[AVX512] Fix v8i1 /v16i1 zext + bitcast lowering pattern. Explicitly zero upper bits.

Differential Revision: http://reviews.llvm.org/D23983

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@280650 91177308-0d34-0410-b5e6-96231b3b80d8

Author: Igor Breger

lib/Target/X86/X86InstrAVX512.td

@@ -2117,15 +2117,15 @@ def : Pat<(i8 (bitconvert (v8i1 VK8:$src))),
           (COPY_TO_REGCLASS VK8:$src, GR8)>;
 
 def : Pat<(i32 (zext (i16 (bitconvert (v16i1 VK16:$src))))),
-          (i32 (SUBREG_TO_REG (i64 0),
-                (i16 (COPY_TO_REGCLASS VK16:$src, GR16)), sub_16bit))>;
+          (KMOVWrk VK16:$src)>;
 def : Pat<(i32 (anyext (i16 (bitconvert (v16i1 VK16:$src))))),
           (i32 (SUBREG_TO_REG (i64 0),
                 (i16 (COPY_TO_REGCLASS VK16:$src, GR16)), sub_16bit))>;
 
 def : Pat<(i32 (zext (i8 (bitconvert (v8i1 VK8:$src))))),
-          (i32 (SUBREG_TO_REG (i64 0),
-                (i8 (COPY_TO_REGCLASS VK8:$src, GR8)), sub_8bit))>;
+          (MOVZX32rr8 (COPY_TO_REGCLASS VK8:$src, GR8))>, Requires<[NoDQI]>;
+def : Pat<(i32 (zext (i8 (bitconvert (v8i1 VK8:$src))))),
+          (KMOVBrk VK8:$src)>, Requires<[HasDQI]>;
 def : Pat<(i32 (anyext (i8 (bitconvert (v8i1 VK8:$src))))),
           (i32 (SUBREG_TO_REG (i64 0),
                 (i8 (COPY_TO_REGCLASS VK8:$src, GR8)), sub_8bit))>;

test/CodeGen/X86/avx512-mask-op.ll

@@ -55,6 +55,7 @@ define i32 @mask8_zext(i8 %x) {
 ; KNL-NEXT:    kmovw %edi, %k0
 ; KNL-NEXT:    knotw %k0, %k0
 ; KNL-NEXT:    kmovw %k0, %eax
+; KNL-NEXT:    movzbl %al, %eax
 ; KNL-NEXT:    retq
 ;
 ; SKX-LABEL: mask8_zext:
@@ -1959,3 +1960,43 @@ define void @store_64i1(<64 x i1>* %a, <64 x i1> %v) {
   store <64 x i1> %v, <64 x i1>* %a
   ret void
 }
+
+define i32 @test_bitcast_v8i1_zext(<16 x i32> %a) {
+; KNL-LABEL: test_bitcast_v8i1_zext:
+; KNL:       ## BB#0:
+; KNL-NEXT:    vpxord %zmm1, %zmm1, %zmm1
+; KNL-NEXT:    vpcmpeqd %zmm1, %zmm0, %k0
+; KNL-NEXT:    kmovw %k0, %eax
+; KNL-NEXT:    movzbl %al, %eax
+; KNL-NEXT:    addl %eax, %eax
+; KNL-NEXT:    retq
+;
+; SKX-LABEL: test_bitcast_v8i1_zext:
+; SKX:       ## BB#0:
+; SKX-NEXT:    vpxord %zmm1, %zmm1, %zmm1
+; SKX-NEXT:    vpcmpeqd %zmm1, %zmm0, %k0
+; SKX-NEXT:    kmovb %k0, %eax
+; SKX-NEXT:    addl %eax, %eax
+; SKX-NEXT:    retq
+   %v1 = icmp eq <16 x i32> %a, zeroinitializer
+   %mask = shufflevector <16 x i1> %v1, <16 x i1> undef, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
+   %mask1 = bitcast <8 x i1> %mask to i8
+   %val = zext i8 %mask1 to i32
+   %val1 = add i32 %val, %val
+   ret i32 %val1
+}
+
+define i32 @test_bitcast_v16i1_zext(<16 x i32> %a) {
+; CHECK-LABEL: test_bitcast_v16i1_zext:
+; CHECK:       ## BB#0:
+; CHECK-NEXT:    vpxord %zmm1, %zmm1, %zmm1
+; CHECK-NEXT:    vpcmpeqd %zmm1, %zmm0, %k0
+; CHECK-NEXT:    kmovw %k0, %eax
+; CHECK-NEXT:    addl %eax, %eax
+; CHECK-NEXT:    retq
+   %v1 = icmp eq <16 x i32> %a, zeroinitializer
+   %mask1 = bitcast <16 x i1> %v1 to i16
+   %val = zext i16 %mask1 to i32
+   %val1 = add i32 %val, %val
+   ret i32 %val1
+}

