[SystemZ] Fix codegen for _[u]128 intrinsics

PR #74625 introduced a regression in the code generated for the following set of intrinsic: vec_add_u128, vec_addc_u128, vec_adde_u128, vec_addec_u128 vec_sub_u128, vec_subc_u128, vec_sube_u128, vec_subec_u128 vec_sum_u128, vec_msum_u128 vec_gfmsum_128, vec_gfmsum_accum_128 This is because the new code incorrectly assumed that a cast from "unsigned __int128" to "vector unsigned char" would simply be a bitcast re-interpretation; instead, this cast actually truncates the __int128 to char and splats the result. Fixed by adding an intermediate cast via a single-element 128-bit integer vector. Fixes: #109113 (cherry picked from commit baf9b7d)
llvm · Oct 11, 2024 · 149884a · 149884a
1 parent f3f4952
commit 149884a
Show file tree

Hide file tree

Showing 2 changed files with 188 additions and 5 deletions.
diff --git a/clang/lib/Headers/vecintrin.h b/clang/lib/Headers/vecintrin.h
@@ -8359,7 +8359,9 @@ vec_min(__vector double __a, __vector double __b) {
 
 static inline __ATTRS_ai __vector unsigned char
 vec_add_u128(__vector unsigned char __a, __vector unsigned char __b) {
-  return (__vector unsigned char)((__int128)__a + (__int128)__b);
+  return (__vector unsigned char)
+         (unsigned __int128 __attribute__((__vector_size__(16))))
+         ((__int128)__a + (__int128)__b);
 }
 
 /*-- vec_addc ---------------------------------------------------------------*/
@@ -8389,6 +8391,7 @@ vec_addc(__vector unsigned long long __a, __vector unsigned long long __b) {
 static inline __ATTRS_ai __vector unsigned char
 vec_addc_u128(__vector unsigned char __a, __vector unsigned char __b) {
   return (__vector unsigned char)
+         (unsigned __int128 __attribute__((__vector_size__(16))))
          __builtin_s390_vaccq((unsigned __int128)__a, (unsigned __int128)__b);
 }
 
@@ -8398,6 +8401,7 @@ static inline __ATTRS_ai __vector unsigned char
 vec_adde_u128(__vector unsigned char __a, __vector unsigned char __b,
               __vector unsigned char __c) {
   return (__vector unsigned char)
+         (unsigned __int128 __attribute__((__vector_size__(16))))
          __builtin_s390_vacq((unsigned __int128)__a, (unsigned __int128)__b,
                              (unsigned __int128)__c);
 }
@@ -8408,6 +8412,7 @@ static inline __ATTRS_ai __vector unsigned char
 vec_addec_u128(__vector unsigned char __a, __vector unsigned char __b,
                __vector unsigned char __c) {
   return (__vector unsigned char)
+         (unsigned __int128 __attribute__((__vector_size__(16))))
          __builtin_s390_vacccq((unsigned __int128)__a, (unsigned __int128)__b,
                                (unsigned __int128)__c);
 }
@@ -8483,7 +8488,9 @@ vec_gfmsum(__vector unsigned int __a, __vector unsigned int __b) {
 static inline __ATTRS_o_ai __vector unsigned char
 vec_gfmsum_128(__vector unsigned long long __a,
                __vector unsigned long long __b) {
-  return (__vector unsigned char)__builtin_s390_vgfmg(__a, __b);
+  return (__vector unsigned char)
+         (unsigned __int128 __attribute__((__vector_size__(16))))
+         __builtin_s390_vgfmg(__a, __b);
 }
 
 /*-- vec_gfmsum_accum -------------------------------------------------------*/
@@ -8513,6 +8520,7 @@ vec_gfmsum_accum_128(__vector unsigned long long __a,
                      __vector unsigned long long __b,
                      __vector unsigned char __c) {
   return (__vector unsigned char)
+         (unsigned __int128 __attribute__((__vector_size__(16))))
          __builtin_s390_vgfmag(__a, __b, (unsigned __int128)__c);
 }
 
