-LABEL: memset2TailV1B15:
+; P8-LE:       # %bb.0: # %entry
+; P8-LE-NEXT:    addis 4, 2, .LCPI9_0@toc@ha
+; P8-LE-NEXT:    addi 4, 4, .LCPI9_0@toc@l
+; P8-LE-NEXT:    lxvd2x 0, 0, 4
+; P8-LE-NEXT:    li 4, 15
+; P8-LE-NEXT:    stxvd2x 0, 3, 4
+; P8-LE-NEXT:    stxvd2x 0, 0, 3
+; P8-LE-NEXT:    blr
+;
+; P9-LE-LABEL: memset2TailV1B15:
+; P9-LE:       # %bb.0: # %entry
+; P9-LE-NEXT:    xxspltib 0, 165
+; P9-LE-NEXT:    li 4, 15
+; P9-LE-NEXT:    stxvx 0, 3, 4
+; P9-LE-NEXT:    stxv 0, 0(3)
+; P9-LE-NEXT:    blr
+;
+; P10-LE-LABEL: memset2TailV1B15:
+; P10-LE:       # %bb.0: # %entry
+; P10-LE-NEXT:    xxspltib 0, 165
+; P10-LE-NEXT:    pstxv 0, 15(3), 0
+; P10-LE-NEXT:    stxv 0, 0(3)
+; P10-LE-NEXT:    blr
+entry:
+  tail call void @llvm.memset.p0.i64(ptr %p, i8 165, i64 31, i1 false)
+  ret void
+}
+
+define dso_local void @memset2TailV1B12(ptr nocapture noundef writeonly %p) local_unnamed_addr {
+; P8-BE-LABEL: memset2TailV1B12:
+; P8-BE:       # %bb.0: # %entry
+; P8-BE-NEXT:    ld 4, L..C1(2) # %const.0
+; P8-BE-NEXT:    lxvw4x 0, 0, 4
+; P8-BE-NEXT:    li 4, 12
+; P8-BE-NEXT:    stxvw4x 0, 3, 4
+; P8-BE-NEXT:    stxvw4x 0, 0, 3
+; P8-BE-NEXT:    blr
+;
+; P9-BE-LABEL: memset2TailV1B12:
+; P9-BE:       # %bb.0: # %entry
+; P9-BE-NEXT:    xxspltib 0, 165
+; P9-BE-NEXT:    li 4, 12
+; P9-BE-NEXT:    stxvx 0, 3, 4
+; P9-BE-NEXT:    stxv 0, 0(3)
+; P9-BE-NEXT:    blr
+;
+; P10-BE-LABEL: memset2TailV1B12:
+; P10-BE:       # %bb.0: # %entry
+; P10-BE-NEXT:    xxspltib 0, 165
+; P10-BE-NEXT:    pstxv 0, 12(3), 0
+; P10-BE-NEXT:    stxv 0, 0(3)
+; P10-BE-NEXT:    blr
+;
+; P8-LE-LABEL: memset2TailV1B12:
+; P8-LE:       # %bb.0: # %entry
+; P8-LE-NEXT:    addis 4, 2, .LCPI10_0@toc@ha
+; P8-LE-NEXT:    addi 4, 4, .LCPI10_0@toc@l
+; P8-LE-NEXT:    lxvd2x 0, 0, 4
+; P8-LE-NEXT:    li 4, 12
+; P8-LE-NEXT:    stxvd2x 0, 3, 4
+; P8-LE-NEXT:    stxvd2x 0, 0, 3
+; P8-LE-NEXT:    blr
+;
+; P9-LE-LABEL: memset2TailV1B12:
+; P9-LE:       # %bb.0: # %entry
+; P9-LE-NEXT:    xxspltib 0, 165
+; P9-LE-NEXT:    li 4, 12
+; P9-LE-NEXT:    stxvx 0, 3, 4
+; P9-LE-NEXT:    stxv 0, 0(3)
+; P9-LE-NEXT:    blr
+;
+; P10-LE-LABEL: memset2TailV1B12:
+; P10-LE:       # %bb.0: # %entry
+; P10-LE-NEXT:    xxspltib 0, 165
+; P10-LE-NEXT:    pstxv 0, 12(3), 0
+; P10-LE-NEXT:    stxv 0, 0(3)
+; P10-LE-NEXT:    blr
+entry:
+  tail call void @llvm.memset.p0.i64(ptr %p, i8 165, i64 28, i1 false)
+  ret void
+}
+
+define dso_local void @memset2TailV1B9(ptr nocapture noundef writeonly %p) local_unnamed_addr {
+; P8-BE-LABEL: memset2TailV1B9:
+; P8-BE:       # %bb.0: # %entry
+; P8-BE-NEXT:    ld 4, L..C2(2) # %const.0
+; P8-BE-NEXT:    lxvw4x 0, 0, 4
+; P8-BE-NEXT:    li 4, 9
+; P8-BE-NEXT:    stxvw4x 0, 3, 4
+; P8-BE-NEXT:    stxvw4x 0, 0, 3
+; P8-BE-NEXT:    blr
+;
+; P9-BE-LABEL: memset2TailV1B9:
+; P9-BE:       # %bb.0: # %entry
+; P9-BE-NEXT:    xxspltib 0, 165
+; P9-BE-NEXT:    li 4, 9
+; P9-BE-NEXT:    stxvx 0, 3, 4
+; P9-BE-NEXT:    stxv 0, 0(3)
+; P9-BE-NEXT:    blr
+;
+; P10-BE-LABEL: memset2TailV1B9:
+; P10-BE:       # %bb.0: # %entry
+; P10-BE-NEXT:    xxspltib 0, 165
+; P10-BE-NEXT:    pstxv 0, 9(3), 0
+; P10-BE-NEXT:    stxv 0, 0(3)
+; P10-BE-NEXT:    blr
+;
+; P8-LE-LABEL: memset2TailV1B9:
+; P8-LE:       # %bb.0: # %entry
+; P8-LE-NEXT:    addis 4, 2, .LCPI11_0@toc@ha
+; P8-LE-NEXT:    addi 4, 4, .LCPI11_0@toc@l
+; P8-LE-NEXT:    lxvd2x 0, 0, 4
+; P8-LE-NEXT:    li 4, 9