test/CodeGen/X86/avx512-mask-zext-bugfix.ll

@@ -0,0 +1,49 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc < %s -O0 -mtriple=x86_64-apple-darwin -mcpu=skx  | FileCheck %s
+
+; ModuleID = 'mask_set.c'
+source_filename = "mask_set.c"
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-linux-gnu"
+
+declare void @llvm.dbg.declare(metadata, metadata, metadata)
+
+; Function Attrs: nounwind uwtable
+declare i64 @calc_expected_mask_val(i8* %valp, i32 %el_size, i32 %length)
+; Function Attrs: nounwind uwtable
+declare i32 @check_mask16(i16 zeroext %res_mask, i16 zeroext %exp_mask, i8* %fname, i8* %input)
+
+; Function Attrs: nounwind uwtable
+define void @test_xmm(i32 %shift, i32 %mulp, <2 x i64> %a,i8* %arraydecay,i8* %fname){
+; CHECK-LABEL: test_xmm:
+; CHECK:       ## BB#0:
+; CHECK:         callq _calc_expected_mask_val
+; CHECK-NEXT:    movl %eax, %edx
+; CHECK-NEXT:    movw %dx, %r9w
+; CHECK-NEXT:    movzwl %r9w, %esi
+; CHECK-NEXT:    kmovw {{[0-9]+}}(%rsp), %k0 ## 2-byte Reload
+; CHECK-NEXT:    kmovb %k0, %edi
+; CHECK-NEXT:    movq {{[0-9]+}}(%rsp), %rdx ## 8-byte Reload
+; CHECK-NEXT:    movq {{[0-9]+}}(%rsp), %rcx ## 8-byte Reload
+; CHECK-NEXT:    callq _check_mask16
+  %d2 = bitcast <2 x i64> %a to <8 x i16>
+  %m2 = call i8 @llvm.x86.avx512.cvtw2mask.128(<8 x i16> %d2)
+  %conv7 = zext i8 %m2 to i16
+  %call9 = call i64 @calc_expected_mask_val(i8* %arraydecay, i32 2, i32 8)
+  %conv10 = trunc i64 %call9 to i16
+  %call12 = call i32 @check_mask16(i16 zeroext %conv7, i16 zeroext %conv10, i8* %fname, i8* %arraydecay)
+  %d3 = bitcast <2 x i64> %a to <4 x i32>
+  %m3 = call i8 @llvm.x86.avx512.cvtd2mask.128(<4 x i32> %d3)
+  %conv14 = zext i8 %m3 to i16
+  %call16 = call i64 @calc_expected_mask_val(i8* %arraydecay, i32 4, i32 4)
+  %conv17 = trunc i64 %call16 to i16
+  %call19 = call i32 @check_mask16(i16 zeroext %conv14, i16 zeroext %conv17, i8* %fname, i8* %arraydecay)
+  ret void
+}
+
+; Function Attrs: nounwind readnone
+declare i8 @llvm.x86.avx512.cvtw2mask.128(<8 x i16>)
+
+; Function Attrs: nounwind readnone
+declare i8 @llvm.x86.avx512.cvtd2mask.128(<4 x i32>)
+