AVX : Fix ISA disabling in case AVX512VL present

[igorb]

Summary

AVX : Fix ISA disabling in case AVX512VL , some instructions should be disabled only if AVX512BW present.
Tests added.

[delena]

lib/Target/X86/X86InstrSSE.td

4036

Please add NoBWI for all relevant templates (i8, i16) and explicit NoVLX for instructions with 32 and 64 bit elements.

test/CodeGen/X86/avx-isa-check.ll

2

Please add a comment what we are checking in this test.

[igorb]

new diff uploaded.
all rejects fixed

[delena]

LGTM

lib/Target/X86/X86InstrInfo.td

def HasPFI       : Predicate<"Subtarget->hasPFI()">,
                      AssemblerPredicate<"FeaturePFI", "AVX-512 PF ISA">;
 def HasERI       : Predicate<"Subtarget->hasERI()">,
                      AssemblerPredicate<"FeatureERI", "AVX-512 ER ISA">;
 def HasDQI       : Predicate<"Subtarget->hasDQI()">,
                      AssemblerPredicate<"FeatureDQI", "AVX-512 DQ ISA">;
 def NoDQI        : Predicate<"!Subtarget->hasDQI()">;
 def HasBWI       : Predicate<"Subtarget->hasBWI()">,
                      AssemblerPredicate<"FeatureBWI", "AVX-512 BW ISA">;
+def NoBWI        : Predicate<"!Subtarget->hasBWI()">;
 def HasVLX       : Predicate<"Subtarget->hasVLX()">,
                      AssemblerPredicate<"FeatureVLX", "AVX-512 VL ISA">;
 def NoVLX        : Predicate<"!Subtarget->hasVLX()">;
 
 def HasPOPCNT    : Predicate<"Subtarget->hasPOPCNT()">;
 def HasAES       : Predicate<"Subtarget->hasAES()">;
 def HasPCLMUL    : Predicate<"Subtarget->hasPCLMUL()">;
 def HasFMA       : Predicate<"Subtarget->hasFMA()">;

lib/Target/X86/X86InstrSSE.td

def rm : PDI<opc, MRMSrcMem, (outs RC:$dst),
                                      (bitconvert (memop_frag addr:$src2)))))],
                                      itins.rm>,
        Sched<[itins.Sched.Folded, ReadAfterLd]>;
 }
 } // ExeDomain = SSEPackedInt
 
 multiclass PDI_binop_all<bits<8> opc, string OpcodeStr, SDNode Opcode,
                          ValueType OpVT128, ValueType OpVT256,
-                         OpndItins itins, bit IsCommutable = 0> {
-let Predicates = [HasAVX, NoVLX] in
+                         OpndItins itins, bit IsCommutable = 0, Predicate prd> {
+let Predicates = [HasAVX, prd] in
   defm V#NAME : PDI_binop_rm<opc, !strconcat("v", OpcodeStr), Opcode, OpVT128,
                     VR128, loadv2i64, i128mem, itins, IsCommutable, 0>, VEX_4V;
 
 let Constraints = "$src1 = $dst" in
   defm NAME : PDI_binop_rm<opc, OpcodeStr, Opcode, OpVT128, VR128,
                            memopv2i64, i128mem, itins, IsCommutable, 1>;
 
-let Predicates = [HasAVX2, NoVLX] in
+let Predicates = [HasAVX2, prd] in
   defm V#NAME#Y : PDI_binop_rm<opc, !strconcat("v", OpcodeStr), Opcode,
                                OpVT256, VR256, loadv4i64, i256mem, itins,
                                IsCommutable, 0>, VEX_4V, VEX_L;
 }
 