+; P8-LE-NEXT:    stxvd2x 0, 3, 4
+; P8-LE-NEXT:    stxvd2x 0, 0, 3
+; P8-LE-NEXT:    blr
+;
+; P9-LE-LABEL: memset2TailV1B9:
+; P9-LE:       # %bb.0: # %entry
+; P9-LE-NEXT:    xxspltib 0, 165
+; P9-LE-NEXT:    li 4, 9
+; P9-LE-NEXT:    stxvx 0, 3, 4
+; P9-LE-NEXT:    stxv 0, 0(3)
+; P9-LE-NEXT:    blr
+;
+; P10-LE-LABEL: memset2TailV1B9:
+; P10-LE:       # %bb.0: # %entry
+; P10-LE-NEXT:    xxspltib 0, 165
+; P10-LE-NEXT:    pstxv 0, 9(3), 0
+; P10-LE-NEXT:    stxv 0, 0(3)
+; P10-LE-NEXT:    blr
+entry:
+  tail call void @llvm.memset.p0.i64(ptr %p, i8 165, i64 25, i1 false)
+  ret void
+}
+
+define dso_local void @memset2TailV1B8(ptr nocapture noundef writeonly %p) local_unnamed_addr {
+; P8-BE-LABEL: memset2TailV1B8:
+; P8-BE:       # %bb.0: # %entry
+; P8-BE-NEXT:    ld 4, L..C3(2) # %const.0
+; P8-BE-NEXT:    lxvw4x 0, 0, 4
+; P8-BE-NEXT:    lis 4, -23131
+; P8-BE-NEXT:    ori 4, 4, 42405
+; P8-BE-NEXT:    rldimi 4, 4, 32, 0
+; P8-BE-NEXT:    stxvw4x 0, 0, 3
+; P8-BE-NEXT:    std 4, 16(3)
+; P8-BE-NEXT:    blr
+;
+; P9-BE-LABEL: memset2TailV1B8:
+; P9-BE:       # %bb.0: # %entry
+; P9-BE-NEXT:    lis 4, -23131
+; P9-BE-NEXT:    xxspltib 0, 165
+; P9-BE-NEXT:    ori 4, 4, 42405
+; P9-BE-NEXT:    stxv 0, 0(3)
+; P9-BE-NEXT:    rldimi 4, 4, 32, 0
+; P9-BE-NEXT:    std 4, 16(3)
+; P9-BE-NEXT:    blr
+;
+; P10-BE-LABEL: memset2TailV1B8:
+; P10-BE:       # %bb.0: # %entry
+; P10-BE-NEXT:    pli 4, 2779096485
+; P10-BE-NEXT:    rldimi 4, 4, 32, 0
+; P10-BE-NEXT:    std 4, 16(3)
+; P10-BE-NEXT:    xxspltib 0, 165
+; P10-BE-NEXT:    stxv 0, 0(3)
+; P10-BE-NEXT:    blr
+;
+; P8-LE-LABEL: memset2TailV1B8:
+; P8-LE:       # %bb.0: # %entry
+; P8-LE-NEXT:    addis 4, 2, .LCPI12_0@toc@ha
+; P8-LE-NEXT:    addi 4, 4, .LCPI12_0@toc@l
+; P8-LE-NEXT:    lxvd2x 0, 0, 4
+; P8-LE-NEXT:    lis 4, -23131
+; P8-LE-NEXT:    ori 4, 4, 42405
+; P8-LE-NEXT:    rldimi 4, 4, 32, 0
+; P8-LE-NEXT:    std 4, 16(3)
+; P8-LE-NEXT:    stxvd2x 0, 0, 3
+; P8-LE-NEXT:    blr
+;
+; P9-LE-LABEL: memset2TailV1B8:
+; P9-LE:       # %bb.0: # %entry
+; P9-LE-NEXT:    lis 4, -23131
+; P9-LE-NEXT:    xxspltib 0, 165
+; P9-LE-NEXT:    ori 4, 4, 42405
+; P9-LE-NEXT:    stxv 0, 0(3)
+; P9-LE-NEXT:    rldimi 4, 4, 32, 0
+; P9-LE-NEXT:    std 4, 16(3)
+; P9-LE-NEXT:    blr
+;
+; P10-LE-LABEL: memset2TailV1B8:
+; P10-LE:       # %bb.0: # %entry
+; P10-LE-NEXT:    pli 4, 2779096485
+; P10-LE-NEXT:    rldimi 4, 4, 32, 0
+; P10-LE-NEXT:    std 4, 16(3)
+; P10-LE-NEXT:    xxspltib 0, 165
+; P10-LE-NEXT:    stxv 0, 0(3)
+; P10-LE-NEXT:    blr
+entry:
+  tail call void @llvm.memset.p0.i64(ptr %p, i8 165, i64 24, i1 false)
+  ret void
+}
+
+define dso_local void @memset2TailV1B7(ptr nocapture noundef writeonly %p) local_unnamed_addr {
+; P8-BE-LABEL: memset2TailV1B7:
+; P8-BE:       # %bb.0: # %entry
+; P8-BE-NEXT:    ld 4, L..C4(2) # %const.0
+; P8-BE-NEXT:    lis 5, -23131
+; P8-BE-NEXT:    lxvw4x 0, 0, 4
+; P8-BE-NEXT:    ori 4, 5, 42405
+; P8-BE-NEXT:    li 5, 15
+; P8-BE-NEXT:    rldimi 4, 4, 32, 0
+; P8-BE-NEXT:    stdx 4, 3, 5
+; P8-BE-NEXT:    stxvw4x 0, 0, 3
+; P8-BE-NEXT:    blr
+;
+; P9-BE-LABEL: memset2TailV1B7:
+; P9-BE:       # %bb.0: # %entry
+; P9-BE-NEXT:    lis 4, -23131
+; P9-BE-NEXT:    li 5, 15
+; P9-BE-NEXT:    ori 4, 4, 42405
+; P9-BE-NEXT:    rldimi 4, 4, 32, 0
+; P9-BE-NEXT:    stdx 4, 3, 5
+; P9-BE-NEXT:    xxspltib 0, 165
+; P9-BE-NEXT:    stxv 0, 0(3)
+; P9-BE-NEXT:    blr
+;
