From ebf9322a50a799aa4bddd97e979fb272fa368cf5 Mon Sep 17 00:00:00 2001
From: Wenju He <wenju.he@intel.com>
Date: Tue, 25 Mar 2025 01:33:59 -0700
Subject: [PATCH 1/2] [SYCL] Move GroupNonUniform SPIR-V built-in declarations
 to clang SPIRVBuiltins.td

This PR changes scope parameter type from enum to int, aligning with
SPV-IR. Mangling becomes more portable for targets that bypass SPIR-V.

This PR constrains Id or Index type to be either uint32_t or uint64_t.
---
 clang/lib/Sema/SPIRVBuiltins.td               | 108 +++++
 ...spirv-builtin-lookup-group-non-uniform.cpp | 387 ++++++++++++++++++
 .../spirv-builtin-lookup-group.cl             |   6 +-
 .../lib/amdgcn-amdhsa/group/group_ballot.cl   |   4 +-
 .../ptx-nvidiacl/group/group_non_uniform.cl   |   8 +-
 sycl/include/sycl/__spirv/spirv_ops.hpp       | 151 -------
 6 files changed, 504 insertions(+), 160 deletions(-)
 create mode 100644 clang/test/CodeGenSPIRV/spirv-builtin-lookup-group-non-uniform.cpp

diff --git a/clang/lib/Sema/SPIRVBuiltins.td b/clang/lib/Sema/SPIRVBuiltins.td
index 2e46b37cf6974..5acfeeb9f1e50 100644
--- a/clang/lib/Sema/SPIRVBuiltins.td
+++ b/clang/lib/Sema/SPIRVBuiltins.td
@@ -358,10 +358,13 @@ def Vec1234     : IntList<"Vec1234", [1, 2, 3, 4]>;
 // Type lists.
 def TLAll           : TypeList<[Char,  UChar, Short,  UShort, Int,  UInt, Long,  ULong, Float, Double, Half]>;
 def TLAllUnsigned   : TypeList<[UChar, UChar, UShort, UShort, UInt, UInt, ULong, ULong, UInt,  ULong,  UShort]>;
+def TLAllWithBool : TypeList<[Bool, Char, UChar, Short, UShort, Int, UInt, Long,
+                              ULong, Float, Double, Half]>;
 def TLFloat         : TypeList<[Float, Double, Half]>;
 // FIXME: handle properly char (signed or unsigned depending on host)
 def TLSignedInts    : TypeList<[Char, Short, Int, Long]>;
 def TLUnsignedInts  : TypeList<[UChar, UShort, UInt, ULong]>;
+def TLUnsignedIntLong : TypeList<[UInt, ULong]>;
 