@@ -8810,14 +8818,17 @@ vec_msum_u128(__vector unsigned long long __a, __vector unsigned long long __b,
 
 #define vec_msum_u128(X, Y, Z, W) \
   ((__typeof__((vec_msum_u128)((X), (Y), (Z), (W)))) \
+   (unsigned __int128 __attribute__((__vector_size__(16)))) \
    __builtin_s390_vmslg((X), (Y), (unsigned __int128)(Z), (W)))
 #endif
 
 /*-- vec_sub_u128 -----------------------------------------------------------*/
 
 static inline __ATTRS_ai __vector unsigned char
 vec_sub_u128(__vector unsigned char __a, __vector unsigned char __b) {
-  return (__vector unsigned char)((__int128)__a - (__int128)__b);
+  return (__vector unsigned char)
+         (unsigned __int128 __attribute__((__vector_size__(16))))
+         ((__int128)__a - (__int128)__b);
 }
 
 /*-- vec_subc ---------------------------------------------------------------*/
@@ -8847,6 +8858,7 @@ vec_subc(__vector unsigned long long __a, __vector unsigned long long __b) {
 static inline __ATTRS_ai __vector unsigned char
 vec_subc_u128(__vector unsigned char __a, __vector unsigned char __b) {
   return (__vector unsigned char)
+         (unsigned __int128 __attribute__((__vector_size__(16))))
          __builtin_s390_vscbiq((unsigned __int128)__a, (unsigned __int128)__b);
 }
 
@@ -8856,6 +8868,7 @@ static inline __ATTRS_ai __vector unsigned char
 vec_sube_u128(__vector unsigned char __a, __vector unsigned char __b,
               __vector unsigned char __c) {
   return (__vector unsigned char)
+         (unsigned __int128 __attribute__((__vector_size__(16))))
          __builtin_s390_vsbiq((unsigned __int128)__a, (unsigned __int128)__b,
                               (unsigned __int128)__c);
 }
@@ -8866,6 +8879,7 @@ static inline __ATTRS_ai __vector unsigned char
 vec_subec_u128(__vector unsigned char __a, __vector unsigned char __b,
                __vector unsigned char __c) {
   return (__vector unsigned char)
+         (unsigned __int128 __attribute__((__vector_size__(16))))
          __builtin_s390_vsbcbiq((unsigned __int128)__a, (unsigned __int128)__b,
                                 (unsigned __int128)__c);
 }
@@ -8886,12 +8900,16 @@ vec_sum2(__vector unsigned int __a, __vector unsigned int __b) {
 
 static inline __ATTRS_o_ai __vector unsigned char
 vec_sum_u128(__vector unsigned int __a, __vector unsigned int __b) {
-  return (__vector unsigned char)__builtin_s390_vsumqf(__a, __b);
+  return (__vector unsigned char)
+         (unsigned __int128 __attribute__((__vector_size__(16))))
+         __builtin_s390_vsumqf(__a, __b);
 }
 
 static inline __ATTRS_o_ai __vector unsigned char
 vec_sum_u128(__vector unsigned long long __a, __vector unsigned long long __b) {
-  return (__vector unsigned char)__builtin_s390_vsumqg(__a, __b);
+  return (__vector unsigned char)
+         (unsigned __int128 __attribute__((__vector_size__(16))))
+         __builtin_s390_vsumqg(__a, __b);
 }
 