+; P10-BE-LABEL: memset2TailV1B7:
+; P10-BE:       # %bb.0: # %entry
+; P10-BE-NEXT:    pli 4, 2779096485
+; P10-BE-NEXT:    rldimi 4, 4, 32, 0
+; P10-BE-NEXT:    pstd 4, 15(3), 0
+; P10-BE-NEXT:    xxspltib 0, 165
+; P10-BE-NEXT:    stxv 0, 0(3)
+; P10-BE-NEXT:    blr
+;
+; P8-LE-LABEL: memset2TailV1B7:
+; P8-LE:       # %bb.0: # %entry
+; P8-LE-NEXT:    addis 4, 2, .LCPI13_0@toc@ha
+; P8-LE-NEXT:    lis 5, -23131
+; P8-LE-NEXT:    addi 4, 4, .LCPI13_0@toc@l
+; P8-LE-NEXT:    lxvd2x 0, 0, 4
+; P8-LE-NEXT:    ori 4, 5, 42405
+; P8-LE-NEXT:    li 5, 15
+; P8-LE-NEXT:    rldimi 4, 4, 32, 0
+; P8-LE-NEXT:    stdx 4, 3, 5
+; P8-LE-NEXT:    stxvd2x 0, 0, 3
+; P8-LE-NEXT:    blr
+;
+; P9-LE-LABEL: memset2TailV1B7:
+; P9-LE:       # %bb.0: # %entry
+; P9-LE-NEXT:    lis 4, -23131
+; P9-LE-NEXT:    li 5, 15
+; P9-LE-NEXT:    ori 4, 4, 42405
+; P9-LE-NEXT:    rldimi 4, 4, 32, 0
+; P9-LE-NEXT:    stdx 4, 3, 5
+; P9-LE-NEXT:    xxspltib 0, 165
+; P9-LE-NEXT:    stxv 0, 0(3)
+; P9-LE-NEXT:    blr
+;
+; P10-LE-LABEL: memset2TailV1B7:
+; P10-LE:       # %bb.0: # %entry
+; P10-LE-NEXT:    pli 4, 2779096485
+; P10-LE-NEXT:    rldimi 4, 4, 32, 0
+; P10-LE-NEXT:    pstd 4, 15(3), 0
+; P10-LE-NEXT:    xxspltib 0, 165
+; P10-LE-NEXT:    stxv 0, 0(3)
+; P10-LE-NEXT:    blr
+entry:
+  tail call void @llvm.memset.p0.i64(ptr %p, i8 165, i64 23, i1 false)
+  ret void
+}
+
+define dso_local void @memset2TailV1B4(ptr nocapture noundef writeonly %p) local_unnamed_addr {
+; P8-BE-LABEL: memset2TailV1B4:
+; P8-BE:       # %bb.0: # %entry
+; P8-BE-NEXT:    ld 4, L..C5(2) # %const.0
+; P8-BE-NEXT:    lxvw4x 0, 0, 4
+; P8-BE-NEXT:    lis 4, -23131
+; P8-BE-NEXT:    ori 4, 4, 42405
+; P8-BE-NEXT:    stw 4, 16(3)
+; P8-BE-NEXT:    stxvw4x 0, 0, 3
+; P8-BE-NEXT:    blr
+;
+; P9-BE-LABEL: memset2TailV1B4:
+; P9-BE:       # %bb.0: # %entry
+; P9-BE-NEXT:    lis 4, -23131
+; P9-BE-NEXT:    ori 4, 4, 42405
+; P9-BE-NEXT:    stw 4, 16(3)
+; P9-BE-NEXT:    xxspltib 0, 165
+; P9-BE-NEXT:    stxv 0, 0(3)
+; P9-BE-NEXT:    blr
+;
+; P10-BE-LABEL: memset2TailV1B4:
+; P10-BE:       # %bb.0: # %entry
+; P10-BE-NEXT:    pli 4, -1515870811
+; P10-BE-NEXT:    stw 4, 16(3)
+; P10-BE-NEXT:    xxspltib 0, 165
+; P10-BE-NEXT:    stxv 0, 0(3)
+; P10-BE-NEXT:    blr
+;
+; P8-LE-LABEL: memset2TailV1B4:
+; P8-LE:       # %bb.0: # %entry
+; P8-LE-NEXT:    addis 4, 2, .LCPI14_0@toc@ha
+; P8-LE-NEXT:    addi 4, 4, .LCPI14_0@toc@l
+; P8-LE-NEXT:    lxvd2x 0, 0, 4
+; P8-LE-NEXT:    lis 4, -23131
+; P8-LE-NEXT:    ori 4, 4, 42405
+; P8-LE-NEXT:    stw 4, 16(3)
+; P8-LE-NEXT:    stxvd2x 0, 0, 3
+; P8-LE-NEXT:    blr
+;
+; P9-LE-LABEL: memset2TailV1B4:
+; P9-LE:       # %bb.0: # %entry
+; P9-LE-NEXT:    lis 4, -23131
+; P9-LE-NEXT:    ori 4, 4, 42405
+; P9-LE-NEXT:    stw 4, 16(3)
+; P9-LE-NEXT:    xxspltib 0, 165
+; P9-LE-NEXT:    stxv 0, 0(3)
+; P9-LE-NEXT:    blr
+;
+; P10-LE-LABEL: memset2TailV1B4:
+; P10-LE:       # %bb.0: # %entry
+; P10-LE-NEXT:    pli 4, -1515870811
+; P10-LE-NEXT:    stw 4, 16(3)
+; P10-LE-NEXT:    xxspltib 0, 165
+; P10-LE-NEXT:    stxv 0, 0(3)
+; P10-LE-NEXT:    blr
+entry:
+  tail call void @llvm.memset.p0.i32(ptr %p, i8 165, i32 20, i1 false)
+  ret void
+}
+
+define dso_local void @memset2TailV1B3(ptr nocapture noundef writeonly %p) local_unnamed_addr {
+; P8-BE-LABEL: memset2TailV1B3:
+; P8-BE:       # %bb.0: # %entry
+; P8-BE-NEXT:    ld 4, L..C6(2) # %const.0
+; P8-BE-NEXT:    lxvw4x 0, 0, 4
