From 62f44d7f20cf939ae0bd487d2e70ac1411e9db88 Mon Sep 17 00:00:00 2001 From: Antoni Boucher Date: Fri, 6 Sep 2024 12:27:20 -0400 Subject: [PATCH] Add more SIMD intrinsics --- src/intrinsic/llvm.rs | 45 +++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 45 insertions(+) diff --git a/src/intrinsic/llvm.rs b/src/intrinsic/llvm.rs index e57dc1cc75f..1a3385a97be 100644 --- a/src/intrinsic/llvm.rs +++ b/src/intrinsic/llvm.rs @@ -418,6 +418,27 @@ pub fn adjust_intrinsic_arguments<'a, 'b, 'gcc, 'tcx>( new_args.push(old_args.swap_remove(0)); args = new_args.into(); } + "__builtin_ia32_addph512_mask_round" + | "__builtin_ia32_subph512_mask_round" + | "__builtin_ia32_mulph512_mask_round" + | "__builtin_ia32_divph512_mask_round" => { + let mut new_args = args.to_vec(); + let last_arg = new_args.pop().expect("last arg"); + + let arg3_type = gcc_func.get_param_type(2); + let vector_type = arg3_type.dyncast_vector().expect("vector type"); + let zero = builder.context.new_rvalue_zero(vector_type.get_element_type()); + let num_units = vector_type.get_num_units(); + let first_arg = + builder.context.new_rvalue_from_vector(None, arg3_type, &vec![zero; num_units]); + new_args.push(first_arg); + + let arg4_type = gcc_func.get_param_type(3); + let minus_one = builder.context.new_rvalue_from_int(arg4_type, -1); + new_args.push(minus_one); + new_args.push(last_arg); + args = new_args.into(); + } _ => (), } } else { @@ -1094,6 +1115,23 @@ pub fn intrinsic<'gcc, 'tcx>(name: &str, cx: &CodegenCx<'gcc, 'tcx>) -> Function "llvm.x86.avx512.mask.load.q.256" => "__builtin_ia32_movdqa64load256_mask", "llvm.x86.avx512fp16.mask.cmp.sh" => "__builtin_ia32_cmpsh_mask_round", "llvm.x86.avx512fp16.vcomi.sh" => "__builtin_ia32_cmpsh_mask_round", + "llvm.x86.avx512fp16.add.ph.512" => "__builtin_ia32_addph512_mask_round", + "llvm.x86.avx512fp16.sub.ph.512" => "__builtin_ia32_subph512_mask_round", + "llvm.x86.avx512fp16.mul.ph.512" => "__builtin_ia32_mulph512_mask_round", + "llvm.x86.avx512fp16.div.ph.512" => "__builtin_ia32_divph512_mask_round", + "llvm.x86.avx512fp16.mask.vfmul.cph.512" => "__builtin_ia32_vfmulcph512_mask_round", + "llvm.x86.avx512fp16.mask.vfmul.csh" => "__builtin_ia32_vfmulcsh_mask_round", + "llvm.x86.avx512fp16.mask.vfcmul.cph.512" => "__builtin_ia32_vfcmulcph512_mask_round", + "llvm.x86.avx512fp16.mask.vfcmul.csh" => "__builtin_ia32_vfcmulcsh_mask_round", + "llvm.x86.avx512fp16.mask.vfmadd.cph.512" => "__builtin_ia32_vfmaddcph512_mask3_round", + "llvm.x86.avx512fp16.maskz.vfmadd.cph.512" => "__builtin_ia32_vfmaddcph512_maskz_round", + "llvm.x86.avx512fp16.mask.vfmadd.csh" => "__builtin_ia32_vfmaddcsh_mask3_round", + "llvm.x86.avx512fp16.maskz.vfmadd.csh" => "__builtin_ia32_vfmaddcsh_maskz_round", + "llvm.x86.avx512fp16.mask.vfcmadd.cph.512" => "__builtin_ia32_vfcmaddcph512_mask3_round", + "llvm.x86.avx512fp16.maskz.vfcmadd.cph.512" => "__builtin_ia32_vfcmaddcph512_maskz_round", + "llvm.x86.avx512fp16.mask.vfcmadd.csh" => "__builtin_ia32_vfcmaddcsh_mask_round", + "llvm.x86.avx512fp16.maskz.vfcmadd.csh" => "__builtin_ia32_vfcmaddcsh_maskz_round", + // TODO: support the tile builtins: "llvm.x86.ldtilecfg" => "__builtin_trap", "llvm.x86.sttilecfg" => "__builtin_trap", @@ -1103,6 +1141,13 @@ pub fn intrinsic<'gcc, 'tcx>(name: &str, cx: &CodegenCx<'gcc, 'tcx>) -> Function "llvm.x86.tileloaddt164" => "__builtin_trap", "llvm.x86.tilezero" => "__builtin_trap", "llvm.x86.tdpbf16ps" => "__builtin_trap", + "llvm.x86.tdpbssd" => "__builtin_trap", + "llvm.x86.tdpbsud" => "__builtin_trap", + "llvm.x86.tdpbusd" => "__builtin_trap", + "llvm.x86.tdpbuud" => "__builtin_trap", + "llvm.x86.tdpfp16ps" => "__builtin_trap", + "llvm.x86.tcmmimfp16ps" => "__builtin_trap", + "llvm.x86.tcmmrlfp16ps" => "__builtin_trap", // NOTE: this file is generated by https://github.com/GuillaumeGomez/llvmint/blob/master/generate_list.py _ => include!("archs.rs"),