[ESIMD] Fix the ambiguous call error for a corner case of gather() (#…

…12506)

Some corner cases with interger scalar operands of gather() may cause
ambiguous call error. Examples of such calls:
  // 0 may be treated as vector `simd_mask` or scalar 'global_offset'
  auto res1 = gather<T, N>(acc, offsets_simd, 0);
  // This case is more tricky, but also can confuse C++ FE.
  auto res2 = gather<T, N>(acc, offsets_simd, 0, mask);

Signed-off-by: Klochkov, Vyacheslav N <vyacheslav.n.klochkov@intel.com>

sycl/include/sycl/ext/intel/esimd/memory.hpp

@@ -2934,15 +2934,21 @@ gather(AccessorT acc, simd<OffsetT, N / VS> byte_offsets,
 /// to 1. This variant is added for convenience and let user omit the template
 /// arguments and call the function as
 /// 'gather(acc, byte_offsets, mask, pass_thru);'.
+// Dev note: the mask type was turned into template parameter `MaskT` to
+// avoid the conflicts of this prototype with the old gather() function
+// accepting a 'global_offset' parameter and avoid 'ambiguous call' errors
+// for calls like this: gather(acc, byte_offsets_simd, 0, mask);
 template <typename T, int N, typename AccessorT, typename OffsetT,
+          typename MaskT,
           typename PropertyListT =
               ext::oneapi::experimental::detail::empty_properties_t>
 __ESIMD_API std::enable_if_t<
     (detail::is_device_accessor_with_v<AccessorT,
                                        detail::accessor_mode_cap::can_read> &&
+     std::is_same_v<MaskT, simd_mask<N>> &&
      ext::oneapi::experimental::is_property_list_v<PropertyListT>),
     simd<T, N>>
-gather(AccessorT acc, simd<OffsetT, N> byte_offsets, simd_mask<N> mask,
+gather(AccessorT acc, simd<OffsetT, N> byte_offsets, MaskT mask,
        simd<T, N> pass_thru, PropertyListT props = {}) {
   return gather<T, N, 1>(acc, byte_offsets, mask, pass_thru, props);
 }
@@ -2954,15 +2960,19 @@ gather(AccessorT acc, simd<OffsetT, N> byte_offsets, simd_mask<N> mask,
 /// This function is identical to (acc-ga-2) except that vector size is fixed
 /// to 1. This variant is added for convenience and let user omit the template
 /// arguments and call the function as 'gather(acc, byte_offsets, mask);'.
+// Dev note: the mask type was turned into template parameter `MaskT` to
+// avoid the conflicts of this prototype with the old gather() function
+// accepting a 'global_offset' parameter and avoid 'ambiguous call' errors
+// for calls like this: gather(acc, byte_offsets_simd, 0);
 template <typename T, int N, typename AccessorT, typename OffsetT,
           typename MaskT,
           typename PropertyListT =
               ext::oneapi::experimental::detail::empty_properties_t>
 __ESIMD_API std::enable_if_t<
     (detail::is_device_accessor_with_v<AccessorT,
                                        detail::accessor_mode_cap::can_read> &&
-     ext::oneapi::experimental::is_property_list_v<PropertyListT> &&
-     std::is_same_v<MaskT, simd_mask<N>>),
+     std::is_same_v<MaskT, simd_mask<N>> &&
+     ext::oneapi::experimental::is_property_list_v<PropertyListT>),
     simd<T, N>>
 gather(AccessorT acc, simd<OffsetT, N> byte_offsets, MaskT mask,
        PropertyListT props = {}) {

sycl/test/esimd/intrins_trans.cpp

@@ -304,14 +304,26 @@ SYCL_ESIMD_FUNCTION SYCL_EXTERNAL simd<float, 16> foo() {
     // CHECK-STATEFUL: call void @llvm.genx.scatter.scaled.v8i1.v8i32.v8i32(<8 x i1> %{{[0-9a-zA-Z_.]+}}, i32 2, i16 0, i32 %[[SI4]], i32 %{{[0-9a-zA-Z_.]+}}, <8 x i32> %{{[0-9a-zA-Z_.]+}}, <8 x i32> %{{[0-9a-zA-Z_.]+}})
     // CHECK-STATELESS: call void @llvm.genx.svm.scatter.v8i1.v8i64.v8i32(<8 x i1> %{{[0-9a-zA-Z_.]+}}, i32 0, <8 x i64> %{{[0-9a-zA-Z_.]+}}, <8 x i32> %{{[0-9a-zA-Z_.]+}})
 