+; P8-BE-NEXT:    lis 4, -23131
+; P8-BE-NEXT:    ori 4, 4, 42405
+; P8-BE-NEXT:    stw 4, 15(3)
+; P8-BE-NEXT:    stxvw4x 0, 0, 3
+; P8-BE-NEXT:    blr
+;
+; P9-BE-LABEL: memset2TailV1B3:
+; P9-BE:       # %bb.0: # %entry
+; P9-BE-NEXT:    lis 4, -23131
+; P9-BE-NEXT:    ori 4, 4, 42405
+; P9-BE-NEXT:    stw 4, 15(3)
+; P9-BE-NEXT:    xxspltib 0, 165
+; P9-BE-NEXT:    stxv 0, 0(3)
+; P9-BE-NEXT:    blr
+;
+; P10-BE-LABEL: memset2TailV1B3:
+; P10-BE:       # %bb.0: # %entry
+; P10-BE-NEXT:    pli 4, -1515870811
+; P10-BE-NEXT:    stw 4, 15(3)
+; P10-BE-NEXT:    xxspltib 0, 165
+; P10-BE-NEXT:    stxv 0, 0(3)
+; P10-BE-NEXT:    blr
+;
+; P8-LE-LABEL: memset2TailV1B3:
+; P8-LE:       # %bb.0: # %entry
+; P8-LE-NEXT:    addis 4, 2, .LCPI15_0@toc@ha
+; P8-LE-NEXT:    addi 4, 4, .LCPI15_0@toc@l
+; P8-LE-NEXT:    lxvd2x 0, 0, 4
+; P8-LE-NEXT:    lis 4, -23131
+; P8-LE-NEXT:    ori 4, 4, 42405
+; P8-LE-NEXT:    stw 4, 15(3)
+; P8-LE-NEXT:    stxvd2x 0, 0, 3
+; P8-LE-NEXT:    blr
+;
+; P9-LE-LABEL: memset2TailV1B3:
+; P9-LE:       # %bb.0: # %entry
+; P9-LE-NEXT:    lis 4, -23131
+; P9-LE-NEXT:    ori 4, 4, 42405
+; P9-LE-NEXT:    stw 4, 15(3)
+; P9-LE-NEXT:    xxspltib 0, 165
+; P9-LE-NEXT:    stxv 0, 0(3)
+; P9-LE-NEXT:    blr
+;
+; P10-LE-LABEL: memset2TailV1B3:
+; P10-LE:       # %bb.0: # %entry
+; P10-LE-NEXT:    pli 4, -1515870811
+; P10-LE-NEXT:    stw 4, 15(3)
+; P10-LE-NEXT:    xxspltib 0, 165
+; P10-LE-NEXT:    stxv 0, 0(3)
+; P10-LE-NEXT:    blr
+entry:
+  tail call void @llvm.memset.p0.i64(ptr %p, i8 165, i64 19, i1 false)
+  ret void
+}
+
+define dso_local void @memset2TailV1B2(ptr nocapture noundef writeonly %p) local_unnamed_addr {
+; P8-BE-LABEL: memset2TailV1B2:
+; P8-BE:       # %bb.0: # %entry
+; P8-BE-NEXT:    ld 4, L..C7(2) # %const.0
+; P8-BE-NEXT:    lxvw4x 0, 0, 4
+; P8-BE-NEXT:    li 4, -23131
+; P8-BE-NEXT:    sth 4, 16(3)
+; P8-BE-NEXT:    stxvw4x 0, 0, 3
+; P8-BE-NEXT:    blr
+;
+; P9-BE-LABEL: memset2TailV1B2:
+; P9-BE:       # %bb.0: # %entry
+; P9-BE-NEXT:    li 4, -23131
+; P9-BE-NEXT:    sth 4, 16(3)
+; P9-BE-NEXT:    xxspltib 0, 165
+; P9-BE-NEXT:    stxv 0, 0(3)
+; P9-BE-NEXT:    blr
+;
+; P10-BE-LABEL: memset2TailV1B2:
+; P10-BE:       # %bb.0: # %entry
+; P10-BE-NEXT:    li 4, -23131
+; P10-BE-NEXT:    sth 4, 16(3)
+; P10-BE-NEXT:    xxspltib 0, 165
+; P10-BE-NEXT:    stxv 0, 0(3)
+; P10-BE-NEXT:    blr
+;
+; P8-LE-LABEL: memset2TailV1B2:
+; P8-LE:       # %bb.0: # %entry
+; P8-LE-NEXT:    addis 4, 2, .LCPI16_0@toc@ha
+; P8-LE-NEXT:    addi 4, 4, .LCPI16_0@toc@l
+; P8-LE-NEXT:    lxvd2x 0, 0, 4
+; P8-LE-NEXT:    li 4, -23131
+; P8-LE-NEXT:    sth 4, 16(3)
+; P8-LE-NEXT:    stxvd2x 0, 0, 3
+; P8-LE-NEXT:    blr
+;
+; P9-LE-LABEL: memset2TailV1B2:
+; P9-LE:       # %bb.0: # %entry
+; P9-LE-NEXT:    li 4, -23131
+; P9-LE-NEXT:    sth 4, 16(3)
+; P9-LE-NEXT:    xxspltib 0, 165
+; P9-LE-NEXT:    stxv 0, 0(3)
+; P9-LE-NEXT:    blr
+;
+; P10-LE-LABEL: memset2TailV1B2:
+; P10-LE:       # %bb.0: # %entry
+; P10-LE-NEXT:    li 4, -23131
+; P10-LE-NEXT:    sth 4, 16(3)
+; P10-LE-NEXT:    xxspltib 0, 165
+; P10-LE-NEXT:    stxv 0, 0(3)
+; P10-LE-NEXT:    blr
+entry:
+  tail call void @llvm.memset.p0.i64(ptr %p, i8 165, i64 18, i1 false)
+  ret void
+}
+
+define dso_local void @memset2TailV1B1(ptr nocapture noundef writeonly %p) local_unnamed_addr {
+; P8-BE-LABEL: memset2TailV1B1:
+; P8-BE:       # %bb.0: # %entry
+; P8-BE-NEXT:    ld 4, L..C8(2) # %const.0
