[X86][SSE] Add selective commutation support for insertps (PR40340)

[RKSimon]

Summary

When we are inserting 1 "inline" element, and zeroing 2 of the other elements then we can safely commute the insertps source inputs to improve memory folding.

[RKSimon]

ping?

[craig.topper]

LGTM

lib/Target/X86/X86InstrAVX512.td

defm : vinsert_for_mask_cast<"VINSERTI64x4Z", v16i16x_info, v32i16_info,
                              v8i64_info, vinsert256_insert,
                              INSERT_get_vinsert256_imm, [HasAVX512]>;
 defm : vinsert_for_mask_cast<"VINSERTI64x4Z", v32i8x_info, v64i8_info,
                              v8i64_info, vinsert256_insert,
                              INSERT_get_vinsert256_imm, [HasAVX512]>;
 
 // vinsertps - insert f32 to XMM
 let ExeDomain = SSEPackedSingle in {
+let isCommutable = 1 in
 def VINSERTPSZrr : AVX512AIi8<0x21, MRMSrcReg, (outs VR128X:$dst),
       (ins VR128X:$src1, VR128X:$src2, u8imm:$src3),
       "vinsertps\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
       [(set VR128X:$dst, (X86insertps VR128X:$src1, VR128X:$src2, imm:$src3))]>,
       EVEX_4V, Sched<[SchedWriteFShuffle.XMM]>;
 def VINSERTPSZrm: AVX512AIi8<0x21, MRMSrcMem, (outs VR128X:$dst),
       (ins VR128X:$src1, f32mem:$src2, u8imm:$src3),
       "vinsertps\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",

lib/Target/X86/X86InstrInfo.cpp

case X86::VPBLENDWYrri:{
     }
     // Only the least significant bits of Imm are used.
     unsigned Imm = MI.getOperand(3).getImm() & Mask;
     auto &WorkingMI = cloneIfNew(MI);
     WorkingMI.getOperand(3).setImm(Mask ^ Imm);
     return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
                                                    OpIdx1, OpIdx2);
   }
+  case X86::INSERTPSrr:
+  case X86::VINSERTPSrr:
+  case X86::VINSERTPSZrr: {
+    unsigned Imm = MI.getOperand(3).getImm();
+    unsigned ZMask = Imm & 15;
+    unsigned DstIdx = (Imm >> 4) & 3;
+    unsigned SrcIdx = (Imm >> 6) & 3;
+
+    // We can commute insertps if we zero 2 of the elements, the insertion is
+    // "inline" and we don't override the insertion with a zero.
+    if (DstIdx == SrcIdx && (ZMask & (1 << DstIdx)) == 0 &&
+        countPopulation(ZMask) == 2) {
+      unsigned AltIdx = findFirstSet((ZMask | (1 << DstIdx)) ^ 15);
+      assert(0 <= AltIdx && AltIdx < 4 && "Illegal insertion index");
+      unsigned AltImm = (AltIdx << 6) | (AltIdx << 4) | ZMask;
+      auto &WorkingMI = cloneIfNew(MI);
+      WorkingMI.getOperand(3).setImm(AltImm);
+      return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
+                                                     OpIdx1, OpIdx2);
+    }
+    return nullptr;
+  }
   case X86::MOVSDrr:
   case X86::MOVSSrr:
   case X86::VMOVSDrr:
   case X86::VMOVSSrr:{
     // On SSE41 or later we can commute a MOVSS/MOVSD to a BLENDPS/BLENDPD.
     assert(Subtarget.hasSSE41() && "Commuting MOVSD/MOVSS requires SSE41!");
 
     unsigned Mask, Opc;

lib/Target/X86/X86InstrSSE.td

 let Constraints = "$src1 = $dst" in
   defm PINSRQ : SS41I_insert64<0x22, "pinsrq">, REX_W;
 
 // insertps has a few different modes, there's the first two here below which
 // are optimized inserts that won't zero arbitrary elements in the destination
 // vector. The next one matches the intrinsic and could zero arbitrary elements
 // in the target vector.
 multiclass SS41I_insertf32<bits<8> opc, string asm, bit Is2Addr = 1> {
+  let isCommutable = 1 in
   def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
       (ins VR128:$src1, VR128:$src2, u8imm:$src3),
       !if(Is2Addr,
         !strconcat(asm, "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
         !strconcat(asm,
                    "\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}")),
       [(set VR128:$dst,
         (X86insertps VR128:$src1, VR128:$src2, imm:$src3))]>,

test/CodeGen/X86/insertps-combine.ll

; AVX-NEXT:    retq
   %ext = extractelement <4 x float> %res, i32 0
   ret float %ext
 }
 
 ; PR40340
 define <4 x float> @commute_load_insertps(<4 x float>, <4 x float>* nocapture readonly) {
 ; SSE-LABEL: commute_load_insertps:
 ; SSE:       # %bb.0:
-; SSE-NEXT:    movaps (%rdi), %xmm1
-; SSE-NEXT:    insertps {{.*#+}} xmm1 = zero,xmm0[1],zero,xmm1[3]
-; SSE-NEXT:    movaps %xmm1, %xmm0
+; SSE-NEXT:    insertps {{.*#+}} xmm0 = zero,xmm0[1],zero,mem[0]
 ; SSE-NEXT:    retq
 ;
 ; AVX-LABEL: commute_load_insertps:
 ; AVX:       # %bb.0:
-; AVX-NEXT:    vmovaps (%rdi), %xmm1
-; AVX-NEXT:    vinsertps {{.*#+}} xmm0 = zero,xmm0[1],zero,xmm1[3]
+; AVX-NEXT:    vinsertps {{.*#+}} xmm0 = zero,xmm0[1],zero,mem[0]
 ; AVX-NEXT:    retq
   %3 = load <4 x float>, <4 x float>* %1
   %4 = tail call <4 x float> @llvm.x86.sse41.insertps(<4 x float> %3, <4 x float> %0, i8 85)
   ret <4 x float> %4
 }
 
 declare <4 x float> @llvm.x86.sse41.insertps(<4 x float>, <4 x float>, i8) nounwind readnone