 /*-- vec_sum4 ---------------------------------------------------------------*/

diff --git a/clang/test/CodeGen/SystemZ/builtins-systemz-i128.c b/clang/test/CodeGen/SystemZ/builtins-systemz-i128.c
@@ -0,0 +1,165 @@
+// NOTE: Assertions have been autogenerated by utils/update_cc_test_checks.py UTC_ARGS: --version 5
+// REQUIRES: systemz-registered-target
+// RUN: %clang_cc1 -target-cpu z14 -triple s390x-linux-gnu \
+// RUN: -O2 -fzvector -flax-vector-conversions=none \
+// RUN: -Wall -Wno-unused -Werror -emit-llvm %s -o - | FileCheck %s
+
+#include <vecintrin.h>
+
+volatile vector unsigned char vuc;
+volatile vector unsigned short vus;
+volatile vector unsigned int vui;
+volatile vector unsigned long long vul;
+
+// CHECK-LABEL: define dso_local void @test(
+// CHECK-SAME: ) local_unnamed_addr #[[ATTR0:[0-9]+]] {
+// CHECK-NEXT:  [[ENTRY:.*:]]
+// CHECK-NEXT:    [[TMP0:%.*]] = load volatile <16 x i8>, ptr @vuc, align 8, !tbaa [[TBAA3:![0-9]+]]
+// CHECK-NEXT:    [[TMP1:%.*]] = load volatile <16 x i8>, ptr @vuc, align 8, !tbaa [[TBAA3]]
+// CHECK-NEXT:    [[TMP2:%.*]] = bitcast <16 x i8> [[TMP0]] to i128
+// CHECK-NEXT:    [[TMP3:%.*]] = bitcast <16 x i8> [[TMP1]] to i128
+// CHECK-NEXT:    [[ADD_I:%.*]] = add nsw i128 [[TMP3]], [[TMP2]]
+// CHECK-NEXT:    [[TMP4:%.*]] = bitcast i128 [[ADD_I]] to <16 x i8>
+// CHECK-NEXT:    store volatile <16 x i8> [[TMP4]], ptr @vuc, align 8, !tbaa [[TBAA3]]
+// CHECK-NEXT:    [[TMP5:%.*]] = load volatile <16 x i8>, ptr @vuc, align 8, !tbaa [[TBAA3]]
+// CHECK-NEXT:    [[TMP6:%.*]] = load volatile <16 x i8>, ptr @vuc, align 8, !tbaa [[TBAA3]]
+// CHECK-NEXT:    [[TMP7:%.*]] = bitcast <16 x i8> [[TMP5]] to i128
+// CHECK-NEXT:    [[TMP8:%.*]] = bitcast <16 x i8> [[TMP6]] to i128
+// CHECK-NEXT:    [[TMP9:%.*]] = tail call i128 @llvm.s390.vaccq(i128 [[TMP7]], i128 [[TMP8]])
+// CHECK-NEXT:    [[TMP10:%.*]] = bitcast i128 [[TMP9]] to <16 x i8>
+// CHECK-NEXT:    store volatile <16 x i8> [[TMP10]], ptr @vuc, align 8, !tbaa [[TBAA3]]
+// CHECK-NEXT:    [[TMP11:%.*]] = load volatile <16 x i8>, ptr @vuc, align 8, !tbaa [[TBAA3]]
+// CHECK-NEXT:    [[TMP12:%.*]] = load volatile <16 x i8>, ptr @vuc, align 8, !tbaa [[TBAA3]]
+// CHECK-NEXT:    [[TMP13:%.*]] = load volatile <16 x i8>, ptr @vuc, align 8, !tbaa [[TBAA3]]
+// CHECK-NEXT:    [[TMP14:%.*]] = bitcast <16 x i8> [[TMP11]] to i128
+// CHECK-NEXT:    [[TMP15:%.*]] = bitcast <16 x i8> [[TMP12]] to i128