+; P8-BE-NEXT:    lxvw4x 0, 0, 4
+; P8-BE-NEXT:    li 4, -91
+; P8-BE-NEXT:    stb 4, 16(3)
+; P8-BE-NEXT:    stxvw4x 0, 0, 3
+; P8-BE-NEXT:    blr
+;
+; P9-BE-LABEL: memset2TailV1B1:
+; P9-BE:       # %bb.0: # %entry
+; P9-BE-NEXT:    li 4, -91
+; P9-BE-NEXT:    stb 4, 16(3)
+; P9-BE-NEXT:    xxspltib 0, 165
+; P9-BE-NEXT:    stxv 0, 0(3)
+; P9-BE-NEXT:    blr
+;
+; P10-BE-LABEL: memset2TailV1B1:
+; P10-BE:       # %bb.0: # %entry
+; P10-BE-NEXT:    li 4, -91
+; P10-BE-NEXT:    stb 4, 16(3)
+; P10-BE-NEXT:    xxspltib 0, 165
+; P10-BE-NEXT:    stxv 0, 0(3)
+; P10-BE-NEXT:    blr
+;
+; P8-LE-LABEL: memset2TailV1B1:
+; P8-LE:       # %bb.0: # %entry
+; P8-LE-NEXT:    addis 4, 2, .LCPI17_0@toc@ha
+; P8-LE-NEXT:    addi 4, 4, .LCPI17_0@toc@l
+; P8-LE-NEXT:    lxvd2x 0, 0, 4
+; P8-LE-NEXT:    li 4, -91
+; P8-LE-NEXT:    stb 4, 16(3)
+; P8-LE-NEXT:    stxvd2x 0, 0, 3
+; P8-LE-NEXT:    blr
+;
+; P9-LE-LABEL: memset2TailV1B1:
+; P9-LE:       # %bb.0: # %entry
+; P9-LE-NEXT:    li 4, -91
+; P9-LE-NEXT:    stb 4, 16(3)
+; P9-LE-NEXT:    xxspltib 0, 165
+; P9-LE-NEXT:    stxv 0, 0(3)
+; P9-LE-NEXT:    blr
+;
+; P10-LE-LABEL: memset2TailV1B1:
+; P10-LE:       # %bb.0: # %entry
+; P10-LE-NEXT:    li 4, -91
+; P10-LE-NEXT:    stb 4, 16(3)
+; P10-LE-NEXT:    xxspltib 0, 165
+; P10-LE-NEXT:    stxv 0, 0(3)
+; P10-LE-NEXT:    blr
+entry:
+  tail call void @llvm.memset.p0.i64(ptr %p, i8 165, i64 17, i1 false)
+  ret void
+}
+
+define dso_local void @memsetTailV0B15(ptr nocapture noundef writeonly %p) local_unnamed_addr {
+; P8-BE-LABEL: memsetTailV0B15:
+; P8-BE:       # %bb.0: # %entry
+; P8-BE-NEXT:    lis 4, 3855
+; P8-BE-NEXT:    li 5, 7
+; P8-BE-NEXT:    ori 4, 4, 3855
+; P8-BE-NEXT:    rldimi 4, 4, 32, 0
+; P8-BE-NEXT:    stdx 4, 3, 5
+; P8-BE-NEXT:    std 4, 0(3)
+; P8-BE-NEXT:    blr
+;
+; P9-BE-LABEL: memsetTailV0B15:
+; P9-BE:       # %bb.0: # %entry
+; P9-BE-NEXT:    lis 4, 3855
+; P9-BE-NEXT:    li 5, 7
+; P9-BE-NEXT:    ori 4, 4, 3855
+; P9-BE-NEXT:    rldimi 4, 4, 32, 0
+; P9-BE-NEXT:    stdx 4, 3, 5
+; P9-BE-NEXT:    std 4, 0(3)
+; P9-BE-NEXT:    blr
+;
+; P10-BE-LABEL: memsetTailV0B15:
+; P10-BE:       # %bb.0: # %entry
+; P10-BE-NEXT:    pli 4, 252645135
+; P10-BE-NEXT:    rldimi 4, 4, 32, 0
+; P10-BE-NEXT:    pstd 4, 7(3), 0
+; P10-BE-NEXT:    std 4, 0(3)
+; P10-BE-NEXT:    blr
+;
+; P8-LE-LABEL: memsetTailV0B15:
+; P8-LE:       # %bb.0: # %entry
+; P8-LE-NEXT:    lis 4, 3855
+; P8-LE-NEXT:    li 5, 7
+; P8-LE-NEXT:    ori 4, 4, 3855
+; P8-LE-NEXT:    rldimi 4, 4, 32, 0
+; P8-LE-NEXT:    stdx 4, 3, 5
+; P8-LE-NEXT:    std 4, 0(3)
+; P8-LE-NEXT:    blr
+;
+; P9-LE-LABEL: memsetTailV0B15:
+; P9-LE:       # %bb.0: # %entry
+; P9-LE-NEXT:    lis 4, 3855
+; P9-LE-NEXT:    li 5, 7
+; P9-LE-NEXT:    ori 4, 4, 3855
+; P9-LE-NEXT:    rldimi 4, 4, 32, 0
+; P9-LE-NEXT:    stdx 4, 3, 5
+; P9-LE-NEXT:    std 4, 0(3)
+; P9-LE-NEXT:    blr
+;
+; P10-LE-LABEL: memsetTailV0B15:
+; P10-LE:       # %bb.0: # %entry
+; P10-LE-NEXT:    pli 4, 252645135
+; P10-LE-NEXT:    rldimi 4, 4, 32, 0
+; P10-LE-NEXT:    pstd 4, 7(3), 0
+; P10-LE-NEXT:    std 4, 0(3)
+; P10-LE-NEXT:    blr
+entry:
+  tail call void @llvm.memset.p0.i64(ptr %p, i8 15, i64 15, i1 false)
+  ret void
+}
+
+define dso_local void @memsetTailV0B12(ptr nocapture noundef writeonly %p) local_unnamed_addr {
+; P8-BE-LABEL: memsetTailV0B12:
+; P8-BE:       # %bb.0: # %entry
+; P8-BE-NEXT:    lis 4, 3855
