Index: include/llvm/IR/IntrinsicsX86.td =================================================================== --- include/llvm/IR/IntrinsicsX86.td +++ include/llvm/IR/IntrinsicsX86.td @@ -13,7 +13,7 @@ //===----------------------------------------------------------------------===// // Interrupt traps let TargetPrefix = "x86" in { // All intrinsics start with "llvm.x86.". - def int_x86_int : Intrinsic<[], [llvm_i8_ty]>; + def int_x86_int : Intrinsic<[], [llvm_i8_ty], [ImmArg<0>]>; } //===----------------------------------------------------------------------===// @@ -203,12 +203,12 @@ let TargetPrefix = "x86" in { // All intrinsics start with "llvm.x86.". def int_x86_sse_cmp_ss : GCCBuiltin<"__builtin_ia32_cmpss">, Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty, - llvm_v4f32_ty, llvm_i8_ty], [IntrNoMem]>; + llvm_v4f32_ty, llvm_i8_ty], [IntrNoMem, ImmArg<2>]>; // NOTE: This comparison intrinsic is not used by clang as long as the // distinction in signaling behaviour is not implemented. def int_x86_sse_cmp_ps : Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty, - llvm_v4f32_ty, llvm_i8_ty], [IntrNoMem]>; + llvm_v4f32_ty, llvm_i8_ty], [IntrNoMem, ImmArg<2>]>; def int_x86_sse_comieq_ss : GCCBuiltin<"__builtin_ia32_comieq">, Intrinsic<[llvm_i32_ty], [llvm_v4f32_ty, llvm_v4f32_ty], [IntrNoMem]>; @@ -311,12 +311,12 @@ let TargetPrefix = "x86" in { // All intrinsics start with "llvm.x86.". def int_x86_sse2_cmp_sd : GCCBuiltin<"__builtin_ia32_cmpsd">, Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty, - llvm_v2f64_ty, llvm_i8_ty], [IntrNoMem]>; + llvm_v2f64_ty, llvm_i8_ty], [IntrNoMem, ImmArg<2>]>; // NOTE: This comparison intrinsic is not used by clang as long as the // distinction in signaling behaviour is not implemented. def int_x86_sse2_cmp_pd : Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty, - llvm_v2f64_ty, llvm_i8_ty], [IntrNoMem]>; + llvm_v2f64_ty, llvm_i8_ty], [IntrNoMem, ImmArg<2>]>; def int_x86_sse2_comieq_sd : GCCBuiltin<"__builtin_ia32_comisdeq">, Intrinsic<[llvm_i32_ty], [llvm_v2f64_ty, llvm_v2f64_ty], [IntrNoMem]>; @@ -398,6 +398,7 @@ Intrinsic<[llvm_v4i32_ty], [llvm_v4i32_ty, llvm_v4i32_ty], [IntrNoMem]>; + // Oddly these don't require an immediate due to a gcc compatibility issue. def int_x86_sse2_pslli_w : GCCBuiltin<"__builtin_ia32_psllwi128">, Intrinsic<[llvm_v8i16_ty], [llvm_v8i16_ty, llvm_i32_ty], [IntrNoMem]>; @@ -603,7 +604,7 @@ llvm_v16i8_ty], [IntrNoMem]>; def int_x86_sse_pshuf_w : GCCBuiltin<"__builtin_ia32_pshufw">, Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty, llvm_i8_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<1>]>; } // Sign ops @@ -649,16 +650,16 @@ let TargetPrefix = "x86" in { // All intrinsics start with "llvm.x86.". def int_x86_sse41_round_ss : GCCBuiltin<"__builtin_ia32_roundss">, Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty, llvm_v4f32_ty, - llvm_i32_ty], [IntrNoMem]>; + llvm_i32_ty], [IntrNoMem, ImmArg<2>]>; def int_x86_sse41_round_ps : GCCBuiltin<"__builtin_ia32_roundps">, Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty, - llvm_i32_ty], [IntrNoMem]>; + llvm_i32_ty], [IntrNoMem, ImmArg<1>]>; def int_x86_sse41_round_sd : GCCBuiltin<"__builtin_ia32_roundsd">, Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty, llvm_v2f64_ty, - llvm_i32_ty], [IntrNoMem]>; + llvm_i32_ty], [IntrNoMem, ImmArg<2>]>; def int_x86_sse41_round_pd : GCCBuiltin<"__builtin_ia32_roundpd">, Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty, - llvm_i32_ty], [IntrNoMem]>; + llvm_i32_ty], [IntrNoMem, ImmArg<1>]>; } // Vector min element @@ -721,20 +722,20 @@ def int_x86_aesni_aeskeygenassist : GCCBuiltin<"__builtin_ia32_aeskeygenassist128">, Intrinsic<[llvm_v2i64_ty], [llvm_v2i64_ty, llvm_i8_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<1>]>; } // PCLMUL instructions let TargetPrefix = "x86" in { // All intrinsics start with "llvm.x86.". def int_x86_pclmulqdq : GCCBuiltin<"__builtin_ia32_pclmulqdq128">, Intrinsic<[llvm_v2i64_ty], [llvm_v2i64_ty, llvm_v2i64_ty, llvm_i8_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<2>]>; def int_x86_pclmulqdq_256 : GCCBuiltin<"__builtin_ia32_pclmulqdq256">, Intrinsic<[llvm_v4i64_ty], [llvm_v4i64_ty, llvm_v4i64_ty, llvm_i8_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<2>]>; def int_x86_pclmulqdq_512 : GCCBuiltin<"__builtin_ia32_pclmulqdq512">, Intrinsic<[llvm_v8i64_ty], [llvm_v8i64_ty, llvm_v8i64_ty, llvm_i8_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<2>]>; } // Vector pack @@ -748,7 +749,7 @@ let TargetPrefix = "x86" in { // All intrinsics start with "llvm.x86.". def int_x86_sse41_insertps : GCCBuiltin<"__builtin_ia32_insertps128">, Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty, llvm_v4f32_ty, llvm_i8_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<2>]>; } // Vector blend @@ -768,17 +769,17 @@ let TargetPrefix = "x86" in { // All intrinsics start with "llvm.x86.". def int_x86_sse41_dppd : GCCBuiltin<"__builtin_ia32_dppd">, Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty, llvm_v2f64_ty, llvm_i8_ty], - [IntrNoMem, Commutative]>; + [IntrNoMem, Commutative, ImmArg<2>]>; def int_x86_sse41_dpps : GCCBuiltin<"__builtin_ia32_dpps">, Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty, llvm_v4f32_ty, llvm_i8_ty], - [IntrNoMem, Commutative]>; + [IntrNoMem, Commutative, ImmArg<2>]>; } // Vector sum of absolute differences let TargetPrefix = "x86" in { // All intrinsics start with "llvm.x86.". def int_x86_sse41_mpsadbw : GCCBuiltin<"__builtin_ia32_mpsadbw128">, Intrinsic<[llvm_v8i16_ty], [llvm_v16i8_ty, llvm_v16i8_ty,llvm_i8_ty], - [IntrNoMem, Commutative]>; + [IntrNoMem, Commutative, ImmArg<2>]>; } // Test instruction with bitwise comparison. @@ -819,66 +820,66 @@ def int_x86_sse42_pcmpistrm128 : GCCBuiltin<"__builtin_ia32_pcmpistrm128">, Intrinsic<[llvm_v16i8_ty], [llvm_v16i8_ty, llvm_v16i8_ty, llvm_i8_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<2>]>; def int_x86_sse42_pcmpistri128 : GCCBuiltin<"__builtin_ia32_pcmpistri128">, Intrinsic<[llvm_i32_ty], [llvm_v16i8_ty, llvm_v16i8_ty, llvm_i8_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<2>]>; def int_x86_sse42_pcmpistria128 : GCCBuiltin<"__builtin_ia32_pcmpistria128">, Intrinsic<[llvm_i32_ty], [llvm_v16i8_ty, llvm_v16i8_ty, llvm_i8_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<2>]>; def int_x86_sse42_pcmpistric128 : GCCBuiltin<"__builtin_ia32_pcmpistric128">, Intrinsic<[llvm_i32_ty], [llvm_v16i8_ty, llvm_v16i8_ty, llvm_i8_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<2>]>; def int_x86_sse42_pcmpistrio128 : GCCBuiltin<"__builtin_ia32_pcmpistrio128">, Intrinsic<[llvm_i32_ty], [llvm_v16i8_ty, llvm_v16i8_ty, llvm_i8_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<2>]>; def int_x86_sse42_pcmpistris128 : GCCBuiltin<"__builtin_ia32_pcmpistris128">, Intrinsic<[llvm_i32_ty], [llvm_v16i8_ty, llvm_v16i8_ty, llvm_i8_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<2>]>; def int_x86_sse42_pcmpistriz128 : GCCBuiltin<"__builtin_ia32_pcmpistriz128">, Intrinsic<[llvm_i32_ty], [llvm_v16i8_ty, llvm_v16i8_ty, llvm_i8_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<2>]>; def int_x86_sse42_pcmpestrm128 : GCCBuiltin<"__builtin_ia32_pcmpestrm128">, Intrinsic<[llvm_v16i8_ty], [llvm_v16i8_ty, llvm_i32_ty, llvm_v16i8_ty, llvm_i32_ty, llvm_i8_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<4>]>; def int_x86_sse42_pcmpestri128 : GCCBuiltin<"__builtin_ia32_pcmpestri128">, Intrinsic<[llvm_i32_ty], [llvm_v16i8_ty, llvm_i32_ty, llvm_v16i8_ty, llvm_i32_ty, llvm_i8_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<4>]>; def int_x86_sse42_pcmpestria128 : GCCBuiltin<"__builtin_ia32_pcmpestria128">, Intrinsic<[llvm_i32_ty], [llvm_v16i8_ty, llvm_i32_ty, llvm_v16i8_ty, llvm_i32_ty, llvm_i8_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<4>]>; def int_x86_sse42_pcmpestric128 : GCCBuiltin<"__builtin_ia32_pcmpestric128">, Intrinsic<[llvm_i32_ty], [llvm_v16i8_ty, llvm_i32_ty, llvm_v16i8_ty, llvm_i32_ty, llvm_i8_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<4>]>; def int_x86_sse42_pcmpestrio128 : GCCBuiltin<"__builtin_ia32_pcmpestrio128">, Intrinsic<[llvm_i32_ty], [llvm_v16i8_ty, llvm_i32_ty, llvm_v16i8_ty, llvm_i32_ty, llvm_i8_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<4>]>; def int_x86_sse42_pcmpestris128 : GCCBuiltin<"__builtin_ia32_pcmpestris128">, Intrinsic<[llvm_i32_ty], [llvm_v16i8_ty, llvm_i32_ty, llvm_v16i8_ty, llvm_i32_ty, llvm_i8_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<4>]>; def int_x86_sse42_pcmpestriz128 : GCCBuiltin<"__builtin_ia32_pcmpestriz128">, Intrinsic<[llvm_i32_ty], [llvm_v16i8_ty, llvm_i32_ty, llvm_v16i8_ty, llvm_i32_ty, llvm_i8_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<4>]>; } //===----------------------------------------------------------------------===// @@ -887,13 +888,14 @@ let TargetPrefix = "x86" in { // All intrinsics start with "llvm.x86.". def int_x86_sse4a_extrqi : GCCBuiltin<"__builtin_ia32_extrqi">, Intrinsic<[llvm_v2i64_ty], [llvm_v2i64_ty, llvm_i8_ty, llvm_i8_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<1>, ImmArg<2>]>; def int_x86_sse4a_extrq : GCCBuiltin<"__builtin_ia32_extrq">, Intrinsic<[llvm_v2i64_ty], [llvm_v2i64_ty, llvm_v16i8_ty], [IntrNoMem]>; def int_x86_sse4a_insertqi : GCCBuiltin<"__builtin_ia32_insertqi">, Intrinsic<[llvm_v2i64_ty], [llvm_v2i64_ty, llvm_v2i64_ty, - llvm_i8_ty, llvm_i8_ty], [IntrNoMem]>; + llvm_i8_ty, llvm_i8_ty], + [IntrNoMem, ImmArg<2>, ImmArg<3>]>; def int_x86_sse4a_insertq : GCCBuiltin<"__builtin_ia32_insertq">, Intrinsic<[llvm_v2i64_ty], [llvm_v2i64_ty, llvm_v2i64_ty], [IntrNoMem]>; } @@ -930,10 +932,10 @@ def int_x86_avx_round_pd_256 : GCCBuiltin<"__builtin_ia32_roundpd256">, Intrinsic<[llvm_v4f64_ty], [llvm_v4f64_ty, - llvm_i32_ty], [IntrNoMem]>; + llvm_i32_ty], [IntrNoMem, ImmArg<1>]>; def int_x86_avx_round_ps_256 : GCCBuiltin<"__builtin_ia32_roundps256">, Intrinsic<[llvm_v8f32_ty], [llvm_v8f32_ty, - llvm_i32_ty], [IntrNoMem]>; + llvm_i32_ty], [IntrNoMem, ImmArg<1>]>; } // Horizontal ops @@ -1085,33 +1087,33 @@ GCCBuiltin<"__builtin_ia32_vgf2p8affineinvqb_v16qi">, Intrinsic<[llvm_v16i8_ty], [llvm_v16i8_ty, llvm_v16i8_ty, llvm_i8_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<2>]>; def int_x86_vgf2p8affineinvqb_256 : GCCBuiltin<"__builtin_ia32_vgf2p8affineinvqb_v32qi">, Intrinsic<[llvm_v32i8_ty], [llvm_v32i8_ty, llvm_v32i8_ty, llvm_i8_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<2>]>; def int_x86_vgf2p8affineinvqb_512 : GCCBuiltin<"__builtin_ia32_vgf2p8affineinvqb_v64qi">, Intrinsic<[llvm_v64i8_ty], [llvm_v64i8_ty, llvm_v64i8_ty, llvm_i8_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<2>]>; def int_x86_vgf2p8affineqb_128 : GCCBuiltin<"__builtin_ia32_vgf2p8affineqb_v16qi">, Intrinsic<[llvm_v16i8_ty], [llvm_v16i8_ty, llvm_v16i8_ty, llvm_i8_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<2>]>; def int_x86_vgf2p8affineqb_256 : GCCBuiltin<"__builtin_ia32_vgf2p8affineqb_v32qi">, Intrinsic<[llvm_v32i8_ty], [llvm_v32i8_ty, llvm_v32i8_ty, llvm_i8_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<2>]>; def int_x86_vgf2p8affineqb_512 : GCCBuiltin<"__builtin_ia32_vgf2p8affineqb_v64qi">, Intrinsic<[llvm_v64i8_ty], [llvm_v64i8_ty, llvm_v64i8_ty, llvm_i8_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<2>]>; def int_x86_vgf2p8mulb_128 : GCCBuiltin<"__builtin_ia32_vgf2p8mulb_v16qi">, @@ -1144,17 +1146,18 @@ let TargetPrefix = "x86" in { // All intrinsics start with "llvm.x86.". def int_x86_avx_dp_ps_256 : GCCBuiltin<"__builtin_ia32_dpps256">, Intrinsic<[llvm_v8f32_ty], [llvm_v8f32_ty, - llvm_v8f32_ty, llvm_i8_ty], [IntrNoMem, Commutative]>; + llvm_v8f32_ty, llvm_i8_ty], + [IntrNoMem, Commutative, ImmArg<2>]>; } // Vector compare let TargetPrefix = "x86" in { // All intrinsics start with "llvm.x86.". def int_x86_avx_cmp_pd_256 : Intrinsic<[llvm_v4f64_ty], [llvm_v4f64_ty, - llvm_v4f64_ty, llvm_i8_ty], [IntrNoMem]>; + llvm_v4f64_ty, llvm_i8_ty], [IntrNoMem, ImmArg<2>]>; def int_x86_avx_cmp_ps_256 : Intrinsic<[llvm_v8f32_ty], [llvm_v8f32_ty, - llvm_v8f32_ty, llvm_i8_ty], [IntrNoMem]>; + llvm_v8f32_ty, llvm_i8_ty], [IntrNoMem, ImmArg<2>]>; } // Vector convert @@ -1221,30 +1224,30 @@ def int_x86_avx512_fpclass_pd_128 : Intrinsic<[llvm_v2i1_ty], [llvm_v2f64_ty, llvm_i32_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<1>]>; def int_x86_avx512_fpclass_pd_256 : Intrinsic<[llvm_v4i1_ty], [llvm_v4f64_ty, llvm_i32_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<1>]>; def int_x86_avx512_fpclass_pd_512 : Intrinsic<[llvm_v8i1_ty], [llvm_v8f64_ty, llvm_i32_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<1>]>; def int_x86_avx512_fpclass_ps_128 : Intrinsic<[llvm_v4i1_ty], [llvm_v4f32_ty, llvm_i32_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<1>]>; def int_x86_avx512_fpclass_ps_256 : Intrinsic<[llvm_v8i1_ty], [llvm_v8f32_ty, llvm_i32_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<1>]>; def int_x86_avx512_fpclass_ps_512 : Intrinsic<[llvm_v16i1_ty], [llvm_v16f32_ty, llvm_i32_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<1>]>; def int_x86_avx512_mask_fpclass_sd : GCCBuiltin<"__builtin_ia32_fpclasssd_mask">, Intrinsic<[llvm_i8_ty], [llvm_v2f64_ty, llvm_i32_ty, llvm_i8_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<1>]>; def int_x86_avx512_mask_fpclass_ss : GCCBuiltin<"__builtin_ia32_fpclassss_mask">, Intrinsic<[llvm_i8_ty], [llvm_v4f32_ty, llvm_i32_ty, llvm_i8_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<1>]>; } // Vector extract sign mask @@ -1359,6 +1362,7 @@ Intrinsic<[llvm_v8i32_ty], [llvm_v8i32_ty, llvm_v4i32_ty], [IntrNoMem]>; + // Oddly these don't require an immediate due to a gcc compatibility issue. def int_x86_avx2_pslli_w : GCCBuiltin<"__builtin_ia32_psllwi256">, Intrinsic<[llvm_v16i16_ty], [llvm_v16i16_ty, llvm_i32_ty], [IntrNoMem]>; @@ -1391,6 +1395,7 @@ Intrinsic<[llvm_v4i64_ty], [llvm_v4i64_ty, llvm_v2i64_ty], [IntrNoMem]>; + // Oddly these don't require an immediate due to a gcc compatibility issue. def int_x86_avx512_psrai_q_128 : GCCBuiltin<"__builtin_ia32_psraqi128">, Intrinsic<[llvm_v2i64_ty], [llvm_v2i64_ty, llvm_i32_ty], [IntrNoMem]>; @@ -1426,6 +1431,7 @@ Intrinsic<[llvm_v8i64_ty], [llvm_v8i64_ty, llvm_v2i64_ty], [IntrNoMem]>; + // Oddly these don't require an immediate due to a gcc compatibility issue. def int_x86_avx512_pslli_w_512 : GCCBuiltin<"__builtin_ia32_psllwi512">, Intrinsic<[llvm_v32i16_ty], [llvm_v32i16_ty, llvm_i32_ty], [IntrNoMem]>; @@ -1752,7 +1758,7 @@ llvm_v32i8_ty], [IntrNoMem]>; def int_x86_avx2_mpsadbw : GCCBuiltin<"__builtin_ia32_mpsadbw256">, Intrinsic<[llvm_v16i16_ty], [llvm_v32i8_ty, llvm_v32i8_ty, - llvm_i8_ty], [IntrNoMem, Commutative]>; + llvm_i8_ty], [IntrNoMem, Commutative, ImmArg<2>]>; } //===----------------------------------------------------------------------===// @@ -1762,32 +1768,32 @@ def int_x86_avx512_vfmadd_pd_512 : Intrinsic<[llvm_v8f64_ty], [llvm_v8f64_ty, llvm_v8f64_ty, llvm_v8f64_ty, llvm_i32_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<3>]>; def int_x86_avx512_vfmadd_ps_512 : Intrinsic<[llvm_v16f32_ty], [llvm_v16f32_ty, llvm_v16f32_ty, llvm_v16f32_ty, llvm_i32_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<3>]>; // TODO: Can we use 2 vfmadds+shufflevector? def int_x86_avx512_vfmaddsub_pd_512 : Intrinsic<[llvm_v8f64_ty], [llvm_v8f64_ty, llvm_v8f64_ty, llvm_v8f64_ty, llvm_i32_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<3>]>; def int_x86_avx512_vfmaddsub_ps_512 : Intrinsic<[llvm_v16f32_ty], [llvm_v16f32_ty, llvm_v16f32_ty, llvm_v16f32_ty, llvm_i32_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<3>]>; def int_x86_avx512_vfmadd_f64 : Intrinsic<[llvm_double_ty], [llvm_double_ty, llvm_double_ty, llvm_double_ty, llvm_i32_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<3>]>; def int_x86_avx512_vfmadd_f32 : Intrinsic<[llvm_float_ty], [llvm_float_ty, llvm_float_ty, llvm_float_ty, llvm_i32_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<3>]>; def int_x86_avx512_vpmadd52h_uq_128 : GCCBuiltin<"__builtin_ia32_vpmadd52huq128">, @@ -1877,23 +1883,23 @@ def int_x86_xop_vpermil2pd : GCCBuiltin<"__builtin_ia32_vpermil2pd">, Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty, llvm_v2f64_ty, llvm_v2i64_ty, llvm_i8_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<3>]>; def int_x86_xop_vpermil2pd_256 : GCCBuiltin<"__builtin_ia32_vpermil2pd256">, Intrinsic<[llvm_v4f64_ty], [llvm_v4f64_ty, llvm_v4f64_ty, llvm_v4i64_ty, llvm_i8_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<3>]>; def int_x86_xop_vpermil2ps : GCCBuiltin<"__builtin_ia32_vpermil2ps">, Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty, llvm_v4f32_ty, llvm_v4i32_ty, llvm_i8_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<3>]>; def int_x86_xop_vpermil2ps_256 : GCCBuiltin<"__builtin_ia32_vpermil2ps256">, Intrinsic<[llvm_v8f32_ty], [llvm_v8f32_ty, llvm_v8f32_ty, llvm_v8i32_ty, llvm_i8_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<3>]>; def int_x86_xop_vfrcz_pd : GCCBuiltin<"__builtin_ia32_vfrczpd">, Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty], [IntrNoMem]>; @@ -2235,6 +2241,7 @@ Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty, llvm_x86mmx_ty], [IntrNoMem]>; + // Oddly these don't require an immediate due to a gcc compatibility issue. def int_x86_mmx_pslli_w : GCCBuiltin<"__builtin_ia32_psllwi">, Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty, llvm_i32_ty], [IntrNoMem]>; @@ -2372,7 +2379,7 @@ def int_x86_mmx_palignr_b : GCCBuiltin<"__builtin_ia32_palignr">, Intrinsic<[llvm_x86mmx_ty], [llvm_x86mmx_ty, - llvm_x86mmx_ty, llvm_i8_ty], [IntrNoMem]>; + llvm_x86mmx_ty, llvm_i8_ty], [IntrNoMem, ImmArg<2>]>; def int_x86_mmx_pextr_w : GCCBuiltin<"__builtin_ia32_vec_ext_v4hi">, Intrinsic<[llvm_i32_ty], [llvm_x86mmx_ty, llvm_i32_ty], @@ -2501,13 +2508,14 @@ Intrinsic<[llvm_v8f32_ty], [llvm_v8i16_ty], [IntrNoMem]>; def int_x86_vcvtps2ph_128 : GCCBuiltin<"__builtin_ia32_vcvtps2ph">, Intrinsic<[llvm_v8i16_ty], [llvm_v4f32_ty, llvm_i32_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<1>]>; def int_x86_vcvtps2ph_256 : GCCBuiltin<"__builtin_ia32_vcvtps2ph256">, Intrinsic<[llvm_v8i16_ty], [llvm_v8f32_ty, llvm_i32_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<1>]>; def int_x86_avx512_mask_vcvtph2ps_512 : GCCBuiltin<"__builtin_ia32_vcvtph2ps512_mask">, Intrinsic<[llvm_v16f32_ty], [llvm_v16i16_ty, llvm_v16f32_ty, - llvm_i16_ty, llvm_i32_ty], [IntrNoMem]>; + llvm_i16_ty, llvm_i32_ty], + [IntrNoMem, ImmArg<3>]>; def int_x86_avx512_mask_vcvtph2ps_256 : GCCBuiltin<"__builtin_ia32_vcvtph2ps256_mask">, Intrinsic<[llvm_v8f32_ty], [llvm_v8i16_ty, llvm_v8f32_ty, llvm_i8_ty], [IntrNoMem]>; @@ -2516,13 +2524,16 @@ llvm_i8_ty], [IntrNoMem]>; def int_x86_avx512_mask_vcvtps2ph_512 : GCCBuiltin<"__builtin_ia32_vcvtps2ph512_mask">, Intrinsic<[llvm_v16i16_ty], [llvm_v16f32_ty, llvm_i32_ty, - llvm_v16i16_ty, llvm_i16_ty], [IntrNoMem]>; + llvm_v16i16_ty, llvm_i16_ty], + [IntrNoMem, ImmArg<1>]>; def int_x86_avx512_mask_vcvtps2ph_256 : GCCBuiltin<"__builtin_ia32_vcvtps2ph256_mask">, Intrinsic<[llvm_v8i16_ty], [llvm_v8f32_ty, llvm_i32_ty, - llvm_v8i16_ty, llvm_i8_ty], [IntrNoMem]>; + llvm_v8i16_ty, llvm_i8_ty], + [IntrNoMem, ImmArg<1>]>; def int_x86_avx512_mask_vcvtps2ph_128 : GCCBuiltin<"__builtin_ia32_vcvtps2ph_mask">, Intrinsic<[llvm_v8i16_ty], [llvm_v4f32_ty, llvm_i32_ty, - llvm_v8i16_ty, llvm_i8_ty], [IntrNoMem]>; + llvm_v8i16_ty, llvm_i8_ty], + [IntrNoMem, ImmArg<1>]>; } //===----------------------------------------------------------------------===// @@ -2530,9 +2541,11 @@ let TargetPrefix = "x86" in { // All intrinsics start with "llvm.x86.". def int_x86_tbm_bextri_u32 : GCCBuiltin<"__builtin_ia32_bextri_u32">, - Intrinsic<[llvm_i32_ty], [llvm_i32_ty, llvm_i32_ty], [IntrNoMem]>; + Intrinsic<[llvm_i32_ty], [llvm_i32_ty, llvm_i32_ty], + [IntrNoMem, ImmArg<1>]>; def int_x86_tbm_bextri_u64 : GCCBuiltin<"__builtin_ia32_bextri_u64">, - Intrinsic<[llvm_i64_ty], [llvm_i64_ty, llvm_i64_ty], [IntrNoMem]>; + Intrinsic<[llvm_i64_ty], [llvm_i64_ty, llvm_i64_ty], + [IntrNoMem, ImmArg<1>]>; } //===----------------------------------------------------------------------===// @@ -2578,7 +2591,7 @@ def int_x86_xend : GCCBuiltin<"__builtin_ia32_xend">, Intrinsic<[], [], []>; def int_x86_xabort : GCCBuiltin<"__builtin_ia32_xabort">, - Intrinsic<[], [llvm_i8_ty], []>; + Intrinsic<[], [llvm_i8_ty], [ImmArg<0>]>; def int_x86_xtest : GCCBuiltin<"__builtin_ia32_xtest">, Intrinsic<[llvm_i32_ty], [], []>; } @@ -2619,55 +2632,71 @@ // Conversion ops let TargetPrefix = "x86" in { // All intrinsics start with "llvm.x86.". def int_x86_avx512_cvttss2si : GCCBuiltin<"__builtin_ia32_vcvttss2si32">, - Intrinsic<[llvm_i32_ty], [llvm_v4f32_ty, llvm_i32_ty], [IntrNoMem]>; + Intrinsic<[llvm_i32_ty], [llvm_v4f32_ty, llvm_i32_ty], + [IntrNoMem, ImmArg<1>]>; def int_x86_avx512_cvttss2si64 : GCCBuiltin<"__builtin_ia32_vcvttss2si64">, - Intrinsic<[llvm_i64_ty], [llvm_v4f32_ty, llvm_i32_ty], [IntrNoMem]>; + Intrinsic<[llvm_i64_ty], [llvm_v4f32_ty, llvm_i32_ty], + [IntrNoMem, ImmArg<1>]>; def int_x86_avx512_cvttss2usi : GCCBuiltin<"__builtin_ia32_vcvttss2usi32">, - Intrinsic<[llvm_i32_ty], [llvm_v4f32_ty, llvm_i32_ty], [IntrNoMem]>; + Intrinsic<[llvm_i32_ty], [llvm_v4f32_ty, llvm_i32_ty], + [IntrNoMem, ImmArg<1>]>; def int_x86_avx512_cvttss2usi64 : GCCBuiltin<"__builtin_ia32_vcvttss2usi64">, - Intrinsic<[llvm_i64_ty], [llvm_v4f32_ty, llvm_i32_ty], [IntrNoMem]>; + Intrinsic<[llvm_i64_ty], [llvm_v4f32_ty, llvm_i32_ty], + [IntrNoMem, ImmArg<1>]>; def int_x86_avx512_cvtusi2ss : GCCBuiltin<"__builtin_ia32_cvtusi2ss32">, Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty, - llvm_i32_ty, llvm_i32_ty], [IntrNoMem]>; + llvm_i32_ty, llvm_i32_ty], [IntrNoMem, ImmArg<2>]>; def int_x86_avx512_cvtusi642ss : GCCBuiltin<"__builtin_ia32_cvtusi2ss64">, Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty, - llvm_i64_ty, llvm_i32_ty], [IntrNoMem]>; + llvm_i64_ty, llvm_i32_ty], [IntrNoMem, ImmArg<2>]>; def int_x86_avx512_cvttsd2si : GCCBuiltin<"__builtin_ia32_vcvttsd2si32">, - Intrinsic<[llvm_i32_ty], [llvm_v2f64_ty, llvm_i32_ty], [IntrNoMem]>; + Intrinsic<[llvm_i32_ty], [llvm_v2f64_ty, llvm_i32_ty], + [IntrNoMem, ImmArg<1>]>; def int_x86_avx512_cvttsd2si64 : GCCBuiltin<"__builtin_ia32_vcvttsd2si64">, - Intrinsic<[llvm_i64_ty], [llvm_v2f64_ty, llvm_i32_ty], [IntrNoMem]>; + Intrinsic<[llvm_i64_ty], [llvm_v2f64_ty, llvm_i32_ty], + [IntrNoMem, ImmArg<1>]>; def int_x86_avx512_cvttsd2usi : GCCBuiltin<"__builtin_ia32_vcvttsd2usi32">, - Intrinsic<[llvm_i32_ty], [llvm_v2f64_ty, llvm_i32_ty], [IntrNoMem]>; + Intrinsic<[llvm_i32_ty], [llvm_v2f64_ty, llvm_i32_ty], + [IntrNoMem, ImmArg<1>]>; def int_x86_avx512_cvttsd2usi64 : GCCBuiltin<"__builtin_ia32_vcvttsd2usi64">, - Intrinsic<[llvm_i64_ty], [llvm_v2f64_ty, llvm_i32_ty], [IntrNoMem]>; + Intrinsic<[llvm_i64_ty], [llvm_v2f64_ty, llvm_i32_ty], + [IntrNoMem, ImmArg<1>]>; def int_x86_avx512_cvtusi642sd : GCCBuiltin<"__builtin_ia32_cvtusi2sd64">, Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty, - llvm_i64_ty, llvm_i32_ty], [IntrNoMem]>; + llvm_i64_ty, llvm_i32_ty], [IntrNoMem, ImmArg<2>]>; def int_x86_avx512_vcvtss2usi32 : GCCBuiltin<"__builtin_ia32_vcvtss2usi32">, - Intrinsic<[llvm_i32_ty], [llvm_v4f32_ty, llvm_i32_ty], [IntrNoMem]>; + Intrinsic<[llvm_i32_ty], [llvm_v4f32_ty, llvm_i32_ty], + [IntrNoMem, ImmArg<1>]>; def int_x86_avx512_vcvtss2usi64 : GCCBuiltin<"__builtin_ia32_vcvtss2usi64">, - Intrinsic<[llvm_i64_ty], [llvm_v4f32_ty, llvm_i32_ty], [IntrNoMem]>; + Intrinsic<[llvm_i64_ty], [llvm_v4f32_ty, llvm_i32_ty], + [IntrNoMem, ImmArg<1>]>; def int_x86_avx512_vcvtss2si32 : GCCBuiltin<"__builtin_ia32_vcvtss2si32">, - Intrinsic<[llvm_i32_ty], [llvm_v4f32_ty, llvm_i32_ty], [IntrNoMem]>; + Intrinsic<[llvm_i32_ty], [llvm_v4f32_ty, llvm_i32_ty], + [IntrNoMem, ImmArg<1>]>; def int_x86_avx512_vcvtss2si64 : GCCBuiltin<"__builtin_ia32_vcvtss2si64">, - Intrinsic<[llvm_i64_ty], [llvm_v4f32_ty, llvm_i32_ty], [IntrNoMem]>; + Intrinsic<[llvm_i64_ty], [llvm_v4f32_ty, llvm_i32_ty], + [IntrNoMem, ImmArg<1>]>; def int_x86_avx512_vcvtsd2usi32 : GCCBuiltin<"__builtin_ia32_vcvtsd2usi32">, - Intrinsic<[llvm_i32_ty], [llvm_v2f64_ty, llvm_i32_ty], [IntrNoMem]>; + Intrinsic<[llvm_i32_ty], [llvm_v2f64_ty, llvm_i32_ty], + [IntrNoMem, ImmArg<1>]>; def int_x86_avx512_vcvtsd2usi64 : GCCBuiltin<"__builtin_ia32_vcvtsd2usi64">, - Intrinsic<[llvm_i64_ty], [llvm_v2f64_ty, llvm_i32_ty], [IntrNoMem]>; + Intrinsic<[llvm_i64_ty], [llvm_v2f64_ty, llvm_i32_ty], + [IntrNoMem, ImmArg<1>]>; def int_x86_avx512_vcvtsd2si32 : GCCBuiltin<"__builtin_ia32_vcvtsd2si32">, - Intrinsic<[llvm_i32_ty], [llvm_v2f64_ty, llvm_i32_ty], [IntrNoMem]>; + Intrinsic<[llvm_i32_ty], [llvm_v2f64_ty, llvm_i32_ty], + [IntrNoMem, ImmArg<1>]>; def int_x86_avx512_vcvtsd2si64 : GCCBuiltin<"__builtin_ia32_vcvtsd2si64">, - Intrinsic<[llvm_i64_ty], [llvm_v2f64_ty, llvm_i32_ty], [IntrNoMem]>; + Intrinsic<[llvm_i64_ty], [llvm_v2f64_ty, llvm_i32_ty], + [IntrNoMem, ImmArg<1>]>; def int_x86_avx512_cvtsi2ss32 : GCCBuiltin<"__builtin_ia32_cvtsi2ss32">, Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty, - llvm_i32_ty, llvm_i32_ty], [IntrNoMem]>; + llvm_i32_ty, llvm_i32_ty], [IntrNoMem, ImmArg<2>]>; def int_x86_avx512_cvtsi2ss64 : GCCBuiltin<"__builtin_ia32_cvtsi2ss64">, Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty, - llvm_i64_ty, llvm_i32_ty], [IntrNoMem]>; + llvm_i64_ty, llvm_i32_ty], [IntrNoMem, ImmArg<2>]>; def int_x86_avx512_cvtsi2sd64 : GCCBuiltin<"__builtin_ia32_cvtsi2sd64">, Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty, - llvm_i64_ty, llvm_i32_ty], [IntrNoMem]>; + llvm_i64_ty, llvm_i32_ty], [IntrNoMem, ImmArg<2>]>; } // Pack ops. @@ -2690,11 +2719,11 @@ let TargetPrefix = "x86" in { // All intrinsics start with "llvm.x86.". def int_x86_avx512_sitofp_round : Intrinsic<[llvm_anyfloat_ty], [llvm_anyint_ty, llvm_i32_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<1>]>; def int_x86_avx512_uitofp_round : Intrinsic<[llvm_anyfloat_ty], [llvm_anyint_ty, llvm_i32_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<1>]>; def int_x86_avx512_mask_cvtpd2dq_128 : GCCBuiltin<"__builtin_ia32_cvtpd2dq128_mask">, @@ -2706,25 +2735,25 @@ GCCBuiltin<"__builtin_ia32_cvtpd2dq512_mask">, Intrinsic<[llvm_v8i32_ty], [llvm_v8f64_ty, llvm_v8i32_ty, llvm_i8_ty, llvm_i32_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<3>]>; def int_x86_avx512_mask_cvtpd2ps_512 : GCCBuiltin<"__builtin_ia32_cvtpd2ps512_mask">, Intrinsic<[llvm_v8f32_ty], [llvm_v8f64_ty, llvm_v8f32_ty, llvm_i8_ty, llvm_i32_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<3>]>; def int_x86_avx512_mask_cvtsd2ss_round : GCCBuiltin<"__builtin_ia32_cvtsd2ss_round_mask">, Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty, llvm_v2f64_ty, llvm_v4f32_ty, llvm_i8_ty, llvm_i32_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<4>]>; def int_x86_avx512_mask_cvtss2sd_round : GCCBuiltin<"__builtin_ia32_cvtss2sd_round_mask">, Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty, llvm_v4f32_ty, llvm_v2f64_ty, llvm_i8_ty, llvm_i32_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<4>]>; def int_x86_avx512_mask_cvtpd2ps : GCCBuiltin<"__builtin_ia32_cvtpd2ps_mask">, @@ -2748,7 +2777,7 @@ GCCBuiltin<"__builtin_ia32_cvtpd2qq512_mask">, Intrinsic<[llvm_v8i64_ty], [llvm_v8f64_ty, llvm_v8i64_ty, llvm_i8_ty, llvm_i32_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<3>]>; def int_x86_avx512_mask_cvtpd2udq_128 : GCCBuiltin<"__builtin_ia32_cvtpd2udq128_mask">, @@ -2766,7 +2795,7 @@ GCCBuiltin<"__builtin_ia32_cvtpd2udq512_mask">, Intrinsic<[llvm_v8i32_ty], [llvm_v8f64_ty, llvm_v8i32_ty, llvm_i8_ty, llvm_i32_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<3>]>; def int_x86_avx512_mask_cvtpd2uqq_128 : GCCBuiltin<"__builtin_ia32_cvtpd2uqq128_mask">, @@ -2784,7 +2813,7 @@ GCCBuiltin<"__builtin_ia32_cvtpd2uqq512_mask">, Intrinsic<[llvm_v8i64_ty], [llvm_v8f64_ty, llvm_v8i64_ty, llvm_i8_ty, llvm_i32_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<3>]>; def int_x86_avx512_mask_cvtps2dq_128 : GCCBuiltin<"__builtin_ia32_cvtps2dq128_mask">, @@ -2802,13 +2831,13 @@ GCCBuiltin<"__builtin_ia32_cvtps2dq512_mask">, Intrinsic<[llvm_v16i32_ty], [llvm_v16f32_ty, llvm_v16i32_ty, llvm_i16_ty, llvm_i32_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<3>]>; def int_x86_avx512_mask_cvtps2pd_512 : GCCBuiltin<"__builtin_ia32_cvtps2pd512_mask">, Intrinsic<[llvm_v8f64_ty], [llvm_v8f32_ty, llvm_v8f64_ty, llvm_i8_ty, llvm_i32_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<3>]>; def int_x86_avx512_mask_cvtps2qq_128 : GCCBuiltin<"__builtin_ia32_cvtps2qq128_mask">, @@ -2826,7 +2855,7 @@ GCCBuiltin<"__builtin_ia32_cvtps2qq512_mask">, Intrinsic<[llvm_v8i64_ty], [llvm_v8f32_ty, llvm_v8i64_ty, llvm_i8_ty, llvm_i32_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<3>]>; def int_x86_avx512_mask_cvtps2udq_128 : GCCBuiltin<"__builtin_ia32_cvtps2udq128_mask">, @@ -2844,7 +2873,7 @@ GCCBuiltin<"__builtin_ia32_cvtps2udq512_mask">, Intrinsic<[llvm_v16i32_ty], [llvm_v16f32_ty, llvm_v16i32_ty, llvm_i16_ty, llvm_i32_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<3>]>; def int_x86_avx512_mask_cvtps2uqq_128 : GCCBuiltin<"__builtin_ia32_cvtps2uqq128_mask">, @@ -2862,7 +2891,7 @@ GCCBuiltin<"__builtin_ia32_cvtps2uqq512_mask">, Intrinsic<[llvm_v8i64_ty], [llvm_v8f32_ty, llvm_v8i64_ty, llvm_i8_ty, llvm_i32_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<3>]>; def int_x86_avx512_mask_cvtqq2ps_128 : GCCBuiltin<"__builtin_ia32_cvtqq2ps128_mask">, @@ -2880,7 +2909,7 @@ GCCBuiltin<"__builtin_ia32_cvttpd2dq512_mask">, Intrinsic<[llvm_v8i32_ty], [llvm_v8f64_ty, llvm_v8i32_ty, llvm_i8_ty, llvm_i32_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<3>]>; def int_x86_avx512_mask_cvttpd2qq_128 : GCCBuiltin<"__builtin_ia32_cvttpd2qq128_mask">, @@ -2898,7 +2927,7 @@ GCCBuiltin<"__builtin_ia32_cvttpd2qq512_mask">, Intrinsic<[llvm_v8i64_ty], [llvm_v8f64_ty, llvm_v8i64_ty, llvm_i8_ty, llvm_i32_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<3>]>; def int_x86_avx512_mask_cvttpd2udq_128 : GCCBuiltin<"__builtin_ia32_cvttpd2udq128_mask">, @@ -2916,7 +2945,7 @@ GCCBuiltin<"__builtin_ia32_cvttpd2udq512_mask">, Intrinsic<[llvm_v8i32_ty], [llvm_v8f64_ty, llvm_v8i32_ty, llvm_i8_ty, llvm_i32_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<3>]>; def int_x86_avx512_mask_cvttpd2uqq_128 : GCCBuiltin<"__builtin_ia32_cvttpd2uqq128_mask">, @@ -2934,13 +2963,13 @@ GCCBuiltin<"__builtin_ia32_cvttpd2uqq512_mask">, Intrinsic<[llvm_v8i64_ty], [llvm_v8f64_ty, llvm_v8i64_ty, llvm_i8_ty, llvm_i32_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<3>]>; def int_x86_avx512_mask_cvttps2dq_512 : GCCBuiltin<"__builtin_ia32_cvttps2dq512_mask">, Intrinsic<[llvm_v16i32_ty], [llvm_v16f32_ty, llvm_v16i32_ty, llvm_i16_ty, llvm_i32_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<3>]>; def int_x86_avx512_mask_cvttps2qq_128 : GCCBuiltin<"__builtin_ia32_cvttps2qq128_mask">, @@ -2958,7 +2987,7 @@ GCCBuiltin<"__builtin_ia32_cvttps2qq512_mask">, Intrinsic<[llvm_v8i64_ty], [llvm_v8f32_ty, llvm_v8i64_ty, llvm_i8_ty, llvm_i32_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<3>]>; def int_x86_avx512_mask_cvttps2udq_128 : GCCBuiltin<"__builtin_ia32_cvttps2udq128_mask">, @@ -2976,7 +3005,7 @@ GCCBuiltin<"__builtin_ia32_cvttps2udq512_mask">, Intrinsic<[llvm_v16i32_ty], [llvm_v16f32_ty, llvm_v16i32_ty, llvm_i16_ty, llvm_i32_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<3>]>; def int_x86_avx512_mask_cvttps2uqq_128 : GCCBuiltin<"__builtin_ia32_cvttps2uqq128_mask">, @@ -2994,7 +3023,7 @@ GCCBuiltin<"__builtin_ia32_cvttps2uqq512_mask">, Intrinsic<[llvm_v8i64_ty], [llvm_v8f32_ty, llvm_v8i64_ty, llvm_i8_ty, llvm_i32_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<3>]>; def int_x86_avx512_mask_cvtuqq2ps_128 : GCCBuiltin<"__builtin_ia32_cvtuqq2ps128_mask">, @@ -3004,58 +3033,76 @@ def int_x86_avx512_mask_rndscale_pd_128 : GCCBuiltin<"__builtin_ia32_rndscalepd_128_mask">, Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty, llvm_i32_ty, - llvm_v2f64_ty, llvm_i8_ty], [IntrNoMem]>; + llvm_v2f64_ty, llvm_i8_ty], + [IntrNoMem, ImmArg<1>]>; def int_x86_avx512_mask_rndscale_pd_256 : GCCBuiltin<"__builtin_ia32_rndscalepd_256_mask">, Intrinsic<[llvm_v4f64_ty], [llvm_v4f64_ty, llvm_i32_ty, - llvm_v4f64_ty, llvm_i8_ty], [IntrNoMem]>; + llvm_v4f64_ty, llvm_i8_ty], + [IntrNoMem, ImmArg<1>]>; def int_x86_avx512_mask_rndscale_pd_512 : GCCBuiltin<"__builtin_ia32_rndscalepd_mask">, Intrinsic<[llvm_v8f64_ty], [llvm_v8f64_ty, llvm_i32_ty, llvm_v8f64_ty, - llvm_i8_ty, llvm_i32_ty], [IntrNoMem]>; + llvm_i8_ty, llvm_i32_ty], + [IntrNoMem, ImmArg<1>, ImmArg<4>]>; def int_x86_avx512_mask_rndscale_ps_128 : GCCBuiltin<"__builtin_ia32_rndscaleps_128_mask">, Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty, llvm_i32_ty, - llvm_v4f32_ty, llvm_i8_ty], [IntrNoMem]>; + llvm_v4f32_ty, llvm_i8_ty], + [IntrNoMem, ImmArg<1>]>; def int_x86_avx512_mask_rndscale_ps_256 : GCCBuiltin<"__builtin_ia32_rndscaleps_256_mask">, Intrinsic<[llvm_v8f32_ty], [llvm_v8f32_ty, llvm_i32_ty, - llvm_v8f32_ty, llvm_i8_ty], [IntrNoMem]>; + llvm_v8f32_ty, llvm_i8_ty], + [IntrNoMem, ImmArg<1>]>; def int_x86_avx512_mask_rndscale_ps_512 : GCCBuiltin<"__builtin_ia32_rndscaleps_mask">, Intrinsic<[llvm_v16f32_ty], [llvm_v16f32_ty, llvm_i32_ty, llvm_v16f32_ty, - llvm_i16_ty, llvm_i32_ty], [IntrNoMem]>; + llvm_i16_ty, llvm_i32_ty], + [IntrNoMem, ImmArg<1>, ImmArg<4>]>; def int_x86_avx512_mask_reduce_pd_128 : GCCBuiltin<"__builtin_ia32_reducepd128_mask">, Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty, llvm_i32_ty, - llvm_v2f64_ty, llvm_i8_ty], [IntrNoMem]>; + llvm_v2f64_ty, llvm_i8_ty], + [IntrNoMem, ImmArg<1>]>; def int_x86_avx512_mask_reduce_pd_256 : GCCBuiltin<"__builtin_ia32_reducepd256_mask">, Intrinsic<[llvm_v4f64_ty], [llvm_v4f64_ty, llvm_i32_ty, - llvm_v4f64_ty, llvm_i8_ty], [IntrNoMem]>; + llvm_v4f64_ty, llvm_i8_ty], + [IntrNoMem, ImmArg<1>]>; def int_x86_avx512_mask_reduce_pd_512 : GCCBuiltin<"__builtin_ia32_reducepd512_mask">, Intrinsic<[llvm_v8f64_ty], [llvm_v8f64_ty, llvm_i32_ty, llvm_v8f64_ty, - llvm_i8_ty, llvm_i32_ty], [IntrNoMem]>; + llvm_i8_ty, llvm_i32_ty], + [IntrNoMem, ImmArg<1>, ImmArg<4>]>; def int_x86_avx512_mask_reduce_ps_128 : GCCBuiltin<"__builtin_ia32_reduceps128_mask">, Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty, llvm_i32_ty, - llvm_v4f32_ty, llvm_i8_ty], [IntrNoMem]>; + llvm_v4f32_ty, llvm_i8_ty], + [IntrNoMem, ImmArg<1>]>; def int_x86_avx512_mask_reduce_ps_256 : GCCBuiltin<"__builtin_ia32_reduceps256_mask">, Intrinsic<[llvm_v8f32_ty], [llvm_v8f32_ty, llvm_i32_ty, - llvm_v8f32_ty, llvm_i8_ty], [IntrNoMem]>; + llvm_v8f32_ty, llvm_i8_ty], + [IntrNoMem, ImmArg<1>]>; def int_x86_avx512_mask_reduce_ps_512 : GCCBuiltin<"__builtin_ia32_reduceps512_mask">, Intrinsic<[llvm_v16f32_ty], [llvm_v16f32_ty, llvm_i32_ty, llvm_v16f32_ty, - llvm_i16_ty, llvm_i32_ty], [IntrNoMem]>; + llvm_i16_ty, llvm_i32_ty], + [IntrNoMem, ImmArg<1>, ImmArg<4>]>; def int_x86_avx512_mask_range_pd_128 : GCCBuiltin<"__builtin_ia32_rangepd128_mask">, Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty, llvm_v2f64_ty, llvm_i32_ty, - llvm_v2f64_ty, llvm_i8_ty], [IntrNoMem]>; + llvm_v2f64_ty, llvm_i8_ty], + [IntrNoMem, ImmArg<2>]>; def int_x86_avx512_mask_range_pd_256 : GCCBuiltin<"__builtin_ia32_rangepd256_mask">, Intrinsic<[llvm_v4f64_ty], [llvm_v4f64_ty, llvm_v4f64_ty, llvm_i32_ty, - llvm_v4f64_ty, llvm_i8_ty], [IntrNoMem]>; + llvm_v4f64_ty, llvm_i8_ty], + [IntrNoMem, ImmArg<2>]>; def int_x86_avx512_mask_range_pd_512 : GCCBuiltin<"__builtin_ia32_rangepd512_mask">, Intrinsic<[llvm_v8f64_ty], [llvm_v8f64_ty, llvm_v8f64_ty, llvm_i32_ty, - llvm_v8f64_ty, llvm_i8_ty, llvm_i32_ty], [IntrNoMem]>; + llvm_v8f64_ty, llvm_i8_ty, llvm_i32_ty], + [IntrNoMem, ImmArg<2>, ImmArg<5>]>; def int_x86_avx512_mask_range_ps_128 : GCCBuiltin<"__builtin_ia32_rangeps128_mask">, Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty, llvm_v4f32_ty, llvm_i32_ty, - llvm_v4f32_ty, llvm_i8_ty], [IntrNoMem]>; + llvm_v4f32_ty, llvm_i8_ty], + [IntrNoMem, ImmArg<2>]>; def int_x86_avx512_mask_range_ps_256 : GCCBuiltin<"__builtin_ia32_rangeps256_mask">, Intrinsic<[llvm_v8f32_ty], [llvm_v8f32_ty, llvm_v8f32_ty, llvm_i32_ty, - llvm_v8f32_ty, llvm_i8_ty], [IntrNoMem]>; + llvm_v8f32_ty, llvm_i8_ty], + [IntrNoMem, ImmArg<2>]>; def int_x86_avx512_mask_range_ps_512 : GCCBuiltin<"__builtin_ia32_rangeps512_mask">, Intrinsic<[llvm_v16f32_ty], [llvm_v16f32_ty, llvm_v16f32_ty, llvm_i32_ty, - llvm_v16f32_ty, llvm_i16_ty, llvm_i32_ty], [IntrNoMem]>; + llvm_v16f32_ty, llvm_i16_ty, llvm_i32_ty], + [IntrNoMem, ImmArg<2>, ImmArg<5>]>; } // Vector load with broadcast @@ -3085,109 +3132,111 @@ def int_x86_avx512_add_ps_512 : GCCBuiltin<"__builtin_ia32_addps512">, Intrinsic<[llvm_v16f32_ty], [llvm_v16f32_ty, llvm_v16f32_ty, - llvm_i32_ty], [IntrNoMem]>; + llvm_i32_ty], [IntrNoMem, ImmArg<2>]>; def int_x86_avx512_add_pd_512 : GCCBuiltin<"__builtin_ia32_addpd512">, Intrinsic<[llvm_v8f64_ty], [llvm_v8f64_ty, llvm_v8f64_ty, - llvm_i32_ty], [IntrNoMem]>; + llvm_i32_ty], [IntrNoMem, ImmArg<2>]>; def int_x86_avx512_sub_ps_512 : GCCBuiltin<"__builtin_ia32_subps512">, Intrinsic<[llvm_v16f32_ty], [llvm_v16f32_ty, llvm_v16f32_ty, - llvm_i32_ty], [IntrNoMem]>; + llvm_i32_ty], [IntrNoMem, ImmArg<2>]>; def int_x86_avx512_sub_pd_512 : GCCBuiltin<"__builtin_ia32_subpd512">, Intrinsic<[llvm_v8f64_ty], [llvm_v8f64_ty, llvm_v8f64_ty, - llvm_i32_ty], [IntrNoMem]>; + llvm_i32_ty], [IntrNoMem, ImmArg<2>]>; def int_x86_avx512_mul_ps_512 : GCCBuiltin<"__builtin_ia32_mulps512">, Intrinsic<[llvm_v16f32_ty], [llvm_v16f32_ty, llvm_v16f32_ty, - llvm_i32_ty], [IntrNoMem]>; + llvm_i32_ty], [IntrNoMem, ImmArg<2>]>; def int_x86_avx512_mul_pd_512 : GCCBuiltin<"__builtin_ia32_mulpd512">, Intrinsic<[llvm_v8f64_ty], [llvm_v8f64_ty, llvm_v8f64_ty, - llvm_i32_ty], [IntrNoMem]>; + llvm_i32_ty], [IntrNoMem, ImmArg<2>]>; def int_x86_avx512_div_ps_512 : GCCBuiltin<"__builtin_ia32_divps512">, Intrinsic<[llvm_v16f32_ty], [llvm_v16f32_ty, llvm_v16f32_ty, - llvm_i32_ty], [IntrNoMem]>; + llvm_i32_ty], [IntrNoMem, ImmArg<2>]>; def int_x86_avx512_div_pd_512 : GCCBuiltin<"__builtin_ia32_divpd512">, Intrinsic<[llvm_v8f64_ty], [llvm_v8f64_ty, llvm_v8f64_ty, - llvm_i32_ty], [IntrNoMem]>; + llvm_i32_ty], [IntrNoMem, ImmArg<2>]>; def int_x86_avx512_max_ps_512 : GCCBuiltin<"__builtin_ia32_maxps512">, Intrinsic<[llvm_v16f32_ty], [llvm_v16f32_ty, llvm_v16f32_ty, - llvm_i32_ty], [IntrNoMem]>; + llvm_i32_ty], [IntrNoMem, ImmArg<2>]>; def int_x86_avx512_max_pd_512 : GCCBuiltin<"__builtin_ia32_maxpd512">, Intrinsic<[llvm_v8f64_ty], [llvm_v8f64_ty, llvm_v8f64_ty, - llvm_i32_ty], [IntrNoMem]>; + llvm_i32_ty], [IntrNoMem, ImmArg<2>]>; def int_x86_avx512_min_ps_512 : GCCBuiltin<"__builtin_ia32_minps512">, Intrinsic<[llvm_v16f32_ty], [llvm_v16f32_ty, llvm_v16f32_ty, - llvm_i32_ty], [IntrNoMem]>; + llvm_i32_ty], [IntrNoMem, ImmArg<2>]>; def int_x86_avx512_min_pd_512 : GCCBuiltin<"__builtin_ia32_minpd512">, Intrinsic<[llvm_v8f64_ty], [llvm_v8f64_ty, llvm_v8f64_ty, - llvm_i32_ty], [IntrNoMem]>; + llvm_i32_ty], [IntrNoMem, ImmArg<2>]>; def int_x86_avx512_mask_add_ss_round : GCCBuiltin<"__builtin_ia32_addss_round_mask">, Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty, llvm_v4f32_ty, - llvm_v4f32_ty, llvm_i8_ty, llvm_i32_ty], [IntrNoMem]>; + llvm_v4f32_ty, llvm_i8_ty, llvm_i32_ty], [IntrNoMem, ImmArg<4>]>; def int_x86_avx512_mask_div_ss_round : GCCBuiltin<"__builtin_ia32_divss_round_mask">, Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty, llvm_v4f32_ty, - llvm_v4f32_ty, llvm_i8_ty, llvm_i32_ty], [IntrNoMem]>; + llvm_v4f32_ty, llvm_i8_ty, llvm_i32_ty], [IntrNoMem, ImmArg<4>]>; def int_x86_avx512_mask_mul_ss_round : GCCBuiltin<"__builtin_ia32_mulss_round_mask">, Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty, llvm_v4f32_ty, - llvm_v4f32_ty, llvm_i8_ty, llvm_i32_ty], [IntrNoMem]>; + llvm_v4f32_ty, llvm_i8_ty, llvm_i32_ty], [IntrNoMem, ImmArg<4>]>; def int_x86_avx512_mask_sub_ss_round : GCCBuiltin<"__builtin_ia32_subss_round_mask">, Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty, llvm_v4f32_ty, - llvm_v4f32_ty, llvm_i8_ty, llvm_i32_ty], [IntrNoMem]>; + llvm_v4f32_ty, llvm_i8_ty, llvm_i32_ty], [IntrNoMem, ImmArg<4>]>; def int_x86_avx512_mask_max_ss_round : GCCBuiltin<"__builtin_ia32_maxss_round_mask">, Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty, llvm_v4f32_ty, - llvm_v4f32_ty, llvm_i8_ty, llvm_i32_ty], [IntrNoMem]>; + llvm_v4f32_ty, llvm_i8_ty, llvm_i32_ty], [IntrNoMem, ImmArg<4>]>; def int_x86_avx512_mask_min_ss_round : GCCBuiltin<"__builtin_ia32_minss_round_mask">, Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty, llvm_v4f32_ty, - llvm_v4f32_ty, llvm_i8_ty, llvm_i32_ty], [IntrNoMem]>; + llvm_v4f32_ty, llvm_i8_ty, llvm_i32_ty], [IntrNoMem, ImmArg<4>]>; def int_x86_avx512_mask_add_sd_round : GCCBuiltin<"__builtin_ia32_addsd_round_mask">, Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty, llvm_v2f64_ty, - llvm_v2f64_ty, llvm_i8_ty, llvm_i32_ty], [IntrNoMem]>; + llvm_v2f64_ty, llvm_i8_ty, llvm_i32_ty], [IntrNoMem, ImmArg<4>]>; def int_x86_avx512_mask_div_sd_round : GCCBuiltin<"__builtin_ia32_divsd_round_mask">, Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty, llvm_v2f64_ty, - llvm_v2f64_ty, llvm_i8_ty, llvm_i32_ty], [IntrNoMem]>; + llvm_v2f64_ty, llvm_i8_ty, llvm_i32_ty], [IntrNoMem, ImmArg<4>]>; def int_x86_avx512_mask_mul_sd_round : GCCBuiltin<"__builtin_ia32_mulsd_round_mask">, Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty, llvm_v2f64_ty, - llvm_v2f64_ty, llvm_i8_ty, llvm_i32_ty], [IntrNoMem]>; + llvm_v2f64_ty, llvm_i8_ty, llvm_i32_ty], [IntrNoMem, ImmArg<4>]>; def int_x86_avx512_mask_sub_sd_round : GCCBuiltin<"__builtin_ia32_subsd_round_mask">, Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty, llvm_v2f64_ty, - llvm_v2f64_ty, llvm_i8_ty, llvm_i32_ty], [IntrNoMem]>; + llvm_v2f64_ty, llvm_i8_ty, llvm_i32_ty], [IntrNoMem, ImmArg<4>]>; def int_x86_avx512_mask_max_sd_round : GCCBuiltin<"__builtin_ia32_maxsd_round_mask">, Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty, llvm_v2f64_ty, - llvm_v2f64_ty, llvm_i8_ty, llvm_i32_ty], [IntrNoMem]>; + llvm_v2f64_ty, llvm_i8_ty, llvm_i32_ty], [IntrNoMem, ImmArg<4>]>; def int_x86_avx512_mask_min_sd_round : GCCBuiltin<"__builtin_ia32_minsd_round_mask">, Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty, llvm_v2f64_ty, - llvm_v2f64_ty, llvm_i8_ty, llvm_i32_ty], [IntrNoMem]>; + llvm_v2f64_ty, llvm_i8_ty, llvm_i32_ty], [IntrNoMem, ImmArg<4>]>; def int_x86_avx512_mask_rndscale_ss : GCCBuiltin<"__builtin_ia32_rndscaless_round_mask">, Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty, llvm_v4f32_ty, llvm_v4f32_ty, llvm_i8_ty, llvm_i32_ty, llvm_i32_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<4>, ImmArg<5>]>; def int_x86_avx512_mask_rndscale_sd : GCCBuiltin<"__builtin_ia32_rndscalesd_round_mask">, Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty, llvm_v2f64_ty, llvm_v2f64_ty, llvm_i8_ty, llvm_i32_ty, llvm_i32_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<4>, ImmArg<5>]>; def int_x86_avx512_mask_range_ss : GCCBuiltin<"__builtin_ia32_rangess128_round_mask">, Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty, llvm_v4f32_ty, llvm_v4f32_ty, llvm_i8_ty, llvm_i32_ty, llvm_i32_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<4>, ImmArg<5>]>; def int_x86_avx512_mask_range_sd : GCCBuiltin<"__builtin_ia32_rangesd128_round_mask">, Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty, llvm_v2f64_ty, llvm_v2f64_ty, llvm_i8_ty, llvm_i32_ty, llvm_i32_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<4>, ImmArg<5>]>; def int_x86_avx512_mask_reduce_ss : GCCBuiltin<"__builtin_ia32_reducess_mask">, Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty, llvm_v4f32_ty, llvm_v4f32_ty, llvm_i8_ty, llvm_i32_ty, llvm_i32_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<4>, ImmArg<5>]>; def int_x86_avx512_mask_reduce_sd : GCCBuiltin<"__builtin_ia32_reducesd_mask">, Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty, llvm_v2f64_ty, llvm_v2f64_ty, llvm_i8_ty, llvm_i32_ty, llvm_i32_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<4>, ImmArg<5>]>; def int_x86_avx512_mask_scalef_sd : GCCBuiltin<"__builtin_ia32_scalefsd_round_mask">, Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty, llvm_v2f64_ty, - llvm_v2f64_ty, llvm_i8_ty, llvm_i32_ty], [IntrNoMem]>; + llvm_v2f64_ty, llvm_i8_ty, llvm_i32_ty], + [IntrNoMem, ImmArg<4>]>; def int_x86_avx512_mask_scalef_ss : GCCBuiltin<"__builtin_ia32_scalefss_round_mask">, Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty, llvm_v4f32_ty, - llvm_v4f32_ty, llvm_i8_ty, llvm_i32_ty], [IntrNoMem]>; + llvm_v4f32_ty, llvm_i8_ty, llvm_i32_ty], + [IntrNoMem, ImmArg<4>]>; def int_x86_avx512_mask_scalef_pd_128 : GCCBuiltin<"__builtin_ia32_scalefpd128_mask">, Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty, llvm_v2f64_ty, llvm_v2f64_ty, llvm_i8_ty], [IntrNoMem]>; @@ -3196,7 +3245,8 @@ llvm_v4f64_ty, llvm_i8_ty],[IntrNoMem]>; def int_x86_avx512_mask_scalef_pd_512 : GCCBuiltin<"__builtin_ia32_scalefpd512_mask">, Intrinsic<[llvm_v8f64_ty], [llvm_v8f64_ty, llvm_v8f64_ty, - llvm_v8f64_ty, llvm_i8_ty, llvm_i32_ty], [IntrNoMem]>; + llvm_v8f64_ty, llvm_i8_ty, llvm_i32_ty], + [IntrNoMem, ImmArg<4>]>; def int_x86_avx512_mask_scalef_ps_128 : GCCBuiltin<"__builtin_ia32_scalefps128_mask">, Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty, llvm_v4f32_ty, llvm_v4f32_ty, llvm_i8_ty], [IntrNoMem]>; @@ -3205,39 +3255,44 @@ llvm_v8f32_ty, llvm_i8_ty], [IntrNoMem]>; def int_x86_avx512_mask_scalef_ps_512 : GCCBuiltin<"__builtin_ia32_scalefps512_mask">, Intrinsic<[llvm_v16f32_ty], [llvm_v16f32_ty, llvm_v16f32_ty, - llvm_v16f32_ty, llvm_i16_ty, llvm_i32_ty], [IntrNoMem]>; + llvm_v16f32_ty, llvm_i16_ty, llvm_i32_ty], + [IntrNoMem, ImmArg<4>]>; def int_x86_avx512_mask_sqrt_ss : Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty, llvm_v4f32_ty, llvm_v4f32_ty, - llvm_i8_ty, llvm_i32_ty], [IntrNoMem]>; + llvm_i8_ty, llvm_i32_ty], + [IntrNoMem, ImmArg<4>]>; def int_x86_avx512_mask_sqrt_sd : Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty, llvm_v2f64_ty, llvm_v2f64_ty, - llvm_i8_ty, llvm_i32_ty], [IntrNoMem]>; + llvm_i8_ty, llvm_i32_ty], + [IntrNoMem, ImmArg<4>]>; def int_x86_avx512_sqrt_pd_512 : - Intrinsic<[llvm_v8f64_ty], [llvm_v8f64_ty, llvm_i32_ty], [IntrNoMem]>; + Intrinsic<[llvm_v8f64_ty], [llvm_v8f64_ty, llvm_i32_ty], + [IntrNoMem, ImmArg<1>]>; def int_x86_avx512_sqrt_ps_512 : - Intrinsic<[llvm_v16f32_ty], [llvm_v16f32_ty, llvm_i32_ty], [IntrNoMem]>; + Intrinsic<[llvm_v16f32_ty], [llvm_v16f32_ty, llvm_i32_ty], + [IntrNoMem, ImmArg<1>]>; def int_x86_avx512_mask_fixupimm_pd_128 : GCCBuiltin<"__builtin_ia32_fixupimmpd128_mask">, Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty, llvm_v2f64_ty, llvm_v2i64_ty, llvm_i32_ty, llvm_i8_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<3>]>; def int_x86_avx512_maskz_fixupimm_pd_128 : GCCBuiltin<"__builtin_ia32_fixupimmpd128_maskz">, Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty, llvm_v2f64_ty, llvm_v2i64_ty, llvm_i32_ty, llvm_i8_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<3>]>; def int_x86_avx512_mask_fixupimm_pd_256 : GCCBuiltin<"__builtin_ia32_fixupimmpd256_mask">, Intrinsic<[llvm_v4f64_ty], [llvm_v4f64_ty, llvm_v4f64_ty, llvm_v4i64_ty, llvm_i32_ty, llvm_i8_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<3>]>; def int_x86_avx512_maskz_fixupimm_pd_256 : GCCBuiltin<"__builtin_ia32_fixupimmpd256_maskz">, Intrinsic<[llvm_v4f64_ty], [llvm_v4f64_ty, llvm_v4f64_ty, llvm_v4i64_ty, llvm_i32_ty, llvm_i8_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<3>]>; def int_x86_avx512_mask_fixupimm_pd_512 : GCCBuiltin<"__builtin_ia32_fixupimmpd512_mask">, Intrinsic<[llvm_v8f64_ty], @@ -3252,52 +3307,52 @@ GCCBuiltin<"__builtin_ia32_fixupimmps128_mask">, Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty, llvm_v4f32_ty, llvm_v4i32_ty, llvm_i32_ty, llvm_i8_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<3>]>; def int_x86_avx512_maskz_fixupimm_ps_128 : GCCBuiltin<"__builtin_ia32_fixupimmps128_maskz">, Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty, llvm_v4f32_ty, llvm_v4i32_ty, llvm_i32_ty, llvm_i8_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<3>]>; def int_x86_avx512_mask_fixupimm_ps_256 : GCCBuiltin<"__builtin_ia32_fixupimmps256_mask">, Intrinsic<[llvm_v8f32_ty], [llvm_v8f32_ty, llvm_v8f32_ty, llvm_v8i32_ty, llvm_i32_ty, llvm_i8_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<3>]>; def int_x86_avx512_maskz_fixupimm_ps_256 : GCCBuiltin<"__builtin_ia32_fixupimmps256_maskz">, Intrinsic<[llvm_v8f32_ty], [llvm_v8f32_ty, llvm_v8f32_ty, llvm_v8i32_ty, llvm_i32_ty, llvm_i8_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<3>]>; def int_x86_avx512_mask_fixupimm_ps_512 : GCCBuiltin<"__builtin_ia32_fixupimmps512_mask">, Intrinsic<[llvm_v16f32_ty], [llvm_v16f32_ty, llvm_v16f32_ty, llvm_v16i32_ty, llvm_i32_ty, - llvm_i16_ty, llvm_i32_ty], [IntrNoMem]>; + llvm_i16_ty, llvm_i32_ty], [IntrNoMem, ImmArg<3>, ImmArg<5>]>; def int_x86_avx512_maskz_fixupimm_ps_512 : GCCBuiltin<"__builtin_ia32_fixupimmps512_maskz">, Intrinsic<[llvm_v16f32_ty], [llvm_v16f32_ty, llvm_v16f32_ty, llvm_v16i32_ty, llvm_i32_ty, - llvm_i16_ty, llvm_i32_ty], [IntrNoMem]>; + llvm_i16_ty, llvm_i32_ty], [IntrNoMem, ImmArg<3>, ImmArg<5>]>; def int_x86_avx512_mask_fixupimm_sd : GCCBuiltin<"__builtin_ia32_fixupimmsd_mask">, Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty, llvm_v2f64_ty, llvm_v2i64_ty, llvm_i32_ty, llvm_i8_ty, - llvm_i32_ty], [IntrNoMem]>; + llvm_i32_ty], [IntrNoMem, ImmArg<3>, ImmArg<5>]>; def int_x86_avx512_maskz_fixupimm_sd : GCCBuiltin<"__builtin_ia32_fixupimmsd_maskz">, Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty, llvm_v2f64_ty, llvm_v2i64_ty, llvm_i32_ty, llvm_i8_ty, - llvm_i32_ty], [IntrNoMem]>; + llvm_i32_ty], [IntrNoMem, ImmArg<3>, ImmArg<5>]>; def int_x86_avx512_mask_fixupimm_ss : GCCBuiltin<"__builtin_ia32_fixupimmss_mask">, Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty, llvm_v4f32_ty, llvm_v4i32_ty, llvm_i32_ty, llvm_i8_ty, - llvm_i32_ty], [IntrNoMem]>; + llvm_i32_ty], [IntrNoMem, ImmArg<3>, ImmArg<5>]>; def int_x86_avx512_maskz_fixupimm_ss : GCCBuiltin<"__builtin_ia32_fixupimmss_maskz">, Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty, llvm_v4f32_ty, llvm_v4i32_ty, llvm_i32_ty, llvm_i8_ty, - llvm_i32_ty], [IntrNoMem]>; + llvm_i32_ty], [IntrNoMem, ImmArg<3>, ImmArg<5>]>; def int_x86_avx512_mask_getexp_pd_128 : GCCBuiltin<"__builtin_ia32_getexppd128_mask">, Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty, llvm_v2f64_ty, llvm_i8_ty], [IntrNoMem]>; @@ -3306,7 +3361,8 @@ llvm_i8_ty], [IntrNoMem]>; def int_x86_avx512_mask_getexp_pd_512 : GCCBuiltin<"__builtin_ia32_getexppd512_mask">, Intrinsic<[llvm_v8f64_ty], [llvm_v8f64_ty, llvm_v8f64_ty, - llvm_i8_ty, llvm_i32_ty], [IntrNoMem]>; + llvm_i8_ty, llvm_i32_ty], + [IntrNoMem, ImmArg<3>]>; def int_x86_avx512_mask_getexp_ps_128 : GCCBuiltin<"__builtin_ia32_getexpps128_mask">, Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty, llvm_v4f32_ty, llvm_i8_ty], [IntrNoMem]>; @@ -3315,62 +3371,65 @@ llvm_i8_ty], [IntrNoMem]>; def int_x86_avx512_mask_getexp_ps_512 : GCCBuiltin<"__builtin_ia32_getexpps512_mask">, Intrinsic<[llvm_v16f32_ty], [llvm_v16f32_ty, llvm_v16f32_ty, - llvm_i16_ty, llvm_i32_ty], [IntrNoMem]>; + llvm_i16_ty, llvm_i32_ty], + [IntrNoMem, ImmArg<3>]>; def int_x86_avx512_mask_getexp_ss : GCCBuiltin<"__builtin_ia32_getexpss128_round_mask">, Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty, llvm_v4f32_ty, llvm_v4f32_ty, - llvm_i8_ty, llvm_i32_ty], [IntrNoMem]>; + llvm_i8_ty, llvm_i32_ty], + [IntrNoMem, ImmArg<4>]>; def int_x86_avx512_mask_getexp_sd : GCCBuiltin<"__builtin_ia32_getexpsd128_round_mask">, Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty, llvm_v2f64_ty, llvm_v2f64_ty, - llvm_i8_ty, llvm_i32_ty], [IntrNoMem]>; + llvm_i8_ty, llvm_i32_ty], + [IntrNoMem, ImmArg<4>]>; def int_x86_avx512_mask_getmant_pd_128 : GCCBuiltin<"__builtin_ia32_getmantpd128_mask">, Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty,llvm_i32_ty, llvm_v2f64_ty, llvm_i8_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<1>]>; def int_x86_avx512_mask_getmant_pd_256 : GCCBuiltin<"__builtin_ia32_getmantpd256_mask">, Intrinsic<[llvm_v4f64_ty], [llvm_v4f64_ty,llvm_i32_ty, llvm_v4f64_ty, llvm_i8_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<1>]>; def int_x86_avx512_mask_getmant_pd_512 : GCCBuiltin<"__builtin_ia32_getmantpd512_mask">, Intrinsic<[llvm_v8f64_ty], [llvm_v8f64_ty,llvm_i32_ty, llvm_v8f64_ty, llvm_i8_ty,llvm_i32_ty ], - [IntrNoMem]>; + [IntrNoMem, ImmArg<1>, ImmArg<4>]>; def int_x86_avx512_mask_getmant_ps_128 : GCCBuiltin<"__builtin_ia32_getmantps128_mask">, Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty, llvm_i32_ty, llvm_v4f32_ty, llvm_i8_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<1>]>; def int_x86_avx512_mask_getmant_ps_256 : GCCBuiltin<"__builtin_ia32_getmantps256_mask">, Intrinsic<[llvm_v8f32_ty], [llvm_v8f32_ty, llvm_i32_ty, llvm_v8f32_ty, llvm_i8_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<1>]>; def int_x86_avx512_mask_getmant_ps_512 : GCCBuiltin<"__builtin_ia32_getmantps512_mask">, Intrinsic<[llvm_v16f32_ty], [llvm_v16f32_ty,llvm_i32_ty, llvm_v16f32_ty,llvm_i16_ty,llvm_i32_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<1>, ImmArg<4>]>; def int_x86_avx512_mask_getmant_ss : GCCBuiltin<"__builtin_ia32_getmantss_round_mask">, Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty, llvm_v4f32_ty, llvm_i32_ty, llvm_v4f32_ty, - llvm_i8_ty, llvm_i32_ty], [IntrNoMem]>; + llvm_i8_ty, llvm_i32_ty], [IntrNoMem, ImmArg<2>, ImmArg<5>]>; def int_x86_avx512_mask_getmant_sd : GCCBuiltin<"__builtin_ia32_getmantsd_round_mask">, Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty, llvm_v2f64_ty, llvm_i32_ty, llvm_v2f64_ty, - llvm_i8_ty, llvm_i32_ty], [IntrNoMem]>; + llvm_i8_ty, llvm_i32_ty], [IntrNoMem, ImmArg<2>, ImmArg<5>]>; def int_x86_avx512_rsqrt14_ss : GCCBuiltin<"__builtin_ia32_rsqrt14ss_mask">, Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty, llvm_v4f32_ty, llvm_v4f32_ty, @@ -3425,41 +3484,41 @@ def int_x86_avx512_rcp28_ps : GCCBuiltin<"__builtin_ia32_rcp28ps_mask">, Intrinsic<[llvm_v16f32_ty], [llvm_v16f32_ty, llvm_v16f32_ty, - llvm_i16_ty, llvm_i32_ty], [IntrNoMem]>; + llvm_i16_ty, llvm_i32_ty], [IntrNoMem, ImmArg<3>]>; def int_x86_avx512_rcp28_pd : GCCBuiltin<"__builtin_ia32_rcp28pd_mask">, Intrinsic<[llvm_v8f64_ty], [llvm_v8f64_ty, llvm_v8f64_ty, - llvm_i8_ty, llvm_i32_ty], [IntrNoMem]>; + llvm_i8_ty, llvm_i32_ty], [IntrNoMem, ImmArg<3>]>; def int_x86_avx512_exp2_ps : GCCBuiltin<"__builtin_ia32_exp2ps_mask">, Intrinsic<[llvm_v16f32_ty], [llvm_v16f32_ty, llvm_v16f32_ty, - llvm_i16_ty, llvm_i32_ty], [IntrNoMem]>; + llvm_i16_ty, llvm_i32_ty], [IntrNoMem, ImmArg<3>]>; def int_x86_avx512_exp2_pd : GCCBuiltin<"__builtin_ia32_exp2pd_mask">, Intrinsic<[llvm_v8f64_ty], [llvm_v8f64_ty, llvm_v8f64_ty, - llvm_i8_ty, llvm_i32_ty], [IntrNoMem]>; + llvm_i8_ty, llvm_i32_ty], [IntrNoMem, ImmArg<3>]>; def int_x86_avx512_rcp28_ss : GCCBuiltin<"__builtin_ia32_rcp28ss_round_mask">, Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty, llvm_v4f32_ty, llvm_v4f32_ty, llvm_i8_ty, llvm_i32_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<4>]>; def int_x86_avx512_rcp28_sd : GCCBuiltin<"__builtin_ia32_rcp28sd_round_mask">, Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty, llvm_v2f64_ty, llvm_v2f64_ty, llvm_i8_ty, llvm_i32_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<4>]>; def int_x86_avx512_rsqrt28_ps : GCCBuiltin<"__builtin_ia32_rsqrt28ps_mask">, Intrinsic<[llvm_v16f32_ty], [llvm_v16f32_ty, llvm_v16f32_ty, llvm_i16_ty, llvm_i32_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<3>]>; def int_x86_avx512_rsqrt28_pd : GCCBuiltin<"__builtin_ia32_rsqrt28pd_mask">, Intrinsic<[llvm_v8f64_ty], [llvm_v8f64_ty, llvm_v8f64_ty, llvm_i8_ty, llvm_i32_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<3>]>; def int_x86_avx512_rsqrt28_ss : GCCBuiltin<"__builtin_ia32_rsqrt28ss_round_mask">, Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty, llvm_v4f32_ty, llvm_v4f32_ty, llvm_i8_ty, llvm_i32_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<4>]>; def int_x86_avx512_rsqrt28_sd : GCCBuiltin<"__builtin_ia32_rsqrt28sd_round_mask">, Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty, llvm_v2f64_ty, llvm_v2f64_ty, llvm_i8_ty, llvm_i32_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<4>]>; def int_x86_avx512_psad_bw_512 : GCCBuiltin<"__builtin_ia32_psadbw512">, Intrinsic<[llvm_v8i64_ty], [llvm_v64i8_ty, llvm_v64i8_ty], [IntrNoMem, Commutative]>; @@ -3482,17 +3541,20 @@ def int_x86_avx512_dbpsadbw_128 : GCCBuiltin<"__builtin_ia32_dbpsadbw128">, Intrinsic<[llvm_v8i16_ty], - [llvm_v16i8_ty, llvm_v16i8_ty, llvm_i32_ty], [IntrNoMem]>; + [llvm_v16i8_ty, llvm_v16i8_ty, llvm_i32_ty], + [IntrNoMem, ImmArg<2>]>; def int_x86_avx512_dbpsadbw_256 : GCCBuiltin<"__builtin_ia32_dbpsadbw256">, Intrinsic<[llvm_v16i16_ty], - [llvm_v32i8_ty, llvm_v32i8_ty, llvm_i32_ty], [IntrNoMem]>; + [llvm_v32i8_ty, llvm_v32i8_ty, llvm_i32_ty], + [IntrNoMem, ImmArg<2>]>; def int_x86_avx512_dbpsadbw_512 : GCCBuiltin<"__builtin_ia32_dbpsadbw512">, Intrinsic<[llvm_v32i16_ty], - [llvm_v64i8_ty, llvm_v64i8_ty, llvm_i32_ty], [IntrNoMem]>; + [llvm_v64i8_ty, llvm_v64i8_ty, llvm_i32_ty], + [IntrNoMem, ImmArg<2>]>; } // Gather and Scatter ops @@ -3501,262 +3563,270 @@ def int_x86_avx512_gather_dpd_512 : Intrinsic<[llvm_v8f64_ty], [llvm_v8f64_ty, llvm_ptr_ty, llvm_v8i32_ty, llvm_i8_ty, llvm_i32_ty], - [IntrReadMem, IntrArgMemOnly]>; + [IntrReadMem, IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_gather_dps_512 : Intrinsic<[llvm_v16f32_ty], [llvm_v16f32_ty, llvm_ptr_ty, llvm_v16i32_ty, llvm_i16_ty, llvm_i32_ty], - [IntrReadMem, IntrArgMemOnly]>; + [IntrReadMem, IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_gather_qpd_512 : Intrinsic<[llvm_v8f64_ty], [llvm_v8f64_ty, llvm_ptr_ty, llvm_v8i64_ty, llvm_i8_ty, llvm_i32_ty], - [IntrReadMem, IntrArgMemOnly]>; + [IntrReadMem, IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_gather_qps_512 : Intrinsic<[llvm_v8f32_ty], [llvm_v8f32_ty, llvm_ptr_ty, llvm_v8i64_ty, llvm_i8_ty, llvm_i32_ty], - [IntrReadMem, IntrArgMemOnly]>; + [IntrReadMem, IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_gather_dpq_512 : Intrinsic<[llvm_v8i64_ty], [llvm_v8i64_ty, llvm_ptr_ty, llvm_v8i32_ty, llvm_i8_ty, llvm_i32_ty], - [IntrReadMem, IntrArgMemOnly]>; + [IntrReadMem, IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_gather_dpi_512 : Intrinsic<[llvm_v16i32_ty], [llvm_v16i32_ty, llvm_ptr_ty, llvm_v16i32_ty, llvm_i16_ty, llvm_i32_ty], - [IntrReadMem, IntrArgMemOnly]>; + [IntrReadMem, IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_gather_qpq_512 : Intrinsic<[llvm_v8i64_ty], [llvm_v8i64_ty, llvm_ptr_ty, llvm_v8i64_ty, llvm_i8_ty, llvm_i32_ty], - [IntrReadMem, IntrArgMemOnly]>; + [IntrReadMem, IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_gather_qpi_512 : Intrinsic<[llvm_v8i32_ty], [llvm_v8i32_ty, llvm_ptr_ty, llvm_v8i64_ty, llvm_i8_ty, llvm_i32_ty], - [IntrReadMem, IntrArgMemOnly]>; + [IntrReadMem, IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_gather3div2_df : Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty, llvm_ptr_ty, llvm_v2i64_ty, llvm_i8_ty, llvm_i32_ty], - [IntrReadMem, IntrArgMemOnly]>; + [IntrReadMem, IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_gather3div2_di : Intrinsic<[llvm_v2i64_ty], [llvm_v2i64_ty, llvm_ptr_ty, llvm_v2i64_ty, llvm_i8_ty, llvm_i32_ty], - [IntrReadMem, IntrArgMemOnly]>; + [IntrReadMem, IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_gather3div4_df : Intrinsic<[llvm_v4f64_ty], [llvm_v4f64_ty, llvm_ptr_ty, llvm_v4i64_ty, llvm_i8_ty, llvm_i32_ty], - [IntrReadMem, IntrArgMemOnly]>; + [IntrReadMem, IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_gather3div4_di : Intrinsic<[llvm_v4i64_ty], [llvm_v4i64_ty, llvm_ptr_ty, llvm_v4i64_ty, llvm_i8_ty, llvm_i32_ty], - [IntrReadMem, IntrArgMemOnly]>; + [IntrReadMem, IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_gather3div4_sf : Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty, llvm_ptr_ty, llvm_v2i64_ty, llvm_i8_ty, llvm_i32_ty], - [IntrReadMem, IntrArgMemOnly]>; + [IntrReadMem, IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_gather3div4_si : Intrinsic<[llvm_v4i32_ty], [llvm_v4i32_ty, llvm_ptr_ty, llvm_v2i64_ty, llvm_i8_ty, llvm_i32_ty], - [IntrReadMem, IntrArgMemOnly]>; + [IntrReadMem, IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_gather3div8_sf : Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty, llvm_ptr_ty, llvm_v4i64_ty, llvm_i8_ty, llvm_i32_ty], - [IntrReadMem, IntrArgMemOnly]>; + [IntrReadMem, IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_gather3div8_si : Intrinsic<[llvm_v4i32_ty], [llvm_v4i32_ty, llvm_ptr_ty, llvm_v4i64_ty, llvm_i8_ty, llvm_i32_ty], - [IntrReadMem, IntrArgMemOnly]>; + [IntrReadMem, IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_gather3siv2_df : Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty, llvm_ptr_ty, llvm_v4i32_ty, llvm_i8_ty, llvm_i32_ty], - [IntrReadMem, IntrArgMemOnly]>; + [IntrReadMem, IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_gather3siv2_di : Intrinsic<[llvm_v2i64_ty], [llvm_v2i64_ty, llvm_ptr_ty, llvm_v4i32_ty, llvm_i8_ty, llvm_i32_ty], - [IntrReadMem, IntrArgMemOnly]>; + [IntrReadMem, IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_gather3siv4_df : Intrinsic<[llvm_v4f64_ty], [llvm_v4f64_ty, llvm_ptr_ty, llvm_v4i32_ty, llvm_i8_ty, llvm_i32_ty], - [IntrReadMem, IntrArgMemOnly]>; + [IntrReadMem, IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_gather3siv4_di : Intrinsic<[llvm_v4i64_ty], [llvm_v4i64_ty, llvm_ptr_ty, llvm_v4i32_ty, llvm_i8_ty, llvm_i32_ty], - [IntrReadMem, IntrArgMemOnly]>; + [IntrReadMem, IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_gather3siv4_sf : Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty, llvm_ptr_ty, llvm_v4i32_ty, llvm_i8_ty, llvm_i32_ty], - [IntrReadMem, IntrArgMemOnly]>; + [IntrReadMem, IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_gather3siv4_si : Intrinsic<[llvm_v4i32_ty], [llvm_v4i32_ty, llvm_ptr_ty, llvm_v4i32_ty, llvm_i8_ty, llvm_i32_ty], - [IntrReadMem, IntrArgMemOnly]>; + [IntrReadMem, IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_gather3siv8_sf : Intrinsic<[llvm_v8f32_ty], [llvm_v8f32_ty, llvm_ptr_ty, llvm_v8i32_ty, llvm_i8_ty, llvm_i32_ty], - [IntrReadMem, IntrArgMemOnly]>; + [IntrReadMem, IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_gather3siv8_si : Intrinsic<[llvm_v8i32_ty], [llvm_v8i32_ty, llvm_ptr_ty, llvm_v8i32_ty, llvm_i8_ty, llvm_i32_ty], - [IntrReadMem, IntrArgMemOnly]>; + [IntrReadMem, IntrArgMemOnly, ImmArg<4>]>; // scatter // NOTE: These are deprecated in favor of the versions that take a vXi1 mask. def int_x86_avx512_scatter_dpd_512 : Intrinsic<[], [llvm_ptr_ty, llvm_i8_ty, llvm_v8i32_ty, llvm_v8f64_ty, llvm_i32_ty], - [IntrArgMemOnly]>; + [IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_scatter_dps_512 : Intrinsic<[], [llvm_ptr_ty, llvm_i16_ty, llvm_v16i32_ty, llvm_v16f32_ty, llvm_i32_ty], - [IntrArgMemOnly]>; + [IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_scatter_qpd_512 : Intrinsic<[], [llvm_ptr_ty, llvm_i8_ty, llvm_v8i64_ty, llvm_v8f64_ty, llvm_i32_ty], - [IntrArgMemOnly]>; + [IntrArgMemOnly, ImmArg<4>, ImmArg<4>]>; def int_x86_avx512_scatter_qps_512 : Intrinsic<[], [llvm_ptr_ty, llvm_i8_ty, llvm_v8i64_ty, llvm_v8f32_ty, llvm_i32_ty], - [IntrArgMemOnly]>; + [IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_scatter_dpq_512 : Intrinsic<[], [llvm_ptr_ty, llvm_i8_ty, llvm_v8i32_ty, llvm_v8i64_ty, llvm_i32_ty], - [IntrArgMemOnly]>; + [IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_scatter_dpi_512 : Intrinsic<[], [llvm_ptr_ty, llvm_i16_ty, llvm_v16i32_ty, llvm_v16i32_ty, llvm_i32_ty], - [IntrArgMemOnly]>; + [IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_scatter_qpq_512 : Intrinsic<[], [llvm_ptr_ty, llvm_i8_ty,llvm_v8i64_ty, llvm_v8i64_ty, llvm_i32_ty], - [IntrArgMemOnly]>; + [IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_scatter_qpi_512 : Intrinsic<[], [llvm_ptr_ty, llvm_i8_ty, llvm_v8i64_ty, llvm_v8i32_ty, llvm_i32_ty], - [IntrArgMemOnly]>; + [IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_scatterdiv2_df : Intrinsic<[], [llvm_ptr_ty, llvm_i8_ty, llvm_v2i64_ty, llvm_v2f64_ty, llvm_i32_ty], - [IntrArgMemOnly]>; + [IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_scatterdiv2_di : Intrinsic<[], [llvm_ptr_ty, llvm_i8_ty, llvm_v2i64_ty, llvm_v2i64_ty, llvm_i32_ty], - [IntrArgMemOnly]>; + [IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_scatterdiv4_df : Intrinsic<[], [llvm_ptr_ty, llvm_i8_ty, llvm_v4i64_ty, llvm_v4f64_ty, llvm_i32_ty], - [IntrArgMemOnly]>; + [IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_scatterdiv4_di : Intrinsic<[], [llvm_ptr_ty, llvm_i8_ty, llvm_v4i64_ty, llvm_v4i64_ty, llvm_i32_ty], - [IntrArgMemOnly]>; + [IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_scatterdiv4_sf : Intrinsic<[], [llvm_ptr_ty, llvm_i8_ty, llvm_v2i64_ty, llvm_v4f32_ty, llvm_i32_ty], - [IntrArgMemOnly]>; + [IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_scatterdiv4_si : Intrinsic<[], [llvm_ptr_ty, llvm_i8_ty, llvm_v2i64_ty, llvm_v4i32_ty, llvm_i32_ty], - [IntrArgMemOnly]>; + [IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_scatterdiv8_sf : Intrinsic<[], [llvm_ptr_ty, llvm_i8_ty, llvm_v4i64_ty, llvm_v4f32_ty, llvm_i32_ty], - [IntrArgMemOnly]>; + [IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_scatterdiv8_si : Intrinsic<[], [llvm_ptr_ty, llvm_i8_ty, llvm_v4i64_ty, llvm_v4i32_ty, llvm_i32_ty], - [IntrArgMemOnly]>; + [IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_scattersiv2_df : Intrinsic<[], [llvm_ptr_ty, llvm_i8_ty, llvm_v4i32_ty, llvm_v2f64_ty, llvm_i32_ty], - [IntrArgMemOnly]>; + [IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_scattersiv2_di : Intrinsic<[], [llvm_ptr_ty, llvm_i8_ty, llvm_v4i32_ty, llvm_v2i64_ty, llvm_i32_ty], - [IntrArgMemOnly]>; + [IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_scattersiv4_df : Intrinsic<[], [llvm_ptr_ty, llvm_i8_ty, llvm_v4i32_ty, llvm_v4f64_ty, llvm_i32_ty], - [IntrArgMemOnly]>; + [IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_scattersiv4_di : Intrinsic<[], [llvm_ptr_ty, llvm_i8_ty, llvm_v4i32_ty, llvm_v4i64_ty, llvm_i32_ty], - [IntrArgMemOnly]>; + [IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_scattersiv4_sf : Intrinsic<[], [llvm_ptr_ty, llvm_i8_ty, llvm_v4i32_ty, llvm_v4f32_ty, llvm_i32_ty], - [IntrArgMemOnly]>; + [IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_scattersiv4_si : Intrinsic<[], [llvm_ptr_ty, llvm_i8_ty, llvm_v4i32_ty, llvm_v4i32_ty, llvm_i32_ty], - [IntrArgMemOnly]>; + [IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_scattersiv8_sf : Intrinsic<[], [llvm_ptr_ty, llvm_i8_ty, llvm_v8i32_ty, llvm_v8f32_ty, llvm_i32_ty], - [IntrArgMemOnly]>; + [IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_scattersiv8_si : Intrinsic<[], [llvm_ptr_ty, llvm_i8_ty, llvm_v8i32_ty, llvm_v8i32_ty, llvm_i32_ty], - [IntrArgMemOnly]>; + [IntrArgMemOnly, ImmArg<4>]>; // gather prefetch def int_x86_avx512_gatherpf_dpd_512 : GCCBuiltin<"__builtin_ia32_gatherpfdpd">, Intrinsic<[], [llvm_i8_ty, llvm_v8i32_ty, llvm_ptr_ty, - llvm_i32_ty, llvm_i32_ty], [IntrArgMemOnly]>; + llvm_i32_ty, llvm_i32_ty], + [IntrArgMemOnly, ImmArg<3>, ImmArg<4>]>; def int_x86_avx512_gatherpf_dps_512 : GCCBuiltin<"__builtin_ia32_gatherpfdps">, Intrinsic<[], [llvm_i16_ty, llvm_v16i32_ty, llvm_ptr_ty, - llvm_i32_ty, llvm_i32_ty], [IntrArgMemOnly]>; + llvm_i32_ty, llvm_i32_ty], + [IntrArgMemOnly, ImmArg<3>, ImmArg<4>]>; def int_x86_avx512_gatherpf_qpd_512 : GCCBuiltin<"__builtin_ia32_gatherpfqpd">, Intrinsic<[], [llvm_i8_ty, llvm_v8i64_ty, llvm_ptr_ty, - llvm_i32_ty, llvm_i32_ty], [IntrArgMemOnly]>; + llvm_i32_ty, llvm_i32_ty], + [IntrArgMemOnly, ImmArg<3>, ImmArg<4>]>; def int_x86_avx512_gatherpf_qps_512 : GCCBuiltin<"__builtin_ia32_gatherpfqps">, Intrinsic<[], [llvm_i8_ty, llvm_v8i64_ty, llvm_ptr_ty, - llvm_i32_ty, llvm_i32_ty], [IntrArgMemOnly]>; + llvm_i32_ty, llvm_i32_ty], + [IntrArgMemOnly, ImmArg<3>, ImmArg<4>]>; // scatter prefetch def int_x86_avx512_scatterpf_dpd_512 : GCCBuiltin<"__builtin_ia32_scatterpfdpd">, Intrinsic<[], [llvm_i8_ty, llvm_v8i32_ty, llvm_ptr_ty, - llvm_i32_ty, llvm_i32_ty], [IntrArgMemOnly]>; + llvm_i32_ty, llvm_i32_ty], + [IntrArgMemOnly, ImmArg<3>, ImmArg<4>]>; def int_x86_avx512_scatterpf_dps_512 : GCCBuiltin<"__builtin_ia32_scatterpfdps">, Intrinsic<[], [llvm_i16_ty, llvm_v16i32_ty, llvm_ptr_ty, - llvm_i32_ty, llvm_i32_ty], [IntrArgMemOnly]>; + llvm_i32_ty, llvm_i32_ty], + [IntrArgMemOnly, ImmArg<3>, ImmArg<4>]>; def int_x86_avx512_scatterpf_qpd_512 : GCCBuiltin<"__builtin_ia32_scatterpfqpd">, Intrinsic<[], [llvm_i8_ty, llvm_v8i64_ty, llvm_ptr_ty, - llvm_i32_ty, llvm_i32_ty], [IntrArgMemOnly]>; + llvm_i32_ty, llvm_i32_ty], + [IntrArgMemOnly, ImmArg<3>, ImmArg<4>]>; def int_x86_avx512_scatterpf_qps_512 : GCCBuiltin<"__builtin_ia32_scatterpfqps">, Intrinsic<[], [llvm_i8_ty, llvm_v8i64_ty, llvm_ptr_ty, - llvm_i32_ty, llvm_i32_ty], [IntrArgMemOnly]>; + llvm_i32_ty, llvm_i32_ty], + [IntrArgMemOnly, ImmArg<3>, ImmArg<4>]>; } // AVX512 gather/scatter intrinsics that use vXi1 masks. @@ -3764,232 +3834,232 @@ def int_x86_avx512_mask_gather_dpd_512 : Intrinsic<[llvm_v8f64_ty], [llvm_v8f64_ty, llvm_ptr_ty, llvm_v8i32_ty, llvm_v8i1_ty, llvm_i32_ty], - [IntrReadMem, IntrArgMemOnly]>; + [IntrReadMem, IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_mask_gather_dps_512 : Intrinsic<[llvm_v16f32_ty], [llvm_v16f32_ty, llvm_ptr_ty, llvm_v16i32_ty, llvm_v16i1_ty, llvm_i32_ty], - [IntrReadMem, IntrArgMemOnly]>; + [IntrReadMem, IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_mask_gather_qpd_512 : Intrinsic<[llvm_v8f64_ty], [llvm_v8f64_ty, llvm_ptr_ty, llvm_v8i64_ty, llvm_v8i1_ty, llvm_i32_ty], - [IntrReadMem, IntrArgMemOnly]>; + [IntrReadMem, IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_mask_gather_qps_512 : Intrinsic<[llvm_v8f32_ty], [llvm_v8f32_ty, llvm_ptr_ty, llvm_v8i64_ty, llvm_v8i1_ty, llvm_i32_ty], - [IntrReadMem, IntrArgMemOnly]>; + [IntrReadMem, IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_mask_gather_dpq_512 : Intrinsic<[llvm_v8i64_ty], [llvm_v8i64_ty, llvm_ptr_ty, llvm_v8i32_ty, llvm_v8i1_ty, llvm_i32_ty], - [IntrReadMem, IntrArgMemOnly]>; + [IntrReadMem, IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_mask_gather_dpi_512 : Intrinsic<[llvm_v16i32_ty], [llvm_v16i32_ty, llvm_ptr_ty, llvm_v16i32_ty, llvm_v16i1_ty, llvm_i32_ty], - [IntrReadMem, IntrArgMemOnly]>; + [IntrReadMem, IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_mask_gather_qpq_512 : Intrinsic<[llvm_v8i64_ty], [llvm_v8i64_ty, llvm_ptr_ty, llvm_v8i64_ty, llvm_v8i1_ty, llvm_i32_ty], - [IntrReadMem, IntrArgMemOnly]>; + [IntrReadMem, IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_mask_gather_qpi_512 : Intrinsic<[llvm_v8i32_ty], [llvm_v8i32_ty, llvm_ptr_ty, llvm_v8i64_ty, llvm_v8i1_ty, llvm_i32_ty], - [IntrReadMem, IntrArgMemOnly]>; + [IntrReadMem, IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_mask_gather3div2_df : Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty, llvm_ptr_ty, llvm_v2i64_ty, llvm_v2i1_ty, llvm_i32_ty], - [IntrReadMem, IntrArgMemOnly]>; + [IntrReadMem, IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_mask_gather3div2_di : Intrinsic<[llvm_v2i64_ty], [llvm_v2i64_ty, llvm_ptr_ty, llvm_v2i64_ty, llvm_v2i1_ty, llvm_i32_ty], - [IntrReadMem, IntrArgMemOnly]>; + [IntrReadMem, IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_mask_gather3div4_df : Intrinsic<[llvm_v4f64_ty], [llvm_v4f64_ty, llvm_ptr_ty, llvm_v4i64_ty, llvm_v4i1_ty, llvm_i32_ty], - [IntrReadMem, IntrArgMemOnly]>; + [IntrReadMem, IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_mask_gather3div4_di : Intrinsic<[llvm_v4i64_ty], [llvm_v4i64_ty, llvm_ptr_ty, llvm_v4i64_ty, llvm_v4i1_ty, llvm_i32_ty], - [IntrReadMem, IntrArgMemOnly]>; + [IntrReadMem, IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_mask_gather3div4_sf : Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty, llvm_ptr_ty, llvm_v2i64_ty, llvm_v2i1_ty, llvm_i32_ty], - [IntrReadMem, IntrArgMemOnly]>; + [IntrReadMem, IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_mask_gather3div4_si : Intrinsic<[llvm_v4i32_ty], [llvm_v4i32_ty, llvm_ptr_ty, llvm_v2i64_ty, llvm_v2i1_ty, llvm_i32_ty], - [IntrReadMem, IntrArgMemOnly]>; + [IntrReadMem, IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_mask_gather3div8_sf : Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty, llvm_ptr_ty, llvm_v4i64_ty, llvm_v4i1_ty, llvm_i32_ty], - [IntrReadMem, IntrArgMemOnly]>; + [IntrReadMem, IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_mask_gather3div8_si : Intrinsic<[llvm_v4i32_ty], [llvm_v4i32_ty, llvm_ptr_ty, llvm_v4i64_ty, llvm_v4i1_ty, llvm_i32_ty], - [IntrReadMem, IntrArgMemOnly]>; + [IntrReadMem, IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_mask_gather3siv2_df : Intrinsic<[llvm_v2f64_ty], [llvm_v2f64_ty, llvm_ptr_ty, llvm_v4i32_ty, llvm_v2i1_ty, llvm_i32_ty], - [IntrReadMem, IntrArgMemOnly]>; + [IntrReadMem, IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_mask_gather3siv2_di : Intrinsic<[llvm_v2i64_ty], [llvm_v2i64_ty, llvm_ptr_ty, llvm_v4i32_ty, llvm_v2i1_ty, llvm_i32_ty], - [IntrReadMem, IntrArgMemOnly]>; + [IntrReadMem, IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_mask_gather3siv4_df : Intrinsic<[llvm_v4f64_ty], [llvm_v4f64_ty, llvm_ptr_ty, llvm_v4i32_ty, llvm_v4i1_ty, llvm_i32_ty], - [IntrReadMem, IntrArgMemOnly]>; + [IntrReadMem, IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_mask_gather3siv4_di : Intrinsic<[llvm_v4i64_ty], [llvm_v4i64_ty, llvm_ptr_ty, llvm_v4i32_ty, llvm_v4i1_ty, llvm_i32_ty], - [IntrReadMem, IntrArgMemOnly]>; + [IntrReadMem, IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_mask_gather3siv4_sf : Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty, llvm_ptr_ty, llvm_v4i32_ty, llvm_v4i1_ty, llvm_i32_ty], - [IntrReadMem, IntrArgMemOnly]>; + [IntrReadMem, IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_mask_gather3siv4_si : Intrinsic<[llvm_v4i32_ty], [llvm_v4i32_ty, llvm_ptr_ty, llvm_v4i32_ty, llvm_v4i1_ty, llvm_i32_ty], - [IntrReadMem, IntrArgMemOnly]>; + [IntrReadMem, IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_mask_gather3siv8_sf : Intrinsic<[llvm_v8f32_ty], [llvm_v8f32_ty, llvm_ptr_ty, llvm_v8i32_ty, llvm_v8i1_ty, llvm_i32_ty], - [IntrReadMem, IntrArgMemOnly]>; + [IntrReadMem, IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_mask_gather3siv8_si : Intrinsic<[llvm_v8i32_ty], [llvm_v8i32_ty, llvm_ptr_ty, llvm_v8i32_ty, llvm_v8i1_ty, llvm_i32_ty], - [IntrReadMem, IntrArgMemOnly]>; + [IntrReadMem, IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_mask_scatter_dpd_512 : Intrinsic<[], [llvm_ptr_ty, llvm_v8i1_ty, llvm_v8i32_ty, llvm_v8f64_ty, llvm_i32_ty], - [IntrArgMemOnly]>; + [IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_mask_scatter_dps_512 : Intrinsic<[], [llvm_ptr_ty, llvm_v16i1_ty, llvm_v16i32_ty, llvm_v16f32_ty, llvm_i32_ty], - [IntrArgMemOnly]>; + [IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_mask_scatter_qpd_512 : Intrinsic<[], [llvm_ptr_ty, llvm_v8i1_ty, llvm_v8i64_ty, llvm_v8f64_ty, llvm_i32_ty], - [IntrArgMemOnly]>; + [IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_mask_scatter_qps_512 : Intrinsic<[], [llvm_ptr_ty, llvm_v8i1_ty, llvm_v8i64_ty, llvm_v8f32_ty, llvm_i32_ty], - [IntrArgMemOnly]>; + [IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_mask_scatter_dpq_512 : Intrinsic<[], [llvm_ptr_ty, llvm_v8i1_ty, llvm_v8i32_ty, llvm_v8i64_ty, llvm_i32_ty], - [IntrArgMemOnly]>; + [IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_mask_scatter_dpi_512 : Intrinsic<[], [llvm_ptr_ty, llvm_v16i1_ty, llvm_v16i32_ty, llvm_v16i32_ty, llvm_i32_ty], - [IntrArgMemOnly]>; + [IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_mask_scatter_qpq_512 : Intrinsic<[], [llvm_ptr_ty, llvm_v8i1_ty,llvm_v8i64_ty, llvm_v8i64_ty, llvm_i32_ty], - [IntrArgMemOnly]>; + [IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_mask_scatter_qpi_512 : Intrinsic<[], [llvm_ptr_ty, llvm_v8i1_ty, llvm_v8i64_ty, llvm_v8i32_ty, llvm_i32_ty], - [IntrArgMemOnly]>; + [IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_mask_scatterdiv2_df : Intrinsic<[], [llvm_ptr_ty, llvm_v2i1_ty, llvm_v2i64_ty, llvm_v2f64_ty, llvm_i32_ty], - [IntrArgMemOnly]>; + [IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_mask_scatterdiv2_di : Intrinsic<[], [llvm_ptr_ty, llvm_v2i1_ty, llvm_v2i64_ty, llvm_v2i64_ty, llvm_i32_ty], - [IntrArgMemOnly]>; + [IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_mask_scatterdiv4_df : Intrinsic<[], [llvm_ptr_ty, llvm_v4i1_ty, llvm_v4i64_ty, llvm_v4f64_ty, llvm_i32_ty], - [IntrArgMemOnly]>; + [IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_mask_scatterdiv4_di : Intrinsic<[], [llvm_ptr_ty, llvm_v4i1_ty, llvm_v4i64_ty, llvm_v4i64_ty, llvm_i32_ty], - [IntrArgMemOnly]>; + [IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_mask_scatterdiv4_sf : Intrinsic<[], [llvm_ptr_ty, llvm_v2i1_ty, llvm_v2i64_ty, llvm_v4f32_ty, llvm_i32_ty], - [IntrArgMemOnly]>; + [IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_mask_scatterdiv4_si : Intrinsic<[], [llvm_ptr_ty, llvm_v2i1_ty, llvm_v2i64_ty, llvm_v4i32_ty, llvm_i32_ty], - [IntrArgMemOnly]>; + [IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_mask_scatterdiv8_sf : Intrinsic<[], [llvm_ptr_ty, llvm_v4i1_ty, llvm_v4i64_ty, llvm_v4f32_ty, llvm_i32_ty], - [IntrArgMemOnly]>; + [IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_mask_scatterdiv8_si : Intrinsic<[], [llvm_ptr_ty, llvm_v4i1_ty, llvm_v4i64_ty, llvm_v4i32_ty, llvm_i32_ty], - [IntrArgMemOnly]>; + [IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_mask_scattersiv2_df : Intrinsic<[], [llvm_ptr_ty, llvm_v2i1_ty, llvm_v4i32_ty, llvm_v2f64_ty, llvm_i32_ty], - [IntrArgMemOnly]>; + [IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_mask_scattersiv2_di : Intrinsic<[], [llvm_ptr_ty, llvm_v2i1_ty, llvm_v4i32_ty, llvm_v2i64_ty, llvm_i32_ty], - [IntrArgMemOnly]>; + [IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_mask_scattersiv4_df : Intrinsic<[], [llvm_ptr_ty, llvm_v4i1_ty, llvm_v4i32_ty, llvm_v4f64_ty, llvm_i32_ty], - [IntrArgMemOnly]>; + [IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_mask_scattersiv4_di : Intrinsic<[], [llvm_ptr_ty, llvm_v4i1_ty, llvm_v4i32_ty, llvm_v4i64_ty, llvm_i32_ty], - [IntrArgMemOnly]>; + [IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_mask_scattersiv4_sf : Intrinsic<[], [llvm_ptr_ty, llvm_v4i1_ty, llvm_v4i32_ty, llvm_v4f32_ty, llvm_i32_ty], - [IntrArgMemOnly]>; + [IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_mask_scattersiv4_si : Intrinsic<[], [llvm_ptr_ty, llvm_v4i1_ty, llvm_v4i32_ty, llvm_v4i32_ty, llvm_i32_ty], - [IntrArgMemOnly]>; + [IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_mask_scattersiv8_sf : Intrinsic<[], [llvm_ptr_ty, llvm_v8i1_ty, llvm_v8i32_ty, llvm_v8f32_ty, llvm_i32_ty], - [IntrArgMemOnly]>; + [IntrArgMemOnly, ImmArg<4>]>; def int_x86_avx512_mask_scattersiv8_si : Intrinsic<[], [llvm_ptr_ty, llvm_v8i1_ty, llvm_v8i32_ty, llvm_v8i32_ty, llvm_i32_ty], - [IntrArgMemOnly]>; + [IntrArgMemOnly, ImmArg<4>]>; } // AVX-512 conflict detection instruction @@ -4021,10 +4091,12 @@ // 512-bit def int_x86_avx512_vcomi_sd : GCCBuiltin<"__builtin_ia32_vcomisd">, Intrinsic<[llvm_i32_ty], [llvm_v2f64_ty, - llvm_v2f64_ty, llvm_i32_ty, llvm_i32_ty], [IntrNoMem]>; + llvm_v2f64_ty, llvm_i32_ty, llvm_i32_ty], + [IntrNoMem, ImmArg<2>, ImmArg<3>]>; def int_x86_avx512_vcomi_ss : GCCBuiltin<"__builtin_ia32_vcomiss">, Intrinsic<[llvm_i32_ty], [llvm_v4f32_ty, - llvm_v4f32_ty, llvm_i32_ty, llvm_i32_ty], [IntrNoMem]>; + llvm_v4f32_ty, llvm_i32_ty, llvm_i32_ty], + [IntrNoMem, ImmArg<2>, ImmArg<3>]>; } // Compress, Expand @@ -4568,37 +4640,37 @@ GCCBuiltin<"__builtin_ia32_pternlogd128">, Intrinsic<[llvm_v4i32_ty], [llvm_v4i32_ty, llvm_v4i32_ty, llvm_v4i32_ty, llvm_i32_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<3>]>; def int_x86_avx512_pternlog_d_256 : GCCBuiltin<"__builtin_ia32_pternlogd256">, Intrinsic<[llvm_v8i32_ty], [llvm_v8i32_ty, llvm_v8i32_ty, llvm_v8i32_ty, llvm_i32_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<3>]>; def int_x86_avx512_pternlog_d_512 : GCCBuiltin<"__builtin_ia32_pternlogd512">, Intrinsic<[llvm_v16i32_ty], [llvm_v16i32_ty, llvm_v16i32_ty, llvm_v16i32_ty, - llvm_i32_ty], [IntrNoMem]>; + llvm_i32_ty], [IntrNoMem, ImmArg<3>]>; def int_x86_avx512_pternlog_q_128 : GCCBuiltin<"__builtin_ia32_pternlogq128">, Intrinsic<[llvm_v2i64_ty], [llvm_v2i64_ty, llvm_v2i64_ty, llvm_v2i64_ty, llvm_i32_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<3>]>; def int_x86_avx512_pternlog_q_256 : GCCBuiltin<"__builtin_ia32_pternlogq256">, Intrinsic<[llvm_v4i64_ty], [llvm_v4i64_ty, llvm_v4i64_ty, llvm_v4i64_ty, llvm_i32_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<3>]>; def int_x86_avx512_pternlog_q_512 : GCCBuiltin<"__builtin_ia32_pternlogq512">, Intrinsic<[llvm_v8i64_ty], [llvm_v8i64_ty, llvm_v8i64_ty, llvm_v8i64_ty, llvm_i32_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<3>]>; } // Misc. @@ -4607,31 +4679,35 @@ // distinction in signaling behaviour is not implemented. def int_x86_avx512_cmp_ps_512 : Intrinsic<[llvm_v16i1_ty], [llvm_v16f32_ty, llvm_v16f32_ty, - llvm_i32_ty, llvm_i32_ty], [IntrNoMem]>; + llvm_i32_ty, llvm_i32_ty], + [IntrNoMem, ImmArg<2>, ImmArg<3>]>; def int_x86_avx512_cmp_pd_512 : Intrinsic<[llvm_v8i1_ty], [llvm_v8f64_ty, llvm_v8f64_ty, - llvm_i32_ty, llvm_i32_ty], [IntrNoMem]>; + llvm_i32_ty, llvm_i32_ty], + [IntrNoMem, ImmArg<2>, ImmArg<3>]>; def int_x86_avx512_cmp_ps_256 : Intrinsic<[llvm_v8i1_ty], [llvm_v8f32_ty, llvm_v8f32_ty, - llvm_i32_ty], [IntrNoMem]>; + llvm_i32_ty], [IntrNoMem, ImmArg<2>]>; def int_x86_avx512_cmp_pd_256 : Intrinsic<[llvm_v4i1_ty], [llvm_v4f64_ty, llvm_v4f64_ty, - llvm_i32_ty], [IntrNoMem]>; + llvm_i32_ty], [IntrNoMem, ImmArg<2>]>; def int_x86_avx512_cmp_ps_128 : Intrinsic<[llvm_v4i1_ty], [llvm_v4f32_ty, llvm_v4f32_ty, - llvm_i32_ty], [IntrNoMem]>; + llvm_i32_ty], [IntrNoMem, ImmArg<2>]>; def int_x86_avx512_cmp_pd_128 : Intrinsic<[llvm_v2i1_ty], [llvm_v2f64_ty, llvm_v2f64_ty, - llvm_i32_ty], [IntrNoMem]>; + llvm_i32_ty], [IntrNoMem, ImmArg<2>]>; def int_x86_avx512_mask_cmp_ss : GCCBuiltin<"__builtin_ia32_cmpss_mask">, Intrinsic<[llvm_i8_ty], [llvm_v4f32_ty, llvm_v4f32_ty, - llvm_i32_ty, llvm_i8_ty, llvm_i32_ty], [IntrNoMem]>; + llvm_i32_ty, llvm_i8_ty, llvm_i32_ty], + [IntrNoMem, ImmArg<2>, ImmArg<4>]>; def int_x86_avx512_mask_cmp_sd : GCCBuiltin<"__builtin_ia32_cmpsd_mask">, Intrinsic<[llvm_i8_ty], [llvm_v2f64_ty, llvm_v2f64_ty, - llvm_i32_ty, llvm_i8_ty, llvm_i32_ty], [IntrNoMem]>; + llvm_i32_ty, llvm_i8_ty, llvm_i32_ty], + [IntrNoMem, ImmArg<2>, ImmArg<4>]>; } //===----------------------------------------------------------------------===// @@ -4639,7 +4715,7 @@ let TargetPrefix = "x86" in { def int_x86_sha1rnds4 : GCCBuiltin<"__builtin_ia32_sha1rnds4">, Intrinsic<[llvm_v4i32_ty], [llvm_v4i32_ty, llvm_v4i32_ty, llvm_i8_ty], - [IntrNoMem]>; + [IntrNoMem, ImmArg<2>]>; def int_x86_sha1nexte : GCCBuiltin<"__builtin_ia32_sha1nexte">, Intrinsic<[llvm_v4i32_ty], [llvm_v4i32_ty, llvm_v4i32_ty], [IntrNoMem]>; def int_x86_sha1msg1 : GCCBuiltin<"__builtin_ia32_sha1msg1">, Index: test/CodeGen/X86/pr33772.ll =================================================================== --- test/CodeGen/X86/pr33772.ll +++ /dev/null @@ -1,15 +0,0 @@ -; RUN: not llc < %s -mcpu=skylake-avx512 2>&1 | FileCheck %s - -target triple = "x86_64-unknown-linux-gnu" - -; make sure we don't crash if scale for gather isn't constant. - -; CHECK: LLVM ERROR: Cannot select: intrinsic %llvm.x86.avx512.gather.dpi.512 -declare <16 x i32> @llvm.x86.avx512.gather.dpi.512(<16 x i32>, i8*, <16 x i32>, i16, i32) - -define internal <16 x i32> @__gather_base_offsets32_i32(i8* readonly %ptr, i32 %offset_scale, <16 x i32> %offsets, <16 x i8> %vecmask) { - %mask_vec_i1.i.i = icmp ne <16 x i8> %vecmask, zeroinitializer - %mask_i16.i = bitcast <16 x i1> %mask_vec_i1.i.i to i16 - %res = tail call <16 x i32> @llvm.x86.avx512.gather.dpi.512(<16 x i32> undef, i8* %ptr, <16 x i32> %offsets, i16 %mask_i16.i, i32 %offset_scale) - ret <16 x i32> %res -} Index: test/Transforms/InstCombine/X86/x86-insertps.ll =================================================================== --- test/Transforms/InstCombine/X86/x86-insertps.ll +++ test/Transforms/InstCombine/X86/x86-insertps.ll @@ -3,18 +3,6 @@ declare <4 x float> @llvm.x86.sse41.insertps(<4 x float>, <4 x float>, i8) nounwind readnone -; This should never happen, but make sure we don't crash handling a non-constant immediate byte. - -define <4 x float> @insertps_non_const_imm(<4 x float> %v1, <4 x float> %v2, i8 %c) { -; CHECK-LABEL: @insertps_non_const_imm( -; CHECK-NEXT: [[RES:%.*]] = call <4 x float> @llvm.x86.sse41.insertps(<4 x float> %v1, <4 x float> %v2, i8 %c) -; CHECK-NEXT: ret <4 x float> [[RES]] -; - %res = call <4 x float> @llvm.x86.sse41.insertps(<4 x float> %v1, <4 x float> %v2, i8 %c) - ret <4 x float> %res - -} - ; If all zero mask bits are set, return a zero regardless of the other control bits. define <4 x float> @insertps_0x0f(<4 x float> %v1, <4 x float> %v2) {