 // Signed to Unsigned conversion
 // FIXME: handle properly char (signed or unsigned depending on host)
@@ -383,6 +386,9 @@ def TLAllInts       : TypeList<[Char, UChar, Short, UShort, Int, UInt, Long, ULo
 def AGenType1              : GenericType<"AGenType1", TLAll, Vec1>;
 def AGenTypeN              : GenericType<"AGenTypeN", TLAll, VecAndScalar>;
 def AGenTypeNNoScalar      : GenericType<"AGenTypeNNoScalar", TLAll, VecNoScalar>;
+def AGenTypeNWithBool
+    : GenericType<"AGenTypeNWithBool", TLAllWithBool, VecAndScalar>;
+def BoolGenTypeN : GenericType<"BoolGenTypeN", TypeList<[Bool]>, VecAndScalar>;
 // All integer
 def AIGenType1             : GenericType<"AIGenType1", TLAllInts, Vec1>;
 def AIGenTypeN             : GenericType<"AIGenTypeN", TLAllInts, VecAndScalar>;
@@ -395,6 +401,7 @@ def AI2UGenTypeN           : GenericType<"AI2UGenTypeN", TLAllUIntsTwice, VecAnd
 def SGenTypeN              : GenericType<"SGenTypeN", TLSignedInts, VecAndScalar>;
 // Unsigned integer
 def UGenTypeN              : GenericType<"UGenTypeN", TLUnsignedInts, VecAndScalar>;
+def UInt4 : GenericType<"UInt4", TypeList<[UInt]>, Vec4>;
 // Float
 def FGenTypeN              : GenericType<"FGenTypeN", TLFloat, VecAndScalar>;
 // (u)int, (u)long, and all floats
@@ -976,3 +983,104 @@ foreach name = ["GroupSMin", "GroupSMax"] in {
 foreach name = ["GroupLogicalAndKHR", "GroupLogicalOrKHR"] in {
   def : SPVBuiltin<name, [Bool, Int, Int, Bool], Attr.Convergent>;
 }
+
+// 3.56.24. Non-Uniform Instructions
+
+foreach name = ["GroupNonUniformElect"] in {
+  def : SPVBuiltin<name, [Bool, Int], Attr.Convergent>;
+}
+
+foreach name = ["GroupNonUniformAll", "GroupNonUniformAny"] in {
+  def : SPVBuiltin<name, [Bool, Int, Bool], Attr.Convergent>;
+}
+
+foreach name = ["GroupNonUniformAllEqual"] in {
+  def : SPVBuiltin<name, [Bool, Int, AGenTypeNWithBool], Attr.Convergent>;
+}
+
+foreach name = ["GroupNonUniformBroadcast",
+                "GroupNonUniformBroadcastFirst"] in {
+  foreach IDType = TLUnsignedIntLong.List in {
+    def : SPVBuiltin<name, [AGenTypeNWithBool, Int, AGenTypeNWithBool, IDType],
+                     Attr.Convergent>;
+  }
+}
+
+foreach name = ["GroupNonUniformBallot"] in {
+  def : SPVBuiltin<name, [UInt4, Int, Bool], Attr.Convergent>;
+}
+
+foreach name = ["GroupNonUniformInverseBallot"] in {
+  def : SPVBuiltin<name, [Bool, Int, UInt4], Attr.Convergent>;
+}
+
+foreach name = ["GroupNonUniformBallotBitExtract"] in {
+  foreach IDType = TLUnsignedIntLong.List in {
+    def : SPVBuiltin<name, [Bool, Int, UInt4, IDType], Attr.Convergent>;
+  }
+}
+
+foreach name = ["GroupNonUniformBallotBitCount", "GroupNonUniformBallotFindLSB",
+                "GroupNonUniformBallotFindMSB"] in {
+  def : SPVBuiltin<name, [UInt, Int, Int, UInt4], Attr.Convergent>;
+}
+
+foreach name = ["GroupNonUniformShuffle", "GroupNonUniformShuffleXor",
+                "GroupNonUniformShuffleUp", "GroupNonUniformShuffleDown"] in {
+  foreach IDType = TLUnsignedIntLong.List in {
+    def : SPVBuiltin<name, [AGenTypeNWithBool, Int, AGenTypeNWithBool, IDType],
+                     Attr.Convergent>;
+  }
+}
+
+foreach name = ["GroupNonUniformIAdd", "GroupNonUniformIMul"] in {
+  def : SPVBuiltin<name, [AIGenTypeN, Int, Int, AIGenTypeN], Attr.Convergent>;
+  foreach IDType = TLUnsignedIntLong.List in {
+    def : SPVBuiltin<name, [AIGenTypeN, Int, Int, AIGenTypeN, IDType],
+                     Attr.Convergent>;
+  }
+}
+
+foreach name = ["GroupNonUniformSMin", "GroupNonUniformSMax"] in {
+  def : SPVBuiltin<name, [ASIGenTypeN, Int, Int, ASIGenTypeN], Attr.Convergent>;
+  foreach IDType = TLUnsignedIntLong.List in {
+    def : SPVBuiltin<name, [ASIGenTypeN, Int, Int, ASIGenTypeN, IDType],
+                     Attr.Convergent>;
+  }
+}
+
+foreach name = ["GroupNonUniformUMin", "GroupNonUniformUMax"] in {
+  def : SPVBuiltin<name, [UGenTypeN, Int, Int, UGenTypeN], Attr.Convergent>;
+  foreach IDType = TLUnsignedIntLong.List in {
+    def : SPVBuiltin<name, [UGenTypeN, Int, Int, UGenTypeN, IDType],
+                     Attr.Convergent>;
+  }
+}
+
+foreach name = ["GroupNonUniformFAdd", "GroupNonUniformFMul",
+                "GroupNonUniformFMin", "GroupNonUniformFMax"] in {
+  def : SPVBuiltin<name, [FGenTypeN, Int, Int, FGenTypeN], Attr.Convergent>;
+  foreach IDType = TLUnsignedIntLong.List in {
+    def : SPVBuiltin<name, [FGenTypeN, Int, Int, FGenTypeN, IDType],
+                     Attr.Convergent>;
+  }
+}
+
+foreach name = ["GroupNonUniformBitwiseAnd", "GroupNonUniformBitwiseOr",
+                "GroupNonUniformBitwiseXor"] in {
+  def : SPVBuiltin<name, [AIGenTypeN, Int, Int, AIGenTypeN], Attr.Convergent>;
+  foreach IDType = TLUnsignedIntLong.List in {
+    def : SPVBuiltin<name, [AIGenTypeN, Int, Int, AIGenTypeN, IDType],
+                     Attr.Convergent>;
+  }
+}
+
+foreach name = ["GroupNonUniformLogicalAnd", "GroupNonUniformLogicalOr",
+                "GroupNonUniformLogicalXor"] in {
+  def : SPVBuiltin<name, [BoolGenTypeN, Int, Int, BoolGenTypeN],
+                   Attr.Convergent>;
+  foreach IDType = TLUnsignedIntLong.List in {
+    def : SPVBuiltin<name, [BoolGenTypeN, Int, Int, BoolGenTypeN, IDType],
+                     Attr.Convergent>;
+  }
+}
diff --git a/clang/test/CodeGenSPIRV/spirv-builtin-lookup-group-non-uniform.cpp b/clang/test/CodeGenSPIRV/spirv-builtin-lookup-group-non-uniform.cpp
new file mode 100644
index 0000000000000..b85265fc4c84f
--- /dev/null
+++ b/clang/test/CodeGenSPIRV/spirv-builtin-lookup-group-non-uniform.cpp
@@ -0,0 +1,387 @@
+// RUN: %clang_cc1 -triple x86_64-unknown-linux-gnu -fdeclare-spirv-builtins -emit-llvm %s -o - | FileCheck %s
+
+bool group_elect() { return __spirv_GroupNonUniformElect(2); }
+
+bool group_all(bool predicate) {
+  return __spirv_GroupNonUniformAll(2, predicate);
+}
+
+bool group_any(bool predicate) {
+  return __spirv_GroupNonUniformAny(2, predicate);
+}
+
+template <class T> bool group_all_equal(T v) {
+  return __spirv_GroupNonUniformAllEqual(2, v);
+}
+
+template <class T> T group_broad_cast(T v, unsigned int id) {
+  return __spirv_GroupNonUniformBroadcast(2, v, id);
+}
+
+template <class T> T group_broad_cast_first(T v, unsigned int id) {
+  return __spirv_GroupNonUniformBroadcastFirst(2, v, id);
+}
+
+typedef unsigned int uint4 __attribute__((ext_vector_type(4)));
+
+uint4 group_ballot(char v) { return __spirv_GroupNonUniformBallot(2, false); }
+
+bool group_inverse_ballot(uint4 v) {
+  return __spirv_GroupNonUniformInverseBallot(2, v);
+}
+
+bool group_ballot_bit_extract(uint4 v, unsigned int id) {
+  return __spirv_GroupNonUniformBallotBitExtract(2, v, id);
+}
+
+unsigned int group_ballot_bit_count(uint4 v) {
+  return __spirv_GroupNonUniformBallotBitCount(2, 0, v);
+}
+
+unsigned int group_ballot_find_lsb(uint4 v) {
+  return __spirv_GroupNonUniformBallotFindLSB(2, 0, v);
+}
+
+unsigned int group_ballot_find_msb(uint4 v) {
+  return __spirv_GroupNonUniformBallotFindMSB(2, 0, v);
+}
+
+template <class T> T group_shuffle(T v, unsigned int id) {
+  return __spirv_GroupNonUniformShuffle(2, v, id);
+}
+
+template <class T> T group_shuffle_xor(T v, unsigned int id) {
+  return __spirv_GroupNonUniformShuffleXor(2, v, id);
+}
+
+template <class T> T group_shuffle_up(T v, unsigned int id) {
+  return __spirv_GroupNonUniformShuffleUp(2, v, id);
+}
+
+template <class T> T group_shuffle_down(T v, unsigned int id) {
+  return __spirv_GroupNonUniformShuffleDown(2, v, id);
+}
+
+template <class T> T group_iadd(T v, unsigned int id) {
+  return __spirv_GroupNonUniformIAdd(2, 0, v) +
+         __spirv_GroupNonUniformIAdd(2, 0, v, id);
+}
+
+template <class T> T group_imul(T v, unsigned int id) {
+  return __spirv_GroupNonUniformIMul(2, 0, v) +
+         __spirv_GroupNonUniformIMul(2, 0, v, id);
+}
+
+template <class T> T group_fadd(T v, unsigned int id) {
+  return __spirv_GroupNonUniformFAdd(2, 1, v) +
+         __spirv_GroupNonUniformFAdd(2, 1, v, id);
+}
+
+template <class T> T group_fmin(T v, unsigned int id) {
+  return __spirv_GroupNonUniformFMin(2, 0, v) +
+         __spirv_GroupNonUniformFMin(2, 0, v, id);
+}
+
+template <class T> T group_fmax(T v, unsigned int id) {
+  return __spirv_GroupNonUniformFMax(2, 0, v) +
+         __spirv_GroupNonUniformFMax(2, 0, v, id);
+}
+
+template <class T> T group_fmul(T v, unsigned int id) {
+  return __spirv_GroupNonUniformFMul(2, 0, v) +
+         __spirv_GroupNonUniformFMul(2, 0, v, id);
+}
+
+template <class T> T group_umin(T v, unsigned int id) {
+  return __spirv_GroupNonUniformUMin(2, 0, v) +
+         __spirv_GroupNonUniformUMin(2, 0, v, id);
+}
+
+template <class T> T group_umax(T v, unsigned int id) {
+  return __spirv_GroupNonUniformUMax(2, 0, v) +
+         __spirv_GroupNonUniformUMax(2, 0, v, id);
+}
+
+template <class T> T group_smin(T v, unsigned int id) {
+  return __spirv_GroupNonUniformSMin(2, 0, v) +
+         __spirv_GroupNonUniformSMin(2, 0, v, id);
+}
+
+template <class T> T group_smax(T v, unsigned int id) {
+  return __spirv_GroupNonUniformSMax(2, 0, v) +
+         __spirv_GroupNonUniformSMax(2, 0, v, id);
+}
+
+template <class T> T group_bitwise_and(T v, unsigned int id) {
+  return __spirv_GroupNonUniformBitwiseAnd(2, 0, v) +
+         __spirv_GroupNonUniformBitwiseAnd(2, 0, v, id);
+}
+
+template <class T> T group_bitwise_or(T v, unsigned int id) {
+  return __spirv_GroupNonUniformBitwiseOr(2, 0, v) +
+         __spirv_GroupNonUniformBitwiseOr(2, 0, v, id);
+}
+
+template <class T> T group_bitwise_xor(T v, unsigned int id) {
+  return __spirv_GroupNonUniformBitwiseXor(2, 0, v) +
+         __spirv_GroupNonUniformBitwiseXor(2, 0, v, id);
+}
+
+template <class T> bool group_logical_and(bool v, unsigned int id) {
+  return __spirv_GroupNonUniformLogicalAnd(2, 0, v) +
+         __spirv_GroupNonUniformLogicalAnd(2, 0, v, id);
+}
+
+template <class T> bool group_logical_or(bool v, unsigned int id) {
+  return __spirv_GroupNonUniformLogicalOr(2, 0, v) +
+         __spirv_GroupNonUniformLogicalOr(2, 0, v, id);
+}
+
+template <class T> bool group_logical_xor(bool v, unsigned int id) {
+  return __spirv_GroupNonUniformLogicalXor(2, 0, v) +
+         __spirv_GroupNonUniformLogicalXor(2, 0, v, id);
+}
+
+template <class T> void test_with_bool() {
+  T v = 0;
+  unsigned int id = 0;
+  group_all_equal<T>(v);
+  group_broad_cast<T>(v, id);
+  group_broad_cast_first<T>(v, id);
+  group_shuffle(v, id);
+  group_shuffle_xor(v, id);
+  group_shuffle_up(v, id);
+  group_shuffle_down(v, id);
+}
+
+template <class T> void test_integer() {
+  T v = 0;
+  unsigned int id = 0;
+  group_iadd<T>(v, id);
+  group_imul<T>(v, id);
+  group_bitwise_and<T>(v, id);
+  group_bitwise_or<T>(v, id);
+  group_bitwise_xor<T>(v, id);
+}
+
+template <class T> void test_signed() {
+  T v = 0;
+  unsigned int id = 0;
+  group_smin<T>(v, id);
+  group_smax<T>(v, id);
+}
+
+template <class T> void test_unsigned() {
+  T v = 0;
+  unsigned int id = 0;
+  group_umin<T>(v, id);
+  group_umax<T>(v, id);
+}
+
+template <class T> void test_logical() {
+  T v = 0;
+  unsigned int id = 0;
+  group_logical_and<T>(v, id);
+  group_logical_or<T>(v, id);
+  group_logical_xor<T>(v, id);
+}
+
+void test() {
+  test_with_bool<bool>();
+  test_with_bool<char>();
+  test_with_bool<unsigned char>();
+  test_with_bool<short>();
+  test_with_bool<unsigned short>();
+  test_with_bool<int>();
+  test_with_bool<unsigned int>();
+  test_with_bool<_Float16>();
+  test_with_bool<float>();
+  test_with_bool<double>();
+
+  test_integer<char>();
+  test_integer<unsigned char>();
+  test_integer<short>();
+  test_integer<unsigned short>();
+  test_integer<int>();
+  test_integer<unsigned int>();
+
+  test_signed<char>();
+  test_signed<short>();
+  test_signed<int>();
+
+  test_unsigned<unsigned char>();
+  test_unsigned<unsigned short>();
+  test_unsigned<unsigned int>();
+
+  test_logical<bool>();
+}
+
+// CHECK: call noundef zeroext i1 @_Z28__spirv_GroupNonUniformElecti
+// CHECK: call noundef zeroext i1 @_Z26__spirv_GroupNonUniformAllib
+// CHECK: call noundef zeroext i1 @_Z26__spirv_GroupNonUniformAnyib
+// CHECK: call noundef <4 x i32> @_Z29__spirv_GroupNonUniformBallotib
+// CHECK: call noundef zeroext i1 @_Z36__spirv_GroupNonUniformInverseBallotiDv4_j