+; P8-BE-NEXT:    ori 4, 4, 3855
+; P8-BE-NEXT:    rldimi 4, 4, 32, 0
+; P8-BE-NEXT:    stw 4, 8(3)
+; P8-BE-NEXT:    std 4, 0(3)
+; P8-BE-NEXT:    blr
+;
+; P9-BE-LABEL: memsetTailV0B12:
+; P9-BE:       # %bb.0: # %entry
+; P9-BE-NEXT:    lis 4, 3855
+; P9-BE-NEXT:    ori 4, 4, 3855
+; P9-BE-NEXT:    rldimi 4, 4, 32, 0
+; P9-BE-NEXT:    stw 4, 8(3)
+; P9-BE-NEXT:    std 4, 0(3)
+; P9-BE-NEXT:    blr
+;
+; P10-BE-LABEL: memsetTailV0B12:
+; P10-BE:       # %bb.0: # %entry
+; P10-BE-NEXT:    pli 4, 252645135
+; P10-BE-NEXT:    rldimi 4, 4, 32, 0
+; P10-BE-NEXT:    stw 4, 8(3)
+; P10-BE-NEXT:    std 4, 0(3)
+; P10-BE-NEXT:    blr
+;
+; P8-LE-LABEL: memsetTailV0B12:
+; P8-LE:       # %bb.0: # %entry
+; P8-LE-NEXT:    lis 4, 3855
+; P8-LE-NEXT:    ori 4, 4, 3855
+; P8-LE-NEXT:    rldimi 4, 4, 32, 0
+; P8-LE-NEXT:    stw 4, 8(3)
+; P8-LE-NEXT:    std 4, 0(3)
+; P8-LE-NEXT:    blr
+;
+; P9-LE-LABEL: memsetTailV0B12:
+; P9-LE:       # %bb.0: # %entry
+; P9-LE-NEXT:    lis 4, 3855
+; P9-LE-NEXT:    ori 4, 4, 3855
+; P9-LE-NEXT:    rldimi 4, 4, 32, 0
+; P9-LE-NEXT:    stw 4, 8(3)
+; P9-LE-NEXT:    std 4, 0(3)
+; P9-LE-NEXT:    blr
+;
+; P10-LE-LABEL: memsetTailV0B12:
+; P10-LE:       # %bb.0: # %entry
+; P10-LE-NEXT:    pli 4, 252645135
+; P10-LE-NEXT:    rldimi 4, 4, 32, 0
+; P10-LE-NEXT:    stw 4, 8(3)
+; P10-LE-NEXT:    std 4, 0(3)
+; P10-LE-NEXT:    blr
+entry:
+  tail call void @llvm.memset.p0.i64(ptr %p, i8 15, i64 12, i1 false)
+  ret void
+}
+
+define dso_local void @memsetTailV0B9(ptr nocapture noundef writeonly %p) local_unnamed_addr {
+; P8-BE-LABEL: memsetTailV0B9:
+; P8-BE:       # %bb.0: # %entry
+; P8-BE-NEXT:    lis 4, 3855
+; P8-BE-NEXT:    li 5, 15
+; P8-BE-NEXT:    ori 4, 4, 3855
+; P8-BE-NEXT:    stb 5, 8(3)
+; P8-BE-NEXT:    rldimi 4, 4, 32, 0
+; P8-BE-NEXT:    std 4, 0(3)
+; P8-BE-NEXT:    blr
+;
+; P9-BE-LABEL: memsetTailV0B9:
+; P9-BE:       # %bb.0: # %entry
+; P9-BE-NEXT:    li 4, 15
+; P9-BE-NEXT:    stb 4, 8(3)
+; P9-BE-NEXT:    lis 4, 3855
+; P9-BE-NEXT:    ori 4, 4, 3855
+; P9-BE-NEXT:    rldimi 4, 4, 32, 0
+; P9-BE-NEXT:    std 4, 0(3)
+; P9-BE-NEXT:    blr
+;
+; P10-BE-LABEL: memsetTailV0B9:
+; P10-BE:       # %bb.0: # %entry
+; P10-BE-NEXT:    li 4, 15
+; P10-BE-NEXT:    stb 4, 8(3)
+; P10-BE-NEXT:    pli 4, 252645135
+; P10-BE-NEXT:    rldimi 4, 4, 32, 0
+; P10-BE-NEXT:    std 4, 0(3)
+; P10-BE-NEXT:    blr
+;
+; P8-LE-LABEL: memsetTailV0B9:
+; P8-LE:       # %bb.0: # %entry
+; P8-LE-NEXT:    lis 4, 3855
+; P8-LE-NEXT:    li 5, 15
+; P8-LE-NEXT:    ori 4, 4, 3855
+; P8-LE-NEXT:    stb 5, 8(3)
+; P8-LE-NEXT:    rldimi 4, 4, 32, 0
+; P8-LE-NEXT:    std 4, 0(3)
+; P8-LE-NEXT:    blr
+;
+; P9-LE-LABEL: memsetTailV0B9:
+; P9-LE:       # %bb.0: # %entry
+; P9-LE-NEXT:    li 4, 15
+; P9-LE-NEXT:    stb 4, 8(3)
+; P9-LE-NEXT:    lis 4, 3855
+; P9-LE-NEXT:    ori 4, 4, 3855
+; P9-LE-NEXT:    rldimi 4, 4, 32, 0
+; P9-LE-NEXT:    std 4, 0(3)
+; P9-LE-NEXT:    blr
+;
+; P10-LE-LABEL: memsetTailV0B9:
+; P10-LE:       # %bb.0: # %entry
+; P10-LE-NEXT:    li 4, 15
+; P10-LE-NEXT:    stb 4, 8(3)
+; P10-LE-NEXT:    pli 4, 252645135
+; P10-LE-NEXT:    rldimi 4, 4, 32, 0
+; P10-LE-NEXT:    std 4, 0(3)
+; P10-LE-NEXT:    blr
+entry:
+  tail call void @llvm.memset.p0.i64(ptr %p, i8 15, i64 9, i1 false)
+  ret void
+}
+
+define dso_local void @memsetTailV0B8(ptr nocapture noundef writeonly %p) local_unnamed_addr {
+; P8-BE-LABEL: memsetTailV0B8:
+; P8-BE:       # %bb.0: # %entry
