[clang] [X86] Add MMX/SSE/AVX PHADD/SUB & HADDPS/D intrinsics to be used in constexpr (PR #156822)

[Simon Pilgrim via cfe-commits]

================
@@ -3674,6 +3750,55 @@ bool InterpretBuiltin(InterpState &S, CodePtr OpPC, const CallExpr *Call,
   case Builtin::BI__builtin_elementwise_min:
     return interp__builtin_elementwise_maxmin(S, OpPC, Call, BuiltinID);
 
+  case clang::X86::BI__builtin_ia32_phaddw128:
+  case clang::X86::BI__builtin_ia32_phaddw256:
+  case clang::X86::BI__builtin_ia32_phaddd128:
+  case clang::X86::BI__builtin_ia32_phaddd256:
+    return interp_builtin_horizontal_int_binop(
+        S, OpPC, Call,
+        [](const APSInt &LHS, const APSInt &RHS) { return LHS + RHS; });
+  case clang::X86::BI__builtin_ia32_phaddsw128:
+  case clang::X86::BI__builtin_ia32_phaddsw256:
+    return interp_builtin_horizontal_int_binop(
+        S, OpPC, Call, [](const APSInt &LHS, const APSInt &RHS) {
+          return LHS.isSigned() ? LHS.sadd_sat(RHS) : LHS.uadd_sat(RHS);
+        });
+  case clang::X86::BI__builtin_ia32_phsubw128:
+  case clang::X86::BI__builtin_ia32_phsubw256:
+  case clang::X86::BI__builtin_ia32_phsubd128:
+  case clang::X86::BI__builtin_ia32_phsubd256:
+    return interp_builtin_horizontal_int_binop(
+        S, OpPC, Call,
+        [](const APSInt &LHS, const APSInt &RHS) { return LHS - RHS; });
+  case clang::X86::BI__builtin_ia32_phsubsw128:
+  case clang::X86::BI__builtin_ia32_phsubsw256:
+    return interp_builtin_horizontal_int_binop(
+        S, OpPC, Call, [](const APSInt &LHS, const APSInt &RHS) {
+          return LHS.isSigned() ? LHS.ssub_sat(RHS) : LHS.usub_sat(RHS);
----------------
RKSimon wrote:

horizontal integer saturated add/sub are all signed operations - you don't need usub_sat at all

https://github.com/llvm/llvm-project/pull/156822