+// CHECK-NEXT:    [[TMP16:%.*]] = bitcast <16 x i8> [[TMP13]] to i128
+// CHECK-NEXT:    [[TMP17:%.*]] = tail call i128 @llvm.s390.vacq(i128 [[TMP14]], i128 [[TMP15]], i128 [[TMP16]])
+// CHECK-NEXT:    [[TMP18:%.*]] = bitcast i128 [[TMP17]] to <16 x i8>
+// CHECK-NEXT:    store volatile <16 x i8> [[TMP18]], ptr @vuc, align 8, !tbaa [[TBAA3]]
+// CHECK-NEXT:    [[TMP19:%.*]] = load volatile <16 x i8>, ptr @vuc, align 8, !tbaa [[TBAA3]]
+// CHECK-NEXT:    [[TMP20:%.*]] = load volatile <16 x i8>, ptr @vuc, align 8, !tbaa [[TBAA3]]
+// CHECK-NEXT:    [[TMP21:%.*]] = load volatile <16 x i8>, ptr @vuc, align 8, !tbaa [[TBAA3]]
+// CHECK-NEXT:    [[TMP22:%.*]] = bitcast <16 x i8> [[TMP19]] to i128
+// CHECK-NEXT:    [[TMP23:%.*]] = bitcast <16 x i8> [[TMP20]] to i128
+// CHECK-NEXT:    [[TMP24:%.*]] = bitcast <16 x i8> [[TMP21]] to i128
+// CHECK-NEXT:    [[TMP25:%.*]] = tail call i128 @llvm.s390.vacccq(i128 [[TMP22]], i128 [[TMP23]], i128 [[TMP24]])
+// CHECK-NEXT:    [[TMP26:%.*]] = bitcast i128 [[TMP25]] to <16 x i8>
+// CHECK-NEXT:    store volatile <16 x i8> [[TMP26]], ptr @vuc, align 8, !tbaa [[TBAA3]]
+// CHECK-NEXT:    [[TMP27:%.*]] = load volatile <16 x i8>, ptr @vuc, align 8, !tbaa [[TBAA3]]
+// CHECK-NEXT:    [[TMP28:%.*]] = load volatile <16 x i8>, ptr @vuc, align 8, !tbaa [[TBAA3]]
+// CHECK-NEXT:    [[TMP29:%.*]] = bitcast <16 x i8> [[TMP27]] to i128
+// CHECK-NEXT:    [[TMP30:%.*]] = bitcast <16 x i8> [[TMP28]] to i128
+// CHECK-NEXT:    [[SUB_I:%.*]] = sub nsw i128 [[TMP29]], [[TMP30]]
+// CHECK-NEXT:    [[TMP31:%.*]] = bitcast i128 [[SUB_I]] to <16 x i8>
+// CHECK-NEXT:    store volatile <16 x i8> [[TMP31]], ptr @vuc, align 8, !tbaa [[TBAA3]]
+// CHECK-NEXT:    [[TMP32:%.*]] = load volatile <16 x i8>, ptr @vuc, align 8, !tbaa [[TBAA3]]
+// CHECK-NEXT:    [[TMP33:%.*]] = load volatile <16 x i8>, ptr @vuc, align 8, !tbaa [[TBAA3]]
+// CHECK-NEXT:    [[TMP34:%.*]] = bitcast <16 x i8> [[TMP32]] to i128
+// CHECK-NEXT:    [[TMP35:%.*]] = bitcast <16 x i8> [[TMP33]] to i128
+// CHECK-NEXT:    [[TMP36:%.*]] = tail call i128 @llvm.s390.vscbiq(i128 [[TMP34]], i128 [[TMP35]])
+// CHECK-NEXT:    [[TMP37:%.*]] = bitcast i128 [[TMP36]] to <16 x i8>
+// CHECK-NEXT:    store volatile <16 x i8> [[TMP37]], ptr @vuc, align 8, !tbaa [[TBAA3]]
+// CHECK-NEXT:    [[TMP38:%.*]] = load volatile <16 x i8>, ptr @vuc, align 8, !tbaa [[TBAA3]]
+// CHECK-NEXT:    [[TMP39:%.*]] = load volatile <16 x i8>, ptr @vuc, align 8, !tbaa [[TBAA3]]
