From: Craig Topper <craig.topper@gmail.com>
Date: Mon, 9 Jan 2017 04:19:34 +0000 (+0000)
Subject: [AVX-512] Change another pattern that was using BLENDM to use masked moves. A future... 
X-Git-Url: https://granicus.if.org/sourcecode?a=commitdiff_plain;h=ca3a0c93a3e5f1f758ad60f103965b2dcbd200c7;p=llvm

[AVX-512] Change another pattern that was using BLENDM to use masked moves. A future patch will conver it back to BLENDM if its beneficial to register allocation.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@291419 91177308-0d34-0410-b5e6-96231b3b80d8
---

diff --git a/lib/Target/X86/X86InstrAVX512.td b/lib/Target/X86/X86InstrAVX512.td
index 75a265bf8a1..d44d1395f24 100644
--- a/lib/Target/X86/X86InstrAVX512.td
+++ b/lib/Target/X86/X86InstrAVX512.td
@@ -1588,21 +1588,6 @@ defm VPBLENDMB : blendmask_bw <0x66, "vpblendmb", avx512vl_i8_info>;
 defm VPBLENDMW : blendmask_bw <0x66, "vpblendmw", avx512vl_i16_info>, VEX_W;
 
 
-let Predicates = [HasAVX512, NoVLX] in {
-def : Pat<(v8f32 (vselect (v8i1 VK8WM:$mask), (v8f32 VR256X:$src1),
-                            (v8f32 VR256X:$src2))),
-            (EXTRACT_SUBREG
-              (v16f32 (VBLENDMPSZrrk (COPY_TO_REGCLASS VK8WM:$mask, VK16WM),
-            (v16f32 (INSERT_SUBREG (IMPLICIT_DEF), VR256X:$src2, sub_ymm)),
-            (v16f32 (INSERT_SUBREG (IMPLICIT_DEF), VR256X:$src1, sub_ymm)))), sub_ymm)>;
-
-def : Pat<(v8i32 (vselect (v8i1 VK8WM:$mask), (v8i32 VR256X:$src1),
-                            (v8i32 VR256X:$src2))),
-            (EXTRACT_SUBREG
-                (v16i32 (VPBLENDMDZrrk (COPY_TO_REGCLASS VK8WM:$mask, VK16WM),
-            (v16i32 (INSERT_SUBREG (IMPLICIT_DEF), VR256X:$src2, sub_ymm)),
-            (v16i32 (INSERT_SUBREG (IMPLICIT_DEF), VR256X:$src1, sub_ymm)))), sub_ymm)>;
-}
 //===----------------------------------------------------------------------===//
 // Compare Instructions
 //===----------------------------------------------------------------------===//
@@ -2978,6 +2963,30 @@ def : Pat<(v16i32 (vselect (xor VK16:$mask, (v16i1 immAllOnesV)),
                            (v16i32 VR512:$src))),
                   (VMOVDQA32Zrrkz VK16WM:$mask, VR512:$src)>;
 
+// Patterns for handling v8i1 selects of 256-bit vectors when VLX isn't
+// available. Use a 512-bit operation and extract.
+let Predicates = [HasAVX512, NoVLX] in {
+def : Pat<(v8f32 (vselect (v8i1 VK8WM:$mask), (v8f32 VR256X:$src1),
+                          (v8f32 VR256X:$src0))),
+          (EXTRACT_SUBREG
+           (v16f32
+            (VMOVAPSZrrk
+             (v16f32 (INSERT_SUBREG (IMPLICIT_DEF), VR256X:$src0, sub_ymm)),
+             (COPY_TO_REGCLASS VK8WM:$mask, VK16WM),
+             (v16f32 (INSERT_SUBREG (IMPLICIT_DEF), VR256X:$src1, sub_ymm)))),
+           sub_ymm)>;
+
+def : Pat<(v8i32 (vselect (v8i1 VK8WM:$mask), (v8i32 VR256X:$src1),
+                          (v8i32 VR256X:$src0))),
+          (EXTRACT_SUBREG
+           (v16i32
+            (VMOVDQA32Zrrk
+             (v16i32 (INSERT_SUBREG (IMPLICIT_DEF), VR256X:$src0, sub_ymm)),
+             (COPY_TO_REGCLASS VK8WM:$mask, VK16WM),
+             (v16i32 (INSERT_SUBREG (IMPLICIT_DEF), VR256X:$src1, sub_ymm)))),
+           sub_ymm)>;
+}
+
 let Predicates = [HasVLX, NoBWI] in {
   // 128-bit load/store without BWI.
