[X86] AVX512: Add non-temporal stores

Note that I followed the AVX2 convention here and didn't add LLVM intrinsics
for stores.  These can be generated with the nontemporal hint on LLVM IR
stores (see new test). The GCC builtins are lowered directly into nontemporal
stores.

<rdar://problem/17082571>

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@211176 91177308-0d34-0410-b5e6-96231b3b80d8

3 files changed, 60 insertions, 0 deletions

diff --git a/lib/Target/X86/X86InstrAVX512.td b/lib/Target/X86/X86InstrAVX512.td
index 618f42abe9..7cac5ebbec 100644
--- a/lib/Target/X86/X86InstrAVX512.td
+++ b/lib/Target/X86/X86InstrAVX512.td
@@ -1800,6 +1800,35 @@ def VMOVNTDQAZrm : AVX5128I<0x2A, MRMSrcMem, (outs VR512:$dst),
                               (int_x86_avx512_movntdqa addr:$src))]>,
                    EVEX, EVEX_V512, EVEX_CD8<64, CD8VF>;
 
+// Prefer non-temporal over temporal versions
+let AddedComplexity = 400, SchedRW = [WriteStore] in {
+
+def VMOVNTPSZmr : AVX512PSI<0x2B, MRMDestMem, (outs),
+                            (ins f512mem:$dst, VR512:$src),
+                            "vmovntps\t{$src, $dst|$dst, $src}",
+                            [(alignednontemporalstore (v16f32 VR512:$src),
+                                                      addr:$dst)],
+                            IIC_SSE_MOVNT>,
+                  EVEX, EVEX_V512, EVEX_CD8<32, CD8VF>;
+
+def VMOVNTPDZmr : AVX512PDI<0x2B, MRMDestMem, (outs),
+                            (ins f512mem:$dst, VR512:$src),
+                            "vmovntpd\t{$src, $dst|$dst, $src}",
+                            [(alignednontemporalstore (v8f64 VR512:$src),
+                                                      addr:$dst)],
+			    IIC_SSE_MOVNT>,
+                  EVEX, EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>;
+
+
+def VMOVNTDQZmr : AVX512BI<0xE7, MRMDestMem, (outs),
+                           (ins i512mem:$dst, VR512:$src),
+                           "vmovntdq\t{$src, $dst|$dst, $src}",
+                           [(alignednontemporalstore (v8i64 VR512:$src),
+                                                     addr:$dst)],
+                           IIC_SSE_MOVNT>,
+                  EVEX, EVEX_V512, EVEX_CD8<64, CD8VF>;
+}
+
 //===----------------------------------------------------------------------===//
 // AVX-512 - Integer arithmetic
 //
diff --git a/test/CodeGen/X86/avx512-nontemporal.ll b/test/CodeGen/X86/avx512-nontemporal.ll
new file mode 100644
index 0000000000..ef50cdb828
--- /dev/null
+++ b/test/CodeGen/X86/avx512-nontemporal.ll
@@ -0,0 +1,19 @@
+; RUN: llc < %s -march=x86-64 -mattr=+avx512f | FileCheck %s
+
+define void @f(<16 x float> %A, <16 x float> %AA, i8* %B, <8 x double> %C, <8 x double> %CC, i32 %D, <8 x i64> %E, <8 x i64> %EE) {
+; CHECK: vmovntps %z
+  %cast = bitcast i8* %B to <16 x float>*
+  %A2 = fadd <16 x float> %A, %AA
+  store <16 x float> %A2, <16 x float>* %cast, align 64, !nontemporal !0
+; CHECK: vmovntdq %z
+  %cast1 = bitcast i8* %B to <8 x i64>*
+  %E2 = add <8 x i64> %E, %EE
+  store <8 x i64> %E2, <8 x i64>* %cast1, align 64, !nontemporal !0
+; CHECK: vmovntpd %z
+  %cast2 = bitcast i8* %B to <8 x double>*
+  %C2 = fadd <8 x double> %C, %CC
+  store <8 x double> %C2, <8 x double>* %cast2, align 64, !nontemporal !0
+  ret void
+}
+
+!0 = metadata !{i32 1}
diff --git a/test/MC/X86/avx512-encodings.s b/test/MC/X86/avx512-encodings.s
index a7a87cd0b4..b9674231d9 100644
--- a/test/MC/X86/avx512-encodings.s
+++ b/test/MC/X86/avx512-encodings.s
@@ -3163,3 +3163,15 @@ vmovntdqa (%r14,%rdx,2), %zmm18
 // CHECK: vmovntdqa
 // CHECK: encoding: [0x62,0xc2,0x7d,0x48,0x2a,0x7c,0x14,0x02]
 vmovntdqa 128(%r12,%rdx), %zmm23
+
+// CHECK: vmovntdq
+// CHECK: encoding: [0x62,0x21,0x7d,0x48,0xe7,0x24,0xa9]
+vmovntdq %zmm28, (%rcx,%r13,4)
+
+// CHECK: vmovntpd
+// CHECK: encoding: [0x62,0xf1,0xfd,0x48,0x2b,0xb2,0x04,0x00,0x00,0x00]
+vmovntpd %zmm6, 4(%rdx)
+
+// CHECK: vmovntps
+// CHECK: encoding: [0x62,0x51,0x7c,0x48,0x2b,0x5c,0x8d,0x00]
+vmovntps %zmm11, (%r13,%rcx,4)