+// CHECK-NEXT:    [[TMP40:%.*]] = load volatile <16 x i8>, ptr @vuc, align 8, !tbaa [[TBAA3]]
+// CHECK-NEXT:    [[TMP41:%.*]] = bitcast <16 x i8> [[TMP38]] to i128
+// CHECK-NEXT:    [[TMP42:%.*]] = bitcast <16 x i8> [[TMP39]] to i128
+// CHECK-NEXT:    [[TMP43:%.*]] = bitcast <16 x i8> [[TMP40]] to i128
+// CHECK-NEXT:    [[TMP44:%.*]] = tail call i128 @llvm.s390.vsbiq(i128 [[TMP41]], i128 [[TMP42]], i128 [[TMP43]])
+// CHECK-NEXT:    [[TMP45:%.*]] = bitcast i128 [[TMP44]] to <16 x i8>
+// CHECK-NEXT:    store volatile <16 x i8> [[TMP45]], ptr @vuc, align 8, !tbaa [[TBAA3]]
+// CHECK-NEXT:    [[TMP46:%.*]] = load volatile <16 x i8>, ptr @vuc, align 8, !tbaa [[TBAA3]]
+// CHECK-NEXT:    [[TMP47:%.*]] = load volatile <16 x i8>, ptr @vuc, align 8, !tbaa [[TBAA3]]
+// CHECK-NEXT:    [[TMP48:%.*]] = load volatile <16 x i8>, ptr @vuc, align 8, !tbaa [[TBAA3]]
+// CHECK-NEXT:    [[TMP49:%.*]] = bitcast <16 x i8> [[TMP46]] to i128
+// CHECK-NEXT:    [[TMP50:%.*]] = bitcast <16 x i8> [[TMP47]] to i128
+// CHECK-NEXT:    [[TMP51:%.*]] = bitcast <16 x i8> [[TMP48]] to i128
+// CHECK-NEXT:    [[TMP52:%.*]] = tail call i128 @llvm.s390.vsbcbiq(i128 [[TMP49]], i128 [[TMP50]], i128 [[TMP51]])
+// CHECK-NEXT:    [[TMP53:%.*]] = bitcast i128 [[TMP52]] to <16 x i8>
+// CHECK-NEXT:    store volatile <16 x i8> [[TMP53]], ptr @vuc, align 8, !tbaa [[TBAA3]]
+// CHECK-NEXT:    [[TMP54:%.*]] = load volatile <4 x i32>, ptr @vui, align 8, !tbaa [[TBAA3]]
+// CHECK-NEXT:    [[TMP55:%.*]] = load volatile <4 x i32>, ptr @vui, align 8, !tbaa [[TBAA3]]
+// CHECK-NEXT:    [[TMP56:%.*]] = tail call i128 @llvm.s390.vsumqf(<4 x i32> [[TMP54]], <4 x i32> [[TMP55]])
+// CHECK-NEXT:    [[TMP57:%.*]] = bitcast i128 [[TMP56]] to <16 x i8>
+// CHECK-NEXT:    store volatile <16 x i8> [[TMP57]], ptr @vuc, align 8, !tbaa [[TBAA3]]
+// CHECK-NEXT:    [[TMP58:%.*]] = load volatile <2 x i64>, ptr @vul, align 8, !tbaa [[TBAA3]]
+// CHECK-NEXT:    [[TMP59:%.*]] = load volatile <2 x i64>, ptr @vul, align 8, !tbaa [[TBAA3]]
+// CHECK-NEXT:    [[TMP60:%.*]] = tail call i128 @llvm.s390.vsumqg(<2 x i64> [[TMP58]], <2 x i64> [[TMP59]])
+// CHECK-NEXT:    [[TMP61:%.*]] = bitcast i128 [[TMP60]] to <16 x i8>
+// CHECK-NEXT:    store volatile <16 x i8> [[TMP61]], ptr @vuc, align 8, !tbaa [[TBAA3]]
+// CHECK-NEXT:    [[TMP62:%.*]] = load volatile <2 x i64>, ptr @vul, align 8, !tbaa [[TBAA3]]
