[clang][bytecode] Implement ia32_select* builtins

[tbaederr]

No description provided.

[llvmbot]

@llvm/pr-subscribers-clang

@llvm/pr-subscribers-backend-x86

Author: Timm Baeder (tbaederr)

Changes

---

Full diff: https://github.com/llvm/llvm-project/pull/154758.diff

2 Files Affected:

• (modified) clang/lib/AST/ByteCode/InterpBuiltin.cpp (+61)
• (modified) clang/test/CodeGen/X86/avx512f-builtins.c (+5)

diff --git a/clang/lib/AST/ByteCode/InterpBuiltin.cpp b/clang/lib/AST/ByteCode/InterpBuiltin.cpp
index 2cbebaf7b630e..c3e765cf99aa3 100644
--- a/clang/lib/AST/ByteCode/InterpBuiltin.cpp
+++ b/clang/lib/AST/ByteCode/InterpBuiltin.cpp
@@ -2778,6 +2778,40 @@ static bool interp__builtin_elementwise_fma(InterpState &S, CodePtr OpPC,
   return true;
 }
 
+/// AVX512 predicated move: "Result = Mask[] ? LHS[] : RHS[]".
+static bool interp__builtin_select(InterpState &S, CodePtr OpPC,
+                                   const CallExpr *Call, unsigned BuiltinID) {
+  const Pointer &RHS = S.Stk.pop<Pointer>();
+  const Pointer &LHS = S.Stk.pop<Pointer>();
+  PrimType MaskT = *S.getContext().classify(Call->getArg(0));
+  APSInt Mask = popToAPSInt(S.Stk, MaskT);
+  const Pointer &Dst = S.Stk.peek<Pointer>();
+
+  assert(LHS.getNumElems() == RHS.getNumElems());
+  assert(LHS.getNumElems() == Dst.getNumElems());
+  unsigned NumElems = LHS.getNumElems();
+  PrimType ElemT = LHS.getFieldDesc()->getPrimType();
+  PrimType DstElemT = Dst.getFieldDesc()->getPrimType();
+
+  for (unsigned I = 0; I != NumElems; ++I) {
+    if (ElemT == PT_Float) {
+      assert(DstElemT == PT_Float);
+      Dst.elem<Floating>(I) =
+          Mask[I] ? LHS.elem<Floating>(I) : RHS.elem<Floating>(I);
+    } else {
+      APSInt Elem;
+      INT_TYPE_SWITCH(ElemT, {
+        Elem = Mask[I] ? LHS.elem<T>(I).toAPSInt() : RHS.elem<T>(I).toAPSInt();
+      });
+      INT_TYPE_SWITCH_NO_BOOL(DstElemT,
+                              { Dst.elem<T>(I) = static_cast<T>(Elem); });
+    }
+  }
+  Dst.initializeAllElements();
+
+  return true;
+}
+
 bool InterpretBuiltin(InterpState &S, CodePtr OpPC, const CallExpr *Call,
                       uint32_t BuiltinID) {
   if (!S.getASTContext().BuiltinInfo.isConstantEvaluated(BuiltinID))
@@ -3210,9 +3244,36 @@ bool InterpretBuiltin(InterpState &S, CodePtr OpPC, const CallExpr *Call,
   case clang::X86::BI__builtin_ia32_pmuludq256:
   case clang::X86::BI__builtin_ia32_pmuludq512:
     return interp__builtin_ia32_pmul(S, OpPC, Call, BuiltinID);
+
   case Builtin::BI__builtin_elementwise_fma:
     return interp__builtin_elementwise_fma(S, OpPC, Call);
 
