[X86] Add MMX nontemporal store pattern

As noticed on #98505 - try to reduce codegen diffs until we're ready to drop MMX entirely

Author: RKSimon

llvm/lib/Target/X86/X86InstrMMX.td

@@ -576,3 +576,13 @@ def : Pat<(x86mmx (MMX_X86movdq2q
                    (bc_v2i64 (v4i32 (X86cvttp2si (v2f64 VR128:$src)))))),
           (MMX_CVTTPD2PIrr VR128:$src)>;
 }
+
+let AddedComplexity = 400 in { // Prefer non-temporal version
+
+// Non-temporal store (no alignment required).
+let Predicates = [HasMMX] in {
+  def : Pat<(nontemporalstore (x86mmx VR64:$src), addr:$dst),
+            (MMX_MOVNTQmr addr:$dst, VR64:$src)>;
+}
+
+} // AddedComplexity

llvm/test/CodeGen/X86/nontemporal.ll

@@ -183,14 +183,14 @@ define void @test_mmx(ptr nocapture %a0, ptr nocapture %a1) {
 ; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
 ; X86-NEXT:    movq (%ecx), %mm0
 ; X86-NEXT:    psrlq $3, %mm0
-; X86-NEXT:    movq %mm0, (%eax)
+; X86-NEXT:    movntq %mm0, (%eax)
 ; X86-NEXT:    retl
 ;
 ; X64-LABEL: test_mmx:
 ; X64:       # %bb.0: # %entry
 ; X64-NEXT:    movq (%rdi), %mm0
 ; X64-NEXT:    psrlq $3, %mm0
-; X64-NEXT:    movq %mm0, (%rsi)
+; X64-NEXT:    movntq %mm0, (%rsi)
 ; X64-NEXT:    retq
 entry:
   %0 = load x86_mmx, ptr %a0