+// CHECK-NEXT:    [[TMP63:%.*]] = load volatile <2 x i64>, ptr @vul, align 8, !tbaa [[TBAA3]]
+// CHECK-NEXT:    [[TMP64:%.*]] = tail call i128 @llvm.s390.vgfmg(<2 x i64> [[TMP62]], <2 x i64> [[TMP63]])
+// CHECK-NEXT:    [[TMP65:%.*]] = bitcast i128 [[TMP64]] to <16 x i8>
+// CHECK-NEXT:    store volatile <16 x i8> [[TMP65]], ptr @vuc, align 8, !tbaa [[TBAA3]]
+// CHECK-NEXT:    [[TMP66:%.*]] = load volatile <2 x i64>, ptr @vul, align 8, !tbaa [[TBAA3]]
+// CHECK-NEXT:    [[TMP67:%.*]] = load volatile <2 x i64>, ptr @vul, align 8, !tbaa [[TBAA3]]
+// CHECK-NEXT:    [[TMP68:%.*]] = load volatile <16 x i8>, ptr @vuc, align 8, !tbaa [[TBAA3]]
+// CHECK-NEXT:    [[TMP69:%.*]] = bitcast <16 x i8> [[TMP68]] to i128
+// CHECK-NEXT:    [[TMP70:%.*]] = tail call i128 @llvm.s390.vgfmag(<2 x i64> [[TMP66]], <2 x i64> [[TMP67]], i128 [[TMP69]])
+// CHECK-NEXT:    [[TMP71:%.*]] = bitcast i128 [[TMP70]] to <16 x i8>
+// CHECK-NEXT:    store volatile <16 x i8> [[TMP71]], ptr @vuc, align 8, !tbaa [[TBAA3]]
+// CHECK-NEXT:    [[TMP72:%.*]] = load volatile <2 x i64>, ptr @vul, align 8, !tbaa [[TBAA3]]
+// CHECK-NEXT:    [[TMP73:%.*]] = load volatile <2 x i64>, ptr @vul, align 8, !tbaa [[TBAA3]]
+// CHECK-NEXT:    [[TMP74:%.*]] = load volatile <16 x i8>, ptr @vuc, align 8, !tbaa [[TBAA3]]
+// CHECK-NEXT:    [[TMP75:%.*]] = bitcast <16 x i8> [[TMP74]] to i128
+// CHECK-NEXT:    [[TMP76:%.*]] = tail call i128 @llvm.s390.vmslg(<2 x i64> [[TMP72]], <2 x i64> [[TMP73]], i128 [[TMP75]], i32 0)
+// CHECK-NEXT:    [[TMP77:%.*]] = bitcast i128 [[TMP76]] to <16 x i8>
+// CHECK-NEXT:    store volatile <16 x i8> [[TMP77]], ptr @vuc, align 8, !tbaa [[TBAA3]]
+// CHECK-NEXT:    [[TMP78:%.*]] = load volatile <2 x i64>, ptr @vul, align 8, !tbaa [[TBAA3]]
+// CHECK-NEXT:    [[TMP79:%.*]] = load volatile <2 x i64>, ptr @vul, align 8, !tbaa [[TBAA3]]
+// CHECK-NEXT:    [[TMP80:%.*]] = load volatile <16 x i8>, ptr @vuc, align 8, !tbaa [[TBAA3]]
+// CHECK-NEXT:    [[TMP81:%.*]] = bitcast <16 x i8> [[TMP80]] to i128
+// CHECK-NEXT:    [[TMP82:%.*]] = tail call i128 @llvm.s390.vmslg(<2 x i64> [[TMP78]], <2 x i64> [[TMP79]], i128 [[TMP81]], i32 4)
+// CHECK-NEXT:    [[TMP83:%.*]] = bitcast i128 [[TMP82]] to <16 x i8>
+// CHECK-NEXT:    store volatile <16 x i8> [[TMP83]], ptr @vuc, align 8, !tbaa [[TBAA3]]
+// CHECK-NEXT:    [[TMP84:%.*]] = load volatile <2 x i64>, ptr @vul, align 8, !tbaa [[TBAA3]]
+// CHECK-NEXT:    [[TMP85:%.*]] = load volatile <2 x i64>, ptr @vul, align 8, !tbaa [[TBAA3]]