+; P8-BE-NEXT:    lis 4, 3855
+; P8-BE-NEXT:    ori 4, 4, 3855
+; P8-BE-NEXT:    rldimi 4, 4, 32, 0
+; P8-BE-NEXT:    std 4, 0(3)
+; P8-BE-NEXT:    blr
+;
+; P9-BE-LABEL: memsetTailV0B8:
+; P9-BE:       # %bb.0: # %entry
+; P9-BE-NEXT:    lis 4, 3855
+; P9-BE-NEXT:    ori 4, 4, 3855
+; P9-BE-NEXT:    rldimi 4, 4, 32, 0
+; P9-BE-NEXT:    std 4, 0(3)
+; P9-BE-NEXT:    blr
+;
+; P10-BE-LABEL: memsetTailV0B8:
+; P10-BE:       # %bb.0: # %entry
+; P10-BE-NEXT:    pli 4, 252645135
+; P10-BE-NEXT:    rldimi 4, 4, 32, 0
+; P10-BE-NEXT:    std 4, 0(3)
+; P10-BE-NEXT:    blr
+;
+; P8-LE-LABEL: memsetTailV0B8:
+; P8-LE:       # %bb.0: # %entry
+; P8-LE-NEXT:    lis 4, 3855
+; P8-LE-NEXT:    ori 4, 4, 3855
+; P8-LE-NEXT:    rldimi 4, 4, 32, 0
+; P8-LE-NEXT:    std 4, 0(3)
+; P8-LE-NEXT:    blr
+;
+; P9-LE-LABEL: memsetTailV0B8:
+; P9-LE:       # %bb.0: # %entry
+; P9-LE-NEXT:    lis 4, 3855
+; P9-LE-NEXT:    ori 4, 4, 3855
+; P9-LE-NEXT:    rldimi 4, 4, 32, 0
+; P9-LE-NEXT:    std 4, 0(3)
+; P9-LE-NEXT:    blr
+;
+; P10-LE-LABEL: memsetTailV0B8:
+; P10-LE:       # %bb.0: # %entry
+; P10-LE-NEXT:    pli 4, 252645135
+; P10-LE-NEXT:    rldimi 4, 4, 32, 0
+; P10-LE-NEXT:    std 4, 0(3)
+; P10-LE-NEXT:    blr
+entry:
+  tail call void @llvm.memset.p0.i64(ptr %p, i8 15, i64 8, i1 false)
+  ret void
+}
+
+define dso_local void @memsetTailV0B7(ptr nocapture noundef writeonly %p) local_unnamed_addr {
+; P8-BE-LABEL: memsetTailV0B7:
+; P8-BE:       # %bb.0: # %entry
+; P8-BE-NEXT:    lis 4, 3855
+; P8-BE-NEXT:    ori 4, 4, 3855
+; P8-BE-NEXT:    stw 4, 3(3)
+; P8-BE-NEXT:    stw 4, 0(3)
+; P8-BE-NEXT:    blr
+;
+; P9-BE-LABEL: memsetTailV0B7:
+; P9-BE:       # %bb.0: # %entry
+; P9-BE-NEXT:    lis 4, 3855
+; P9-BE-NEXT:    ori 4, 4, 3855
+; P9-BE-NEXT:    stw 4, 3(3)
+; P9-BE-NEXT:    stw 4, 0(3)
+; P9-BE-NEXT:    blr
+;
+; P10-BE-LABEL: memsetTailV0B7:
+; P10-BE:       # %bb.0: # %entry
+; P10-BE-NEXT:    pli 4, 252645135
+; P10-BE-NEXT:    stw 4, 3(3)
+; P10-BE-NEXT:    stw 4, 0(3)
+; P10-BE-NEXT:    blr
+;
+; P8-LE-LABEL: memsetTailV0B7:
+; P8-LE:       # %bb.0: # %entry
+; P8-LE-NEXT:    lis 4, 3855
+; P8-LE-NEXT:    ori 4, 4, 3855
+; P8-LE-NEXT:    stw 4, 3(3)
+; P8-LE-NEXT:    stw 4, 0(3)
+; P8-LE-NEXT:    blr
+;
+; P9-LE-LABEL: memsetTailV0B7:
+; P9-LE:       # %bb.0: # %entry
+; P9-LE-NEXT:    lis 4, 3855
+; P9-LE-NEXT:    ori 4, 4, 3855
+; P9-LE-NEXT:    stw 4, 3(3)
+; P9-LE-NEXT:    stw 4, 0(3)
+; P9-LE-NEXT:    blr
+;
+; P10-LE-LABEL: memsetTailV0B7:
+; P10-LE:       # %bb.0: # %entry
+; P10-LE-NEXT:    pli 4, 252645135
+; P10-LE-NEXT:    stw 4, 3(3)
+; P10-LE-NEXT:    stw 4, 0(3)
+; P10-LE-NEXT:    blr
+entry:
+  tail call void @llvm.memset.p0.i64(ptr %p, i8 15, i64 7, i1 false)
+  ret void
+}
+
+define dso_local void @memsetTailV0B5(ptr nocapture noundef writeonly %p) local_unnamed_addr {
+; P8-BE-LABEL: memsetTailV0B5:
+; P8-BE:       # %bb.0: # %entry
+; P8-BE-NEXT:    lis 4, 3855
+; P8-BE-NEXT:    li 5, 15
+; P8-BE-NEXT:    ori 4, 4, 3855
+; P8-BE-NEXT:    stb 5, 4(3)
+; P8-BE-NEXT:    stw 4, 0(3)
+; P8-BE-NEXT:    blr
+;
+; P9-BE-LABEL: memsetTailV0B5:
+; P9-BE:       # %bb.0: # %entry
+; P9-BE-NEXT:    li 4, 15
+; P9-BE-NEXT:    stb 4, 4(3)
+; P9-BE-NEXT:    lis 4, 3855
+; P9-BE-NEXT:    ori 4, 4, 3855
+; P9-BE-NEXT:    stw 4, 0(3)
+; P9-BE-NEXT:    blr
+;
+; P10-BE-LABEL: memsetTailV0B5:
+; P10-BE:       # %bb.0: # %entry
+; P10-BE-NEXT:    li 4, 15