+// CHECK: call noundef zeroext i1 @_Z39__spirv_GroupNonUniformBallotBitExtractiDv4_jj
+// CHECK: call noundef i32 @_Z37__spirv_GroupNonUniformBallotBitCountiiDv4_j
+// CHECK: call noundef i32 @_Z36__spirv_GroupNonUniformBallotFindLSBiiDv4_j
+// CHECK: call noundef i32 @_Z36__spirv_GroupNonUniformBallotFindMSBiiDv4_j
+// CHECK: call noundef zeroext i1 @_Z31__spirv_GroupNonUniformAllEqualib
+// CHECK: call noundef zeroext i1 @_Z32__spirv_GroupNonUniformBroadcastibj
+// CHECK: call noundef zeroext i1 @_Z37__spirv_GroupNonUniformBroadcastFirstibj
+// CHECK: call noundef zeroext i1 @_Z30__spirv_GroupNonUniformShuffleibj
+// CHECK: call noundef zeroext i1 @_Z33__spirv_GroupNonUniformShuffleXoribj
+// CHECK: call noundef zeroext i1 @_Z32__spirv_GroupNonUniformShuffleUpibj
+// CHECK: call noundef zeroext i1 @_Z34__spirv_GroupNonUniformShuffleDownibj
+// CHECK: call noundef zeroext i1 @_Z31__spirv_GroupNonUniformAllEqualii
+// CHECK: call noundef i32 @_Z32__spirv_GroupNonUniformBroadcastiij
+// CHECK: call noundef i32 @_Z37__spirv_GroupNonUniformBroadcastFirstiij
+// CHECK: call noundef i32 @_Z30__spirv_GroupNonUniformShuffleiij
+// CHECK: call noundef i32 @_Z33__spirv_GroupNonUniformShuffleXoriij
+// CHECK: call noundef i32 @_Z32__spirv_GroupNonUniformShuffleUpiij
+// CHECK: call noundef i32 @_Z34__spirv_GroupNonUniformShuffleDowniij
+// CHECK: call noundef zeroext i1 @_Z31__spirv_GroupNonUniformAllEqualih
+// CHECK: call noundef zeroext i8 @_Z32__spirv_GroupNonUniformBroadcastihj
+// CHECK: call noundef zeroext i8 @_Z37__spirv_GroupNonUniformBroadcastFirstihj
+// CHECK: call noundef zeroext i8 @_Z30__spirv_GroupNonUniformShuffleihj
+// CHECK: call noundef zeroext i8 @_Z33__spirv_GroupNonUniformShuffleXorihj
+// CHECK: call noundef zeroext i8 @_Z32__spirv_GroupNonUniformShuffleUpihj
+// CHECK: call noundef zeroext i8 @_Z34__spirv_GroupNonUniformShuffleDownihj
+// CHECK: call noundef zeroext i1 @_Z31__spirv_GroupNonUniformAllEqualis
+// CHECK: call noundef signext i16 @_Z32__spirv_GroupNonUniformBroadcastisj
+// CHECK: call noundef signext i16 @_Z37__spirv_GroupNonUniformBroadcastFirstisj
+// CHECK: call noundef signext i16 @_Z30__spirv_GroupNonUniformShuffleisj
+// CHECK: call noundef signext i16 @_Z33__spirv_GroupNonUniformShuffleXorisj
+// CHECK: call noundef signext i16 @_Z32__spirv_GroupNonUniformShuffleUpisj
+// CHECK: call noundef signext i16 @_Z34__spirv_GroupNonUniformShuffleDownisj
+// CHECK: call noundef zeroext i1 @_Z31__spirv_GroupNonUniformAllEqualit
+// CHECK: call noundef zeroext i16 @_Z32__spirv_GroupNonUniformBroadcastitj
+// CHECK: call noundef zeroext i16 @_Z37__spirv_GroupNonUniformBroadcastFirstitj
+// CHECK: call noundef zeroext i16 @_Z30__spirv_GroupNonUniformShuffleitj
+// CHECK: call noundef zeroext i16 @_Z33__spirv_GroupNonUniformShuffleXoritj
+// CHECK: call noundef zeroext i16 @_Z32__spirv_GroupNonUniformShuffleUpitj
+// CHECK: call noundef zeroext i16 @_Z34__spirv_GroupNonUniformShuffleDownitj
+// CHECK: call noundef zeroext i1 @_Z31__spirv_GroupNonUniformAllEqualii
+// CHECK: call noundef i32 @_Z32__spirv_GroupNonUniformBroadcastiij
+// CHECK: call noundef i32 @_Z37__spirv_GroupNonUniformBroadcastFirstiij
+// CHECK: call noundef i32 @_Z30__spirv_GroupNonUniformShuffleiij
+// CHECK: call noundef i32 @_Z33__spirv_GroupNonUniformShuffleXoriij
+// CHECK: call noundef i32 @_Z32__spirv_GroupNonUniformShuffleUpiij
+// CHECK: call noundef i32 @_Z34__spirv_GroupNonUniformShuffleDowniij
+// CHECK: call noundef zeroext i1 @_Z31__spirv_GroupNonUniformAllEqualij
+// CHECK: call noundef i32 @_Z32__spirv_GroupNonUniformBroadcastijj
+// CHECK: call noundef i32 @_Z37__spirv_GroupNonUniformBroadcastFirstijj
+// CHECK: call noundef i32 @_Z30__spirv_GroupNonUniformShuffleijj
+// CHECK: call noundef i32 @_Z33__spirv_GroupNonUniformShuffleXorijj
+// CHECK: call noundef i32 @_Z32__spirv_GroupNonUniformShuffleUpijj
+// CHECK: call noundef i32 @_Z34__spirv_GroupNonUniformShuffleDownijj
+// CHECK: call noundef zeroext i1 @_Z31__spirv_GroupNonUniformAllEqualiDF16_
+// CHECK: call noundef half @_Z32__spirv_GroupNonUniformBroadcastiDF16_j
+// CHECK: call noundef half @_Z37__spirv_GroupNonUniformBroadcastFirstiDF16_j
+// CHECK: call noundef half @_Z30__spirv_GroupNonUniformShuffleiDF16_j
+// CHECK: call noundef half @_Z33__spirv_GroupNonUniformShuffleXoriDF16_j
+// CHECK: call noundef half @_Z32__spirv_GroupNonUniformShuffleUpiDF16_j
+// CHECK: call noundef half @_Z34__spirv_GroupNonUniformShuffleDowniDF16_j
+// CHECK: call noundef zeroext i1 @_Z31__spirv_GroupNonUniformAllEqualif
+// CHECK: call noundef float @_Z32__spirv_GroupNonUniformBroadcastifj
+// CHECK: call noundef float @_Z37__spirv_GroupNonUniformBroadcastFirstifj
+// CHECK: call noundef float @_Z30__spirv_GroupNonUniformShuffleifj
+// CHECK: call noundef float @_Z33__spirv_GroupNonUniformShuffleXorifj
+// CHECK: call noundef float @_Z32__spirv_GroupNonUniformShuffleUpifj
+// CHECK: call noundef float @_Z34__spirv_GroupNonUniformShuffleDownifj
+// CHECK: call noundef zeroext i1 @_Z31__spirv_GroupNonUniformAllEqualid
+// CHECK: call noundef double @_Z32__spirv_GroupNonUniformBroadcastidj
+// CHECK: call noundef double @_Z37__spirv_GroupNonUniformBroadcastFirstidj
+// CHECK: call noundef double @_Z30__spirv_GroupNonUniformShuffleidj
+// CHECK: call noundef double @_Z33__spirv_GroupNonUniformShuffleXoridj
+// CHECK: call noundef double @_Z32__spirv_GroupNonUniformShuffleUpidj
+// CHECK: call noundef double @_Z34__spirv_GroupNonUniformShuffleDownidj
+// CHECK: call noundef i32 @_Z27__spirv_GroupNonUniformIAddiii
+// CHECK: call noundef i32 @_Z27__spirv_GroupNonUniformIAddiiij
+// CHECK: call noundef i32 @_Z27__spirv_GroupNonUniformIMuliii
+// CHECK: call noundef i32 @_Z27__spirv_GroupNonUniformIMuliiij
+// CHECK: call noundef i32 @_Z33__spirv_GroupNonUniformBitwiseAndiii
+// CHECK: call noundef i32 @_Z33__spirv_GroupNonUniformBitwiseAndiiij
+// CHECK: call noundef i32 @_Z32__spirv_GroupNonUniformBitwiseOriii
+// CHECK: call noundef i32 @_Z32__spirv_GroupNonUniformBitwiseOriiij
+// CHECK: call noundef i32 @_Z33__spirv_GroupNonUniformBitwiseXoriii
+// CHECK: call noundef i32 @_Z33__spirv_GroupNonUniformBitwiseXoriiij
+// CHECK: call noundef zeroext i8 @_Z27__spirv_GroupNonUniformIAddiih
+// CHECK: call noundef zeroext i8 @_Z27__spirv_GroupNonUniformIAddiihj
+// CHECK: call noundef zeroext i8 @_Z27__spirv_GroupNonUniformIMuliih
+// CHECK: call noundef zeroext i8 @_Z27__spirv_GroupNonUniformIMuliihj
+// CHECK: call noundef zeroext i8 @_Z33__spirv_GroupNonUniformBitwiseAndiih
+// CHECK: call noundef zeroext i8 @_Z33__spirv_GroupNonUniformBitwiseAndiihj
+// CHECK: call noundef zeroext i8 @_Z32__spirv_GroupNonUniformBitwiseOriih
+// CHECK: call noundef zeroext i8 @_Z32__spirv_GroupNonUniformBitwiseOriihj
+// CHECK: call noundef zeroext i8 @_Z33__spirv_GroupNonUniformBitwiseXoriih
+// CHECK: call noundef zeroext i8 @_Z33__spirv_GroupNonUniformBitwiseXoriihj
+// CHECK: call noundef signext i16 @_Z27__spirv_GroupNonUniformIAddiis
+// CHECK: call noundef signext i16 @_Z27__spirv_GroupNonUniformIAddiisj
+// CHECK: call noundef signext i16 @_Z27__spirv_GroupNonUniformIMuliis
+// CHECK: call noundef signext i16 @_Z27__spirv_GroupNonUniformIMuliisj
+// CHECK: call noundef signext i16 @_Z33__spirv_GroupNonUniformBitwiseAndiis
+// CHECK: call noundef signext i16 @_Z33__spirv_GroupNonUniformBitwiseAndiisj
+// CHECK: call noundef signext i16 @_Z32__spirv_GroupNonUniformBitwiseOriis
+// CHECK: call noundef signext i16 @_Z32__spirv_GroupNonUniformBitwiseOriisj
+// CHECK: call noundef signext i16 @_Z33__spirv_GroupNonUniformBitwiseXoriis
+// CHECK: call noundef signext i16 @_Z33__spirv_GroupNonUniformBitwiseXoriisj
+// CHECK: call noundef zeroext i16 @_Z27__spirv_GroupNonUniformIAddiit
+// CHECK: call noundef zeroext i16 @_Z27__spirv_GroupNonUniformIAddiitj
+// CHECK: call noundef zeroext i16 @_Z27__spirv_GroupNonUniformIMuliit
+// CHECK: call noundef zeroext i16 @_Z27__spirv_GroupNonUniformIMuliitj
+// CHECK: call noundef zeroext i16 @_Z33__spirv_GroupNonUniformBitwiseAndiit
+// CHECK: call noundef zeroext i16 @_Z33__spirv_GroupNonUniformBitwiseAndiitj
+// CHECK: call noundef zeroext i16 @_Z32__spirv_GroupNonUniformBitwiseOriit
+// CHECK: call noundef zeroext i16 @_Z32__spirv_GroupNonUniformBitwiseOriitj
+// CHECK: call noundef zeroext i16 @_Z33__spirv_GroupNonUniformBitwiseXoriit
+// CHECK: call noundef zeroext i16 @_Z33__spirv_GroupNonUniformBitwiseXoriitj
+// CHECK: call noundef i32 @_Z27__spirv_GroupNonUniformIAddiii
+// CHECK: call noundef i32 @_Z27__spirv_GroupNonUniformIAddiiij
+// CHECK: call noundef i32 @_Z27__spirv_GroupNonUniformIMuliii
+// CHECK: call noundef i32 @_Z27__spirv_GroupNonUniformIMuliiij
+// CHECK: call noundef i32 @_Z33__spirv_GroupNonUniformBitwiseAndiii
+// CHECK: call noundef i32 @_Z33__spirv_GroupNonUniformBitwiseAndiiij
+// CHECK: call noundef i32 @_Z32__spirv_GroupNonUniformBitwiseOriii
+// CHECK: call noundef i32 @_Z32__spirv_GroupNonUniformBitwiseOriiij
+// CHECK: call noundef i32 @_Z33__spirv_GroupNonUniformBitwiseXoriii
+// CHECK: call noundef i32 @_Z33__spirv_GroupNonUniformBitwiseXoriiij
+// CHECK: call noundef i32 @_Z27__spirv_GroupNonUniformIAddiij
+// CHECK: call noundef i32 @_Z27__spirv_GroupNonUniformIAddiijj
+// CHECK: call noundef i32 @_Z27__spirv_GroupNonUniformIMuliij
+// CHECK: call noundef i32 @_Z27__spirv_GroupNonUniformIMuliijj
+// CHECK: call noundef i32 @_Z33__spirv_GroupNonUniformBitwiseAndiij
+// CHECK: call noundef i32 @_Z33__spirv_GroupNonUniformBitwiseAndiijj
+// CHECK: call noundef i32 @_Z32__spirv_GroupNonUniformBitwiseOriij
+// CHECK: call noundef i32 @_Z32__spirv_GroupNonUniformBitwiseOriijj
+// CHECK: call noundef i32 @_Z33__spirv_GroupNonUniformBitwiseXoriij
+// CHECK: call noundef i32 @_Z33__spirv_GroupNonUniformBitwiseXoriijj
+// CHECK: call noundef i32 @_Z27__spirv_GroupNonUniformSMiniii
+// CHECK: call noundef i32 @_Z27__spirv_GroupNonUniformSMiniiij
+// CHECK: call noundef i32 @_Z27__spirv_GroupNonUniformSMaxiii
+// CHECK: call noundef i32 @_Z27__spirv_GroupNonUniformSMaxiiij
+// CHECK: call noundef signext i16 @_Z27__spirv_GroupNonUniformSMiniis
+// CHECK: call noundef signext i16 @_Z27__spirv_GroupNonUniformSMiniisj
+// CHECK: call noundef signext i16 @_Z27__spirv_GroupNonUniformSMaxiis
+// CHECK: call noundef signext i16 @_Z27__spirv_GroupNonUniformSMaxiisj
+// CHECK: call noundef i32 @_Z27__spirv_GroupNonUniformSMiniii
+// CHECK: call noundef i32 @_Z27__spirv_GroupNonUniformSMiniiij
+// CHECK: call noundef i32 @_Z27__spirv_GroupNonUniformSMaxiii
+// CHECK: call noundef i32 @_Z27__spirv_GroupNonUniformSMaxiiij
+// CHECK: call noundef zeroext i8 @_Z27__spirv_GroupNonUniformUMiniih
+// CHECK: call noundef zeroext i8 @_Z27__spirv_GroupNonUniformUMiniihj
+// CHECK: call noundef zeroext i8 @_Z27__spirv_GroupNonUniformUMaxiih
+// CHECK: call noundef zeroext i8 @_Z27__spirv_GroupNonUniformUMaxiihj
+// CHECK: call noundef zeroext i16 @_Z27__spirv_GroupNonUniformUMiniit
+// CHECK: call noundef zeroext i16 @_Z27__spirv_GroupNonUniformUMiniitj
+// CHECK: call noundef zeroext i16 @_Z27__spirv_GroupNonUniformUMaxiit
+// CHECK: call noundef zeroext i16 @_Z27__spirv_GroupNonUniformUMaxiitj
+// CHECK: call noundef i32 @_Z27__spirv_GroupNonUniformUMiniij
+// CHECK: call noundef i32 @_Z27__spirv_GroupNonUniformUMiniijj
+// CHECK: call noundef i32 @_Z27__spirv_GroupNonUniformUMaxiij
+// CHECK: call noundef i32 @_Z27__spirv_GroupNonUniformUMaxiijj
+// CHECK: call noundef zeroext i1 @_Z33__spirv_GroupNonUniformLogicalAndiib
+// CHECK: call noundef zeroext i1 @_Z33__spirv_GroupNonUniformLogicalAndiibj
+// CHECK: call noundef zeroext i1 @_Z32__spirv_GroupNonUniformLogicalOriib
+// CHECK: call noundef zeroext i1 @_Z32__spirv_GroupNonUniformLogicalOriibj
+// CHECK: call noundef zeroext i1 @_Z33__spirv_GroupNonUniformLogicalXoriib
+// CHECK: call noundef zeroext i1 @_Z33__spirv_GroupNonUniformLogicalXoriibj
diff --git a/clang/test/CodeGenSPIRV/spirv-builtin-lookup-group.cl b/clang/test/CodeGenSPIRV/spirv-builtin-lookup-group.cl
index 0a31736f04122..52872bf7ac3b9 100644
--- a/clang/test/CodeGenSPIRV/spirv-builtin-lookup-group.cl
+++ b/clang/test/CodeGenSPIRV/spirv-builtin-lookup-group.cl
@@ -43,7 +43,7 @@ int group_imul_khr(short a) {
   return __spirv_GroupIMulKHR(2, 0, a);
 }
 