+// CHECK-NEXT:    [[TMP86:%.*]] = load volatile <16 x i8>, ptr @vuc, align 8, !tbaa [[TBAA3]]
+// CHECK-NEXT:    [[TMP87:%.*]] = bitcast <16 x i8> [[TMP86]] to i128
+// CHECK-NEXT:    [[TMP88:%.*]] = tail call i128 @llvm.s390.vmslg(<2 x i64> [[TMP84]], <2 x i64> [[TMP85]], i128 [[TMP87]], i32 8)
+// CHECK-NEXT:    [[TMP89:%.*]] = bitcast i128 [[TMP88]] to <16 x i8>
+// CHECK-NEXT:    store volatile <16 x i8> [[TMP89]], ptr @vuc, align 8, !tbaa [[TBAA3]]
+// CHECK-NEXT:    [[TMP90:%.*]] = load volatile <2 x i64>, ptr @vul, align 8, !tbaa [[TBAA3]]
+// CHECK-NEXT:    [[TMP91:%.*]] = load volatile <2 x i64>, ptr @vul, align 8, !tbaa [[TBAA3]]
+// CHECK-NEXT:    [[TMP92:%.*]] = load volatile <16 x i8>, ptr @vuc, align 8, !tbaa [[TBAA3]]
+// CHECK-NEXT:    [[TMP93:%.*]] = bitcast <16 x i8> [[TMP92]] to i128
+// CHECK-NEXT:    [[TMP94:%.*]] = tail call i128 @llvm.s390.vmslg(<2 x i64> [[TMP90]], <2 x i64> [[TMP91]], i128 [[TMP93]], i32 12)
+// CHECK-NEXT:    [[TMP95:%.*]] = bitcast i128 [[TMP94]] to <16 x i8>
+// CHECK-NEXT:    store volatile <16 x i8> [[TMP95]], ptr @vuc, align 8, !tbaa [[TBAA3]]
+// CHECK-NEXT:    [[TMP96:%.*]] = load volatile <16 x i8>, ptr @vuc, align 8, !tbaa [[TBAA3]]
+// CHECK-NEXT:    [[TMP97:%.*]] = load volatile <16 x i8>, ptr @vuc, align 8, !tbaa [[TBAA3]]
+// CHECK-NEXT:    [[TMP98:%.*]] = tail call <2 x i64> @llvm.s390.vbperm(<16 x i8> [[TMP96]], <16 x i8> [[TMP97]])
+// CHECK-NEXT:    store volatile <2 x i64> [[TMP98]], ptr @vul, align 8, !tbaa [[TBAA3]]
+// CHECK-NEXT:    ret void
+//
+void test(void) {
+  vuc = vec_add_u128(vuc, vuc);
+  vuc = vec_addc_u128(vuc, vuc);
+  vuc = vec_adde_u128(vuc, vuc, vuc);
+  vuc = vec_addec_u128(vuc, vuc, vuc);
+
+  vuc = vec_sub_u128(vuc, vuc);
+  vuc = vec_subc_u128(vuc, vuc);
+  vuc = vec_sube_u128(vuc, vuc, vuc);
+  vuc = vec_subec_u128(vuc, vuc, vuc);
+
+  vuc = vec_sum_u128(vui, vui);
+  vuc = vec_sum_u128(vul, vul);
+
+  vuc = vec_gfmsum_128(vul, vul);
+  vuc = vec_gfmsum_accum_128(vul, vul, vuc);
+
+  vuc = vec_msum_u128(vul, vul, vuc, 0);
+  vuc = vec_msum_u128(vul, vul, vuc, 4);
+  vuc = vec_msum_u128(vul, vul, vuc, 8);
+  vuc = vec_msum_u128(vul, vul, vuc, 12);
+
+  vul = vec_bperm_u128(vuc, vuc);
+}
+//.
+// CHECK: [[TBAA3]] = !{[[META4:![0-9]+]], [[META4]], i64 0}
+// CHECK: [[META4]] = !{!"omnipotent char", [[META5:![0-9]+]], i64 0}
+// CHECK: [[META5]] = !{!"Simple C/C++ TBAA"}
+//.