 // These are ordered here for pattern ordering requirements with the fp versions
 
 defm PAND  : PDI_binop_all<0xDB, "pand", and, v2i64, v4i64,
-                           SSE_VEC_BIT_ITINS_P, 1>;
+                           SSE_VEC_BIT_ITINS_P, 1, NoVLX>;
 defm POR   : PDI_binop_all<0xEB, "por", or, v2i64, v4i64,
-                           SSE_VEC_BIT_ITINS_P, 1>;
+                           SSE_VEC_BIT_ITINS_P, 1, NoVLX>;
 defm PXOR  : PDI_binop_all<0xEF, "pxor", xor, v2i64, v4i64,
-                           SSE_VEC_BIT_ITINS_P, 1>;
+                           SSE_VEC_BIT_ITINS_P, 1, NoVLX>;
 defm PANDN : PDI_binop_all<0xDF, "pandn", X86andnp, v2i64, v4i64,
-                           SSE_VEC_BIT_ITINS_P, 0>;
+                           SSE_VEC_BIT_ITINS_P, 0, NoVLX>;
 
 //===----------------------------------------------------------------------===//
 // SSE 1 & 2 - Logical Instructions
 //===----------------------------------------------------------------------===//
 
 // Multiclass for scalars using the X86 logical operation aliases for FP.
 multiclass sse12_fp_packed_scalar_logical_alias<
     bits<8> opc, string OpcodeStr, SDNode OpNode, OpndItins itins> {
def rm : PDI<opc, MRMSrcMem, (outs RC:$dst),
            !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
        [(set RC:$dst, (DstVT (OpNode (SrcVT RC:$src1),
                                      (bitconvert (memop_frag addr:$src2)))))]>,
        Sched<[itins.Sched.Folded, ReadAfterLd]>;
 }
 } // ExeDomain = SSEPackedInt
 
 defm PADDB   : PDI_binop_all<0xFC, "paddb", add, v16i8, v32i8,
-                             SSE_INTALU_ITINS_P, 1>;
+                             SSE_INTALU_ITINS_P, 1, NoBWI>;
 defm PADDW   : PDI_binop_all<0xFD, "paddw", add, v8i16, v16i16,
-                             SSE_INTALU_ITINS_P, 1>;
+                             SSE_INTALU_ITINS_P, 1, NoBWI>;
 defm PADDD   : PDI_binop_all<0xFE, "paddd", add, v4i32, v8i32,
-                             SSE_INTALU_ITINS_P, 1>;
+                             SSE_INTALU_ITINS_P, 1, NoVLX>;
 defm PADDQ   : PDI_binop_all<0xD4, "paddq", add, v2i64, v4i64,
-                             SSE_INTALUQ_ITINS_P, 1>;
+                             SSE_INTALUQ_ITINS_P, 1, NoVLX>;
 defm PMULLW  : PDI_binop_all<0xD5, "pmullw", mul, v8i16, v16i16,
-                             SSE_INTMUL_ITINS_P, 1>;
+                             SSE_INTMUL_ITINS_P, 1, NoBWI>;
 defm PMULHUW : PDI_binop_all<0xE4, "pmulhuw", mulhu, v8i16, v16i16,
-                             SSE_INTMUL_ITINS_P, 1>;
+                             SSE_INTMUL_ITINS_P, 1, NoBWI>;
 defm PMULHW  : PDI_binop_all<0xE5, "pmulhw", mulhs, v8i16, v16i16,
-                             SSE_INTMUL_ITINS_P, 1>;
+                             SSE_INTMUL_ITINS_P, 1, NoBWI>;
 defm PSUBB   : PDI_binop_all<0xF8, "psubb", sub, v16i8, v32i8,
-                             SSE_INTALU_ITINS_P, 0>;
+                             SSE_INTALU_ITINS_P, 0, NoBWI>;
 defm PSUBW   : PDI_binop_all<0xF9, "psubw", sub, v8i16, v16i16,
-                             SSE_INTALU_ITINS_P, 0>;
+                             SSE_INTALU_ITINS_P, 0, NoBWI>;
 defm PSUBD   : PDI_binop_all<0xFA, "psubd", sub, v4i32, v8i32,
-                             SSE_INTALU_ITINS_P, 0>;
+                             SSE_INTALU_ITINS_P, 0, NoVLX>;
 defm PSUBQ   : PDI_binop_all<0xFB, "psubq", sub, v2i64, v4i64,
-                             SSE_INTALUQ_ITINS_P, 0>;
+                             SSE_INTALUQ_ITINS_P, 0, NoVLX>;
 defm PSUBUSB : PDI_binop_all<0xD8, "psubusb", X86subus, v16i8, v32i8,
-                             SSE_INTALU_ITINS_P, 0>;
+                             SSE_INTALU_ITINS_P, 0, NoBWI>;
 defm PSUBUSW : PDI_binop_all<0xD9, "psubusw", X86subus, v8i16, v16i16,
-                             SSE_INTALU_ITINS_P, 0>;
+                             SSE_INTALU_ITINS_P, 0, NoBWI>;
 defm PMINUB  : PDI_binop_all<0xDA, "pminub", umin, v16i8, v32i8,
-                             SSE_INTALU_ITINS_P, 1>;
+                             SSE_INTALU_ITINS_P, 1, NoBWI>;
 defm PMINSW  : PDI_binop_all<0xEA, "pminsw", smin, v8i16, v16i16,
-                             SSE_INTALU_ITINS_P, 1>;
+                             SSE_INTALU_ITINS_P, 1, NoBWI>;
 defm PMAXUB  : PDI_binop_all<0xDE, "pmaxub", umax, v16i8, v32i8,
-                             SSE_INTALU_ITINS_P, 1>;
+                             SSE_INTALU_ITINS_P, 1, NoBWI>;
 defm PMAXSW  : PDI_binop_all<0xEE, "pmaxsw", smax, v8i16, v16i16,
-                             SSE_INTALU_ITINS_P, 1>;
+                             SSE_INTALU_ITINS_P, 1, NoBWI>;
 
