191     // %vec.phi.row = phi <256 x i32> [ zeroinitializer, %entry ], [ %ResVec,  in createTileLoadStoreLoops()
195     PHINode *VecCPhiRowLoop = B.CreatePHI(V256I32Ty, 2, "vec.phi.row");  in createTileLoadStoreLoops()
199     // %vec.phi = phi <256 x i32> [ %vec.phi.row, %tileload.scalarize.rows.body  in createTileLoadStoreLoops()
202     PHINode *VecPhi = B.CreatePHI(V256I32Ty, 2, "vec.phi");  in createTileLoadStoreLoops()
209     // %ResVec = insertelement <256 x i32> %vec.phi, i32 %elt, i16 %idxvec  in createTileLoadStoreLoops()
219     Value *Vec = BitCast->getOperand(0);  in createTileLoadStoreLoops()  local
220     assert(isV256I32Ty(Vec->getType()) && "bitcast from non-v256i32 to x86amx");  in createTileLoadStoreLoops()
224     // %vec = extractelement <16 x i32> %vec, i16 %idx  in createTileLoadStoreLoops()
225     // store i32 %vec, i32* %ptr  in createTileLoadStoreLoops()
227     Value *Elt = B.CreateExtractElement(Vec, Idx);  in createTileLoadStoreLoops()
315   // %vec.c.phi.row = phi <256 x i32> [ %VecC, %continue ], [ %NewVecC,  in createTileDPLoops()
318   // %vec.d.phi.row = phi <256 x i32> [ zeroinitializer, %continue ], [  in createTileDPLoops()
321   PHINode *VecCPhiRowLoop = B.CreatePHI(V256I32Ty, 2, "vec.c.phi.row");  in createTileDPLoops()
324   PHINode *VecDPhiRowLoop = B.CreatePHI(V256I32Ty, 2, "vec.d.phi.row");  in createTileDPLoops()
328   // %vec.c.phi.col = phi <256 x i32> [ %vec.c.phi.row,  in createTileDPLoops()
332   // %vec.d.phi.col = phi <256 x i32> [  in createTileDPLoops()
333   // %vec.d.phi.row, %tiledpbssd.scalarize.rows.body ], [ %NewVecD,  in createTileDPLoops()
338   PHINode *VecCPhiColLoop = B.CreatePHI(V256I32Ty, 2, "vec.c.phi.col");  in createTileDPLoops()
340   PHINode *VecDPhiColLoop = B.CreatePHI(V256I32Ty, 2, "vec.d.phi.col");  in createTileDPLoops()
346   // %vec.c.inner.phi = phi <256 x i32> [ %vec.c.phi.col,  in createTileDPLoops()
351   PHINode *VecCPhi = B.CreatePHI(V256I32Ty, 2, "vec.c.inner.phi");  in createTileDPLoops()
364     // %eltc = extractelement <256 x i32> %vec.c.inner.phi, i16 %idxc  in createTileDPLoops()
374     // %NewVecC = insertelement <256 x i32> %vec.c.inner.phi, i32 %neweltc,  in createTileDPLoops()
411     // %eltc = extractelement <256 x i32> %vec.c.inner.phi, i16 %idxc  in createTileDPLoops()
427     // %NewVecC = insertelement <256 x i32> %vec.c.inner.phi, i32 %neweltc,  in createTileDPLoops()
429     // %NewVecD = insertelement <256 x i32> %vec.d.inner.phi, i32 %neweltc,  in createTileDPLoops()
452   // %NewEltC = extractelement <256 x i32> %vec.c.phi.col, i16 %idxc  in createTileDPLoops()
453   // %NewVecD = insertelement <256 x i32> %vec.d.phi.col, i32 %NewEltC,  in createTileDPLoops()
501     Value *Vec;  in lowerTileDP()  local
502     if (match(I, m_BitCast(m_Value(Vec)))) {  in lowerTileDP()
545       Value *Vec;  in lowerTileLoadStore()  local
546       if (match(I, m_BitCast(m_Value(Vec)))) {  in lowerTileLoadStore()
563     Value *Vec;  in lowerTileZero()  local
564     if (match(I, m_BitCast(m_Value(Vec)))) {  in lowerTileZero()