AVX-512: Fixed a bug, add not materializable instructions.

[igorb]

Summary

Bug https://llvm.org/bugs/show_bug.cgi?id=25270

Add AVX-512 not materializable instructions.
Otherwise value can be reused , despite its value could be changed - produces incorrect assembler.

example

define void @bar__512(<16 x i32>* %var) {

%var_load_load = load <16 x i32>, <16 x i32>* %var, align 1
store <16 x i32> <i32 4, i32 4, i32 4, i32 4, i32 4, i32 4, i32 4, i32 4, i32 4, i32 4, i32 4, i32 4, i32 4, i32 4, i32 4, i32 4>, <16 x i32>* %var, align 64  
call void @Print__512(<16 x i32> %var_load_load)
call void @Print__512(<16 x i32> %var_load_load)  
store <16 x i32> <i32 7, i32 7, i32 7, i32 7, i32 7, i32 7, i32 7, i32 7, i32 7, i32 7, i32 7, i32 7, i32 7, i32 7, i32 7, i32 7>, <16 x i32>* %var, align 64
ret void

}

Befor the fix
movq %rdi, %rbx
vmovdqu32 (%rbx), %zmm0
vpbroadcastd .LCPI0_0(%rip), %zmm1
vmovdqa32 %zmm1, (%rbx) -- change value
callq Print512
vmovdqu32 (%rbx), %zmm0 -- incorrect , value has been changed ,
callq Print512

After fix
movq %rdi, %rbx
vmovdqu32 (%rbx), %zmm0
vmovups %zmm0, (%rsp) # 64-byte Spill
vpbroadcastd .LCPI0_0(%rip), %zmm1
vmovdqa32 %zmm1, (%rbx)
callq Print512
vmovups (%rsp), %zmm0 # 64-byte Reload
callq Print512

[delena]

LGTM

llvm/trunk/lib/Target/X86/X86InstrInfo.cpp

X86InstrInfo::isReallyTriviallyReMaterializable(const MachineInstr *MI,
   case X86::VMOVAPDYrm:
   case X86::VMOVDQAYrm:
   case X86::VMOVDQUYrm:
   case X86::MMX_MOVD64rm:
   case X86::MMX_MOVQ64rm:
   case X86::FsVMOVAPSrm:
   case X86::FsVMOVAPDrm:
   case X86::FsMOVAPSrm:
-  case X86::FsMOVAPDrm: {
+  case X86::FsMOVAPDrm:
+  // AVX-512
+  case X86::VMOVAPDZ128rm:
+  case X86::VMOVAPDZ256rm:
+  case X86::VMOVAPDZrm:
+  case X86::VMOVAPSZ128rm:
+  case X86::VMOVAPSZ256rm:
+  case X86::VMOVAPSZrm:
+  case X86::VMOVDQA32Z128rm:
+  case X86::VMOVDQA32Z256rm:
+  case X86::VMOVDQA32Zrm:
+  case X86::VMOVDQA64Z128rm:
+  case X86::VMOVDQA64Z256rm:
+  case X86::VMOVDQA64Zrm:
+  case X86::VMOVDQU16Z128rm:
+  case X86::VMOVDQU16Z256rm:
+  case X86::VMOVDQU16Zrm:
+  case X86::VMOVDQU32Z128rm:
+  case X86::VMOVDQU32Z256rm:
+  case X86::VMOVDQU32Zrm:
+  case X86::VMOVDQU64Z128rm:
+  case X86::VMOVDQU64Z256rm:
+  case X86::VMOVDQU64Zrm:
+  case X86::VMOVDQU8Z128rm:
+  case X86::VMOVDQU8Z256rm:
+  case X86::VMOVDQU8Zrm:
+  case X86::VMOVUPSZ128rm:
+  case X86::VMOVUPSZ256rm:
+  case X86::VMOVUPSZrm: {
     // Loads from constant pools are trivially rematerializable.
     if (MI->getOperand(1+X86::AddrBaseReg).isReg() &&
         MI->getOperand(1+X86::AddrScaleAmt).isImm() &&
         MI->getOperand(1+X86::AddrIndexReg).isReg() &&
         MI->getOperand(1+X86::AddrIndexReg).getReg() == 0 &&
         MI->isInvariantLoad(AA)) {
       unsigned BaseReg = MI->getOperand(1+X86::AddrBaseReg).getReg();
       if (BaseReg == 0 || BaseReg == X86::RIP)

llvm/trunk/test/CodeGen/X86/avx512-bugfix-25270.ll

+; RUN: llc < %s -mtriple=x86_64-apple-darwin -mcpu=knl | FileCheck %s
+
+declare void @Print__512(<16 x i32>) #0
+
+define void @bar__512(<16 x i32>* %var) #0 {
+; CHECK-LABEL: bar__512:
+; CHECK:       ## BB#0: ## %allocas
+; CHECK-NEXT:    pushq %rbx
+; CHECK-NEXT:    subq $112, %rsp
+; CHECK-NEXT:    movq %rdi, %rbx
+; CHECK-NEXT:    vmovdqu32 (%rbx), %zmm0
+; CHECK-NEXT:    vmovups %zmm0, (%rsp) ## 64-byte Spill
+; CHECK-NEXT:    vpbroadcastd {{.*}}(%rip), %zmm1
+; CHECK-NEXT:    vmovdqa32 %zmm1, (%rbx)
+; CHECK-NEXT:    callq _Print__512
+; CHECK-NEXT:    vmovups (%rsp), %zmm0 ## 64-byte Reload
+; CHECK-NEXT:    callq _Print__512
+; CHECK-NEXT:    vpbroadcastd {{.*}}(%rip), %zmm0
+; CHECK-NEXT:    vmovdqa32 %zmm0, (%rbx)
+; CHECK-NEXT:    addq $112, %rsp
+; CHECK-NEXT:    popq %rbx
+; CHECK-NEXT:    retq
+allocas:
+  %var_load_load = load <16 x i32>, <16 x i32>* %var, align 1
+  store <16 x i32> <i32 4, i32 4, i32 4, i32 4, i32 4, i32 4, i32 4, i32 4, i32 4, i32 4, i32 4, i32 4, i32 4, i32 4, i32 4, i32 4>, <16 x i32>* %var, align 64
+  call void @Print__512(<16 x i32> %var_load_load)
+ ; %var_load_load value should be reloaded
+  call void @Print__512(<16 x i32> %var_load_load)
+  store <16 x i32> <i32 7, i32 7, i32 7, i32 7, i32 7, i32 7, i32 7, i32 7, i32 7, i32 7, i32 7, i32 7, i32 7, i32 7, i32 7, i32 7>, <16 x i32>* %var, align 64
+  ret void
+}
+
+
+attributes #0 = { nounwind }