 // Intrinsic forms
 defm PSUBSB  : PDI_binop_all_int<0xE8, "psubsb", int_x86_sse2_psubs_b,
                                  int_x86_avx2_psubs_b, SSE_INTALU_ITINS_P, 0>;
 defm PSUBSW  : PDI_binop_all_int<0xE9, "psubsw" , int_x86_sse2_psubs_w,
                                  int_x86_avx2_psubs_w, SSE_INTALU_ITINS_P, 0>;
 defm PADDSB  : PDI_binop_all_int<0xEC, "paddsb" , int_x86_sse2_padds_b,
                                  int_x86_avx2_padds_b, SSE_INTALU_ITINS_P, 1>;

 }
 } // Constraints = "$src1 = $dst"
 
 //===---------------------------------------------------------------------===//
 // SSE2 - Packed Integer Comparison Instructions
 //===---------------------------------------------------------------------===//
 
 defm PCMPEQB : PDI_binop_all<0x74, "pcmpeqb", X86pcmpeq, v16i8, v32i8,
-                             SSE_INTALU_ITINS_P, 1>;
+                             SSE_INTALU_ITINS_P, 1, NoBWI>;
 defm PCMPEQW : PDI_binop_all<0x75, "pcmpeqw", X86pcmpeq, v8i16, v16i16,
-                             SSE_INTALU_ITINS_P, 1>;
+                             SSE_INTALU_ITINS_P, 1, NoBWI>;
 defm PCMPEQD : PDI_binop_all<0x76, "pcmpeqd", X86pcmpeq, v4i32, v8i32,
-                             SSE_INTALU_ITINS_P, 1>;
+                             SSE_INTALU_ITINS_P, 1, NoVLX>;
 defm PCMPGTB : PDI_binop_all<0x64, "pcmpgtb", X86pcmpgt, v16i8, v32i8,
-                             SSE_INTALU_ITINS_P, 0>;
+                             SSE_INTALU_ITINS_P, 0, NoBWI>;
 defm PCMPGTW : PDI_binop_all<0x65, "pcmpgtw", X86pcmpgt, v8i16, v16i16,
-                             SSE_INTALU_ITINS_P, 0>;
+                             SSE_INTALU_ITINS_P, 0, NoBWI>;
 defm PCMPGTD : PDI_binop_all<0x66, "pcmpgtd", X86pcmpgt, v4i32, v8i32,
-                             SSE_INTALU_ITINS_P, 0>;
+                             SSE_INTALU_ITINS_P, 0, NoVLX>;
 