-bool group_bitwise_or_khr(long a) {
+long group_bitwise_or_khr(long a) {
   // CHECK-LABEL: @group_bitwise_or_khr(
   // CHECK: call spir_func i64 @_Z25__spirv_GroupBitwiseOrKHRiil(
   return __spirv_GroupBitwiseOrKHR(2, 0, a);
@@ -67,8 +67,8 @@ float group_fmax(float a) {
   return __spirv_GroupFMax(2, 2, a);
 }
 
-float group_ful_khr(float a) {
-  // CHECK-LABEL: @group_ful_khr(
+float group_fmul_khr(float a) {
+  // CHECK-LABEL: @group_fmul_khr(
   // CHECK: call spir_func float @_Z20__spirv_GroupFMulKHRiif(
   return __spirv_GroupFMulKHR(2, 2, a);
 }
diff --git a/libclc/libspirv/lib/amdgcn-amdhsa/group/group_ballot.cl b/libclc/libspirv/lib/amdgcn-amdhsa/group/group_ballot.cl
index 49fa34a8cd7ab..f3cfa38615510 100644
--- a/libclc/libspirv/lib/amdgcn-amdhsa/group/group_ballot.cl
+++ b/libclc/libspirv/lib/amdgcn-amdhsa/group/group_ballot.cl
@@ -12,8 +12,8 @@
 // from llvm/include/llvm/IR/InstrTypes.h
 #define ICMP_NE 33
 