+; P10-BE-NEXT:    stb 4, 4(3)
+; P10-BE-NEXT:    pli 4, 252645135
+; P10-BE-NEXT:    stw 4, 0(3)
+; P10-BE-NEXT:    blr
+;
+; P8-LE-LABEL: memsetTailV0B5:
+; P8-LE:       # %bb.0: # %entry
+; P8-LE-NEXT:    lis 4, 3855
+; P8-LE-NEXT:    li 5, 15
+; P8-LE-NEXT:    ori 4, 4, 3855
+; P8-LE-NEXT:    stb 5, 4(3)
+; P8-LE-NEXT:    stw 4, 0(3)
+; P8-LE-NEXT:    blr
+;
+; P9-LE-LABEL: memsetTailV0B5:
+; P9-LE:       # %bb.0: # %entry
+; P9-LE-NEXT:    li 4, 15
+; P9-LE-NEXT:    stb 4, 4(3)
+; P9-LE-NEXT:    lis 4, 3855
+; P9-LE-NEXT:    ori 4, 4, 3855
+; P9-LE-NEXT:    stw 4, 0(3)
+; P9-LE-NEXT:    blr
+;
+; P10-LE-LABEL: memsetTailV0B5:
+; P10-LE:       # %bb.0: # %entry
+; P10-LE-NEXT:    li 4, 15
+; P10-LE-NEXT:    stb 4, 4(3)
+; P10-LE-NEXT:    pli 4, 252645135
+; P10-LE-NEXT:    stw 4, 0(3)
+; P10-LE-NEXT:    blr
+entry:
+  tail call void @llvm.memset.p0.i32(ptr %p, i8 15, i32 5, i1 false)
+  ret void
+}
+
+define dso_local void @memsetTailV0B4(ptr nocapture noundef writeonly %p) local_unnamed_addr {
+; P8-BE-LABEL: memsetTailV0B4:
+; P8-BE:       # %bb.0: # %entry
+; P8-BE-NEXT:    lis 4, 3855
+; P8-BE-NEXT:    ori 4, 4, 3855
+; P8-BE-NEXT:    stw 4, 0(3)
+; P8-BE-NEXT:    blr
+;
+; P9-BE-LABEL: memsetTailV0B4:
+; P9-BE:       # %bb.0: # %entry
+; P9-BE-NEXT:    lis 4, 3855
+; P9-BE-NEXT:    ori 4, 4, 3855
+; P9-BE-NEXT:    stw 4, 0(3)
+; P9-BE-NEXT:    blr
+;
+; P10-BE-LABEL: memsetTailV0B4:
+; P10-BE:       # %bb.0: # %entry
+; P10-BE-NEXT:    pli 4, 252645135
+; P10-BE-NEXT:    stw 4, 0(3)
+; P10-BE-NEXT:    blr
+;
+; P8-LE-LABEL: memsetTailV0B4:
+; P8-LE:       # %bb.0: # %entry
+; P8-LE-NEXT:    lis 4, 3855
+; P8-LE-NEXT:    ori 4, 4, 3855
+; P8-LE-NEXT:    stw 4, 0(3)
+; P8-LE-NEXT:    blr
+;
+; P9-LE-LABEL: memsetTailV0B4:
+; P9-LE:       # %bb.0: # %entry
+; P9-LE-NEXT:    lis 4, 3855
+; P9-LE-NEXT:    ori 4, 4, 3855
+; P9-LE-NEXT:    stw 4, 0(3)
+; P9-LE-NEXT:    blr
+;
+; P10-LE-LABEL: memsetTailV0B4:
+; P10-LE:       # %bb.0: # %entry
+; P10-LE-NEXT:    pli 4, 252645135
+; P10-LE-NEXT:    stw 4, 0(3)
+; P10-LE-NEXT:    blr
+entry:
+  tail call void @llvm.memset.p0.i32(ptr %p, i8 15, i32 4, i1 false)
+  ret void
+}
+
+define dso_local void @memsetTailV0B1(ptr nocapture noundef writeonly %p) local_unnamed_addr {
+; P8-BE-LABEL: memsetTailV0B1:
+; P8-BE:       # %bb.0: # %entry
+; P8-BE-NEXT:    li 4, -91
+; P8-BE-NEXT:    stb 4, 0(3)
+; P8-BE-NEXT:    blr
+;
+; P9-BE-LABEL: memsetTailV0B1:
+; P9-BE:       # %bb.0: # %entry
+; P9-BE-NEXT:    li 4, -91
+; P9-BE-NEXT:    stb 4, 0(3)
+; P9-BE-NEXT:    blr
+;
+; P10-BE-LABEL: memsetTailV0B1:
+; P10-BE:       # %bb.0: # %entry
+; P10-BE-NEXT:    li 4, -91
+; P10-BE-NEXT:    stb 4, 0(3)
+; P10-BE-NEXT:    blr
+;
+; P8-LE-LABEL: memsetTailV0B1:
+; P8-LE:       # %bb.0: # %entry
+; P8-LE-NEXT:    li 4, -91
+; P8-LE-NEXT:    stb 4, 0(3)
+; P8-LE-NEXT:    blr
+;
+; P9-LE-LABEL: memsetTailV0B1:
+; P9-LE:       # %bb.0: # %entry
+; P9-LE-NEXT:    li 4, -91
+; P9-LE-NEXT:    stb 4, 0(3)
+; P9-LE-NEXT:    blr
+;
+; P10-LE-LABEL: memsetTailV0B1:
+; P10-LE:       # %bb.0: # %entry
+; P10-LE-NEXT:    li 4, -91
+; P10-LE-NEXT:    stb 4, 0(3)
+; P10-LE-NEXT:    blr
+entry:
+  tail call void @llvm.memset.p0.i64(ptr %p, i8 165, i64 1, i1 false)
+  ret void
+}
+
+declare void @llvm.memset.p0.i32(ptr nocapture writeonly, i8, i32, i1 immarg)
+declare void @llvm.memset.p0.i64(ptr nocapture writeonly, i8, i64, i1 immarg)