 //===---------------------------------------------------------------------===//
 // SSE2 - Packed Integer Shuffle Instructions
 //===---------------------------------------------------------------------===//
 
 let ExeDomain = SSEPackedInt in {
 multiclass sse2_pshuffle<string OpcodeStr, ValueType vt128, ValueType vt256,
                          SDNode OpNode> {

test/CodeGen/X86/avx-isa-check.ll

+; check AVX2 instructions that are disabled in case avx512VL/avx512BW present
+   
+; RUN: llc < %s -mtriple=x86_64-apple-darwin -show-mc-encoding -mcpu=core-avx2 -mattr=+avx2                 -o /dev/null
+; RUN: llc < %s -mtriple=x86_64-apple-darwin -show-mc-encoding -mcpu=knl                                    -o /dev/null
+; RUN: llc < %s -mtriple=x86_64-apple-darwin -show-mc-encoding -mcpu=knl  -mattr=+avx512vl                  -o /dev/null
+; RUN: llc < %s -mtriple=x86_64-apple-darwin -show-mc-encoding -mcpu=knl  -mattr=+avx512vl -mattr=+avx512bw -o /dev/null
+; RUN: llc < %s -mtriple=x86_64-apple-darwin -show-mc-encoding -mcpu=skx                                    -o /dev/null
+
+define <4 x i64> @vpand_256(<4 x i64> %a, <4 x i64> %b) nounwind uwtable readnone ssp {
+  ; Force the execution domain with an add.
+  %a2 = add <4 x i64> %a, <i64 1, i64 1, i64 1, i64 1>
+  %x = and <4 x i64> %a2, %b
+  ret <4 x i64> %x
+}
+
+define <2 x i64> @vpand_128(<2 x i64> %a, <2 x i64> %b) nounwind uwtable readnone ssp {
+  ; Force the execution domain with an add.
+  %a2 = add <2 x i64> %a, <i64 1, i64 1>
+  %x = and <2 x i64> %a2, %b
+  ret <2 x i64> %x
+}
+
+define <4 x i64> @vpandn_256(<4 x i64> %a, <4 x i64> %b) nounwind uwtable readnone ssp {
+  ; Force the execution domain with an add.
+  %a2 = add <4 x i64> %a, <i64 1, i64 1, i64 1, i64 1>
+  %y = xor <4 x i64> %a2, <i64 -1, i64 -1, i64 -1, i64 -1>
+  %x = and <4 x i64> %a, %y
+  ret <4 x i64> %x
+}
+
+define <2 x i64> @vpandn_128(<2 x i64> %a, <2 x i64> %b) nounwind uwtable readnone ssp {
+  ; Force the execution domain with an add.
+  %a2 = add <2 x i64> %a, <i64 1, i64 1>
+  %y = xor <2 x i64> %a2, <i64 -1, i64 -1>
+  %x = and <2 x i64> %a, %y
+  ret <2 x i64> %x
+}
+
+define <4 x i64> @vpor_256(<4 x i64> %a, <4 x i64> %b) nounwind uwtable readnone ssp {
+  ; Force the execution domain with an add.
+  %a2 = add <4 x i64> %a, <i64 1, i64 1, i64 1, i64 1>
+  %x = or <4 x i64> %a2, %b
+  ret <4 x i64> %x
+}
+
+define <4 x i64> @vpxor_256(<4 x i64> %a, <4 x i64> %b) nounwind uwtable readnone ssp {
