[AVX-512] Add isel patterns to turn native masked scalar add/sub/mul/div into masked...

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290564 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/X86/X86InstrAVX512.td b/lib/Target/X86/X86InstrAVX512.td
index bd6ae0cb731f32598bc2db653606fb25959b5e8c..da7437ea0ccb2d9110982ed58e79a910d175ac35 100644 (file)
--- a/lib/Target/X86/X86InstrAVX512.td
+++ b/lib/Target/X86/X86InstrAVX512.td
@@ -9117,6 +9117,17 @@ multiclass AVX512_scalar_math_f32_patterns<SDNode Op, string OpcPrefix> {
     def : Pat<(v4f32 (X86Blendi (v4f32 VR128X:$dst),
           (Op (v4f32 VR128X:$dst), (v4f32 VR128X:$src)), (i8 1))),
       (!cast<I>("V"#OpcPrefix#SSZrr_Int) v4f32:$dst, v4f32:$src)>;
+
+    // extracted masked scalar math op with insert via movss
+    def : Pat<(X86Movss (v4f32 VR128X:$src1),
+               (scalar_to_vector
+                (X86selects VK1WM:$mask,
+                            (Op (f32 (extractelt (v4f32 VR128X:$src1), (iPTR 0))),
+                                FR32X:$src2),
+                            FR32X:$src0))),
+      (!cast<I>("V"#OpcPrefix#SSZrr_Intk) (COPY_TO_REGCLASS FR32X:$src0, VR128X),
+          VK1WM:$mask, v4f32:$src1,
+          (COPY_TO_REGCLASS FR32X:$src2, VR128X))>;
   }
 }
 
@@ -9150,6 +9161,17 @@ multiclass AVX512_scalar_math_f64_patterns<SDNode Op, string OpcPrefix> {
     def : Pat<(v2f64 (X86Blendi (v2f64 VR128X:$dst),
           (Op (v2f64 VR128X:$dst), (v2f64 VR128X:$src)), (i8 1))),
       (!cast<I>("V"#OpcPrefix#SDZrr_Int) v2f64:$dst, v2f64:$src)>;
+
+    // extracted masked scalar math op with insert via movss
+    def : Pat<(X86Movsd (v2f64 VR128X:$src1),
+               (scalar_to_vector
+                (X86selects VK1WM:$mask,
+                            (Op (f64 (extractelt (v2f64 VR128X:$src1), (iPTR 0))),
+                                FR64X:$src2),
+                            FR64X:$src0))),
+      (!cast<I>("V"#OpcPrefix#SDZrr_Intk) (COPY_TO_REGCLASS FR64X:$src0, VR128X),
+          VK1WM:$mask, v2f64:$src1,
+          (COPY_TO_REGCLASS FR64X:$src2, VR128X))>;
   }
 }
 

diff --git a/test/CodeGen/X86/sse-scalar-fp-arith.ll b/test/CodeGen/X86/sse-scalar-fp-arith.ll
index 4b2af6fce8de7b99f5bfc6fcb05223b9ed442e22..f711dc615742c39669abb9822d894ca9aed0db7e 100644 (file)
--- a/test/CodeGen/X86/sse-scalar-fp-arith.ll
+++ b/test/CodeGen/X86/sse-scalar-fp-arith.ll
@@ -1119,9 +1119,9 @@ define <4 x float> @add_ss_mask(<4 x float> %a, <4 x float> %b, <4 x float> %c,
 ;
 ; AVX512-LABEL: add_ss_mask:
 ; AVX512:       # BB#0:
-; AVX512-NEXT:    vaddss %xmm1, %xmm0, %xmm1
+; AVX512-NEXT:    andl $1, %edi
 ; AVX512-NEXT:    kmovw %edi, %k1
-; AVX512-NEXT:    vmovss %xmm1, %xmm0, %xmm2 {%k1}
+; AVX512-NEXT:    vaddss %xmm1, %xmm0, %xmm2 {%k1}
 ; AVX512-NEXT:    vmovaps %xmm2, %xmm0
 ; AVX512-NEXT:    retq
   %1 = extractelement <4 x float> %a, i64 0
@@ -1174,9 +1174,9 @@ define <2 x double> @add_sd_mask(<2 x double> %a, <2 x double> %b, <2 x double>
 ;
 ; AVX512-LABEL: add_sd_mask:
 ; AVX512:       # BB#0:
-; AVX512-NEXT:    vaddsd %xmm1, %xmm0, %xmm1
+; AVX512-NEXT:    andl $1, %edi
 ; AVX512-NEXT:    kmovw %edi, %k1
-; AVX512-NEXT:    vmovsd %xmm1, %xmm0, %xmm2 {%k1}
+; AVX512-NEXT:    vaddsd %xmm1, %xmm0, %xmm2 {%k1}
 ; AVX512-NEXT:    vmovapd %xmm2, %xmm0
 ; AVX512-NEXT:    retq
   %1 = extractelement <2 x double> %a, i64 0