   def : Pat<(alignedstore (v8i16 VR128X:$src), addr:$dst),
diff --git a/test/CodeGen/X86/avx512-ext.ll b/test/CodeGen/X86/avx512-ext.ll
index 358225b6fe0..03d6127ae5d 100644
--- a/test/CodeGen/X86/avx512-ext.ll
+++ b/test/CodeGen/X86/avx512-ext.ll
@@ -345,9 +345,9 @@ define <8 x i32> @zext_8x8mem_to_8x32(<8 x i8> *%i , <8 x i1> %mask) nounwind re
 ; KNL-NEXT:    vpmovsxwq %xmm0, %zmm0
 ; KNL-NEXT:    vpsllq $63, %zmm0, %zmm0
 ; KNL-NEXT:    vptestmq %zmm0, %zmm0, %k1
-; KNL-NEXT:    vpmovzxbd {{.*#+}} ymm0 = mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,zero,zero,mem[3],zero,zero,zero,mem[4],zero,zero,zero,mem[5],zero,zero,zero,mem[6],zero,zero,zero,mem[7],zero,zero,zero
-; KNL-NEXT:    vpxor %ymm1, %ymm1, %ymm1
-; KNL-NEXT:    vpblendmd %zmm0, %zmm1, %zmm0 {%k1}
+; KNL-NEXT:    vpmovzxbd {{.*#+}} ymm1 = mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,zero,zero,mem[3],zero,zero,zero,mem[4],zero,zero,zero,mem[5],zero,zero,zero,mem[6],zero,zero,zero,mem[7],zero,zero,zero
+; KNL-NEXT:    vpxor %ymm0, %ymm0, %ymm0
+; KNL-NEXT:    vmovdqa32 %zmm1, %zmm0 {%k1}
 ; KNL-NEXT:    ## kill: %YMM0<def> %YMM0<kill> %ZMM0<kill>
 ; KNL-NEXT:    retq
 ;
@@ -369,9 +369,9 @@ define <8 x i32> @sext_8x8mem_to_8x32(<8 x i8> *%i , <8 x i1> %mask) nounwind re
 ; KNL-NEXT:    vpmovsxwq %xmm0, %zmm0
 ; KNL-NEXT:    vpsllq $63, %zmm0, %zmm0
 ; KNL-NEXT:    vptestmq %zmm0, %zmm0, %k1
-; KNL-NEXT:    vpmovsxbd (%rdi), %ymm0
-; KNL-NEXT:    vpxor %ymm1, %ymm1, %ymm1
-; KNL-NEXT:    vpblendmd %zmm0, %zmm1, %zmm0 {%k1}
+; KNL-NEXT:    vpmovsxbd (%rdi), %ymm1
+; KNL-NEXT:    vpxor %ymm0, %ymm0, %ymm0
+; KNL-NEXT:    vmovdqa32 %zmm1, %zmm0 {%k1}
 ; KNL-NEXT:    ## kill: %YMM0<def> %YMM0<kill> %ZMM0<kill>
 ; KNL-NEXT:    retq
 ;
@@ -704,9 +704,9 @@ define <8 x i32> @zext_8x16mem_to_8x32(<8 x i16> *%i , <8 x i1> %mask) nounwind
 ; KNL-NEXT:    vpmovsxwq %xmm0, %zmm0
 ; KNL-NEXT:    vpsllq $63, %zmm0, %zmm0
 ; KNL-NEXT:    vptestmq %zmm0, %zmm0, %k1
-; KNL-NEXT:    vpmovzxwd {{.*#+}} ymm0 = mem[0],zero,mem[1],zero,mem[2],zero,mem[3],zero,mem[4],zero,mem[5],zero,mem[6],zero,mem[7],zero
-; KNL-NEXT:    vpxor %ymm1, %ymm1, %ymm1
-; KNL-NEXT:    vpblendmd %zmm0, %zmm1, %zmm0 {%k1}
+; KNL-NEXT:    vpmovzxwd {{.*#+}} ymm1 = mem[0],zero,mem[1],zero,mem[2],zero,mem[3],zero,mem[4],zero,mem[5],zero,mem[6],zero,mem[7],zero
+; KNL-NEXT:    vpxor %ymm0, %ymm0, %ymm0
+; KNL-NEXT:    vmovdqa32 %zmm1, %zmm0 {%k1}
 ; KNL-NEXT:    ## kill: %YMM0<def> %YMM0<kill> %ZMM0<kill>
 ; KNL-NEXT:    retq
 ;
@@ -728,9 +728,9 @@ define <8 x i32> @sext_8x16mem_to_8x32mask(<8 x i16> *%i , <8 x i1> %mask) nounw
 ; KNL-NEXT:    vpmovsxwq %xmm0, %zmm0
 ; KNL-NEXT:    vpsllq $63, %zmm0, %zmm0
 ; KNL-NEXT:    vptestmq %zmm0, %zmm0, %k1
-; KNL-NEXT:    vpmovsxwd (%rdi), %ymm0
-; KNL-NEXT:    vpxor %ymm1, %ymm1, %ymm1
-; KNL-NEXT:    vpblendmd %zmm0, %zmm1, %zmm0 {%k1}
+; KNL-NEXT:    vpmovsxwd (%rdi), %ymm1
+; KNL-NEXT:    vpxor %ymm0, %ymm0, %ymm0