+  ; Force the execution domain with an add.
+  %a2 = add <4 x i64> %a, <i64 1, i64 1, i64 1, i64 1>
+  %x = xor <4 x i64> %a2, %b
+  ret <4 x i64> %x
+}
+
+define <2 x i64> @vpor_128(<2 x i64> %a, <2 x i64> %b) nounwind uwtable readnone ssp {
+  ; Force the execution domain with an add.
+  %a2 = add <2 x i64> %a, <i64 1, i64 1>
+  %x = or <2 x i64> %a2, %b
+  ret <2 x i64> %x
+}
+
+define <2 x i64> @vpxor_128(<2 x i64> %a, <2 x i64> %b) nounwind uwtable readnone ssp {
+  ; Force the execution domain with an add.
+  %a2 = add <2 x i64> %a, <i64 1, i64 1>
+  %x = xor <2 x i64> %a2, %b
+  ret <2 x i64> %x
+}
+
+define <4 x i64> @test_vpaddq_256(<4 x i64> %i, <4 x i64> %j) nounwind readnone {
+  %x = add <4 x i64> %i, %j
+  ret <4 x i64> %x
+}
+
+define <8 x i32> @test_vpaddd_256(<8 x i32> %i, <8 x i32> %j) nounwind readnone {
+  %x = add <8 x i32> %i, %j
+  ret <8 x i32> %x
+}
+
+define <16 x i16> @test_vpaddw_256(<16 x i16> %i, <16 x i16> %j) nounwind readnone {
+  %x = add <16 x i16> %i, %j
+  ret <16 x i16> %x
+}
+
+define <32 x i8> @test_vpaddb_256(<32 x i8> %i, <32 x i8> %j) nounwind readnone {
+  %x = add <32 x i8> %i, %j
+  ret <32 x i8> %x
+}
+
+define <4 x i64> @test_vpsubq_256(<4 x i64> %i, <4 x i64> %j) nounwind readnone {
+  %x = sub <4 x i64> %i, %j
+  ret <4 x i64> %x
+}
+
+define <8 x i32> @test_vpsubd_256(<8 x i32> %i, <8 x i32> %j) nounwind readnone {
+  %x = sub <8 x i32> %i, %j
+  ret <8 x i32> %x
+}
+
+define <16 x i16> @test_vpsubw_256(<16 x i16> %i, <16 x i16> %j) nounwind readnone {
+  %x = sub <16 x i16> %i, %j
+  ret <16 x i16> %x
+}
+
+define <32 x i8> @test_vpsubb_256(<32 x i8> %i, <32 x i8> %j) nounwind readnone {
+  %x = sub <32 x i8> %i, %j
+  ret <32 x i8> %x
+}
+
+define <16 x i16> @test_vpmullw_256(<16 x i16> %i, <16 x i16> %j) nounwind readnone {
+  %x = mul <16 x i16> %i, %j
+  ret <16 x i16> %x
+}
+
+define <8 x i32> @test_vpcmpgtd_256(<8 x i32> %i, <8 x i32> %j) nounwind readnone {
+  %bincmp = icmp slt <8 x i32> %i, %j
+  %x = sext <8 x i1> %bincmp to <8 x i32>
+  ret <8 x i32> %x
+}
+
+define <32 x i8> @test_vpcmpeqb_256(<32 x i8> %i, <32 x i8> %j) nounwind readnone {
+  %bincmp = icmp eq <32 x i8> %i, %j
+  %x = sext <32 x i1> %bincmp to <32 x i8>
+  ret <32 x i8> %x
+}
+
+define <16 x i16> @test_vpcmpeqw_256(<16 x i16> %i, <16 x i16> %j) nounwind readnone {
+  %bincmp = icmp eq <16 x i16> %i, %j
+  %x = sext <16 x i1> %bincmp to <16 x i16>
+  ret <16 x i16> %x
+}
+
+define <32 x i8> @test_vpcmpgtb_256(<32 x i8> %i, <32 x i8> %j) nounwind readnone {
+  %bincmp = icmp slt <32 x i8> %i, %j
+  %x = sext <32 x i1> %bincmp to <32 x i8>
+  ret <32 x i8> %x
+}
+
+define <16 x i16> @test_vpcmpgtw_256(<16 x i16> %i, <16 x i16> %j) nounwind readnone {
+  %bincmp = icmp slt <16 x i16> %i, %j
+  %x = sext <16 x i1> %bincmp to <16 x i16>
+  ret <16 x i16> %x
+}
+
+define <8 x i32> @test_vpcmpeqd_256(<8 x i32> %i, <8 x i32> %j) nounwind readnone {
+  %bincmp = icmp eq <8 x i32> %i, %j
+  %x = sext <8 x i1> %bincmp to <8 x i32>
+  ret <8 x i32> %x
+}
+
+define <2 x i64> @test_vpaddq_128(<2 x i64> %i, <2 x i64> %j) nounwind readnone {
+  %x = add <2 x i64> %i, %j
+  ret <2 x i64> %x
+}
+
+define <4 x i32> @test_vpaddd_128(<4 x i32> %i, <4 x i32> %j) nounwind readnone {
+  %x = add <4 x i32> %i, %j
+  ret <4 x i32> %x
+}
+
+define <8 x i16> @test_vpaddw_128(<8 x i16> %i, <8 x i16> %j) nounwind readnone {
+  %x = add <8 x i16> %i, %j
+  ret <8 x i16> %x
+}
+
+define <16 x i8> @test_vpaddb_128(<16 x i8> %i, <16 x i8> %j) nounwind readnone {
+  %x = add <16 x i8> %i, %j
+  ret <16 x i8> %x
+}
+
+define <2 x i64> @test_vpsubq_128(<2 x i64> %i, <2 x i64> %j) nounwind readnone {
+  %x = sub <2 x i64> %i, %j
+  ret <2 x i64> %x
+}
+
+define <4 x i32> @test_vpsubd_128(<4 x i32> %i, <4 x i32> %j) nounwind readnone {
+  %x = sub <4 x i32> %i, %j
+  ret <4 x i32> %x
+}
+
+define <8 x i16> @test_vpsubw_128(<8 x i16> %i, <8 x i16> %j) nounwind readnone {
+  %x = sub <8 x i16> %i, %j
+  ret <8 x i16> %x
+}
+
+define <16 x i8> @test_vpsubb_128(<16 x i8> %i, <16 x i8> %j) nounwind readnone {
+  %x = sub <16 x i8> %i, %j
+  ret <16 x i8> %x
+}
+
+define <8 x i16> @test_vpmullw_128(<8 x i16> %i, <8 x i16> %j) nounwind readnone {
+  %x = mul <8 x i16> %i, %j
+  ret <8 x i16> %x
+}
+
+define <8 x i16> @test_vpcmpgtw_128(<8 x i16> %i, <8 x i16> %j) nounwind readnone {
+  %bincmp = icmp slt <8 x i16> %i, %j
+  %x = sext <8 x i1> %bincmp to <8 x i16>
+  ret <8 x i16> %x
+}
+
+define <16 x i8> @test_vpcmpgtb_128(<16 x i8> %i, <16 x i8> %j) nounwind readnone {
+  %bincmp = icmp slt <16 x i8> %i, %j
+  %x = sext <16 x i1> %bincmp to <16 x i8>
+  ret <16 x i8> %x
+}
+
+define <8 x i16> @test_vpcmpeqw_128(<8 x i16> %i, <8 x i16> %j) nounwind readnone {
+  %bincmp = icmp eq <8 x i16> %i, %j
+  %x = sext <8 x i1> %bincmp to <8 x i16>
+  ret <8 x i16> %x
+}
+
+define <16 x i8> @test_vpcmpeqb_128(<16 x i8> %i, <16 x i8> %j) nounwind readnone {
+  %bincmp = icmp eq <16 x i8> %i, %j
+  %x = sext <16 x i1> %bincmp to <16 x i8>
+  ret <16 x i8> %x
+}