-_CLC_DEF _CLC_CONVERGENT __clc_vec4_uint32_t
-_Z29__spirv_GroupNonUniformBallotib(int flag, bool predicate) {
+_CLC_OVERLOAD _CLC_DEF _CLC_CONVERGENT __clc_vec4_uint32_t
+__spirv_GroupNonUniformBallot(int flag, bool predicate) {
   // only support subgroup for now
   if (flag != Subgroup) {
     __builtin_trap();
diff --git a/libclc/libspirv/lib/ptx-nvidiacl/group/group_non_uniform.cl b/libclc/libspirv/lib/ptx-nvidiacl/group/group_non_uniform.cl
index 3474f17666080..afc1f3d8df81f 100644
--- a/libclc/libspirv/lib/ptx-nvidiacl/group/group_non_uniform.cl
+++ b/libclc/libspirv/lib/ptx-nvidiacl/group/group_non_uniform.cl
@@ -12,8 +12,8 @@
 #include <libspirv/spirv.h>
 #include <libspirv/spirv_types.h>
 
-_CLC_DEF _CLC_CONVERGENT __clc_vec4_uint32_t
-_Z29__spirv_GroupNonUniformBallotib(int flag, bool predicate) {
+_CLC_OVERLOAD _CLC_DEF _CLC_CONVERGENT __clc_vec4_uint32_t
+__spirv_GroupNonUniformBallot(int flag, bool predicate) {
   // only support subgroup for now
   if (flag != Subgroup) {
     __builtin_trap();
@@ -36,8 +36,8 @@ _Z29__spirv_GroupNonUniformBallotib(int flag, bool predicate) {
   return res;
 }
 
-_CLC_DEF _CLC_CONVERGENT uint
-_Z37__spirv_GroupNonUniformBallotBitCountN5__spv5Scope4FlagEiDv4_j(
+_CLC_OVERLOAD _CLC_DEF _CLC_CONVERGENT uint
+__spirv_GroupNonUniformBallotBitCount(
     int scope, int flag, __clc_vec4_uint32_t mask) {
   // here we assume scope == __spv::Scope::Subgroup
   // flag == InclusiveScan is not yet implemented
diff --git a/sycl/include/sycl/__spirv/spirv_ops.hpp b/sycl/include/sycl/__spirv/spirv_ops.hpp
index 04cfec68e7679..4003b02485b87 100644
--- a/sycl/include/sycl/__spirv/spirv_ops.hpp
+++ b/sycl/include/sycl/__spirv/spirv_ops.hpp
@@ -881,157 +881,6 @@ template <int N>
 extern __DPCPP_SYCL_EXTERNAL __ocl_vec_t<uint16_t, N>
     __spirv_ConvertFToBF16INTEL(__ocl_vec_t<float, N>) noexcept;
 
-__SYCL_CONVERGENT__ extern __DPCPP_SYCL_EXTERNAL
-    __SYCL_EXPORT __ocl_vec_t<uint32_t, 4>
-    __spirv_GroupNonUniformBallot(int32_t Execution, bool Predicate) noexcept;
-
-// TODO: I'm not 100% sure that these NonUniform instructions should be
-// convergent Following precedent set for GroupNonUniformBallot above
-__SYCL_CONVERGENT__ extern __DPCPP_SYCL_EXTERNAL __SYCL_EXPORT uint32_t
-__spirv_GroupNonUniformBallotBitCount(__spv::Scope::Flag, int,
-                                      __ocl_vec_t<uint32_t, 4>) noexcept;
-
-__SYCL_CONVERGENT__ extern __DPCPP_SYCL_EXTERNAL __SYCL_EXPORT int
-    __spirv_GroupNonUniformBallotFindLSB(__spv::Scope::Flag,
-                                         __ocl_vec_t<uint32_t, 4>) noexcept;
-
-template <typename ValueT, typename IdT>
-__SYCL_CONVERGENT__ extern __DPCPP_SYCL_EXTERNAL __SYCL_EXPORT ValueT
-    __spirv_GroupNonUniformBroadcast(__spv::Scope::Flag, ValueT, IdT);
-
-template <typename ValueT, typename IdT>
-__SYCL_CONVERGENT__ extern __DPCPP_SYCL_EXTERNAL __SYCL_EXPORT ValueT
-    __spirv_GroupNonUniformShuffle(__spv::Scope::Flag, ValueT, IdT) noexcept;
-
-__SYCL_CONVERGENT__ extern __DPCPP_SYCL_EXTERNAL __SYCL_EXPORT bool
-__spirv_GroupNonUniformAll(__spv::Scope::Flag, bool);
-
-__SYCL_CONVERGENT__ extern __DPCPP_SYCL_EXTERNAL __SYCL_EXPORT bool
-__spirv_GroupNonUniformAny(__spv::Scope::Flag, bool);
-
-template <typename ValueT>
-__SYCL_CONVERGENT__ extern __DPCPP_SYCL_EXTERNAL __SYCL_EXPORT ValueT
-__spirv_GroupNonUniformSMin(__spv::Scope::Flag, int, ValueT);
-
-template <typename ValueT>
-__SYCL_CONVERGENT__ extern __DPCPP_SYCL_EXTERNAL __SYCL_EXPORT ValueT
-__spirv_GroupNonUniformUMin(__spv::Scope::Flag, int, ValueT);
-
-template <typename ValueT>
-__SYCL_CONVERGENT__ extern __DPCPP_SYCL_EXTERNAL __SYCL_EXPORT ValueT
-__spirv_GroupNonUniformFMin(__spv::Scope::Flag, int, ValueT);
-
-template <typename ValueT>
-__SYCL_CONVERGENT__ extern __DPCPP_SYCL_EXTERNAL __SYCL_EXPORT ValueT
-__spirv_GroupNonUniformSMax(__spv::Scope::Flag, int, ValueT);
-
-template <typename ValueT>
-__SYCL_CONVERGENT__ extern __DPCPP_SYCL_EXTERNAL __SYCL_EXPORT ValueT
-__spirv_GroupNonUniformUMax(__spv::Scope::Flag, int, ValueT);
-
-template <typename ValueT>
-__SYCL_CONVERGENT__ extern __DPCPP_SYCL_EXTERNAL __SYCL_EXPORT ValueT
-__spirv_GroupNonUniformFMax(__spv::Scope::Flag, int, ValueT);
-
-template <typename ValueT>
-__SYCL_CONVERGENT__ extern __DPCPP_SYCL_EXTERNAL __SYCL_EXPORT ValueT
-__spirv_GroupNonUniformIAdd(__spv::Scope::Flag, int, ValueT);
-
-template <typename ValueT>
-__SYCL_CONVERGENT__ extern __DPCPP_SYCL_EXTERNAL __SYCL_EXPORT ValueT
-__spirv_GroupNonUniformFAdd(__spv::Scope::Flag, int, ValueT);
-
-template <typename ValueT>
-__SYCL_CONVERGENT__ extern __DPCPP_SYCL_EXTERNAL __SYCL_EXPORT ValueT
-__spirv_GroupNonUniformIMul(__spv::Scope::Flag, int, ValueT);
-
-template <typename ValueT>
-__SYCL_CONVERGENT__ extern __DPCPP_SYCL_EXTERNAL __SYCL_EXPORT ValueT
-__spirv_GroupNonUniformFMul(__spv::Scope::Flag, int, ValueT);
-
-template <typename ValueT>
-__SYCL_CONVERGENT__ extern __DPCPP_SYCL_EXTERNAL __SYCL_EXPORT ValueT
-__spirv_GroupNonUniformBitwiseOr(__spv::Scope::Flag, int, ValueT);
-
-template <typename ValueT>
-__SYCL_CONVERGENT__ extern __DPCPP_SYCL_EXTERNAL __SYCL_EXPORT ValueT
-__spirv_GroupNonUniformBitwiseXor(__spv::Scope::Flag, int, ValueT);
-
-template <typename ValueT>
-__SYCL_CONVERGENT__ extern __DPCPP_SYCL_EXTERNAL __SYCL_EXPORT ValueT
-__spirv_GroupNonUniformBitwiseAnd(__spv::Scope::Flag, int, ValueT);
-
-template <typename ValueT>
-__SYCL_CONVERGENT__ extern __DPCPP_SYCL_EXTERNAL __SYCL_EXPORT ValueT
-__spirv_GroupNonUniformLogicalOr(__spv::Scope::Flag, int, ValueT);
-
-template <typename ValueT>
-__SYCL_CONVERGENT__ extern __DPCPP_SYCL_EXTERNAL __SYCL_EXPORT ValueT
-__spirv_GroupNonUniformLogicalAnd(__spv::Scope::Flag, int, ValueT);
-
-template <typename ValueT>
-__SYCL_CONVERGENT__ extern __DPCPP_SYCL_EXTERNAL __SYCL_EXPORT ValueT
-__spirv_GroupNonUniformSMin(__spv::Scope::Flag, int, ValueT, unsigned int);
-
-template <typename ValueT>
-__SYCL_CONVERGENT__ extern __DPCPP_SYCL_EXTERNAL __SYCL_EXPORT ValueT
-__spirv_GroupNonUniformUMin(__spv::Scope::Flag, int, ValueT, unsigned int);
-
-template <typename ValueT>
-__SYCL_CONVERGENT__ extern __DPCPP_SYCL_EXTERNAL __SYCL_EXPORT ValueT
-__spirv_GroupNonUniformFMin(__spv::Scope::Flag, int, ValueT, unsigned int);
-
-template <typename ValueT>
-__SYCL_CONVERGENT__ extern __DPCPP_SYCL_EXTERNAL __SYCL_EXPORT ValueT
-__spirv_GroupNonUniformSMax(__spv::Scope::Flag, int, ValueT, unsigned int);
-
-template <typename ValueT>
-__SYCL_CONVERGENT__ extern __DPCPP_SYCL_EXTERNAL __SYCL_EXPORT ValueT
-__spirv_GroupNonUniformUMax(__spv::Scope::Flag, int, ValueT, unsigned int);
-
-template <typename ValueT>
-__SYCL_CONVERGENT__ extern __DPCPP_SYCL_EXTERNAL __SYCL_EXPORT ValueT
-__spirv_GroupNonUniformFMax(__spv::Scope::Flag, int, ValueT, unsigned int);
-
-template <typename ValueT>
-__SYCL_CONVERGENT__ extern __DPCPP_SYCL_EXTERNAL __SYCL_EXPORT ValueT
-__spirv_GroupNonUniformIAdd(__spv::Scope::Flag, int, ValueT, unsigned int);
-
-template <typename ValueT>
-__SYCL_CONVERGENT__ extern __DPCPP_SYCL_EXTERNAL __SYCL_EXPORT ValueT
-__spirv_GroupNonUniformFAdd(__spv::Scope::Flag, int, ValueT, unsigned int);
-
-template <typename ValueT>
-__SYCL_CONVERGENT__ extern __DPCPP_SYCL_EXTERNAL __SYCL_EXPORT ValueT
-__spirv_GroupNonUniformIMul(__spv::Scope::Flag, int, ValueT, unsigned int);
-
-template <typename ValueT>
-__SYCL_CONVERGENT__ extern __DPCPP_SYCL_EXTERNAL __SYCL_EXPORT ValueT
-__spirv_GroupNonUniformFMul(__spv::Scope::Flag, int, ValueT, unsigned int);
-
-template <typename ValueT>
-__SYCL_CONVERGENT__ extern __DPCPP_SYCL_EXTERNAL __SYCL_EXPORT ValueT
-__spirv_GroupNonUniformBitwiseOr(__spv::Scope::Flag, int, ValueT, unsigned int);
-
-template <typename ValueT>
-__SYCL_CONVERGENT__ extern __DPCPP_SYCL_EXTERNAL __SYCL_EXPORT ValueT
-__spirv_GroupNonUniformBitwiseXor(__spv::Scope::Flag, int, ValueT,
-                                  unsigned int);
-
-template <typename ValueT>
-__SYCL_CONVERGENT__ extern __DPCPP_SYCL_EXTERNAL __SYCL_EXPORT ValueT
-__spirv_GroupNonUniformBitwiseAnd(__spv::Scope::Flag, int, ValueT,
-                                  unsigned int);
-
-template <typename ValueT>
-__SYCL_CONVERGENT__ extern __DPCPP_SYCL_EXTERNAL __SYCL_EXPORT ValueT
-__spirv_GroupNonUniformLogicalOr(__spv::Scope::Flag, int, ValueT, unsigned int);
-
-template <typename ValueT>
-__SYCL_CONVERGENT__ extern __DPCPP_SYCL_EXTERNAL __SYCL_EXPORT ValueT
-__spirv_GroupNonUniformLogicalAnd(__spv::Scope::Flag, int, ValueT,
-                                  unsigned int);
-
 extern __DPCPP_SYCL_EXTERNAL __SYCL_EXPORT void
 __clc_BarrierInitialize(int64_t *state, int32_t expected_count) noexcept;
 

From 4155e5413e5a876ab89a213786c6c12b4473e2e1 Mon Sep 17 00:00:00 2001
From: Wenju He <wenju.he@intel.com>
Date: Tue, 25 Mar 2025 18:06:01 -0700
Subject: [PATCH 2/2] remove ULong variant for Delta/ClusterSize/Index
 parameters, only keep ULong for Id type in GroupNonUniformBroadcast

---
 clang/lib/Sema/SPIRVBuiltins.td               | 60 ++++++++-----------
 ...spirv-builtin-lookup-group-non-uniform.cpp | 26 ++++----
 2 files changed, 37 insertions(+), 49 deletions(-)

diff --git a/clang/lib/Sema/SPIRVBuiltins.td b/clang/lib/Sema/SPIRVBuiltins.td
index 5acfeeb9f1e50..bba559f0f255a 100644
--- a/clang/lib/Sema/SPIRVBuiltins.td
+++ b/clang/lib/Sema/SPIRVBuiltins.td
@@ -364,7 +364,6 @@ def TLFloat         : TypeList<[Float, Double, Half]>;
 // FIXME: handle properly char (signed or unsigned depending on host)
 def TLSignedInts    : TypeList<[Char, Short, Int, Long]>;
 def TLUnsignedInts  : TypeList<[UChar, UShort, UInt, ULong]>;
-def TLUnsignedIntLong : TypeList<[UInt, ULong]>;
 
 // Signed to Unsigned conversion
 // FIXME: handle properly char (signed or unsigned depending on host)
@@ -388,6 +387,7 @@ def AGenTypeN              : GenericType<"AGenTypeN", TLAll, VecAndScalar>;
 def AGenTypeNNoScalar      : GenericType<"AGenTypeNNoScalar", TLAll, VecNoScalar>;
 def AGenTypeNWithBool
     : GenericType<"AGenTypeNWithBool", TLAllWithBool, VecAndScalar>;
+// Bool
 def BoolGenTypeN : GenericType<"BoolGenTypeN", TypeList<[Bool]>, VecAndScalar>;
 // All integer
 def AIGenType1             : GenericType<"AIGenType1", TLAllInts, Vec1>;
@@ -998,14 +998,18 @@ foreach name = ["GroupNonUniformAllEqual"] in {
   def : SPVBuiltin<name, [Bool, Int, AGenTypeNWithBool], Attr.Convergent>;
 }
 
-foreach name = ["GroupNonUniformBroadcast",
-                "GroupNonUniformBroadcastFirst"] in {
-  foreach IDType = TLUnsignedIntLong.List in {
+foreach name = ["GroupNonUniformBroadcast"] in {
+  foreach IDType = [UInt, ULong] in {
     def : SPVBuiltin<name, [AGenTypeNWithBool, Int, AGenTypeNWithBool, IDType],
                      Attr.Convergent>;
   }
 }
 
+foreach name = ["GroupNonUniformBroadcastFirst"] in {
+  def : SPVBuiltin<name, [AGenTypeNWithBool, Int, AGenTypeNWithBool],
+                   Attr.Convergent>;
+}
+
 foreach name = ["GroupNonUniformBallot"] in {
   def : SPVBuiltin<name, [UInt4, Int, Bool], Attr.Convergent>;
 }
@@ -1015,9 +1019,7 @@ foreach name = ["GroupNonUniformInverseBallot"] in {
 }
 
 foreach name = ["GroupNonUniformBallotBitExtract"] in {
-  foreach IDType = TLUnsignedIntLong.List in {
-    def : SPVBuiltin<name, [Bool, Int, UInt4, IDType], Attr.Convergent>;
-  }
+  def : SPVBuiltin<name, [Bool, Int, UInt4, UInt], Attr.Convergent>;
 }
 
 foreach name = ["GroupNonUniformBallotBitCount", "GroupNonUniformBallotFindLSB",
@@ -1027,60 +1029,46 @@ foreach name = ["GroupNonUniformBallotBitCount", "GroupNonUniformBallotFindLSB",
 
 foreach name = ["GroupNonUniformShuffle", "GroupNonUniformShuffleXor",
                 "GroupNonUniformShuffleUp", "GroupNonUniformShuffleDown"] in {
-  foreach IDType = TLUnsignedIntLong.List in {
-    def : SPVBuiltin<name, [AGenTypeNWithBool, Int, AGenTypeNWithBool, IDType],
-                     Attr.Convergent>;
-  }
+  def : SPVBuiltin<name, [AGenTypeNWithBool, Int, AGenTypeNWithBool, UInt],
+                   Attr.Convergent>;
 }
 
 foreach name = ["GroupNonUniformIAdd", "GroupNonUniformIMul"] in {
   def : SPVBuiltin<name, [AIGenTypeN, Int, Int, AIGenTypeN], Attr.Convergent>;
-  foreach IDType = TLUnsignedIntLong.List in {
-    def : SPVBuiltin<name, [AIGenTypeN, Int, Int, AIGenTypeN, IDType],
-                     Attr.Convergent>;
-  }
+  def : SPVBuiltin<name, [AIGenTypeN, Int, Int, AIGenTypeN, UInt],
+                   Attr.Convergent>;
 }
 
 foreach name = ["GroupNonUniformSMin", "GroupNonUniformSMax"] in {
-  def : SPVBuiltin<name, [ASIGenTypeN, Int, Int, ASIGenTypeN], Attr.Convergent>;
-  foreach IDType = TLUnsignedIntLong.List in {
-    def : SPVBuiltin<name, [ASIGenTypeN, Int, Int, ASIGenTypeN, IDType],
-                     Attr.Convergent>;
-  }
+  def : SPVBuiltin<name, [SGenTypeN, Int, Int, SGenTypeN], Attr.Convergent>;
+  def : SPVBuiltin<name, [SGenTypeN, Int, Int, SGenTypeN, UInt],
+                   Attr.Convergent>;
 }
 
 foreach name = ["GroupNonUniformUMin", "GroupNonUniformUMax"] in {
   def : SPVBuiltin<name, [UGenTypeN, Int, Int, UGenTypeN], Attr.Convergent>;
-  foreach IDType = TLUnsignedIntLong.List in {
-    def : SPVBuiltin<name, [UGenTypeN, Int, Int, UGenTypeN, IDType],
-                     Attr.Convergent>;
-  }
+  def : SPVBuiltin<name, [UGenTypeN, Int, Int, UGenTypeN, UInt],
+                   Attr.Convergent>;
 }
 
 foreach name = ["GroupNonUniformFAdd", "GroupNonUniformFMul",
                 "GroupNonUniformFMin", "GroupNonUniformFMax"] in {
   def : SPVBuiltin<name, [FGenTypeN, Int, Int, FGenTypeN], Attr.Convergent>;
-  foreach IDType = TLUnsignedIntLong.List in {
-    def : SPVBuiltin<name, [FGenTypeN, Int, Int, FGenTypeN, IDType],
-                     Attr.Convergent>;
-  }
+  def : SPVBuiltin<name, [FGenTypeN, Int, Int, FGenTypeN, UInt],
+                   Attr.Convergent>;
 }
 
 foreach name = ["GroupNonUniformBitwiseAnd", "GroupNonUniformBitwiseOr",
                 "GroupNonUniformBitwiseXor"] in {
   def : SPVBuiltin<name, [AIGenTypeN, Int, Int, AIGenTypeN], Attr.Convergent>;
-  foreach IDType = TLUnsignedIntLong.List in {
-    def : SPVBuiltin<name, [AIGenTypeN, Int, Int, AIGenTypeN, IDType],
-                     Attr.Convergent>;
-  }
+  def : SPVBuiltin<name, [AIGenTypeN, Int, Int, AIGenTypeN, UInt],
+                   Attr.Convergent>;
 }
 
 foreach name = ["GroupNonUniformLogicalAnd", "GroupNonUniformLogicalOr",
                 "GroupNonUniformLogicalXor"] in {
   def : SPVBuiltin<name, [BoolGenTypeN, Int, Int, BoolGenTypeN],
                    Attr.Convergent>;
-  foreach IDType = TLUnsignedIntLong.List in {
-    def : SPVBuiltin<name, [BoolGenTypeN, Int, Int, BoolGenTypeN, IDType],
-                     Attr.Convergent>;
-  }
+  def : SPVBuiltin<name, [BoolGenTypeN, Int, Int, BoolGenTypeN, UInt],
+                   Attr.Convergent>;
 }
diff --git a/clang/test/CodeGenSPIRV/spirv-builtin-lookup-group-non-uniform.cpp b/clang/test/CodeGenSPIRV/spirv-builtin-lookup-group-non-uniform.cpp
index b85265fc4c84f..08001cdda8bab 100644
--- a/clang/test/CodeGenSPIRV/spirv-builtin-lookup-group-non-uniform.cpp
+++ b/clang/test/CodeGenSPIRV/spirv-builtin-lookup-group-non-uniform.cpp
@@ -18,8 +18,8 @@ template <class T> T group_broad_cast(T v, unsigned int id) {
   return __spirv_GroupNonUniformBroadcast(2, v, id);
 }
 
-template <class T> T group_broad_cast_first(T v, unsigned int id) {
-  return __spirv_GroupNonUniformBroadcastFirst(2, v, id);
+template <class T> T group_broad_cast_first(T v) {
+  return __spirv_GroupNonUniformBroadcastFirst(2, v);
 }
 
 typedef unsigned int uint4 __attribute__((ext_vector_type(4)));
@@ -147,7 +147,7 @@ template <class T> void test_with_bool() {
   unsigned int id = 0;
   group_all_equal<T>(v);
   group_broad_cast<T>(v, id);
-  group_broad_cast_first<T>(v, id);
+  group_broad_cast_first<T>(v);
   group_shuffle(v, id);
   group_shuffle_xor(v, id);
   group_shuffle_up(v, id);
@@ -227,70 +227,70 @@ void test() {
 // CHECK: call noundef i32 @_Z36__spirv_GroupNonUniformBallotFindMSBiiDv4_j
 // CHECK: call noundef zeroext i1 @_Z31__spirv_GroupNonUniformAllEqualib
 // CHECK: call noundef zeroext i1 @_Z32__spirv_GroupNonUniformBroadcastibj
-// CHECK: call noundef zeroext i1 @_Z37__spirv_GroupNonUniformBroadcastFirstibj
+// CHECK: call noundef zeroext i1 @_Z37__spirv_GroupNonUniformBroadcastFirstib
 // CHECK: call noundef zeroext i1 @_Z30__spirv_GroupNonUniformShuffleibj
 // CHECK: call noundef zeroext i1 @_Z33__spirv_GroupNonUniformShuffleXoribj
 // CHECK: call noundef zeroext i1 @_Z32__spirv_GroupNonUniformShuffleUpibj
 // CHECK: call noundef zeroext i1 @_Z34__spirv_GroupNonUniformShuffleDownibj
 // CHECK: call noundef zeroext i1 @_Z31__spirv_GroupNonUniformAllEqualii
 // CHECK: call noundef i32 @_Z32__spirv_GroupNonUniformBroadcastiij
-// CHECK: call noundef i32 @_Z37__spirv_GroupNonUniformBroadcastFirstiij
+// CHECK: call noundef i32 @_Z37__spirv_GroupNonUniformBroadcastFirstii
 // CHECK: call noundef i32 @_Z30__spirv_GroupNonUniformShuffleiij
 // CHECK: call noundef i32 @_Z33__spirv_GroupNonUniformShuffleXoriij
 // CHECK: call noundef i32 @_Z32__spirv_GroupNonUniformShuffleUpiij
 // CHECK: call noundef i32 @_Z34__spirv_GroupNonUniformShuffleDowniij
 // CHECK: call noundef zeroext i1 @_Z31__spirv_GroupNonUniformAllEqualih
 // CHECK: call noundef zeroext i8 @_Z32__spirv_GroupNonUniformBroadcastihj
-// CHECK: call noundef zeroext i8 @_Z37__spirv_GroupNonUniformBroadcastFirstihj
+// CHECK: call noundef zeroext i8 @_Z37__spirv_GroupNonUniformBroadcastFirstih
 // CHECK: call noundef zeroext i8 @_Z30__spirv_GroupNonUniformShuffleihj
 // CHECK: call noundef zeroext i8 @_Z33__spirv_GroupNonUniformShuffleXorihj
 // CHECK: call noundef zeroext i8 @_Z32__spirv_GroupNonUniformShuffleUpihj
 // CHECK: call noundef zeroext i8 @_Z34__spirv_GroupNonUniformShuffleDownihj
 // CHECK: call noundef zeroext i1 @_Z31__spirv_GroupNonUniformAllEqualis
 // CHECK: call noundef signext i16 @_Z32__spirv_GroupNonUniformBroadcastisj
-// CHECK: call noundef signext i16 @_Z37__spirv_GroupNonUniformBroadcastFirstisj
+// CHECK: call noundef signext i16 @_Z37__spirv_GroupNonUniformBroadcastFirstis
 // CHECK: call noundef signext i16 @_Z30__spirv_GroupNonUniformShuffleisj
 // CHECK: call noundef signext i16 @_Z33__spirv_GroupNonUniformShuffleXorisj
 // CHECK: call noundef signext i16 @_Z32__spirv_GroupNonUniformShuffleUpisj
 // CHECK: call noundef signext i16 @_Z34__spirv_GroupNonUniformShuffleDownisj
 // CHECK: call noundef zeroext i1 @_Z31__spirv_GroupNonUniformAllEqualit
 // CHECK: call noundef zeroext i16 @_Z32__spirv_GroupNonUniformBroadcastitj
-// CHECK: call noundef zeroext i16 @_Z37__spirv_GroupNonUniformBroadcastFirstitj
+// CHECK: call noundef zeroext i16 @_Z37__spirv_GroupNonUniformBroadcastFirstit
 // CHECK: call noundef zeroext i16 @_Z30__spirv_GroupNonUniformShuffleitj
 // CHECK: call noundef zeroext i16 @_Z33__spirv_GroupNonUniformShuffleXoritj
 // CHECK: call noundef zeroext i16 @_Z32__spirv_GroupNonUniformShuffleUpitj
 // CHECK: call noundef zeroext i16 @_Z34__spirv_GroupNonUniformShuffleDownitj
 // CHECK: call noundef zeroext i1 @_Z31__spirv_GroupNonUniformAllEqualii
 // CHECK: call noundef i32 @_Z32__spirv_GroupNonUniformBroadcastiij
-// CHECK: call noundef i32 @_Z37__spirv_GroupNonUniformBroadcastFirstiij
+// CHECK: call noundef i32 @_Z37__spirv_GroupNonUniformBroadcastFirstii
 // CHECK: call noundef i32 @_Z30__spirv_GroupNonUniformShuffleiij
 // CHECK: call noundef i32 @_Z33__spirv_GroupNonUniformShuffleXoriij
 // CHECK: call noundef i32 @_Z32__spirv_GroupNonUniformShuffleUpiij
 // CHECK: call noundef i32 @_Z34__spirv_GroupNonUniformShuffleDowniij
 // CHECK: call noundef zeroext i1 @_Z31__spirv_GroupNonUniformAllEqualij
 // CHECK: call noundef i32 @_Z32__spirv_GroupNonUniformBroadcastijj
-// CHECK: call noundef i32 @_Z37__spirv_GroupNonUniformBroadcastFirstijj
+// CHECK: call noundef i32 @_Z37__spirv_GroupNonUniformBroadcastFirstij
 // CHECK: call noundef i32 @_Z30__spirv_GroupNonUniformShuffleijj
 // CHECK: call noundef i32 @_Z33__spirv_GroupNonUniformShuffleXorijj
 // CHECK: call noundef i32 @_Z32__spirv_GroupNonUniformShuffleUpijj
 // CHECK: call noundef i32 @_Z34__spirv_GroupNonUniformShuffleDownijj
 // CHECK: call noundef zeroext i1 @_Z31__spirv_GroupNonUniformAllEqualiDF16_
 // CHECK: call noundef half @_Z32__spirv_GroupNonUniformBroadcastiDF16_j
-// CHECK: call noundef half @_Z37__spirv_GroupNonUniformBroadcastFirstiDF16_j
+// CHECK: call noundef half @_Z37__spirv_GroupNonUniformBroadcastFirstiDF16_
 // CHECK: call noundef half @_Z30__spirv_GroupNonUniformShuffleiDF16_j
 // CHECK: call noundef half @_Z33__spirv_GroupNonUniformShuffleXoriDF16_j
 // CHECK: call noundef half @_Z32__spirv_GroupNonUniformShuffleUpiDF16_j
 // CHECK: call noundef half @_Z34__spirv_GroupNonUniformShuffleDowniDF16_j
 // CHECK: call noundef zeroext i1 @_Z31__spirv_GroupNonUniformAllEqualif
 // CHECK: call noundef float @_Z32__spirv_GroupNonUniformBroadcastifj
-// CHECK: call noundef float @_Z37__spirv_GroupNonUniformBroadcastFirstifj
+// CHECK: call noundef float @_Z37__spirv_GroupNonUniformBroadcastFirstif
 // CHECK: call noundef float @_Z30__spirv_GroupNonUniformShuffleifj
 // CHECK: call noundef float @_Z33__spirv_GroupNonUniformShuffleXorifj
 // CHECK: call noundef float @_Z32__spirv_GroupNonUniformShuffleUpifj
 // CHECK: call noundef float @_Z34__spirv_GroupNonUniformShuffleDownifj
 // CHECK: call noundef zeroext i1 @_Z31__spirv_GroupNonUniformAllEqualid
 // CHECK: call noundef double @_Z32__spirv_GroupNonUniformBroadcastidj
-// CHECK: call noundef double @_Z37__spirv_GroupNonUniformBroadcastFirstidj
+// CHECK: call noundef double @_Z37__spirv_GroupNonUniformBroadcastFirstid
 // CHECK: call noundef double @_Z30__spirv_GroupNonUniformShuffleidj
 // CHECK: call noundef double @_Z33__spirv_GroupNonUniformShuffleXoridj
 // CHECK: call noundef double @_Z32__spirv_GroupNonUniformShuffleUpidj