-    // 1-byte element gather
+    // 1-byte element gather: same code with and without mask
     simd<unsigned char, 8> v1 = gather<unsigned char, 8>(acc, offsets, 100);
     // CHECK-STATEFUL: %[[SI5_VAL:[0-9a-zA-Z_.]+]] = call spir_func noundef i32 @_Z21__spirv_ConvertPtrToU{{.*}}(ptr addrspace(1) noundef %{{[0-9a-zA-Z_.]+}})
     // CHECK-STATEFUL: store i32 %[[SI5_VAL]], ptr addrspace(4) %[[SI5_ADDR:[0-9a-zA-Z_.]+]]
     // CHECK-STATEFUL: %[[SI5:[0-9a-zA-Z_.]+]] = load i32, ptr addrspace(4) %[[SI5_ADDR]]
     // CHECK-STATEFUL: call <8 x i32> @llvm.genx.gather.masked.scaled2.v8i32.v8i32.v8i1(i32 0, i16 0, i32 %[[SI5]], i32 %{{[0-9a-zA-Z_.]+}}, <8 x i32> %{{[0-9a-zA-Z_.]+}}, <8 x i1> %{{[0-9a-zA-Z_.]+}})
     // CHECK-STATELESS: call <32 x i8> @llvm.genx.svm.gather.v32i8.v8i1.v8i64(<8 x i1> %{{[0-9a-zA-Z_.]+}}, i32 0, <8 x i64> %{{[0-9a-zA-Z_.]+}}, <32 x i8> undef)
 
+    // 1-byte element gather using the mask
+    v1 = gather<unsigned char, 8>(acc, offsets, 100, pred);
+    // CHECK-STATEFUL: call <8 x i32> @llvm.genx.gather.masked.scaled2.v8i32.v8i32.v8i1(i32 0, i16 0, i32 {{[^)]+}}, i32 {{[^)]+}}, <8 x i32> {{[^)]+}}, <8 x i1> {{[^)]+}})
+    // CHECK-STATELESS: call <32 x i8> @llvm.genx.svm.gather.v32i8.v8i1.v8i64(<8 x i1> {{[^)]+}}, i32 0, <8 x i64> {{[^)]+}}, <32 x i8> undef)
+
+    // 1-byte element gather using the mask - the mask is signed, which may
+    // expose different issues/conflicts in gather API.
+    simd<int32_t, 8> ioffsets = 1;
+    v1 = gather<unsigned char, 8>(acc, ioffsets, 0, pred);
+    // CHECK-STATEFUL: call <8 x i32> @llvm.genx.gather.masked.scaled2.v8i32.v8i32.v8i1(i32 0, i16 0, i32 {{[^)]+}}, i32 {{[^)]+}}, <8 x i32> {{[^)]+}}, <8 x i1> {{[^)]+}})
+    // CHECK-STATELESS: call <32 x i8> @llvm.genx.svm.gather.v32i8.v8i1.v8i64(<8 x i1> {{[^)]+}}, i32 0, <8 x i64> {{[^)]+}}, <32 x i8> undef)
+
     // 1-byte element scatter
     scatter<unsigned char, 8>(acc, offsets, v1, 100, pred);
     // CHECK-STATEFUL: %[[SI6_VAL:[0-9a-zA-Z_.]+]] = call spir_func noundef i32 @_Z21__spirv_ConvertPtrToU{{.*}}(ptr addrspace(1) noundef %{{[0-9a-zA-Z_.]+}})