+  case X86::BI__builtin_ia32_selectb_128:
+  case X86::BI__builtin_ia32_selectb_256:
+  case X86::BI__builtin_ia32_selectb_512:
+  case X86::BI__builtin_ia32_selectw_128:
+  case X86::BI__builtin_ia32_selectw_256:
+  case X86::BI__builtin_ia32_selectw_512:
+  case X86::BI__builtin_ia32_selectd_128:
+  case X86::BI__builtin_ia32_selectd_256:
+  case X86::BI__builtin_ia32_selectd_512:
+  case X86::BI__builtin_ia32_selectq_128:
+  case X86::BI__builtin_ia32_selectq_256:
+  case X86::BI__builtin_ia32_selectq_512:
+  case X86::BI__builtin_ia32_selectph_128:
+  case X86::BI__builtin_ia32_selectph_256:
+  case X86::BI__builtin_ia32_selectph_512:
+  case X86::BI__builtin_ia32_selectpbf_128:
+  case X86::BI__builtin_ia32_selectpbf_256:
+  case X86::BI__builtin_ia32_selectpbf_512:
+  case X86::BI__builtin_ia32_selectps_128:
+  case X86::BI__builtin_ia32_selectps_256:
+  case X86::BI__builtin_ia32_selectps_512:
+  case X86::BI__builtin_ia32_selectpd_128:
+  case X86::BI__builtin_ia32_selectpd_256:
+  case X86::BI__builtin_ia32_selectpd_512:
+    return interp__builtin_select(S, OpPC, Call, BuiltinID);
+
   default:
     S.FFDiag(S.Current->getLocation(OpPC),
              diag::note_invalid_subexpr_in_const_expr)
diff --git a/clang/test/CodeGen/X86/avx512f-builtins.c b/clang/test/CodeGen/X86/avx512f-builtins.c
index 14a4bbf712f8c..29916be23cbae 100644
--- a/clang/test/CodeGen/X86/avx512f-builtins.c
+++ b/clang/test/CodeGen/X86/avx512f-builtins.c
@@ -3,6 +3,11 @@
 // RUN: %clang_cc1 -x c++ -flax-vector-conversions=none -ffreestanding %s -triple=x86_64-apple-darwin -target-feature +avx512f -emit-llvm -o - -Wall -Werror -Wsign-conversion | FileCheck %s
 // RUN: %clang_cc1 -x c++ -flax-vector-conversions=none -fms-extensions -fms-compatibility -ffreestanding %s -triple=x86_64-windows-msvc -target-feature +avx512f -emit-llvm -o - -Wall -Werror -Wsign-conversion | FileCheck %s
 
+// RUN: %clang_cc1 -x c -flax-vector-conversions=none -ffreestanding %s -triple=x86_64-apple-darwin -target-feature +avx512f -emit-llvm -o - -Wall -Werror -Wsign-conversion -fexperimental-new-constant-interpreter | FileCheck %s
+// RUN: %clang_cc1 -x c -flax-vector-conversions=none -fms-extensions -fms-compatibility -ffreestanding %s -triple=x86_64-windows-msvc -target-feature +avx512f -emit-llvm -o - -Wall -Werror -Wsign-conversion -fexperimental-new-constant-interpreter | FileCheck %s
+// RUN: %clang_cc1 -x c++ -flax-vector-conversions=none -ffreestanding %s -triple=x86_64-apple-darwin -target-feature +avx512f -emit-llvm -o - -Wall -Werror -Wsign-conversion -fexperimental-new-constant-interpreter | FileCheck %s
+// RUN: %clang_cc1 -x c++ -flax-vector-conversions=none -fms-extensions -fms-compatibility -ffreestanding %s -triple=x86_64-windows-msvc -target-feature +avx512f -emit-llvm -o - -Wall -Werror -Wsign-conversion -fexperimental-new-constant-interpreter | FileCheck %s
+
 #include <immintrin.h>
 #include "builtin_test_helpers.h"
 

[RKSimon]

@tbaederr for the new interpreter is the plan to always have target builtins in the same file and switch statement - or should we split them off into their own files like CodeGen did with EmitTargetArchBuiltinExpr?

[tbaederr]

I don't see a good reason for different implementation files right now, InterpBuiltin.cpp isn't even that long. That might change in the future though.

[RKSimon]

There are a decent number of x86 intrinsics I'm hoping to support in constexpr - we'll have to see if we step over any kind of tolerance limit :/

[shafik]

assert(Call->getNumArgs() == 1);?

[brandonxin]

Hello, I am new to LLVM/Clang. I want to ask why there is an INT_TYPE_SWITCH and then another INT_TYPE_SWITCH_NO_BOOL. Is it possible that ElemT is PT_Bool? I checked BuiltinsX86.td and all these select buitins require vectors of integers as their parameters.

[tbaederr]

I don't remember anymore, but looking at this now, I'd probably use one only one of those switches. If it can't be bool, there should be an assert somewhere.