+; KNL-NEXT:    vmovdqa32 %zmm1, %zmm0 {%k1}
 ; KNL-NEXT:    ## kill: %YMM0<def> %YMM0<kill> %ZMM0<kill>
 ; KNL-NEXT:    retq
 ;
@@ -762,9 +762,9 @@ define <8 x i32> @zext_8x16_to_8x32mask(<8 x i16> %a , <8 x i1> %mask) nounwind
 ; KNL-NEXT:    vpmovsxwq %xmm1, %zmm1
 ; KNL-NEXT:    vpsllq $63, %zmm1, %zmm1
 ; KNL-NEXT:    vptestmq %zmm1, %zmm1, %k1
-; KNL-NEXT:    vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
-; KNL-NEXT:    vpxor %ymm1, %ymm1, %ymm1
-; KNL-NEXT:    vpblendmd %zmm0, %zmm1, %zmm0 {%k1}
+; KNL-NEXT:    vpmovzxwd {{.*#+}} ymm1 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
+; KNL-NEXT:    vpxor %ymm0, %ymm0, %ymm0
+; KNL-NEXT:    vmovdqa32 %zmm1, %zmm0 {%k1}
 ; KNL-NEXT:    ## kill: %YMM0<def> %YMM0<kill> %ZMM0<kill>
 ; KNL-NEXT:    retq
 ;
diff --git a/test/CodeGen/X86/avx512-vec-cmp.ll b/test/CodeGen/X86/avx512-vec-cmp.ll
index 391d2a38853..361ee1ddbf9 100644
--- a/test/CodeGen/X86/avx512-vec-cmp.ll
+++ b/test/CodeGen/X86/avx512-vec-cmp.ll
@@ -122,8 +122,8 @@ define <8 x i32> @test9(<8 x i32> %x, <8 x i32> %y) nounwind {
 ; KNL-NEXT:    ## kill: %YMM1<def> %YMM1<kill> %ZMM1<def>
 ; KNL-NEXT:    ## kill: %YMM0<def> %YMM0<kill> %ZMM0<def>
 ; KNL-NEXT:    vpcmpeqd %zmm1, %zmm0, %k1
-; KNL-NEXT:    vpblendmd %zmm0, %zmm1, %zmm0 {%k1}
-; KNL-NEXT:    ## kill: %YMM0<def> %YMM0<kill> %ZMM0<kill>
+; KNL-NEXT:    vmovdqa32 %zmm0, %zmm1 {%k1}
+; KNL-NEXT:    vmovdqa %ymm1, %ymm0
 ; KNL-NEXT:    retq
 ;
 ; SKX-LABEL: test9:
@@ -143,8 +143,8 @@ define <8 x float> @test10(<8 x float> %x, <8 x float> %y) nounwind {
 ; KNL-NEXT:    ## kill: %YMM1<def> %YMM1<kill> %ZMM1<def>
 ; KNL-NEXT:    ## kill: %YMM0<def> %YMM0<kill> %ZMM0<def>
 ; KNL-NEXT:    vcmpeqps %zmm1, %zmm0, %k1
-; KNL-NEXT:    vblendmps %zmm0, %zmm1, %zmm0 {%k1}
-; KNL-NEXT:    ## kill: %YMM0<def> %YMM0<kill> %ZMM0<kill>
+; KNL-NEXT:    vmovaps %zmm0, %zmm1 {%k1}
+; KNL-NEXT:    vmovaps %ymm1, %ymm0
 ; KNL-NEXT:    retq
 ;
 ; SKX-LABEL: test10:
@@ -1020,8 +1020,8 @@ define <8 x float> @test35(<8 x float> %x, <8 x float> %x1, <8 x float>* %yp) no
 ; KNL-NEXT:    ## kill: %YMM0<def> %YMM0<kill> %ZMM0<def>
 ; KNL-NEXT:    vmovups (%rdi), %ymm2
 ; KNL-NEXT:    vcmpltps %zmm2, %zmm0, %k1
-; KNL-NEXT:    vblendmps %zmm0, %zmm1, %zmm0 {%k1}
-; KNL-NEXT:    ## kill: %YMM0<def> %YMM0<kill> %ZMM0<kill>
+; KNL-NEXT:    vmovaps %zmm0, %zmm1 {%k1}
+; KNL-NEXT:    vmovaps %ymm1, %ymm0
 ; KNL-NEXT:    retq
 ;
 ; SKX-LABEL: test35:
@@ -1140,8 +1140,8 @@ define <8  x float> @test41(<8  x float> %x, <8  x float> %x1, float* %ptr) noun
 ; KNL-NEXT:    ## kill: %YMM0<def> %YMM0<kill> %ZMM0<def>
 ; KNL-NEXT:    vbroadcastss (%rdi), %ymm2
 ; KNL-NEXT:    vcmpltps %zmm2, %zmm0, %k1
-; KNL-NEXT:    vblendmps %zmm0, %zmm1, %zmm0 {%k1}
-; KNL-NEXT:    ## kill: %YMM0<def> %YMM0<kill> %ZMM0<kill>
+; KNL-NEXT:    vmovaps %zmm0, %zmm1 {%k1}
+; KNL-NEXT:    vmovaps %ymm1, %ymm0
 ; KNL-NEXT:    retq
 ;
 ; SKX-LABEL: test41: