Index: llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp =================================================================== --- llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp +++ llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp @@ -2176,9 +2176,51 @@ LLT DstTy = MRI.getType(Dst); LLT SrcTy = MRI.getType(Src); - if (SrcTy.isVector() && !DstTy.isVector()) { + if (SrcTy.isVector()) { + LLT SrcEltTy = SrcTy.getElementType(); SmallVector SrcRegs; - getUnmergePieces(SrcRegs, MIRBuilder, Src, SrcTy.getElementType()); + + if (DstTy.isVector()) { + int NumDstElt = DstTy.getNumElements(); + int NumSrcElt = SrcTy.getNumElements(); + + LLT DstEltTy = DstTy.getElementType(); + LLT DstCastTy = DstEltTy; // Intermediate bitcast result type + LLT SrcPartTy = SrcEltTy; // Original unmerge result type. + + // If there's an element size mismatch, insert intermediate casts to match + // the result element type. + if (NumSrcElt < NumDstElt) { // Source element type is larger. + // %1:_(<4 x s8>) = G_BITCAST %0:_(<2 x s16>) + // + // => + // + // %2:_(s16), %3:_(s16) = G_UNMERGE_VALUES %0 + // %3:_(<2 x s8>) = G_BITCAST %2 + // %4:_(<2 x s8>) = G_BITCAST %3 + // %1:_(<4 x s16>) = G_CONCAT_VECTORS %3, %4 + DstCastTy = LLT::vector(NumDstElt / NumSrcElt, DstEltTy); + SrcPartTy = SrcEltTy; + } else if (NumSrcElt > NumDstElt) { // Source element type is smaller. + // + // %1:_(<2 x s16>) = G_BITCAST %0:_(<4 x s8>) + // + // => + // + // %2:_(<2 x s8>), %3:_(<2 x s8>) = G_UNMERGE_VALUES %0 + // %3:_(s16) = G_BITCAST %2 + // %4:_(s16) = G_BITCAST %3 + // %1:_(<2 x s16>) = G_BUILD_VECTOR %3, %4 + SrcPartTy = LLT::vector(NumSrcElt / NumDstElt, SrcEltTy); + DstCastTy = DstEltTy; + } + + getUnmergePieces(SrcRegs, MIRBuilder, Src, SrcPartTy); + for (Register &SrcReg : SrcRegs) + SrcReg = MIRBuilder.buildBitcast(DstCastTy, SrcReg).getReg(0); + } else + getUnmergePieces(SrcRegs, MIRBuilder, Src, SrcElt); + MIRBuilder.buildMerge(Dst, SrcRegs); MI.eraseFromParent(); return Legalized; Index: llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-bitcast.mir =================================================================== --- llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-bitcast.mir +++ llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-bitcast.mir @@ -481,3 +481,250 @@ %3:_(s32) = G_ANYEXT %2 $vgpr0 = COPY %3 ... + +--- +name: test_bitcast_v2s16_to_v4s8 +body: | + bb.0: + liveins: $vgpr0 + + ; CHECK-LABEL: name: test_bitcast_v2s16_to_v4s8 + ; CHECK: [[COPY:%[0-9]+]]:_(<2 x s16>) = COPY $vgpr0 + ; CHECK: [[BITCAST:%[0-9]+]]:_(s32) = G_BITCAST [[COPY]](<2 x s16>) + ; CHECK: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST]](s32) + ; CHECK: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 16 + ; CHECK: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST]], [[C]](s32) + ; CHECK: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR]](s32) + ; CHECK: [[C1:%[0-9]+]]:_(s16) = G_CONSTANT i16 8 + ; CHECK: [[LSHR1:%[0-9]+]]:_(s16) = G_LSHR [[TRUNC]], [[C1]](s16) + ; CHECK: [[LSHR2:%[0-9]+]]:_(s16) = G_LSHR [[TRUNC1]], [[C1]](s16) + ; CHECK: [[COPY1:%[0-9]+]]:_(s16) = COPY [[TRUNC]](s16) + ; CHECK: [[COPY2:%[0-9]+]]:_(s16) = COPY [[TRUNC]](s16) + ; CHECK: [[ADD:%[0-9]+]]:_(s16) = G_ADD [[COPY1]], [[COPY2]] + ; CHECK: [[COPY3:%[0-9]+]]:_(s16) = COPY [[LSHR1]](s16) + ; CHECK: [[COPY4:%[0-9]+]]:_(s16) = COPY [[LSHR1]](s16) + ; CHECK: [[ADD1:%[0-9]+]]:_(s16) = G_ADD [[COPY3]], [[COPY4]] + ; CHECK: [[COPY5:%[0-9]+]]:_(s16) = COPY [[TRUNC1]](s16) + ; CHECK: [[COPY6:%[0-9]+]]:_(s16) = COPY [[TRUNC1]](s16) + ; CHECK: [[ADD2:%[0-9]+]]:_(s16) = G_ADD [[COPY5]], [[COPY6]] + ; CHECK: [[COPY7:%[0-9]+]]:_(s16) = COPY [[LSHR2]](s16) + ; CHECK: [[COPY8:%[0-9]+]]:_(s16) = COPY [[LSHR2]](s16) + ; CHECK: [[ADD3:%[0-9]+]]:_(s16) = G_ADD [[COPY7]], [[COPY8]] + ; CHECK: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[ADD]](s16) + ; CHECK: [[ANYEXT1:%[0-9]+]]:_(s32) = G_ANYEXT [[ADD1]](s16) + ; CHECK: [[ANYEXT2:%[0-9]+]]:_(s32) = G_ANYEXT [[ADD2]](s16) + ; CHECK: [[ANYEXT3:%[0-9]+]]:_(s32) = G_ANYEXT [[ADD3]](s16) + ; CHECK: [[COPY9:%[0-9]+]]:_(s32) = COPY [[ANYEXT]](s32) + ; CHECK: [[COPY10:%[0-9]+]]:_(s32) = COPY [[ANYEXT1]](s32) + ; CHECK: [[COPY11:%[0-9]+]]:_(s32) = COPY [[ANYEXT2]](s32) + ; CHECK: [[COPY12:%[0-9]+]]:_(s32) = COPY [[ANYEXT3]](s32) + ; CHECK: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY9]](s32), [[COPY10]](s32), [[COPY11]](s32), [[COPY12]](s32) + ; CHECK: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BUILD_VECTOR]](<4 x s32>) + %0:_(<2 x s16>) = COPY $vgpr0 + %1:_(<4 x s8>) = G_BITCAST %0 + %2:_(<4 x s8>) = G_ADD %1, %1 + %3:_(<4 x s32>) = G_ANYEXT %2 + $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %3 +... + + +--- +name: test_bitcast_v4s8_to_v2s16 +body: | + bb.0: + liveins: $vgpr0_vgpr1_vgpr2_vgpr3 + + ; CHECK-LABEL: name: test_bitcast_v4s8_to_v2s16 + ; CHECK: [[COPY:%[0-9]+]]:_(<4 x s32>) = COPY $vgpr0_vgpr1_vgpr2_vgpr3 + ; CHECK: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32), [[UV2:%[0-9]+]]:_(s32), [[UV3:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[COPY]](<4 x s32>) + ; CHECK: [[UV4:%[0-9]+]]:_(s32), [[UV5:%[0-9]+]]:_(s32), [[UV6:%[0-9]+]]:_(s32), [[UV7:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[COPY]](<4 x s32>) + ; CHECK: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[UV]](s32) + ; CHECK: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[UV4]](s32) + ; CHECK: [[ADD:%[0-9]+]]:_(s16) = G_ADD [[TRUNC]], [[TRUNC1]] + ; CHECK: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[UV1]](s32) + ; CHECK: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[UV5]](s32) + ; CHECK: [[ADD1:%[0-9]+]]:_(s16) = G_ADD [[TRUNC2]], [[TRUNC3]] + ; CHECK: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[UV2]](s32) + ; CHECK: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[UV6]](s32) + ; CHECK: [[ADD2:%[0-9]+]]:_(s16) = G_ADD [[TRUNC4]], [[TRUNC5]] + ; CHECK: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[UV3]](s32) + ; CHECK: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[UV7]](s32) + ; CHECK: [[ADD3:%[0-9]+]]:_(s16) = G_ADD [[TRUNC6]], [[TRUNC7]] + ; CHECK: [[C:%[0-9]+]]:_(s16) = G_CONSTANT i16 255 + ; CHECK: [[COPY1:%[0-9]+]]:_(s16) = COPY [[ADD]](s16) + ; CHECK: [[AND:%[0-9]+]]:_(s16) = G_AND [[COPY1]], [[C]] + ; CHECK: [[COPY2:%[0-9]+]]:_(s16) = COPY [[ADD1]](s16) + ; CHECK: [[AND1:%[0-9]+]]:_(s16) = G_AND [[COPY2]], [[C]] + ; CHECK: [[C1:%[0-9]+]]:_(s16) = G_CONSTANT i16 8 + ; CHECK: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C1]](s16) + ; CHECK: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]] + ; CHECK: [[COPY3:%[0-9]+]]:_(s16) = COPY [[ADD2]](s16) + ; CHECK: [[AND2:%[0-9]+]]:_(s16) = G_AND [[COPY3]], [[C]] + ; CHECK: [[COPY4:%[0-9]+]]:_(s16) = COPY [[ADD3]](s16) + ; CHECK: [[AND3:%[0-9]+]]:_(s16) = G_AND [[COPY4]], [[C]] + ; CHECK: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[AND3]], [[C1]](s16) + ; CHECK: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[SHL1]] + ; CHECK: [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[OR]](s16) + ; CHECK: [[ZEXT1:%[0-9]+]]:_(s32) = G_ZEXT [[OR1]](s16) + ; CHECK: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 16 + ; CHECK: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[ZEXT1]], [[C2]](s32) + ; CHECK: [[OR2:%[0-9]+]]:_(s32) = G_OR [[ZEXT]], [[SHL2]] + ; CHECK: [[BITCAST:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR2]](s32) + ; CHECK: $vgpr0 = COPY [[BITCAST]](<2 x s16>) + %0:_(<4 x s32>) = COPY $vgpr0_vgpr1_vgpr2_vgpr3 + %1:_(<4 x s8>) = G_TRUNC %0 + %2:_(<4 x s8>) = G_ADD %1, %1 + %3:_(<2 x s16>) = G_BITCAST %2 + $vgpr0 = COPY %3 +... + +--- +name: test_bitcast_v2s16_to_v8s4 +body: | + bb.0: + liveins: $vgpr0 + + ; CHECK-LABEL: name: test_bitcast_v2s16_to_v8s4 + ; CHECK: [[COPY:%[0-9]+]]:_(<2 x s16>) = COPY $vgpr0 + ; CHECK: [[BITCAST:%[0-9]+]]:_(s32) = G_BITCAST [[COPY]](<2 x s16>) + ; CHECK: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST]](s32) + ; CHECK: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 16 + ; CHECK: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST]], [[C]](s32) + ; CHECK: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR]](s32) + ; CHECK: [[C1:%[0-9]+]]:_(s16) = G_CONSTANT i16 4 + ; CHECK: [[LSHR1:%[0-9]+]]:_(s16) = G_LSHR [[TRUNC]], [[C1]](s16) + ; CHECK: [[C2:%[0-9]+]]:_(s16) = G_CONSTANT i16 8 + ; CHECK: [[LSHR2:%[0-9]+]]:_(s16) = G_LSHR [[TRUNC]], [[C2]](s16) + ; CHECK: [[C3:%[0-9]+]]:_(s16) = G_CONSTANT i16 12 + ; CHECK: [[LSHR3:%[0-9]+]]:_(s16) = G_LSHR [[TRUNC]], [[C3]](s16) + ; CHECK: [[LSHR4:%[0-9]+]]:_(s16) = G_LSHR [[TRUNC1]], [[C1]](s16) + ; CHECK: [[LSHR5:%[0-9]+]]:_(s16) = G_LSHR [[TRUNC1]], [[C2]](s16) + ; CHECK: [[LSHR6:%[0-9]+]]:_(s16) = G_LSHR [[TRUNC1]], [[C3]](s16) + ; CHECK: [[COPY1:%[0-9]+]]:_(s16) = COPY [[TRUNC]](s16) + ; CHECK: [[COPY2:%[0-9]+]]:_(s16) = COPY [[TRUNC]](s16) + ; CHECK: [[ADD:%[0-9]+]]:_(s16) = G_ADD [[COPY1]], [[COPY2]] + ; CHECK: [[COPY3:%[0-9]+]]:_(s16) = COPY [[LSHR1]](s16) + ; CHECK: [[COPY4:%[0-9]+]]:_(s16) = COPY [[LSHR1]](s16) + ; CHECK: [[ADD1:%[0-9]+]]:_(s16) = G_ADD [[COPY3]], [[COPY4]] + ; CHECK: [[COPY5:%[0-9]+]]:_(s16) = COPY [[LSHR2]](s16) + ; CHECK: [[COPY6:%[0-9]+]]:_(s16) = COPY [[LSHR2]](s16) + ; CHECK: [[ADD2:%[0-9]+]]:_(s16) = G_ADD [[COPY5]], [[COPY6]] + ; CHECK: [[COPY7:%[0-9]+]]:_(s16) = COPY [[LSHR3]](s16) + ; CHECK: [[COPY8:%[0-9]+]]:_(s16) = COPY [[LSHR3]](s16) + ; CHECK: [[ADD3:%[0-9]+]]:_(s16) = G_ADD [[COPY7]], [[COPY8]] + ; CHECK: [[COPY9:%[0-9]+]]:_(s16) = COPY [[TRUNC1]](s16) + ; CHECK: [[COPY10:%[0-9]+]]:_(s16) = COPY [[TRUNC1]](s16) + ; CHECK: [[ADD4:%[0-9]+]]:_(s16) = G_ADD [[COPY9]], [[COPY10]] + ; CHECK: [[COPY11:%[0-9]+]]:_(s16) = COPY [[LSHR4]](s16) + ; CHECK: [[COPY12:%[0-9]+]]:_(s16) = COPY [[LSHR4]](s16) + ; CHECK: [[ADD5:%[0-9]+]]:_(s16) = G_ADD [[COPY11]], [[COPY12]] + ; CHECK: [[COPY13:%[0-9]+]]:_(s16) = COPY [[LSHR5]](s16) + ; CHECK: [[COPY14:%[0-9]+]]:_(s16) = COPY [[LSHR5]](s16) + ; CHECK: [[ADD6:%[0-9]+]]:_(s16) = G_ADD [[COPY13]], [[COPY14]] + ; CHECK: [[COPY15:%[0-9]+]]:_(s16) = COPY [[LSHR6]](s16) + ; CHECK: [[COPY16:%[0-9]+]]:_(s16) = COPY [[LSHR6]](s16) + ; CHECK: [[ADD7:%[0-9]+]]:_(s16) = G_ADD [[COPY15]], [[COPY16]] + ; CHECK: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[ADD]](s16) + ; CHECK: [[ANYEXT1:%[0-9]+]]:_(s32) = G_ANYEXT [[ADD1]](s16) + ; CHECK: [[ANYEXT2:%[0-9]+]]:_(s32) = G_ANYEXT [[ADD2]](s16) + ; CHECK: [[ANYEXT3:%[0-9]+]]:_(s32) = G_ANYEXT [[ADD3]](s16) + ; CHECK: [[ANYEXT4:%[0-9]+]]:_(s32) = G_ANYEXT [[ADD4]](s16) + ; CHECK: [[ANYEXT5:%[0-9]+]]:_(s32) = G_ANYEXT [[ADD5]](s16) + ; CHECK: [[ANYEXT6:%[0-9]+]]:_(s32) = G_ANYEXT [[ADD6]](s16) + ; CHECK: [[ANYEXT7:%[0-9]+]]:_(s32) = G_ANYEXT [[ADD7]](s16) + ; CHECK: [[COPY17:%[0-9]+]]:_(s32) = COPY [[ANYEXT]](s32) + ; CHECK: [[COPY18:%[0-9]+]]:_(s32) = COPY [[ANYEXT1]](s32) + ; CHECK: [[COPY19:%[0-9]+]]:_(s32) = COPY [[ANYEXT2]](s32) + ; CHECK: [[COPY20:%[0-9]+]]:_(s32) = COPY [[ANYEXT3]](s32) + ; CHECK: [[COPY21:%[0-9]+]]:_(s32) = COPY [[ANYEXT4]](s32) + ; CHECK: [[COPY22:%[0-9]+]]:_(s32) = COPY [[ANYEXT5]](s32) + ; CHECK: [[COPY23:%[0-9]+]]:_(s32) = COPY [[ANYEXT6]](s32) + ; CHECK: [[COPY24:%[0-9]+]]:_(s32) = COPY [[ANYEXT7]](s32) + ; CHECK: [[BUILD_VECTOR:%[0-9]+]]:_(<8 x s32>) = G_BUILD_VECTOR [[COPY17]](s32), [[COPY18]](s32), [[COPY19]](s32), [[COPY20]](s32), [[COPY21]](s32), [[COPY22]](s32), [[COPY23]](s32), [[COPY24]](s32) + ; CHECK: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[BUILD_VECTOR]](<8 x s32>) + %0:_(<2 x s16>) = COPY $vgpr0 + %1:_(<8 x s4>) = G_BITCAST %0 + %2:_(<8 x s4>) = G_ADD %1, %1 + %3:_(<8 x s32>) = G_ANYEXT %2 + $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY %3 +... + +--- +name: test_bitcast_v8s4_to_v2s16 +body: | + bb.0: + liveins: $vgpr0_vgpr1_vgpr2_vgpr3 + + ; CHECK-LABEL: name: test_bitcast_v8s4_to_v2s16 + ; CHECK: [[COPY:%[0-9]+]]:_(<8 x s32>) = COPY $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 + ; CHECK: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32), [[UV2:%[0-9]+]]:_(s32), [[UV3:%[0-9]+]]:_(s32), [[UV4:%[0-9]+]]:_(s32), [[UV5:%[0-9]+]]:_(s32), [[UV6:%[0-9]+]]:_(s32), [[UV7:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[COPY]](<8 x s32>) + ; CHECK: [[UV8:%[0-9]+]]:_(s32), [[UV9:%[0-9]+]]:_(s32), [[UV10:%[0-9]+]]:_(s32), [[UV11:%[0-9]+]]:_(s32), [[UV12:%[0-9]+]]:_(s32), [[UV13:%[0-9]+]]:_(s32), [[UV14:%[0-9]+]]:_(s32), [[UV15:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[COPY]](<8 x s32>) + ; CHECK: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[UV]](s32) + ; CHECK: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[UV8]](s32) + ; CHECK: [[ADD:%[0-9]+]]:_(s16) = G_ADD [[TRUNC]], [[TRUNC1]] + ; CHECK: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[UV1]](s32) + ; CHECK: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[UV9]](s32) + ; CHECK: [[ADD1:%[0-9]+]]:_(s16) = G_ADD [[TRUNC2]], [[TRUNC3]] + ; CHECK: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[UV2]](s32) + ; CHECK: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[UV10]](s32) + ; CHECK: [[ADD2:%[0-9]+]]:_(s16) = G_ADD [[TRUNC4]], [[TRUNC5]] + ; CHECK: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[UV3]](s32) + ; CHECK: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[UV11]](s32) + ; CHECK: [[ADD3:%[0-9]+]]:_(s16) = G_ADD [[TRUNC6]], [[TRUNC7]] + ; CHECK: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[UV4]](s32) + ; CHECK: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[UV12]](s32) + ; CHECK: [[ADD4:%[0-9]+]]:_(s16) = G_ADD [[TRUNC8]], [[TRUNC9]] + ; CHECK: [[TRUNC10:%[0-9]+]]:_(s16) = G_TRUNC [[UV5]](s32) + ; CHECK: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[UV13]](s32) + ; CHECK: [[ADD5:%[0-9]+]]:_(s16) = G_ADD [[TRUNC10]], [[TRUNC11]] + ; CHECK: [[TRUNC12:%[0-9]+]]:_(s16) = G_TRUNC [[UV6]](s32) + ; CHECK: [[TRUNC13:%[0-9]+]]:_(s16) = G_TRUNC [[UV14]](s32) + ; CHECK: [[ADD6:%[0-9]+]]:_(s16) = G_ADD [[TRUNC12]], [[TRUNC13]] + ; CHECK: [[TRUNC14:%[0-9]+]]:_(s16) = G_TRUNC [[UV7]](s32) + ; CHECK: [[TRUNC15:%[0-9]+]]:_(s16) = G_TRUNC [[UV15]](s32) + ; CHECK: [[ADD7:%[0-9]+]]:_(s16) = G_ADD [[TRUNC14]], [[TRUNC15]] + ; CHECK: [[C:%[0-9]+]]:_(s16) = G_CONSTANT i16 15 + ; CHECK: [[COPY1:%[0-9]+]]:_(s16) = COPY [[ADD]](s16) + ; CHECK: [[AND:%[0-9]+]]:_(s16) = G_AND [[COPY1]], [[C]] + ; CHECK: [[COPY2:%[0-9]+]]:_(s16) = COPY [[ADD1]](s16) + ; CHECK: [[AND1:%[0-9]+]]:_(s16) = G_AND [[COPY2]], [[C]] + ; CHECK: [[C1:%[0-9]+]]:_(s16) = G_CONSTANT i16 4 + ; CHECK: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C1]](s16) + ; CHECK: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]] + ; CHECK: [[COPY3:%[0-9]+]]:_(s16) = COPY [[ADD2]](s16) + ; CHECK: [[AND2:%[0-9]+]]:_(s16) = G_AND [[COPY3]], [[C]] + ; CHECK: [[C2:%[0-9]+]]:_(s16) = G_CONSTANT i16 8 + ; CHECK: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[AND2]], [[C2]](s16) + ; CHECK: [[OR1:%[0-9]+]]:_(s16) = G_OR [[OR]], [[SHL1]] + ; CHECK: [[COPY4:%[0-9]+]]:_(s16) = COPY [[ADD3]](s16) + ; CHECK: [[AND3:%[0-9]+]]:_(s16) = G_AND [[COPY4]], [[C]] + ; CHECK: [[C3:%[0-9]+]]:_(s16) = G_CONSTANT i16 12 + ; CHECK: [[SHL2:%[0-9]+]]:_(s16) = G_SHL [[AND3]], [[C3]](s16) + ; CHECK: [[OR2:%[0-9]+]]:_(s16) = G_OR [[OR1]], [[SHL2]] + ; CHECK: [[COPY5:%[0-9]+]]:_(s16) = COPY [[ADD4]](s16) + ; CHECK: [[AND4:%[0-9]+]]:_(s16) = G_AND [[COPY5]], [[C]] + ; CHECK: [[COPY6:%[0-9]+]]:_(s16) = COPY [[ADD5]](s16) + ; CHECK: [[AND5:%[0-9]+]]:_(s16) = G_AND [[COPY6]], [[C]] + ; CHECK: [[SHL3:%[0-9]+]]:_(s16) = G_SHL [[AND5]], [[C1]](s16) + ; CHECK: [[OR3:%[0-9]+]]:_(s16) = G_OR [[AND4]], [[SHL3]] + ; CHECK: [[COPY7:%[0-9]+]]:_(s16) = COPY [[ADD6]](s16) + ; CHECK: [[AND6:%[0-9]+]]:_(s16) = G_AND [[COPY7]], [[C]] + ; CHECK: [[SHL4:%[0-9]+]]:_(s16) = G_SHL [[AND6]], [[C2]](s16) + ; CHECK: [[OR4:%[0-9]+]]:_(s16) = G_OR [[OR3]], [[SHL4]] + ; CHECK: [[COPY8:%[0-9]+]]:_(s16) = COPY [[ADD7]](s16) + ; CHECK: [[AND7:%[0-9]+]]:_(s16) = G_AND [[COPY8]], [[C]] + ; CHECK: [[SHL5:%[0-9]+]]:_(s16) = G_SHL [[AND7]], [[C3]](s16) + ; CHECK: [[OR5:%[0-9]+]]:_(s16) = G_OR [[OR4]], [[SHL5]] + ; CHECK: [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[OR2]](s16) + ; CHECK: [[ZEXT1:%[0-9]+]]:_(s32) = G_ZEXT [[OR5]](s16) + ; CHECK: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 16 + ; CHECK: [[SHL6:%[0-9]+]]:_(s32) = G_SHL [[ZEXT1]], [[C4]](s32) + ; CHECK: [[OR6:%[0-9]+]]:_(s32) = G_OR [[ZEXT]], [[SHL6]] + ; CHECK: [[BITCAST:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR6]](s32) + ; CHECK: $vgpr0 = COPY [[BITCAST]](<2 x s16>) + %0:_(<8 x s32>) = COPY $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 + %1:_(<8 x s4>) = G_TRUNC %0 + %2:_(<8 x s4>) = G_ADD %1, %1 + %3:_(<2 x s16>) = G_BITCAST %2 + $vgpr0 = COPY %3 +... Index: llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-load-constant.mir =================================================================== --- llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-load-constant.mir +++ llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-load-constant.mir @@ -4941,28 +4941,146 @@ ; CI-LABEL: name: test_load_constant_v8s8_align8 ; CI: [[COPY:%[0-9]+]]:_(p4) = COPY $vgpr0_vgpr1 ; CI: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p4) :: (load 8, addrspace 4) - ; CI: [[BITCAST:%[0-9]+]]:_(<8 x s8>) = G_BITCAST [[LOAD]](<2 x s32>) - ; CI: $vgpr0_vgpr1 = COPY [[BITCAST]](<8 x s8>) + ; CI: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[LOAD]](<2 x s32>) + ; CI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8 + ; CI: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C]](s32) + ; CI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16 + ; CI: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C1]](s32) + ; CI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24 + ; CI: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C2]](s32) + ; CI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[UV]](s32) + ; CI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32) + ; CI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32) + ; CI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32) + ; CI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32) + ; CI: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>) + ; CI: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C]](s32) + ; CI: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C1]](s32) + ; CI: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C2]](s32) + ; CI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[UV1]](s32) + ; CI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32) + ; CI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32) + ; CI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32) + ; CI: [[BUILD_VECTOR1:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY5]](s32), [[COPY6]](s32), [[COPY7]](s32), [[COPY8]](s32) + ; CI: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR1]](<4 x s32>) + ; CI: [[CONCAT_VECTORS:%[0-9]+]]:_(<8 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>) + ; CI: $vgpr0_vgpr1 = COPY [[CONCAT_VECTORS]](<8 x s8>) ; VI-LABEL: name: test_load_constant_v8s8_align8 ; VI: [[COPY:%[0-9]+]]:_(p4) = COPY $vgpr0_vgpr1 ; VI: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p4) :: (load 8, addrspace 4) - ; VI: [[BITCAST:%[0-9]+]]:_(<8 x s8>) = G_BITCAST [[LOAD]](<2 x s32>) - ; VI: $vgpr0_vgpr1 = COPY [[BITCAST]](<8 x s8>) + ; VI: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[LOAD]](<2 x s32>) + ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8 + ; VI: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C]](s32) + ; VI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16 + ; VI: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C1]](s32) + ; VI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24 + ; VI: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C2]](s32) + ; VI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[UV]](s32) + ; VI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32) + ; VI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32) + ; VI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32) + ; VI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32) + ; VI: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>) + ; VI: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C]](s32) + ; VI: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C1]](s32) + ; VI: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C2]](s32) + ; VI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[UV1]](s32) + ; VI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32) + ; VI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32) + ; VI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32) + ; VI: [[BUILD_VECTOR1:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY5]](s32), [[COPY6]](s32), [[COPY7]](s32), [[COPY8]](s32) + ; VI: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR1]](<4 x s32>) + ; VI: [[CONCAT_VECTORS:%[0-9]+]]:_(<8 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>) + ; VI: $vgpr0_vgpr1 = COPY [[CONCAT_VECTORS]](<8 x s8>) ; GFX9-LABEL: name: test_load_constant_v8s8_align8 ; GFX9: [[COPY:%[0-9]+]]:_(p4) = COPY $vgpr0_vgpr1 ; GFX9: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p4) :: (load 8, addrspace 4) - ; GFX9: [[BITCAST:%[0-9]+]]:_(<8 x s8>) = G_BITCAST [[LOAD]](<2 x s32>) - ; GFX9: $vgpr0_vgpr1 = COPY [[BITCAST]](<8 x s8>) + ; GFX9: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[LOAD]](<2 x s32>) + ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8 + ; GFX9: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C]](s32) + ; GFX9: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16 + ; GFX9: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C1]](s32) + ; GFX9: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24 + ; GFX9: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C2]](s32) + ; GFX9: [[COPY1:%[0-9]+]]:_(s32) = COPY [[UV]](s32) + ; GFX9: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32) + ; GFX9: [[BUILD_VECTOR_TRUNC:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY1]](s32), [[COPY2]](s32) + ; GFX9: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32) + ; GFX9: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32) + ; GFX9: [[BUILD_VECTOR_TRUNC1:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY3]](s32), [[COPY4]](s32) + ; GFX9: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC]](<2 x s16>), [[BUILD_VECTOR_TRUNC1]](<2 x s16>) + ; GFX9: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS]](<4 x s16>) + ; GFX9: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C]](s32) + ; GFX9: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C1]](s32) + ; GFX9: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C2]](s32) + ; GFX9: [[COPY5:%[0-9]+]]:_(s32) = COPY [[UV1]](s32) + ; GFX9: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32) + ; GFX9: [[BUILD_VECTOR_TRUNC2:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY5]](s32), [[COPY6]](s32) + ; GFX9: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32) + ; GFX9: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32) + ; GFX9: [[BUILD_VECTOR_TRUNC3:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY7]](s32), [[COPY8]](s32) + ; GFX9: [[CONCAT_VECTORS1:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC2]](<2 x s16>), [[BUILD_VECTOR_TRUNC3]](<2 x s16>) + ; GFX9: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS1]](<4 x s16>) + ; GFX9: [[CONCAT_VECTORS2:%[0-9]+]]:_(<8 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>) + ; GFX9: $vgpr0_vgpr1 = COPY [[CONCAT_VECTORS2]](<8 x s8>) ; CI-MESA-LABEL: name: test_load_constant_v8s8_align8 ; CI-MESA: [[COPY:%[0-9]+]]:_(p4) = COPY $vgpr0_vgpr1 ; CI-MESA: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p4) :: (load 8, addrspace 4) - ; CI-MESA: [[BITCAST:%[0-9]+]]:_(<8 x s8>) = G_BITCAST [[LOAD]](<2 x s32>) - ; CI-MESA: $vgpr0_vgpr1 = COPY [[BITCAST]](<8 x s8>) + ; CI-MESA: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[LOAD]](<2 x s32>) + ; CI-MESA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8 + ; CI-MESA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C]](s32) + ; CI-MESA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16 + ; CI-MESA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C1]](s32) + ; CI-MESA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24 + ; CI-MESA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C2]](s32) + ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[UV]](s32) + ; CI-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32) + ; CI-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32) + ; CI-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32) + ; CI-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32) + ; CI-MESA: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>) + ; CI-MESA: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C]](s32) + ; CI-MESA: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C1]](s32) + ; CI-MESA: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C2]](s32) + ; CI-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[UV1]](s32) + ; CI-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32) + ; CI-MESA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32) + ; CI-MESA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32) + ; CI-MESA: [[BUILD_VECTOR1:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY5]](s32), [[COPY6]](s32), [[COPY7]](s32), [[COPY8]](s32) + ; CI-MESA: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR1]](<4 x s32>) + ; CI-MESA: [[CONCAT_VECTORS:%[0-9]+]]:_(<8 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>) + ; CI-MESA: $vgpr0_vgpr1 = COPY [[CONCAT_VECTORS]](<8 x s8>) ; GFX9-MESA-LABEL: name: test_load_constant_v8s8_align8 ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p4) = COPY $vgpr0_vgpr1 ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p4) :: (load 8, addrspace 4) - ; GFX9-MESA: [[BITCAST:%[0-9]+]]:_(<8 x s8>) = G_BITCAST [[LOAD]](<2 x s32>) - ; GFX9-MESA: $vgpr0_vgpr1 = COPY [[BITCAST]](<8 x s8>) + ; GFX9-MESA: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[LOAD]](<2 x s32>) + ; GFX9-MESA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8 + ; GFX9-MESA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C]](s32) + ; GFX9-MESA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16 + ; GFX9-MESA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C1]](s32) + ; GFX9-MESA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24 + ; GFX9-MESA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C2]](s32) + ; GFX9-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[UV]](s32) + ; GFX9-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY1]](s32), [[COPY2]](s32) + ; GFX9-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32) + ; GFX9-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC1:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY3]](s32), [[COPY4]](s32) + ; GFX9-MESA: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC]](<2 x s16>), [[BUILD_VECTOR_TRUNC1]](<2 x s16>) + ; GFX9-MESA: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS]](<4 x s16>) + ; GFX9-MESA: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C]](s32) + ; GFX9-MESA: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C1]](s32) + ; GFX9-MESA: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C2]](s32) + ; GFX9-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[UV1]](s32) + ; GFX9-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC2:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY5]](s32), [[COPY6]](s32) + ; GFX9-MESA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32) + ; GFX9-MESA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC3:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY7]](s32), [[COPY8]](s32) + ; GFX9-MESA: [[CONCAT_VECTORS1:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC2]](<2 x s16>), [[BUILD_VECTOR_TRUNC3]](<2 x s16>) + ; GFX9-MESA: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS1]](<4 x s16>) + ; GFX9-MESA: [[CONCAT_VECTORS2:%[0-9]+]]:_(<8 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>) + ; GFX9-MESA: $vgpr0_vgpr1 = COPY [[CONCAT_VECTORS2]](<8 x s8>) %0:_(p4) = COPY $vgpr0_vgpr1 %1:_(<8 x s8>) = G_LOAD %0 :: (load 8, align 8, addrspace 4) $vgpr0_vgpr1 = COPY %1 @@ -4977,28 +5095,244 @@ ; CI-LABEL: name: test_load_constant_v16s8_align16 ; CI: [[COPY:%[0-9]+]]:_(p4) = COPY $vgpr0_vgpr1 ; CI: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p4) :: (load 16, addrspace 4) - ; CI: [[BITCAST:%[0-9]+]]:_(<16 x s8>) = G_BITCAST [[LOAD]](<4 x s32>) - ; CI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](<16 x s8>) + ; CI: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32), [[UV2:%[0-9]+]]:_(s32), [[UV3:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[LOAD]](<4 x s32>) + ; CI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8 + ; CI: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C]](s32) + ; CI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16 + ; CI: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C1]](s32) + ; CI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24 + ; CI: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C2]](s32) + ; CI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[UV]](s32) + ; CI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32) + ; CI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32) + ; CI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32) + ; CI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32) + ; CI: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>) + ; CI: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C]](s32) + ; CI: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C1]](s32) + ; CI: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C2]](s32) + ; CI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[UV1]](s32) + ; CI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32) + ; CI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32) + ; CI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32) + ; CI: [[BUILD_VECTOR1:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY5]](s32), [[COPY6]](s32), [[COPY7]](s32), [[COPY8]](s32) + ; CI: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR1]](<4 x s32>) + ; CI: [[LSHR6:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C]](s32) + ; CI: [[LSHR7:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C1]](s32) + ; CI: [[LSHR8:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C2]](s32) + ; CI: [[COPY9:%[0-9]+]]:_(s32) = COPY [[UV2]](s32) + ; CI: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LSHR6]](s32) + ; CI: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LSHR7]](s32) + ; CI: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LSHR8]](s32) + ; CI: [[BUILD_VECTOR2:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY9]](s32), [[COPY10]](s32), [[COPY11]](s32), [[COPY12]](s32) + ; CI: [[TRUNC2:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR2]](<4 x s32>) + ; CI: [[LSHR9:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C]](s32) + ; CI: [[LSHR10:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C1]](s32) + ; CI: [[LSHR11:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C2]](s32) + ; CI: [[COPY13:%[0-9]+]]:_(s32) = COPY [[UV3]](s32) + ; CI: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LSHR9]](s32) + ; CI: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LSHR10]](s32) + ; CI: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LSHR11]](s32) + ; CI: [[BUILD_VECTOR3:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY13]](s32), [[COPY14]](s32), [[COPY15]](s32), [[COPY16]](s32) + ; CI: [[TRUNC3:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR3]](<4 x s32>) + ; CI: [[CONCAT_VECTORS:%[0-9]+]]:_(<16 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>), [[TRUNC2]](<4 x s8>), [[TRUNC3]](<4 x s8>) + ; CI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[CONCAT_VECTORS]](<16 x s8>) ; VI-LABEL: name: test_load_constant_v16s8_align16 ; VI: [[COPY:%[0-9]+]]:_(p4) = COPY $vgpr0_vgpr1 ; VI: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p4) :: (load 16, addrspace 4) - ; VI: [[BITCAST:%[0-9]+]]:_(<16 x s8>) = G_BITCAST [[LOAD]](<4 x s32>) - ; VI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](<16 x s8>) + ; VI: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32), [[UV2:%[0-9]+]]:_(s32), [[UV3:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[LOAD]](<4 x s32>) + ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8 + ; VI: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C]](s32) + ; VI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16 + ; VI: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C1]](s32) + ; VI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24 + ; VI: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C2]](s32) + ; VI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[UV]](s32) + ; VI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32) + ; VI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32) + ; VI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32) + ; VI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32) + ; VI: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>) + ; VI: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C]](s32) + ; VI: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C1]](s32) + ; VI: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C2]](s32) + ; VI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[UV1]](s32) + ; VI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32) + ; VI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32) + ; VI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32) + ; VI: [[BUILD_VECTOR1:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY5]](s32), [[COPY6]](s32), [[COPY7]](s32), [[COPY8]](s32) + ; VI: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR1]](<4 x s32>) + ; VI: [[LSHR6:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C]](s32) + ; VI: [[LSHR7:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C1]](s32) + ; VI: [[LSHR8:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C2]](s32) + ; VI: [[COPY9:%[0-9]+]]:_(s32) = COPY [[UV2]](s32) + ; VI: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LSHR6]](s32) + ; VI: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LSHR7]](s32) + ; VI: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LSHR8]](s32) + ; VI: [[BUILD_VECTOR2:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY9]](s32), [[COPY10]](s32), [[COPY11]](s32), [[COPY12]](s32) + ; VI: [[TRUNC2:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR2]](<4 x s32>) + ; VI: [[LSHR9:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C]](s32) + ; VI: [[LSHR10:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C1]](s32) + ; VI: [[LSHR11:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C2]](s32) + ; VI: [[COPY13:%[0-9]+]]:_(s32) = COPY [[UV3]](s32) + ; VI: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LSHR9]](s32) + ; VI: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LSHR10]](s32) + ; VI: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LSHR11]](s32) + ; VI: [[BUILD_VECTOR3:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY13]](s32), [[COPY14]](s32), [[COPY15]](s32), [[COPY16]](s32) + ; VI: [[TRUNC3:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR3]](<4 x s32>) + ; VI: [[CONCAT_VECTORS:%[0-9]+]]:_(<16 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>), [[TRUNC2]](<4 x s8>), [[TRUNC3]](<4 x s8>) + ; VI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[CONCAT_VECTORS]](<16 x s8>) ; GFX9-LABEL: name: test_load_constant_v16s8_align16 ; GFX9: [[COPY:%[0-9]+]]:_(p4) = COPY $vgpr0_vgpr1 ; GFX9: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p4) :: (load 16, addrspace 4) - ; GFX9: [[BITCAST:%[0-9]+]]:_(<16 x s8>) = G_BITCAST [[LOAD]](<4 x s32>) - ; GFX9: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](<16 x s8>) + ; GFX9: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32), [[UV2:%[0-9]+]]:_(s32), [[UV3:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[LOAD]](<4 x s32>) + ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8 + ; GFX9: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C]](s32) + ; GFX9: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16 + ; GFX9: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C1]](s32) + ; GFX9: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24 + ; GFX9: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C2]](s32) + ; GFX9: [[COPY1:%[0-9]+]]:_(s32) = COPY [[UV]](s32) + ; GFX9: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32) + ; GFX9: [[BUILD_VECTOR_TRUNC:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY1]](s32), [[COPY2]](s32) + ; GFX9: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32) + ; GFX9: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32) + ; GFX9: [[BUILD_VECTOR_TRUNC1:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY3]](s32), [[COPY4]](s32) + ; GFX9: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC]](<2 x s16>), [[BUILD_VECTOR_TRUNC1]](<2 x s16>) + ; GFX9: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS]](<4 x s16>) + ; GFX9: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C]](s32) + ; GFX9: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C1]](s32) + ; GFX9: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C2]](s32) + ; GFX9: [[COPY5:%[0-9]+]]:_(s32) = COPY [[UV1]](s32) + ; GFX9: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32) + ; GFX9: [[BUILD_VECTOR_TRUNC2:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY5]](s32), [[COPY6]](s32) + ; GFX9: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32) + ; GFX9: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32) + ; GFX9: [[BUILD_VECTOR_TRUNC3:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY7]](s32), [[COPY8]](s32) + ; GFX9: [[CONCAT_VECTORS1:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC2]](<2 x s16>), [[BUILD_VECTOR_TRUNC3]](<2 x s16>) + ; GFX9: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS1]](<4 x s16>) + ; GFX9: [[LSHR6:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C]](s32) + ; GFX9: [[LSHR7:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C1]](s32) + ; GFX9: [[LSHR8:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C2]](s32) + ; GFX9: [[COPY9:%[0-9]+]]:_(s32) = COPY [[UV2]](s32) + ; GFX9: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LSHR6]](s32) + ; GFX9: [[BUILD_VECTOR_TRUNC4:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY9]](s32), [[COPY10]](s32) + ; GFX9: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LSHR7]](s32) + ; GFX9: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LSHR8]](s32) + ; GFX9: [[BUILD_VECTOR_TRUNC5:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY11]](s32), [[COPY12]](s32) + ; GFX9: [[CONCAT_VECTORS2:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC4]](<2 x s16>), [[BUILD_VECTOR_TRUNC5]](<2 x s16>) + ; GFX9: [[TRUNC2:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS2]](<4 x s16>) + ; GFX9: [[LSHR9:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C]](s32) + ; GFX9: [[LSHR10:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C1]](s32) + ; GFX9: [[LSHR11:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C2]](s32) + ; GFX9: [[COPY13:%[0-9]+]]:_(s32) = COPY [[UV3]](s32) + ; GFX9: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LSHR9]](s32) + ; GFX9: [[BUILD_VECTOR_TRUNC6:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY13]](s32), [[COPY14]](s32) + ; GFX9: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LSHR10]](s32) + ; GFX9: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LSHR11]](s32) + ; GFX9: [[BUILD_VECTOR_TRUNC7:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY15]](s32), [[COPY16]](s32) + ; GFX9: [[CONCAT_VECTORS3:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC6]](<2 x s16>), [[BUILD_VECTOR_TRUNC7]](<2 x s16>) + ; GFX9: [[TRUNC3:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS3]](<4 x s16>) + ; GFX9: [[CONCAT_VECTORS4:%[0-9]+]]:_(<16 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>), [[TRUNC2]](<4 x s8>), [[TRUNC3]](<4 x s8>) + ; GFX9: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[CONCAT_VECTORS4]](<16 x s8>) ; CI-MESA-LABEL: name: test_load_constant_v16s8_align16 ; CI-MESA: [[COPY:%[0-9]+]]:_(p4) = COPY $vgpr0_vgpr1 ; CI-MESA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p4) :: (load 16, addrspace 4) - ; CI-MESA: [[BITCAST:%[0-9]+]]:_(<16 x s8>) = G_BITCAST [[LOAD]](<4 x s32>) - ; CI-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](<16 x s8>) + ; CI-MESA: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32), [[UV2:%[0-9]+]]:_(s32), [[UV3:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[LOAD]](<4 x s32>) + ; CI-MESA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8 + ; CI-MESA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C]](s32) + ; CI-MESA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16 + ; CI-MESA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C1]](s32) + ; CI-MESA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24 + ; CI-MESA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C2]](s32) + ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[UV]](s32) + ; CI-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32) + ; CI-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32) + ; CI-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32) + ; CI-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32) + ; CI-MESA: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>) + ; CI-MESA: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C]](s32) + ; CI-MESA: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C1]](s32) + ; CI-MESA: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C2]](s32) + ; CI-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[UV1]](s32) + ; CI-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32) + ; CI-MESA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32) + ; CI-MESA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32) + ; CI-MESA: [[BUILD_VECTOR1:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY5]](s32), [[COPY6]](s32), [[COPY7]](s32), [[COPY8]](s32) + ; CI-MESA: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR1]](<4 x s32>) + ; CI-MESA: [[LSHR6:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C]](s32) + ; CI-MESA: [[LSHR7:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C1]](s32) + ; CI-MESA: [[LSHR8:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C2]](s32) + ; CI-MESA: [[COPY9:%[0-9]+]]:_(s32) = COPY [[UV2]](s32) + ; CI-MESA: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LSHR6]](s32) + ; CI-MESA: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LSHR7]](s32) + ; CI-MESA: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LSHR8]](s32) + ; CI-MESA: [[BUILD_VECTOR2:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY9]](s32), [[COPY10]](s32), [[COPY11]](s32), [[COPY12]](s32) + ; CI-MESA: [[TRUNC2:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR2]](<4 x s32>) + ; CI-MESA: [[LSHR9:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C]](s32) + ; CI-MESA: [[LSHR10:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C1]](s32) + ; CI-MESA: [[LSHR11:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C2]](s32) + ; CI-MESA: [[COPY13:%[0-9]+]]:_(s32) = COPY [[UV3]](s32) + ; CI-MESA: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LSHR9]](s32) + ; CI-MESA: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LSHR10]](s32) + ; CI-MESA: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LSHR11]](s32) + ; CI-MESA: [[BUILD_VECTOR3:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY13]](s32), [[COPY14]](s32), [[COPY15]](s32), [[COPY16]](s32) + ; CI-MESA: [[TRUNC3:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR3]](<4 x s32>) + ; CI-MESA: [[CONCAT_VECTORS:%[0-9]+]]:_(<16 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>), [[TRUNC2]](<4 x s8>), [[TRUNC3]](<4 x s8>) + ; CI-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[CONCAT_VECTORS]](<16 x s8>) ; GFX9-MESA-LABEL: name: test_load_constant_v16s8_align16 ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p4) = COPY $vgpr0_vgpr1 ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p4) :: (load 16, addrspace 4) - ; GFX9-MESA: [[BITCAST:%[0-9]+]]:_(<16 x s8>) = G_BITCAST [[LOAD]](<4 x s32>) - ; GFX9-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](<16 x s8>) + ; GFX9-MESA: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32), [[UV2:%[0-9]+]]:_(s32), [[UV3:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[LOAD]](<4 x s32>) + ; GFX9-MESA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8 + ; GFX9-MESA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C]](s32) + ; GFX9-MESA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16 + ; GFX9-MESA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C1]](s32) + ; GFX9-MESA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24 + ; GFX9-MESA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C2]](s32) + ; GFX9-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[UV]](s32) + ; GFX9-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY1]](s32), [[COPY2]](s32) + ; GFX9-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32) + ; GFX9-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC1:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY3]](s32), [[COPY4]](s32) + ; GFX9-MESA: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC]](<2 x s16>), [[BUILD_VECTOR_TRUNC1]](<2 x s16>) + ; GFX9-MESA: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS]](<4 x s16>) + ; GFX9-MESA: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C]](s32) + ; GFX9-MESA: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C1]](s32) + ; GFX9-MESA: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C2]](s32) + ; GFX9-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[UV1]](s32) + ; GFX9-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC2:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY5]](s32), [[COPY6]](s32) + ; GFX9-MESA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32) + ; GFX9-MESA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC3:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY7]](s32), [[COPY8]](s32) + ; GFX9-MESA: [[CONCAT_VECTORS1:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC2]](<2 x s16>), [[BUILD_VECTOR_TRUNC3]](<2 x s16>) + ; GFX9-MESA: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS1]](<4 x s16>) + ; GFX9-MESA: [[LSHR6:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C]](s32) + ; GFX9-MESA: [[LSHR7:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C1]](s32) + ; GFX9-MESA: [[LSHR8:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C2]](s32) + ; GFX9-MESA: [[COPY9:%[0-9]+]]:_(s32) = COPY [[UV2]](s32) + ; GFX9-MESA: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LSHR6]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC4:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY9]](s32), [[COPY10]](s32) + ; GFX9-MESA: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LSHR7]](s32) + ; GFX9-MESA: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LSHR8]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC5:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY11]](s32), [[COPY12]](s32) + ; GFX9-MESA: [[CONCAT_VECTORS2:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC4]](<2 x s16>), [[BUILD_VECTOR_TRUNC5]](<2 x s16>) + ; GFX9-MESA: [[TRUNC2:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS2]](<4 x s16>) + ; GFX9-MESA: [[LSHR9:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C]](s32) + ; GFX9-MESA: [[LSHR10:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C1]](s32) + ; GFX9-MESA: [[LSHR11:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C2]](s32) + ; GFX9-MESA: [[COPY13:%[0-9]+]]:_(s32) = COPY [[UV3]](s32) + ; GFX9-MESA: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LSHR9]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC6:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY13]](s32), [[COPY14]](s32) + ; GFX9-MESA: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LSHR10]](s32) + ; GFX9-MESA: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LSHR11]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC7:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY15]](s32), [[COPY16]](s32) + ; GFX9-MESA: [[CONCAT_VECTORS3:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC6]](<2 x s16>), [[BUILD_VECTOR_TRUNC7]](<2 x s16>) + ; GFX9-MESA: [[TRUNC3:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS3]](<4 x s16>) + ; GFX9-MESA: [[CONCAT_VECTORS4:%[0-9]+]]:_(<16 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>), [[TRUNC2]](<4 x s8>), [[TRUNC3]](<4 x s8>) + ; GFX9-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[CONCAT_VECTORS4]](<16 x s8>) %0:_(p4) = COPY $vgpr0_vgpr1 %1:_(<16 x s8>) = G_LOAD %0 :: (load 16, align 16, addrspace 4) $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %1 @@ -5013,28 +5347,440 @@ ; CI-LABEL: name: test_load_constant_v32s8_align32 ; CI: [[COPY:%[0-9]+]]:_(p4) = COPY $vgpr0_vgpr1 ; CI: [[LOAD:%[0-9]+]]:_(<8 x s32>) = G_LOAD [[COPY]](p4) :: (load 32, addrspace 4) - ; CI: [[BITCAST:%[0-9]+]]:_(<32 x s8>) = G_BITCAST [[LOAD]](<8 x s32>) - ; CI: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[BITCAST]](<32 x s8>) + ; CI: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32), [[UV2:%[0-9]+]]:_(s32), [[UV3:%[0-9]+]]:_(s32), [[UV4:%[0-9]+]]:_(s32), [[UV5:%[0-9]+]]:_(s32), [[UV6:%[0-9]+]]:_(s32), [[UV7:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[LOAD]](<8 x s32>) + ; CI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8 + ; CI: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C]](s32) + ; CI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16 + ; CI: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C1]](s32) + ; CI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24 + ; CI: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C2]](s32) + ; CI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[UV]](s32) + ; CI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32) + ; CI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32) + ; CI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32) + ; CI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32) + ; CI: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>) + ; CI: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C]](s32) + ; CI: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C1]](s32) + ; CI: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C2]](s32) + ; CI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[UV1]](s32) + ; CI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32) + ; CI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32) + ; CI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32) + ; CI: [[BUILD_VECTOR1:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY5]](s32), [[COPY6]](s32), [[COPY7]](s32), [[COPY8]](s32) + ; CI: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR1]](<4 x s32>) + ; CI: [[LSHR6:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C]](s32) + ; CI: [[LSHR7:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C1]](s32) + ; CI: [[LSHR8:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C2]](s32) + ; CI: [[COPY9:%[0-9]+]]:_(s32) = COPY [[UV2]](s32) + ; CI: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LSHR6]](s32) + ; CI: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LSHR7]](s32) + ; CI: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LSHR8]](s32) + ; CI: [[BUILD_VECTOR2:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY9]](s32), [[COPY10]](s32), [[COPY11]](s32), [[COPY12]](s32) + ; CI: [[TRUNC2:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR2]](<4 x s32>) + ; CI: [[LSHR9:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C]](s32) + ; CI: [[LSHR10:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C1]](s32) + ; CI: [[LSHR11:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C2]](s32) + ; CI: [[COPY13:%[0-9]+]]:_(s32) = COPY [[UV3]](s32) + ; CI: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LSHR9]](s32) + ; CI: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LSHR10]](s32) + ; CI: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LSHR11]](s32) + ; CI: [[BUILD_VECTOR3:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY13]](s32), [[COPY14]](s32), [[COPY15]](s32), [[COPY16]](s32) + ; CI: [[TRUNC3:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR3]](<4 x s32>) + ; CI: [[LSHR12:%[0-9]+]]:_(s32) = G_LSHR [[UV4]], [[C]](s32) + ; CI: [[LSHR13:%[0-9]+]]:_(s32) = G_LSHR [[UV4]], [[C1]](s32) + ; CI: [[LSHR14:%[0-9]+]]:_(s32) = G_LSHR [[UV4]], [[C2]](s32) + ; CI: [[COPY17:%[0-9]+]]:_(s32) = COPY [[UV4]](s32) + ; CI: [[COPY18:%[0-9]+]]:_(s32) = COPY [[LSHR12]](s32) + ; CI: [[COPY19:%[0-9]+]]:_(s32) = COPY [[LSHR13]](s32) + ; CI: [[COPY20:%[0-9]+]]:_(s32) = COPY [[LSHR14]](s32) + ; CI: [[BUILD_VECTOR4:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY17]](s32), [[COPY18]](s32), [[COPY19]](s32), [[COPY20]](s32) + ; CI: [[TRUNC4:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR4]](<4 x s32>) + ; CI: [[LSHR15:%[0-9]+]]:_(s32) = G_LSHR [[UV5]], [[C]](s32) + ; CI: [[LSHR16:%[0-9]+]]:_(s32) = G_LSHR [[UV5]], [[C1]](s32) + ; CI: [[LSHR17:%[0-9]+]]:_(s32) = G_LSHR [[UV5]], [[C2]](s32) + ; CI: [[COPY21:%[0-9]+]]:_(s32) = COPY [[UV5]](s32) + ; CI: [[COPY22:%[0-9]+]]:_(s32) = COPY [[LSHR15]](s32) + ; CI: [[COPY23:%[0-9]+]]:_(s32) = COPY [[LSHR16]](s32) + ; CI: [[COPY24:%[0-9]+]]:_(s32) = COPY [[LSHR17]](s32) + ; CI: [[BUILD_VECTOR5:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY21]](s32), [[COPY22]](s32), [[COPY23]](s32), [[COPY24]](s32) + ; CI: [[TRUNC5:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR5]](<4 x s32>) + ; CI: [[LSHR18:%[0-9]+]]:_(s32) = G_LSHR [[UV6]], [[C]](s32) + ; CI: [[LSHR19:%[0-9]+]]:_(s32) = G_LSHR [[UV6]], [[C1]](s32) + ; CI: [[LSHR20:%[0-9]+]]:_(s32) = G_LSHR [[UV6]], [[C2]](s32) + ; CI: [[COPY25:%[0-9]+]]:_(s32) = COPY [[UV6]](s32) + ; CI: [[COPY26:%[0-9]+]]:_(s32) = COPY [[LSHR18]](s32) + ; CI: [[COPY27:%[0-9]+]]:_(s32) = COPY [[LSHR19]](s32) + ; CI: [[COPY28:%[0-9]+]]:_(s32) = COPY [[LSHR20]](s32) + ; CI: [[BUILD_VECTOR6:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY25]](s32), [[COPY26]](s32), [[COPY27]](s32), [[COPY28]](s32) + ; CI: [[TRUNC6:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR6]](<4 x s32>) + ; CI: [[LSHR21:%[0-9]+]]:_(s32) = G_LSHR [[UV7]], [[C]](s32) + ; CI: [[LSHR22:%[0-9]+]]:_(s32) = G_LSHR [[UV7]], [[C1]](s32) + ; CI: [[LSHR23:%[0-9]+]]:_(s32) = G_LSHR [[UV7]], [[C2]](s32) + ; CI: [[COPY29:%[0-9]+]]:_(s32) = COPY [[UV7]](s32) + ; CI: [[COPY30:%[0-9]+]]:_(s32) = COPY [[LSHR21]](s32) + ; CI: [[COPY31:%[0-9]+]]:_(s32) = COPY [[LSHR22]](s32) + ; CI: [[COPY32:%[0-9]+]]:_(s32) = COPY [[LSHR23]](s32) + ; CI: [[BUILD_VECTOR7:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY29]](s32), [[COPY30]](s32), [[COPY31]](s32), [[COPY32]](s32) + ; CI: [[TRUNC7:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR7]](<4 x s32>) + ; CI: [[CONCAT_VECTORS:%[0-9]+]]:_(<32 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>), [[TRUNC2]](<4 x s8>), [[TRUNC3]](<4 x s8>), [[TRUNC4]](<4 x s8>), [[TRUNC5]](<4 x s8>), [[TRUNC6]](<4 x s8>), [[TRUNC7]](<4 x s8>) + ; CI: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[CONCAT_VECTORS]](<32 x s8>) ; VI-LABEL: name: test_load_constant_v32s8_align32 ; VI: [[COPY:%[0-9]+]]:_(p4) = COPY $vgpr0_vgpr1 ; VI: [[LOAD:%[0-9]+]]:_(<8 x s32>) = G_LOAD [[COPY]](p4) :: (load 32, addrspace 4) - ; VI: [[BITCAST:%[0-9]+]]:_(<32 x s8>) = G_BITCAST [[LOAD]](<8 x s32>) - ; VI: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[BITCAST]](<32 x s8>) + ; VI: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32), [[UV2:%[0-9]+]]:_(s32), [[UV3:%[0-9]+]]:_(s32), [[UV4:%[0-9]+]]:_(s32), [[UV5:%[0-9]+]]:_(s32), [[UV6:%[0-9]+]]:_(s32), [[UV7:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[LOAD]](<8 x s32>) + ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8 + ; VI: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C]](s32) + ; VI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16 + ; VI: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C1]](s32) + ; VI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24 + ; VI: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C2]](s32) + ; VI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[UV]](s32) + ; VI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32) + ; VI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32) + ; VI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32) + ; VI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32) + ; VI: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>) + ; VI: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C]](s32) + ; VI: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C1]](s32) + ; VI: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C2]](s32) + ; VI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[UV1]](s32) + ; VI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32) + ; VI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32) + ; VI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32) + ; VI: [[BUILD_VECTOR1:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY5]](s32), [[COPY6]](s32), [[COPY7]](s32), [[COPY8]](s32) + ; VI: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR1]](<4 x s32>) + ; VI: [[LSHR6:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C]](s32) + ; VI: [[LSHR7:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C1]](s32) + ; VI: [[LSHR8:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C2]](s32) + ; VI: [[COPY9:%[0-9]+]]:_(s32) = COPY [[UV2]](s32) + ; VI: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LSHR6]](s32) + ; VI: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LSHR7]](s32) + ; VI: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LSHR8]](s32) + ; VI: [[BUILD_VECTOR2:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY9]](s32), [[COPY10]](s32), [[COPY11]](s32), [[COPY12]](s32) + ; VI: [[TRUNC2:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR2]](<4 x s32>) + ; VI: [[LSHR9:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C]](s32) + ; VI: [[LSHR10:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C1]](s32) + ; VI: [[LSHR11:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C2]](s32) + ; VI: [[COPY13:%[0-9]+]]:_(s32) = COPY [[UV3]](s32) + ; VI: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LSHR9]](s32) + ; VI: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LSHR10]](s32) + ; VI: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LSHR11]](s32) + ; VI: [[BUILD_VECTOR3:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY13]](s32), [[COPY14]](s32), [[COPY15]](s32), [[COPY16]](s32) + ; VI: [[TRUNC3:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR3]](<4 x s32>) + ; VI: [[LSHR12:%[0-9]+]]:_(s32) = G_LSHR [[UV4]], [[C]](s32) + ; VI: [[LSHR13:%[0-9]+]]:_(s32) = G_LSHR [[UV4]], [[C1]](s32) + ; VI: [[LSHR14:%[0-9]+]]:_(s32) = G_LSHR [[UV4]], [[C2]](s32) + ; VI: [[COPY17:%[0-9]+]]:_(s32) = COPY [[UV4]](s32) + ; VI: [[COPY18:%[0-9]+]]:_(s32) = COPY [[LSHR12]](s32) + ; VI: [[COPY19:%[0-9]+]]:_(s32) = COPY [[LSHR13]](s32) + ; VI: [[COPY20:%[0-9]+]]:_(s32) = COPY [[LSHR14]](s32) + ; VI: [[BUILD_VECTOR4:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY17]](s32), [[COPY18]](s32), [[COPY19]](s32), [[COPY20]](s32) + ; VI: [[TRUNC4:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR4]](<4 x s32>) + ; VI: [[LSHR15:%[0-9]+]]:_(s32) = G_LSHR [[UV5]], [[C]](s32) + ; VI: [[LSHR16:%[0-9]+]]:_(s32) = G_LSHR [[UV5]], [[C1]](s32) + ; VI: [[LSHR17:%[0-9]+]]:_(s32) = G_LSHR [[UV5]], [[C2]](s32) + ; VI: [[COPY21:%[0-9]+]]:_(s32) = COPY [[UV5]](s32) + ; VI: [[COPY22:%[0-9]+]]:_(s32) = COPY [[LSHR15]](s32) + ; VI: [[COPY23:%[0-9]+]]:_(s32) = COPY [[LSHR16]](s32) + ; VI: [[COPY24:%[0-9]+]]:_(s32) = COPY [[LSHR17]](s32) + ; VI: [[BUILD_VECTOR5:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY21]](s32), [[COPY22]](s32), [[COPY23]](s32), [[COPY24]](s32) + ; VI: [[TRUNC5:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR5]](<4 x s32>) + ; VI: [[LSHR18:%[0-9]+]]:_(s32) = G_LSHR [[UV6]], [[C]](s32) + ; VI: [[LSHR19:%[0-9]+]]:_(s32) = G_LSHR [[UV6]], [[C1]](s32) + ; VI: [[LSHR20:%[0-9]+]]:_(s32) = G_LSHR [[UV6]], [[C2]](s32) + ; VI: [[COPY25:%[0-9]+]]:_(s32) = COPY [[UV6]](s32) + ; VI: [[COPY26:%[0-9]+]]:_(s32) = COPY [[LSHR18]](s32) + ; VI: [[COPY27:%[0-9]+]]:_(s32) = COPY [[LSHR19]](s32) + ; VI: [[COPY28:%[0-9]+]]:_(s32) = COPY [[LSHR20]](s32) + ; VI: [[BUILD_VECTOR6:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY25]](s32), [[COPY26]](s32), [[COPY27]](s32), [[COPY28]](s32) + ; VI: [[TRUNC6:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR6]](<4 x s32>) + ; VI: [[LSHR21:%[0-9]+]]:_(s32) = G_LSHR [[UV7]], [[C]](s32) + ; VI: [[LSHR22:%[0-9]+]]:_(s32) = G_LSHR [[UV7]], [[C1]](s32) + ; VI: [[LSHR23:%[0-9]+]]:_(s32) = G_LSHR [[UV7]], [[C2]](s32) + ; VI: [[COPY29:%[0-9]+]]:_(s32) = COPY [[UV7]](s32) + ; VI: [[COPY30:%[0-9]+]]:_(s32) = COPY [[LSHR21]](s32) + ; VI: [[COPY31:%[0-9]+]]:_(s32) = COPY [[LSHR22]](s32) + ; VI: [[COPY32:%[0-9]+]]:_(s32) = COPY [[LSHR23]](s32) + ; VI: [[BUILD_VECTOR7:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY29]](s32), [[COPY30]](s32), [[COPY31]](s32), [[COPY32]](s32) + ; VI: [[TRUNC7:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR7]](<4 x s32>) + ; VI: [[CONCAT_VECTORS:%[0-9]+]]:_(<32 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>), [[TRUNC2]](<4 x s8>), [[TRUNC3]](<4 x s8>), [[TRUNC4]](<4 x s8>), [[TRUNC5]](<4 x s8>), [[TRUNC6]](<4 x s8>), [[TRUNC7]](<4 x s8>) + ; VI: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[CONCAT_VECTORS]](<32 x s8>) ; GFX9-LABEL: name: test_load_constant_v32s8_align32 ; GFX9: [[COPY:%[0-9]+]]:_(p4) = COPY $vgpr0_vgpr1 ; GFX9: [[LOAD:%[0-9]+]]:_(<8 x s32>) = G_LOAD [[COPY]](p4) :: (load 32, addrspace 4) - ; GFX9: [[BITCAST:%[0-9]+]]:_(<32 x s8>) = G_BITCAST [[LOAD]](<8 x s32>) - ; GFX9: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[BITCAST]](<32 x s8>) + ; GFX9: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32), [[UV2:%[0-9]+]]:_(s32), [[UV3:%[0-9]+]]:_(s32), [[UV4:%[0-9]+]]:_(s32), [[UV5:%[0-9]+]]:_(s32), [[UV6:%[0-9]+]]:_(s32), [[UV7:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[LOAD]](<8 x s32>) + ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8 + ; GFX9: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C]](s32) + ; GFX9: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16 + ; GFX9: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C1]](s32) + ; GFX9: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24 + ; GFX9: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C2]](s32) + ; GFX9: [[COPY1:%[0-9]+]]:_(s32) = COPY [[UV]](s32) + ; GFX9: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32) + ; GFX9: [[BUILD_VECTOR_TRUNC:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY1]](s32), [[COPY2]](s32) + ; GFX9: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32) + ; GFX9: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32) + ; GFX9: [[BUILD_VECTOR_TRUNC1:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY3]](s32), [[COPY4]](s32) + ; GFX9: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC]](<2 x s16>), [[BUILD_VECTOR_TRUNC1]](<2 x s16>) + ; GFX9: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS]](<4 x s16>) + ; GFX9: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C]](s32) + ; GFX9: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C1]](s32) + ; GFX9: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C2]](s32) + ; GFX9: [[COPY5:%[0-9]+]]:_(s32) = COPY [[UV1]](s32) + ; GFX9: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32) + ; GFX9: [[BUILD_VECTOR_TRUNC2:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY5]](s32), [[COPY6]](s32) + ; GFX9: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32) + ; GFX9: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32) + ; GFX9: [[BUILD_VECTOR_TRUNC3:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY7]](s32), [[COPY8]](s32) + ; GFX9: [[CONCAT_VECTORS1:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC2]](<2 x s16>), [[BUILD_VECTOR_TRUNC3]](<2 x s16>) + ; GFX9: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS1]](<4 x s16>) + ; GFX9: [[LSHR6:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C]](s32) + ; GFX9: [[LSHR7:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C1]](s32) + ; GFX9: [[LSHR8:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C2]](s32) + ; GFX9: [[COPY9:%[0-9]+]]:_(s32) = COPY [[UV2]](s32) + ; GFX9: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LSHR6]](s32) + ; GFX9: [[BUILD_VECTOR_TRUNC4:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY9]](s32), [[COPY10]](s32) + ; GFX9: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LSHR7]](s32) + ; GFX9: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LSHR8]](s32) + ; GFX9: [[BUILD_VECTOR_TRUNC5:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY11]](s32), [[COPY12]](s32) + ; GFX9: [[CONCAT_VECTORS2:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC4]](<2 x s16>), [[BUILD_VECTOR_TRUNC5]](<2 x s16>) + ; GFX9: [[TRUNC2:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS2]](<4 x s16>) + ; GFX9: [[LSHR9:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C]](s32) + ; GFX9: [[LSHR10:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C1]](s32) + ; GFX9: [[LSHR11:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C2]](s32) + ; GFX9: [[COPY13:%[0-9]+]]:_(s32) = COPY [[UV3]](s32) + ; GFX9: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LSHR9]](s32) + ; GFX9: [[BUILD_VECTOR_TRUNC6:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY13]](s32), [[COPY14]](s32) + ; GFX9: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LSHR10]](s32) + ; GFX9: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LSHR11]](s32) + ; GFX9: [[BUILD_VECTOR_TRUNC7:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY15]](s32), [[COPY16]](s32) + ; GFX9: [[CONCAT_VECTORS3:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC6]](<2 x s16>), [[BUILD_VECTOR_TRUNC7]](<2 x s16>) + ; GFX9: [[TRUNC3:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS3]](<4 x s16>) + ; GFX9: [[LSHR12:%[0-9]+]]:_(s32) = G_LSHR [[UV4]], [[C]](s32) + ; GFX9: [[LSHR13:%[0-9]+]]:_(s32) = G_LSHR [[UV4]], [[C1]](s32) + ; GFX9: [[LSHR14:%[0-9]+]]:_(s32) = G_LSHR [[UV4]], [[C2]](s32) + ; GFX9: [[COPY17:%[0-9]+]]:_(s32) = COPY [[UV4]](s32) + ; GFX9: [[COPY18:%[0-9]+]]:_(s32) = COPY [[LSHR12]](s32) + ; GFX9: [[BUILD_VECTOR_TRUNC8:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY17]](s32), [[COPY18]](s32) + ; GFX9: [[COPY19:%[0-9]+]]:_(s32) = COPY [[LSHR13]](s32) + ; GFX9: [[COPY20:%[0-9]+]]:_(s32) = COPY [[LSHR14]](s32) + ; GFX9: [[BUILD_VECTOR_TRUNC9:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY19]](s32), [[COPY20]](s32) + ; GFX9: [[CONCAT_VECTORS4:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC8]](<2 x s16>), [[BUILD_VECTOR_TRUNC9]](<2 x s16>) + ; GFX9: [[TRUNC4:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS4]](<4 x s16>) + ; GFX9: [[LSHR15:%[0-9]+]]:_(s32) = G_LSHR [[UV5]], [[C]](s32) + ; GFX9: [[LSHR16:%[0-9]+]]:_(s32) = G_LSHR [[UV5]], [[C1]](s32) + ; GFX9: [[LSHR17:%[0-9]+]]:_(s32) = G_LSHR [[UV5]], [[C2]](s32) + ; GFX9: [[COPY21:%[0-9]+]]:_(s32) = COPY [[UV5]](s32) + ; GFX9: [[COPY22:%[0-9]+]]:_(s32) = COPY [[LSHR15]](s32) + ; GFX9: [[BUILD_VECTOR_TRUNC10:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY21]](s32), [[COPY22]](s32) + ; GFX9: [[COPY23:%[0-9]+]]:_(s32) = COPY [[LSHR16]](s32) + ; GFX9: [[COPY24:%[0-9]+]]:_(s32) = COPY [[LSHR17]](s32) + ; GFX9: [[BUILD_VECTOR_TRUNC11:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY23]](s32), [[COPY24]](s32) + ; GFX9: [[CONCAT_VECTORS5:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC10]](<2 x s16>), [[BUILD_VECTOR_TRUNC11]](<2 x s16>) + ; GFX9: [[TRUNC5:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS5]](<4 x s16>) + ; GFX9: [[LSHR18:%[0-9]+]]:_(s32) = G_LSHR [[UV6]], [[C]](s32) + ; GFX9: [[LSHR19:%[0-9]+]]:_(s32) = G_LSHR [[UV6]], [[C1]](s32) + ; GFX9: [[LSHR20:%[0-9]+]]:_(s32) = G_LSHR [[UV6]], [[C2]](s32) + ; GFX9: [[COPY25:%[0-9]+]]:_(s32) = COPY [[UV6]](s32) + ; GFX9: [[COPY26:%[0-9]+]]:_(s32) = COPY [[LSHR18]](s32) + ; GFX9: [[BUILD_VECTOR_TRUNC12:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY25]](s32), [[COPY26]](s32) + ; GFX9: [[COPY27:%[0-9]+]]:_(s32) = COPY [[LSHR19]](s32) + ; GFX9: [[COPY28:%[0-9]+]]:_(s32) = COPY [[LSHR20]](s32) + ; GFX9: [[BUILD_VECTOR_TRUNC13:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY27]](s32), [[COPY28]](s32) + ; GFX9: [[CONCAT_VECTORS6:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC12]](<2 x s16>), [[BUILD_VECTOR_TRUNC13]](<2 x s16>) + ; GFX9: [[TRUNC6:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS6]](<4 x s16>) + ; GFX9: [[LSHR21:%[0-9]+]]:_(s32) = G_LSHR [[UV7]], [[C]](s32) + ; GFX9: [[LSHR22:%[0-9]+]]:_(s32) = G_LSHR [[UV7]], [[C1]](s32) + ; GFX9: [[LSHR23:%[0-9]+]]:_(s32) = G_LSHR [[UV7]], [[C2]](s32) + ; GFX9: [[COPY29:%[0-9]+]]:_(s32) = COPY [[UV7]](s32) + ; GFX9: [[COPY30:%[0-9]+]]:_(s32) = COPY [[LSHR21]](s32) + ; GFX9: [[BUILD_VECTOR_TRUNC14:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY29]](s32), [[COPY30]](s32) + ; GFX9: [[COPY31:%[0-9]+]]:_(s32) = COPY [[LSHR22]](s32) + ; GFX9: [[COPY32:%[0-9]+]]:_(s32) = COPY [[LSHR23]](s32) + ; GFX9: [[BUILD_VECTOR_TRUNC15:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY31]](s32), [[COPY32]](s32) + ; GFX9: [[CONCAT_VECTORS7:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC14]](<2 x s16>), [[BUILD_VECTOR_TRUNC15]](<2 x s16>) + ; GFX9: [[TRUNC7:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS7]](<4 x s16>) + ; GFX9: [[CONCAT_VECTORS8:%[0-9]+]]:_(<32 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>), [[TRUNC2]](<4 x s8>), [[TRUNC3]](<4 x s8>), [[TRUNC4]](<4 x s8>), [[TRUNC5]](<4 x s8>), [[TRUNC6]](<4 x s8>), [[TRUNC7]](<4 x s8>) + ; GFX9: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[CONCAT_VECTORS8]](<32 x s8>) ; CI-MESA-LABEL: name: test_load_constant_v32s8_align32 ; CI-MESA: [[COPY:%[0-9]+]]:_(p4) = COPY $vgpr0_vgpr1 ; CI-MESA: [[LOAD:%[0-9]+]]:_(<8 x s32>) = G_LOAD [[COPY]](p4) :: (load 32, addrspace 4) - ; CI-MESA: [[BITCAST:%[0-9]+]]:_(<32 x s8>) = G_BITCAST [[LOAD]](<8 x s32>) - ; CI-MESA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[BITCAST]](<32 x s8>) + ; CI-MESA: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32), [[UV2:%[0-9]+]]:_(s32), [[UV3:%[0-9]+]]:_(s32), [[UV4:%[0-9]+]]:_(s32), [[UV5:%[0-9]+]]:_(s32), [[UV6:%[0-9]+]]:_(s32), [[UV7:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[LOAD]](<8 x s32>) + ; CI-MESA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8 + ; CI-MESA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C]](s32) + ; CI-MESA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16 + ; CI-MESA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C1]](s32) + ; CI-MESA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24 + ; CI-MESA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C2]](s32) + ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[UV]](s32) + ; CI-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32) + ; CI-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32) + ; CI-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32) + ; CI-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32) + ; CI-MESA: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>) + ; CI-MESA: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C]](s32) + ; CI-MESA: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C1]](s32) + ; CI-MESA: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C2]](s32) + ; CI-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[UV1]](s32) + ; CI-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32) + ; CI-MESA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32) + ; CI-MESA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32) + ; CI-MESA: [[BUILD_VECTOR1:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY5]](s32), [[COPY6]](s32), [[COPY7]](s32), [[COPY8]](s32) + ; CI-MESA: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR1]](<4 x s32>) + ; CI-MESA: [[LSHR6:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C]](s32) + ; CI-MESA: [[LSHR7:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C1]](s32) + ; CI-MESA: [[LSHR8:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C2]](s32) + ; CI-MESA: [[COPY9:%[0-9]+]]:_(s32) = COPY [[UV2]](s32) + ; CI-MESA: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LSHR6]](s32) + ; CI-MESA: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LSHR7]](s32) + ; CI-MESA: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LSHR8]](s32) + ; CI-MESA: [[BUILD_VECTOR2:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY9]](s32), [[COPY10]](s32), [[COPY11]](s32), [[COPY12]](s32) + ; CI-MESA: [[TRUNC2:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR2]](<4 x s32>) + ; CI-MESA: [[LSHR9:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C]](s32) + ; CI-MESA: [[LSHR10:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C1]](s32) + ; CI-MESA: [[LSHR11:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C2]](s32) + ; CI-MESA: [[COPY13:%[0-9]+]]:_(s32) = COPY [[UV3]](s32) + ; CI-MESA: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LSHR9]](s32) + ; CI-MESA: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LSHR10]](s32) + ; CI-MESA: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LSHR11]](s32) + ; CI-MESA: [[BUILD_VECTOR3:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY13]](s32), [[COPY14]](s32), [[COPY15]](s32), [[COPY16]](s32) + ; CI-MESA: [[TRUNC3:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR3]](<4 x s32>) + ; CI-MESA: [[LSHR12:%[0-9]+]]:_(s32) = G_LSHR [[UV4]], [[C]](s32) + ; CI-MESA: [[LSHR13:%[0-9]+]]:_(s32) = G_LSHR [[UV4]], [[C1]](s32) + ; CI-MESA: [[LSHR14:%[0-9]+]]:_(s32) = G_LSHR [[UV4]], [[C2]](s32) + ; CI-MESA: [[COPY17:%[0-9]+]]:_(s32) = COPY [[UV4]](s32) + ; CI-MESA: [[COPY18:%[0-9]+]]:_(s32) = COPY [[LSHR12]](s32) + ; CI-MESA: [[COPY19:%[0-9]+]]:_(s32) = COPY [[LSHR13]](s32) + ; CI-MESA: [[COPY20:%[0-9]+]]:_(s32) = COPY [[LSHR14]](s32) + ; CI-MESA: [[BUILD_VECTOR4:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY17]](s32), [[COPY18]](s32), [[COPY19]](s32), [[COPY20]](s32) + ; CI-MESA: [[TRUNC4:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR4]](<4 x s32>) + ; CI-MESA: [[LSHR15:%[0-9]+]]:_(s32) = G_LSHR [[UV5]], [[C]](s32) + ; CI-MESA: [[LSHR16:%[0-9]+]]:_(s32) = G_LSHR [[UV5]], [[C1]](s32) + ; CI-MESA: [[LSHR17:%[0-9]+]]:_(s32) = G_LSHR [[UV5]], [[C2]](s32) + ; CI-MESA: [[COPY21:%[0-9]+]]:_(s32) = COPY [[UV5]](s32) + ; CI-MESA: [[COPY22:%[0-9]+]]:_(s32) = COPY [[LSHR15]](s32) + ; CI-MESA: [[COPY23:%[0-9]+]]:_(s32) = COPY [[LSHR16]](s32) + ; CI-MESA: [[COPY24:%[0-9]+]]:_(s32) = COPY [[LSHR17]](s32) + ; CI-MESA: [[BUILD_VECTOR5:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY21]](s32), [[COPY22]](s32), [[COPY23]](s32), [[COPY24]](s32) + ; CI-MESA: [[TRUNC5:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR5]](<4 x s32>) + ; CI-MESA: [[LSHR18:%[0-9]+]]:_(s32) = G_LSHR [[UV6]], [[C]](s32) + ; CI-MESA: [[LSHR19:%[0-9]+]]:_(s32) = G_LSHR [[UV6]], [[C1]](s32) + ; CI-MESA: [[LSHR20:%[0-9]+]]:_(s32) = G_LSHR [[UV6]], [[C2]](s32) + ; CI-MESA: [[COPY25:%[0-9]+]]:_(s32) = COPY [[UV6]](s32) + ; CI-MESA: [[COPY26:%[0-9]+]]:_(s32) = COPY [[LSHR18]](s32) + ; CI-MESA: [[COPY27:%[0-9]+]]:_(s32) = COPY [[LSHR19]](s32) + ; CI-MESA: [[COPY28:%[0-9]+]]:_(s32) = COPY [[LSHR20]](s32) + ; CI-MESA: [[BUILD_VECTOR6:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY25]](s32), [[COPY26]](s32), [[COPY27]](s32), [[COPY28]](s32) + ; CI-MESA: [[TRUNC6:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR6]](<4 x s32>) + ; CI-MESA: [[LSHR21:%[0-9]+]]:_(s32) = G_LSHR [[UV7]], [[C]](s32) + ; CI-MESA: [[LSHR22:%[0-9]+]]:_(s32) = G_LSHR [[UV7]], [[C1]](s32) + ; CI-MESA: [[LSHR23:%[0-9]+]]:_(s32) = G_LSHR [[UV7]], [[C2]](s32) + ; CI-MESA: [[COPY29:%[0-9]+]]:_(s32) = COPY [[UV7]](s32) + ; CI-MESA: [[COPY30:%[0-9]+]]:_(s32) = COPY [[LSHR21]](s32) + ; CI-MESA: [[COPY31:%[0-9]+]]:_(s32) = COPY [[LSHR22]](s32) + ; CI-MESA: [[COPY32:%[0-9]+]]:_(s32) = COPY [[LSHR23]](s32) + ; CI-MESA: [[BUILD_VECTOR7:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY29]](s32), [[COPY30]](s32), [[COPY31]](s32), [[COPY32]](s32) + ; CI-MESA: [[TRUNC7:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR7]](<4 x s32>) + ; CI-MESA: [[CONCAT_VECTORS:%[0-9]+]]:_(<32 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>), [[TRUNC2]](<4 x s8>), [[TRUNC3]](<4 x s8>), [[TRUNC4]](<4 x s8>), [[TRUNC5]](<4 x s8>), [[TRUNC6]](<4 x s8>), [[TRUNC7]](<4 x s8>) + ; CI-MESA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[CONCAT_VECTORS]](<32 x s8>) ; GFX9-MESA-LABEL: name: test_load_constant_v32s8_align32 ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p4) = COPY $vgpr0_vgpr1 ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<8 x s32>) = G_LOAD [[COPY]](p4) :: (load 32, addrspace 4) - ; GFX9-MESA: [[BITCAST:%[0-9]+]]:_(<32 x s8>) = G_BITCAST [[LOAD]](<8 x s32>) - ; GFX9-MESA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[BITCAST]](<32 x s8>) + ; GFX9-MESA: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32), [[UV2:%[0-9]+]]:_(s32), [[UV3:%[0-9]+]]:_(s32), [[UV4:%[0-9]+]]:_(s32), [[UV5:%[0-9]+]]:_(s32), [[UV6:%[0-9]+]]:_(s32), [[UV7:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[LOAD]](<8 x s32>) + ; GFX9-MESA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8 + ; GFX9-MESA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C]](s32) + ; GFX9-MESA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16 + ; GFX9-MESA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C1]](s32) + ; GFX9-MESA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24 + ; GFX9-MESA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C2]](s32) + ; GFX9-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[UV]](s32) + ; GFX9-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY1]](s32), [[COPY2]](s32) + ; GFX9-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32) + ; GFX9-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC1:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY3]](s32), [[COPY4]](s32) + ; GFX9-MESA: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC]](<2 x s16>), [[BUILD_VECTOR_TRUNC1]](<2 x s16>) + ; GFX9-MESA: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS]](<4 x s16>) + ; GFX9-MESA: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C]](s32) + ; GFX9-MESA: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C1]](s32) + ; GFX9-MESA: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C2]](s32) + ; GFX9-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[UV1]](s32) + ; GFX9-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC2:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY5]](s32), [[COPY6]](s32) + ; GFX9-MESA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32) + ; GFX9-MESA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC3:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY7]](s32), [[COPY8]](s32) + ; GFX9-MESA: [[CONCAT_VECTORS1:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC2]](<2 x s16>), [[BUILD_VECTOR_TRUNC3]](<2 x s16>) + ; GFX9-MESA: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS1]](<4 x s16>) + ; GFX9-MESA: [[LSHR6:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C]](s32) + ; GFX9-MESA: [[LSHR7:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C1]](s32) + ; GFX9-MESA: [[LSHR8:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C2]](s32) + ; GFX9-MESA: [[COPY9:%[0-9]+]]:_(s32) = COPY [[UV2]](s32) + ; GFX9-MESA: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LSHR6]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC4:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY9]](s32), [[COPY10]](s32) + ; GFX9-MESA: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LSHR7]](s32) + ; GFX9-MESA: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LSHR8]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC5:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY11]](s32), [[COPY12]](s32) + ; GFX9-MESA: [[CONCAT_VECTORS2:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC4]](<2 x s16>), [[BUILD_VECTOR_TRUNC5]](<2 x s16>) + ; GFX9-MESA: [[TRUNC2:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS2]](<4 x s16>) + ; GFX9-MESA: [[LSHR9:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C]](s32) + ; GFX9-MESA: [[LSHR10:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C1]](s32) + ; GFX9-MESA: [[LSHR11:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C2]](s32) + ; GFX9-MESA: [[COPY13:%[0-9]+]]:_(s32) = COPY [[UV3]](s32) + ; GFX9-MESA: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LSHR9]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC6:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY13]](s32), [[COPY14]](s32) + ; GFX9-MESA: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LSHR10]](s32) + ; GFX9-MESA: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LSHR11]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC7:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY15]](s32), [[COPY16]](s32) + ; GFX9-MESA: [[CONCAT_VECTORS3:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC6]](<2 x s16>), [[BUILD_VECTOR_TRUNC7]](<2 x s16>) + ; GFX9-MESA: [[TRUNC3:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS3]](<4 x s16>) + ; GFX9-MESA: [[LSHR12:%[0-9]+]]:_(s32) = G_LSHR [[UV4]], [[C]](s32) + ; GFX9-MESA: [[LSHR13:%[0-9]+]]:_(s32) = G_LSHR [[UV4]], [[C1]](s32) + ; GFX9-MESA: [[LSHR14:%[0-9]+]]:_(s32) = G_LSHR [[UV4]], [[C2]](s32) + ; GFX9-MESA: [[COPY17:%[0-9]+]]:_(s32) = COPY [[UV4]](s32) + ; GFX9-MESA: [[COPY18:%[0-9]+]]:_(s32) = COPY [[LSHR12]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC8:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY17]](s32), [[COPY18]](s32) + ; GFX9-MESA: [[COPY19:%[0-9]+]]:_(s32) = COPY [[LSHR13]](s32) + ; GFX9-MESA: [[COPY20:%[0-9]+]]:_(s32) = COPY [[LSHR14]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC9:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY19]](s32), [[COPY20]](s32) + ; GFX9-MESA: [[CONCAT_VECTORS4:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC8]](<2 x s16>), [[BUILD_VECTOR_TRUNC9]](<2 x s16>) + ; GFX9-MESA: [[TRUNC4:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS4]](<4 x s16>) + ; GFX9-MESA: [[LSHR15:%[0-9]+]]:_(s32) = G_LSHR [[UV5]], [[C]](s32) + ; GFX9-MESA: [[LSHR16:%[0-9]+]]:_(s32) = G_LSHR [[UV5]], [[C1]](s32) + ; GFX9-MESA: [[LSHR17:%[0-9]+]]:_(s32) = G_LSHR [[UV5]], [[C2]](s32) + ; GFX9-MESA: [[COPY21:%[0-9]+]]:_(s32) = COPY [[UV5]](s32) + ; GFX9-MESA: [[COPY22:%[0-9]+]]:_(s32) = COPY [[LSHR15]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC10:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY21]](s32), [[COPY22]](s32) + ; GFX9-MESA: [[COPY23:%[0-9]+]]:_(s32) = COPY [[LSHR16]](s32) + ; GFX9-MESA: [[COPY24:%[0-9]+]]:_(s32) = COPY [[LSHR17]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC11:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY23]](s32), [[COPY24]](s32) + ; GFX9-MESA: [[CONCAT_VECTORS5:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC10]](<2 x s16>), [[BUILD_VECTOR_TRUNC11]](<2 x s16>) + ; GFX9-MESA: [[TRUNC5:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS5]](<4 x s16>) + ; GFX9-MESA: [[LSHR18:%[0-9]+]]:_(s32) = G_LSHR [[UV6]], [[C]](s32) + ; GFX9-MESA: [[LSHR19:%[0-9]+]]:_(s32) = G_LSHR [[UV6]], [[C1]](s32) + ; GFX9-MESA: [[LSHR20:%[0-9]+]]:_(s32) = G_LSHR [[UV6]], [[C2]](s32) + ; GFX9-MESA: [[COPY25:%[0-9]+]]:_(s32) = COPY [[UV6]](s32) + ; GFX9-MESA: [[COPY26:%[0-9]+]]:_(s32) = COPY [[LSHR18]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC12:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY25]](s32), [[COPY26]](s32) + ; GFX9-MESA: [[COPY27:%[0-9]+]]:_(s32) = COPY [[LSHR19]](s32) + ; GFX9-MESA: [[COPY28:%[0-9]+]]:_(s32) = COPY [[LSHR20]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC13:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY27]](s32), [[COPY28]](s32) + ; GFX9-MESA: [[CONCAT_VECTORS6:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC12]](<2 x s16>), [[BUILD_VECTOR_TRUNC13]](<2 x s16>) + ; GFX9-MESA: [[TRUNC6:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS6]](<4 x s16>) + ; GFX9-MESA: [[LSHR21:%[0-9]+]]:_(s32) = G_LSHR [[UV7]], [[C]](s32) + ; GFX9-MESA: [[LSHR22:%[0-9]+]]:_(s32) = G_LSHR [[UV7]], [[C1]](s32) + ; GFX9-MESA: [[LSHR23:%[0-9]+]]:_(s32) = G_LSHR [[UV7]], [[C2]](s32) + ; GFX9-MESA: [[COPY29:%[0-9]+]]:_(s32) = COPY [[UV7]](s32) + ; GFX9-MESA: [[COPY30:%[0-9]+]]:_(s32) = COPY [[LSHR21]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC14:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY29]](s32), [[COPY30]](s32) + ; GFX9-MESA: [[COPY31:%[0-9]+]]:_(s32) = COPY [[LSHR22]](s32) + ; GFX9-MESA: [[COPY32:%[0-9]+]]:_(s32) = COPY [[LSHR23]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC15:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY31]](s32), [[COPY32]](s32) + ; GFX9-MESA: [[CONCAT_VECTORS7:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC14]](<2 x s16>), [[BUILD_VECTOR_TRUNC15]](<2 x s16>) + ; GFX9-MESA: [[TRUNC7:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS7]](<4 x s16>) + ; GFX9-MESA: [[CONCAT_VECTORS8:%[0-9]+]]:_(<32 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>), [[TRUNC2]](<4 x s8>), [[TRUNC3]](<4 x s8>), [[TRUNC4]](<4 x s8>), [[TRUNC5]](<4 x s8>), [[TRUNC6]](<4 x s8>), [[TRUNC7]](<4 x s8>) + ; GFX9-MESA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[CONCAT_VECTORS8]](<32 x s8>) %0:_(p4) = COPY $vgpr0_vgpr1 %1:_(<32 x s8>) = G_LOAD %0 :: (load 32, align 32, addrspace 4) $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY %1 Index: llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-load-flat.mir =================================================================== --- llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-load-flat.mir +++ llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-load-flat.mir @@ -4961,28 +4961,146 @@ ; CI-LABEL: name: test_load_flat_v8s8_align8 ; CI: [[COPY:%[0-9]+]]:_(p0) = COPY $vgpr0_vgpr1 ; CI: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p0) :: (load 8) - ; CI: [[BITCAST:%[0-9]+]]:_(<8 x s8>) = G_BITCAST [[LOAD]](<2 x s32>) - ; CI: $vgpr0_vgpr1 = COPY [[BITCAST]](<8 x s8>) + ; CI: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[LOAD]](<2 x s32>) + ; CI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8 + ; CI: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C]](s32) + ; CI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16 + ; CI: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C1]](s32) + ; CI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24 + ; CI: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C2]](s32) + ; CI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[UV]](s32) + ; CI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32) + ; CI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32) + ; CI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32) + ; CI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32) + ; CI: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>) + ; CI: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C]](s32) + ; CI: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C1]](s32) + ; CI: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C2]](s32) + ; CI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[UV1]](s32) + ; CI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32) + ; CI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32) + ; CI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32) + ; CI: [[BUILD_VECTOR1:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY5]](s32), [[COPY6]](s32), [[COPY7]](s32), [[COPY8]](s32) + ; CI: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR1]](<4 x s32>) + ; CI: [[CONCAT_VECTORS:%[0-9]+]]:_(<8 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>) + ; CI: $vgpr0_vgpr1 = COPY [[CONCAT_VECTORS]](<8 x s8>) ; VI-LABEL: name: test_load_flat_v8s8_align8 ; VI: [[COPY:%[0-9]+]]:_(p0) = COPY $vgpr0_vgpr1 ; VI: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p0) :: (load 8) - ; VI: [[BITCAST:%[0-9]+]]:_(<8 x s8>) = G_BITCAST [[LOAD]](<2 x s32>) - ; VI: $vgpr0_vgpr1 = COPY [[BITCAST]](<8 x s8>) + ; VI: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[LOAD]](<2 x s32>) + ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8 + ; VI: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C]](s32) + ; VI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16 + ; VI: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C1]](s32) + ; VI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24 + ; VI: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C2]](s32) + ; VI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[UV]](s32) + ; VI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32) + ; VI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32) + ; VI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32) + ; VI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32) + ; VI: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>) + ; VI: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C]](s32) + ; VI: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C1]](s32) + ; VI: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C2]](s32) + ; VI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[UV1]](s32) + ; VI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32) + ; VI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32) + ; VI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32) + ; VI: [[BUILD_VECTOR1:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY5]](s32), [[COPY6]](s32), [[COPY7]](s32), [[COPY8]](s32) + ; VI: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR1]](<4 x s32>) + ; VI: [[CONCAT_VECTORS:%[0-9]+]]:_(<8 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>) + ; VI: $vgpr0_vgpr1 = COPY [[CONCAT_VECTORS]](<8 x s8>) ; GFX9-LABEL: name: test_load_flat_v8s8_align8 ; GFX9: [[COPY:%[0-9]+]]:_(p0) = COPY $vgpr0_vgpr1 ; GFX9: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p0) :: (load 8) - ; GFX9: [[BITCAST:%[0-9]+]]:_(<8 x s8>) = G_BITCAST [[LOAD]](<2 x s32>) - ; GFX9: $vgpr0_vgpr1 = COPY [[BITCAST]](<8 x s8>) + ; GFX9: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[LOAD]](<2 x s32>) + ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8 + ; GFX9: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C]](s32) + ; GFX9: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16 + ; GFX9: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C1]](s32) + ; GFX9: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24 + ; GFX9: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C2]](s32) + ; GFX9: [[COPY1:%[0-9]+]]:_(s32) = COPY [[UV]](s32) + ; GFX9: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32) + ; GFX9: [[BUILD_VECTOR_TRUNC:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY1]](s32), [[COPY2]](s32) + ; GFX9: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32) + ; GFX9: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32) + ; GFX9: [[BUILD_VECTOR_TRUNC1:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY3]](s32), [[COPY4]](s32) + ; GFX9: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC]](<2 x s16>), [[BUILD_VECTOR_TRUNC1]](<2 x s16>) + ; GFX9: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS]](<4 x s16>) + ; GFX9: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C]](s32) + ; GFX9: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C1]](s32) + ; GFX9: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C2]](s32) + ; GFX9: [[COPY5:%[0-9]+]]:_(s32) = COPY [[UV1]](s32) + ; GFX9: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32) + ; GFX9: [[BUILD_VECTOR_TRUNC2:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY5]](s32), [[COPY6]](s32) + ; GFX9: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32) + ; GFX9: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32) + ; GFX9: [[BUILD_VECTOR_TRUNC3:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY7]](s32), [[COPY8]](s32) + ; GFX9: [[CONCAT_VECTORS1:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC2]](<2 x s16>), [[BUILD_VECTOR_TRUNC3]](<2 x s16>) + ; GFX9: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS1]](<4 x s16>) + ; GFX9: [[CONCAT_VECTORS2:%[0-9]+]]:_(<8 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>) + ; GFX9: $vgpr0_vgpr1 = COPY [[CONCAT_VECTORS2]](<8 x s8>) ; CI-MESA-LABEL: name: test_load_flat_v8s8_align8 ; CI-MESA: [[COPY:%[0-9]+]]:_(p0) = COPY $vgpr0_vgpr1 ; CI-MESA: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p0) :: (load 8) - ; CI-MESA: [[BITCAST:%[0-9]+]]:_(<8 x s8>) = G_BITCAST [[LOAD]](<2 x s32>) - ; CI-MESA: $vgpr0_vgpr1 = COPY [[BITCAST]](<8 x s8>) + ; CI-MESA: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[LOAD]](<2 x s32>) + ; CI-MESA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8 + ; CI-MESA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C]](s32) + ; CI-MESA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16 + ; CI-MESA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C1]](s32) + ; CI-MESA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24 + ; CI-MESA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C2]](s32) + ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[UV]](s32) + ; CI-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32) + ; CI-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32) + ; CI-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32) + ; CI-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32) + ; CI-MESA: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>) + ; CI-MESA: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C]](s32) + ; CI-MESA: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C1]](s32) + ; CI-MESA: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C2]](s32) + ; CI-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[UV1]](s32) + ; CI-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32) + ; CI-MESA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32) + ; CI-MESA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32) + ; CI-MESA: [[BUILD_VECTOR1:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY5]](s32), [[COPY6]](s32), [[COPY7]](s32), [[COPY8]](s32) + ; CI-MESA: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR1]](<4 x s32>) + ; CI-MESA: [[CONCAT_VECTORS:%[0-9]+]]:_(<8 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>) + ; CI-MESA: $vgpr0_vgpr1 = COPY [[CONCAT_VECTORS]](<8 x s8>) ; GFX9-MESA-LABEL: name: test_load_flat_v8s8_align8 ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p0) = COPY $vgpr0_vgpr1 ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p0) :: (load 8) - ; GFX9-MESA: [[BITCAST:%[0-9]+]]:_(<8 x s8>) = G_BITCAST [[LOAD]](<2 x s32>) - ; GFX9-MESA: $vgpr0_vgpr1 = COPY [[BITCAST]](<8 x s8>) + ; GFX9-MESA: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[LOAD]](<2 x s32>) + ; GFX9-MESA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8 + ; GFX9-MESA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C]](s32) + ; GFX9-MESA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16 + ; GFX9-MESA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C1]](s32) + ; GFX9-MESA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24 + ; GFX9-MESA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C2]](s32) + ; GFX9-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[UV]](s32) + ; GFX9-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY1]](s32), [[COPY2]](s32) + ; GFX9-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32) + ; GFX9-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC1:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY3]](s32), [[COPY4]](s32) + ; GFX9-MESA: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC]](<2 x s16>), [[BUILD_VECTOR_TRUNC1]](<2 x s16>) + ; GFX9-MESA: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS]](<4 x s16>) + ; GFX9-MESA: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C]](s32) + ; GFX9-MESA: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C1]](s32) + ; GFX9-MESA: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C2]](s32) + ; GFX9-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[UV1]](s32) + ; GFX9-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC2:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY5]](s32), [[COPY6]](s32) + ; GFX9-MESA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32) + ; GFX9-MESA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC3:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY7]](s32), [[COPY8]](s32) + ; GFX9-MESA: [[CONCAT_VECTORS1:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC2]](<2 x s16>), [[BUILD_VECTOR_TRUNC3]](<2 x s16>) + ; GFX9-MESA: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS1]](<4 x s16>) + ; GFX9-MESA: [[CONCAT_VECTORS2:%[0-9]+]]:_(<8 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>) + ; GFX9-MESA: $vgpr0_vgpr1 = COPY [[CONCAT_VECTORS2]](<8 x s8>) %0:_(p0) = COPY $vgpr0_vgpr1 %1:_(<8 x s8>) = G_LOAD %0 :: (load 8, align 8, addrspace 0) $vgpr0_vgpr1 = COPY %1 @@ -4997,28 +5115,244 @@ ; CI-LABEL: name: test_load_flat_v16s8_align16 ; CI: [[COPY:%[0-9]+]]:_(p0) = COPY $vgpr0_vgpr1 ; CI: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p0) :: (load 16) - ; CI: [[BITCAST:%[0-9]+]]:_(<16 x s8>) = G_BITCAST [[LOAD]](<4 x s32>) - ; CI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](<16 x s8>) + ; CI: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32), [[UV2:%[0-9]+]]:_(s32), [[UV3:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[LOAD]](<4 x s32>) + ; CI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8 + ; CI: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C]](s32) + ; CI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16 + ; CI: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C1]](s32) + ; CI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24 + ; CI: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C2]](s32) + ; CI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[UV]](s32) + ; CI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32) + ; CI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32) + ; CI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32) + ; CI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32) + ; CI: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>) + ; CI: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C]](s32) + ; CI: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C1]](s32) + ; CI: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C2]](s32) + ; CI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[UV1]](s32) + ; CI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32) + ; CI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32) + ; CI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32) + ; CI: [[BUILD_VECTOR1:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY5]](s32), [[COPY6]](s32), [[COPY7]](s32), [[COPY8]](s32) + ; CI: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR1]](<4 x s32>) + ; CI: [[LSHR6:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C]](s32) + ; CI: [[LSHR7:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C1]](s32) + ; CI: [[LSHR8:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C2]](s32) + ; CI: [[COPY9:%[0-9]+]]:_(s32) = COPY [[UV2]](s32) + ; CI: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LSHR6]](s32) + ; CI: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LSHR7]](s32) + ; CI: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LSHR8]](s32) + ; CI: [[BUILD_VECTOR2:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY9]](s32), [[COPY10]](s32), [[COPY11]](s32), [[COPY12]](s32) + ; CI: [[TRUNC2:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR2]](<4 x s32>) + ; CI: [[LSHR9:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C]](s32) + ; CI: [[LSHR10:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C1]](s32) + ; CI: [[LSHR11:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C2]](s32) + ; CI: [[COPY13:%[0-9]+]]:_(s32) = COPY [[UV3]](s32) + ; CI: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LSHR9]](s32) + ; CI: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LSHR10]](s32) + ; CI: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LSHR11]](s32) + ; CI: [[BUILD_VECTOR3:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY13]](s32), [[COPY14]](s32), [[COPY15]](s32), [[COPY16]](s32) + ; CI: [[TRUNC3:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR3]](<4 x s32>) + ; CI: [[CONCAT_VECTORS:%[0-9]+]]:_(<16 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>), [[TRUNC2]](<4 x s8>), [[TRUNC3]](<4 x s8>) + ; CI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[CONCAT_VECTORS]](<16 x s8>) ; VI-LABEL: name: test_load_flat_v16s8_align16 ; VI: [[COPY:%[0-9]+]]:_(p0) = COPY $vgpr0_vgpr1 ; VI: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p0) :: (load 16) - ; VI: [[BITCAST:%[0-9]+]]:_(<16 x s8>) = G_BITCAST [[LOAD]](<4 x s32>) - ; VI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](<16 x s8>) + ; VI: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32), [[UV2:%[0-9]+]]:_(s32), [[UV3:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[LOAD]](<4 x s32>) + ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8 + ; VI: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C]](s32) + ; VI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16 + ; VI: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C1]](s32) + ; VI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24 + ; VI: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C2]](s32) + ; VI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[UV]](s32) + ; VI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32) + ; VI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32) + ; VI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32) + ; VI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32) + ; VI: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>) + ; VI: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C]](s32) + ; VI: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C1]](s32) + ; VI: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C2]](s32) + ; VI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[UV1]](s32) + ; VI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32) + ; VI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32) + ; VI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32) + ; VI: [[BUILD_VECTOR1:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY5]](s32), [[COPY6]](s32), [[COPY7]](s32), [[COPY8]](s32) + ; VI: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR1]](<4 x s32>) + ; VI: [[LSHR6:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C]](s32) + ; VI: [[LSHR7:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C1]](s32) + ; VI: [[LSHR8:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C2]](s32) + ; VI: [[COPY9:%[0-9]+]]:_(s32) = COPY [[UV2]](s32) + ; VI: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LSHR6]](s32) + ; VI: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LSHR7]](s32) + ; VI: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LSHR8]](s32) + ; VI: [[BUILD_VECTOR2:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY9]](s32), [[COPY10]](s32), [[COPY11]](s32), [[COPY12]](s32) + ; VI: [[TRUNC2:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR2]](<4 x s32>) + ; VI: [[LSHR9:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C]](s32) + ; VI: [[LSHR10:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C1]](s32) + ; VI: [[LSHR11:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C2]](s32) + ; VI: [[COPY13:%[0-9]+]]:_(s32) = COPY [[UV3]](s32) + ; VI: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LSHR9]](s32) + ; VI: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LSHR10]](s32) + ; VI: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LSHR11]](s32) + ; VI: [[BUILD_VECTOR3:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY13]](s32), [[COPY14]](s32), [[COPY15]](s32), [[COPY16]](s32) + ; VI: [[TRUNC3:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR3]](<4 x s32>) + ; VI: [[CONCAT_VECTORS:%[0-9]+]]:_(<16 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>), [[TRUNC2]](<4 x s8>), [[TRUNC3]](<4 x s8>) + ; VI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[CONCAT_VECTORS]](<16 x s8>) ; GFX9-LABEL: name: test_load_flat_v16s8_align16 ; GFX9: [[COPY:%[0-9]+]]:_(p0) = COPY $vgpr0_vgpr1 ; GFX9: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p0) :: (load 16) - ; GFX9: [[BITCAST:%[0-9]+]]:_(<16 x s8>) = G_BITCAST [[LOAD]](<4 x s32>) - ; GFX9: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](<16 x s8>) + ; GFX9: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32), [[UV2:%[0-9]+]]:_(s32), [[UV3:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[LOAD]](<4 x s32>) + ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8 + ; GFX9: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C]](s32) + ; GFX9: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16 + ; GFX9: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C1]](s32) + ; GFX9: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24 + ; GFX9: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C2]](s32) + ; GFX9: [[COPY1:%[0-9]+]]:_(s32) = COPY [[UV]](s32) + ; GFX9: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32) + ; GFX9: [[BUILD_VECTOR_TRUNC:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY1]](s32), [[COPY2]](s32) + ; GFX9: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32) + ; GFX9: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32) + ; GFX9: [[BUILD_VECTOR_TRUNC1:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY3]](s32), [[COPY4]](s32) + ; GFX9: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC]](<2 x s16>), [[BUILD_VECTOR_TRUNC1]](<2 x s16>) + ; GFX9: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS]](<4 x s16>) + ; GFX9: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C]](s32) + ; GFX9: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C1]](s32) + ; GFX9: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C2]](s32) + ; GFX9: [[COPY5:%[0-9]+]]:_(s32) = COPY [[UV1]](s32) + ; GFX9: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32) + ; GFX9: [[BUILD_VECTOR_TRUNC2:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY5]](s32), [[COPY6]](s32) + ; GFX9: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32) + ; GFX9: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32) + ; GFX9: [[BUILD_VECTOR_TRUNC3:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY7]](s32), [[COPY8]](s32) + ; GFX9: [[CONCAT_VECTORS1:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC2]](<2 x s16>), [[BUILD_VECTOR_TRUNC3]](<2 x s16>) + ; GFX9: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS1]](<4 x s16>) + ; GFX9: [[LSHR6:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C]](s32) + ; GFX9: [[LSHR7:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C1]](s32) + ; GFX9: [[LSHR8:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C2]](s32) + ; GFX9: [[COPY9:%[0-9]+]]:_(s32) = COPY [[UV2]](s32) + ; GFX9: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LSHR6]](s32) + ; GFX9: [[BUILD_VECTOR_TRUNC4:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY9]](s32), [[COPY10]](s32) + ; GFX9: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LSHR7]](s32) + ; GFX9: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LSHR8]](s32) + ; GFX9: [[BUILD_VECTOR_TRUNC5:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY11]](s32), [[COPY12]](s32) + ; GFX9: [[CONCAT_VECTORS2:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC4]](<2 x s16>), [[BUILD_VECTOR_TRUNC5]](<2 x s16>) + ; GFX9: [[TRUNC2:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS2]](<4 x s16>) + ; GFX9: [[LSHR9:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C]](s32) + ; GFX9: [[LSHR10:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C1]](s32) + ; GFX9: [[LSHR11:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C2]](s32) + ; GFX9: [[COPY13:%[0-9]+]]:_(s32) = COPY [[UV3]](s32) + ; GFX9: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LSHR9]](s32) + ; GFX9: [[BUILD_VECTOR_TRUNC6:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY13]](s32), [[COPY14]](s32) + ; GFX9: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LSHR10]](s32) + ; GFX9: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LSHR11]](s32) + ; GFX9: [[BUILD_VECTOR_TRUNC7:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY15]](s32), [[COPY16]](s32) + ; GFX9: [[CONCAT_VECTORS3:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC6]](<2 x s16>), [[BUILD_VECTOR_TRUNC7]](<2 x s16>) + ; GFX9: [[TRUNC3:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS3]](<4 x s16>) + ; GFX9: [[CONCAT_VECTORS4:%[0-9]+]]:_(<16 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>), [[TRUNC2]](<4 x s8>), [[TRUNC3]](<4 x s8>) + ; GFX9: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[CONCAT_VECTORS4]](<16 x s8>) ; CI-MESA-LABEL: name: test_load_flat_v16s8_align16 ; CI-MESA: [[COPY:%[0-9]+]]:_(p0) = COPY $vgpr0_vgpr1 ; CI-MESA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p0) :: (load 16) - ; CI-MESA: [[BITCAST:%[0-9]+]]:_(<16 x s8>) = G_BITCAST [[LOAD]](<4 x s32>) - ; CI-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](<16 x s8>) + ; CI-MESA: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32), [[UV2:%[0-9]+]]:_(s32), [[UV3:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[LOAD]](<4 x s32>) + ; CI-MESA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8 + ; CI-MESA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C]](s32) + ; CI-MESA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16 + ; CI-MESA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C1]](s32) + ; CI-MESA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24 + ; CI-MESA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C2]](s32) + ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[UV]](s32) + ; CI-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32) + ; CI-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32) + ; CI-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32) + ; CI-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32) + ; CI-MESA: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>) + ; CI-MESA: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C]](s32) + ; CI-MESA: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C1]](s32) + ; CI-MESA: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C2]](s32) + ; CI-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[UV1]](s32) + ; CI-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32) + ; CI-MESA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32) + ; CI-MESA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32) + ; CI-MESA: [[BUILD_VECTOR1:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY5]](s32), [[COPY6]](s32), [[COPY7]](s32), [[COPY8]](s32) + ; CI-MESA: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR1]](<4 x s32>) + ; CI-MESA: [[LSHR6:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C]](s32) + ; CI-MESA: [[LSHR7:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C1]](s32) + ; CI-MESA: [[LSHR8:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C2]](s32) + ; CI-MESA: [[COPY9:%[0-9]+]]:_(s32) = COPY [[UV2]](s32) + ; CI-MESA: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LSHR6]](s32) + ; CI-MESA: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LSHR7]](s32) + ; CI-MESA: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LSHR8]](s32) + ; CI-MESA: [[BUILD_VECTOR2:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY9]](s32), [[COPY10]](s32), [[COPY11]](s32), [[COPY12]](s32) + ; CI-MESA: [[TRUNC2:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR2]](<4 x s32>) + ; CI-MESA: [[LSHR9:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C]](s32) + ; CI-MESA: [[LSHR10:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C1]](s32) + ; CI-MESA: [[LSHR11:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C2]](s32) + ; CI-MESA: [[COPY13:%[0-9]+]]:_(s32) = COPY [[UV3]](s32) + ; CI-MESA: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LSHR9]](s32) + ; CI-MESA: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LSHR10]](s32) + ; CI-MESA: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LSHR11]](s32) + ; CI-MESA: [[BUILD_VECTOR3:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY13]](s32), [[COPY14]](s32), [[COPY15]](s32), [[COPY16]](s32) + ; CI-MESA: [[TRUNC3:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR3]](<4 x s32>) + ; CI-MESA: [[CONCAT_VECTORS:%[0-9]+]]:_(<16 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>), [[TRUNC2]](<4 x s8>), [[TRUNC3]](<4 x s8>) + ; CI-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[CONCAT_VECTORS]](<16 x s8>) ; GFX9-MESA-LABEL: name: test_load_flat_v16s8_align16 ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p0) = COPY $vgpr0_vgpr1 ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p0) :: (load 16) - ; GFX9-MESA: [[BITCAST:%[0-9]+]]:_(<16 x s8>) = G_BITCAST [[LOAD]](<4 x s32>) - ; GFX9-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](<16 x s8>) + ; GFX9-MESA: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32), [[UV2:%[0-9]+]]:_(s32), [[UV3:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[LOAD]](<4 x s32>) + ; GFX9-MESA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8 + ; GFX9-MESA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C]](s32) + ; GFX9-MESA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16 + ; GFX9-MESA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C1]](s32) + ; GFX9-MESA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24 + ; GFX9-MESA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C2]](s32) + ; GFX9-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[UV]](s32) + ; GFX9-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY1]](s32), [[COPY2]](s32) + ; GFX9-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32) + ; GFX9-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC1:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY3]](s32), [[COPY4]](s32) + ; GFX9-MESA: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC]](<2 x s16>), [[BUILD_VECTOR_TRUNC1]](<2 x s16>) + ; GFX9-MESA: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS]](<4 x s16>) + ; GFX9-MESA: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C]](s32) + ; GFX9-MESA: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C1]](s32) + ; GFX9-MESA: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C2]](s32) + ; GFX9-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[UV1]](s32) + ; GFX9-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC2:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY5]](s32), [[COPY6]](s32) + ; GFX9-MESA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32) + ; GFX9-MESA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC3:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY7]](s32), [[COPY8]](s32) + ; GFX9-MESA: [[CONCAT_VECTORS1:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC2]](<2 x s16>), [[BUILD_VECTOR_TRUNC3]](<2 x s16>) + ; GFX9-MESA: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS1]](<4 x s16>) + ; GFX9-MESA: [[LSHR6:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C]](s32) + ; GFX9-MESA: [[LSHR7:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C1]](s32) + ; GFX9-MESA: [[LSHR8:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C2]](s32) + ; GFX9-MESA: [[COPY9:%[0-9]+]]:_(s32) = COPY [[UV2]](s32) + ; GFX9-MESA: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LSHR6]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC4:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY9]](s32), [[COPY10]](s32) + ; GFX9-MESA: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LSHR7]](s32) + ; GFX9-MESA: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LSHR8]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC5:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY11]](s32), [[COPY12]](s32) + ; GFX9-MESA: [[CONCAT_VECTORS2:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC4]](<2 x s16>), [[BUILD_VECTOR_TRUNC5]](<2 x s16>) + ; GFX9-MESA: [[TRUNC2:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS2]](<4 x s16>) + ; GFX9-MESA: [[LSHR9:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C]](s32) + ; GFX9-MESA: [[LSHR10:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C1]](s32) + ; GFX9-MESA: [[LSHR11:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C2]](s32) + ; GFX9-MESA: [[COPY13:%[0-9]+]]:_(s32) = COPY [[UV3]](s32) + ; GFX9-MESA: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LSHR9]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC6:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY13]](s32), [[COPY14]](s32) + ; GFX9-MESA: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LSHR10]](s32) + ; GFX9-MESA: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LSHR11]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC7:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY15]](s32), [[COPY16]](s32) + ; GFX9-MESA: [[CONCAT_VECTORS3:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC6]](<2 x s16>), [[BUILD_VECTOR_TRUNC7]](<2 x s16>) + ; GFX9-MESA: [[TRUNC3:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS3]](<4 x s16>) + ; GFX9-MESA: [[CONCAT_VECTORS4:%[0-9]+]]:_(<16 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>), [[TRUNC2]](<4 x s8>), [[TRUNC3]](<4 x s8>) + ; GFX9-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[CONCAT_VECTORS4]](<16 x s8>) %0:_(p0) = COPY $vgpr0_vgpr1 %1:_(<16 x s8>) = G_LOAD %0 :: (load 16, align 16, addrspace 0) $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %1 @@ -5036,45 +5370,457 @@ ; CI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 16 ; CI: [[PTR_ADD:%[0-9]+]]:_(p0) = G_PTR_ADD [[COPY]], [[C]](s64) ; CI: [[LOAD1:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[PTR_ADD]](p0) :: (load 16 + 16) - ; CI: [[CONCAT_VECTORS:%[0-9]+]]:_(<8 x s32>) = G_CONCAT_VECTORS [[LOAD]](<4 x s32>), [[LOAD1]](<4 x s32>) - ; CI: [[BITCAST:%[0-9]+]]:_(<32 x s8>) = G_BITCAST [[CONCAT_VECTORS]](<8 x s32>) - ; CI: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[BITCAST]](<32 x s8>) + ; CI: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32), [[UV2:%[0-9]+]]:_(s32), [[UV3:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[LOAD]](<4 x s32>) + ; CI: [[UV4:%[0-9]+]]:_(s32), [[UV5:%[0-9]+]]:_(s32), [[UV6:%[0-9]+]]:_(s32), [[UV7:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[LOAD1]](<4 x s32>) + ; CI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8 + ; CI: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C1]](s32) + ; CI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 16 + ; CI: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C2]](s32) + ; CI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 24 + ; CI: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C3]](s32) + ; CI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[UV]](s32) + ; CI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32) + ; CI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32) + ; CI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32) + ; CI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32) + ; CI: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>) + ; CI: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C1]](s32) + ; CI: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C2]](s32) + ; CI: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C3]](s32) + ; CI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[UV1]](s32) + ; CI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32) + ; CI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32) + ; CI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32) + ; CI: [[BUILD_VECTOR1:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY5]](s32), [[COPY6]](s32), [[COPY7]](s32), [[COPY8]](s32) + ; CI: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR1]](<4 x s32>) + ; CI: [[LSHR6:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C1]](s32) + ; CI: [[LSHR7:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C2]](s32) + ; CI: [[LSHR8:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C3]](s32) + ; CI: [[COPY9:%[0-9]+]]:_(s32) = COPY [[UV2]](s32) + ; CI: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LSHR6]](s32) + ; CI: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LSHR7]](s32) + ; CI: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LSHR8]](s32) + ; CI: [[BUILD_VECTOR2:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY9]](s32), [[COPY10]](s32), [[COPY11]](s32), [[COPY12]](s32) + ; CI: [[TRUNC2:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR2]](<4 x s32>) + ; CI: [[LSHR9:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C1]](s32) + ; CI: [[LSHR10:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C2]](s32) + ; CI: [[LSHR11:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C3]](s32) + ; CI: [[COPY13:%[0-9]+]]:_(s32) = COPY [[UV3]](s32) + ; CI: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LSHR9]](s32) + ; CI: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LSHR10]](s32) + ; CI: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LSHR11]](s32) + ; CI: [[BUILD_VECTOR3:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY13]](s32), [[COPY14]](s32), [[COPY15]](s32), [[COPY16]](s32) + ; CI: [[TRUNC3:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR3]](<4 x s32>) + ; CI: [[LSHR12:%[0-9]+]]:_(s32) = G_LSHR [[UV4]], [[C1]](s32) + ; CI: [[LSHR13:%[0-9]+]]:_(s32) = G_LSHR [[UV4]], [[C2]](s32) + ; CI: [[LSHR14:%[0-9]+]]:_(s32) = G_LSHR [[UV4]], [[C3]](s32) + ; CI: [[COPY17:%[0-9]+]]:_(s32) = COPY [[UV4]](s32) + ; CI: [[COPY18:%[0-9]+]]:_(s32) = COPY [[LSHR12]](s32) + ; CI: [[COPY19:%[0-9]+]]:_(s32) = COPY [[LSHR13]](s32) + ; CI: [[COPY20:%[0-9]+]]:_(s32) = COPY [[LSHR14]](s32) + ; CI: [[BUILD_VECTOR4:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY17]](s32), [[COPY18]](s32), [[COPY19]](s32), [[COPY20]](s32) + ; CI: [[TRUNC4:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR4]](<4 x s32>) + ; CI: [[LSHR15:%[0-9]+]]:_(s32) = G_LSHR [[UV5]], [[C1]](s32) + ; CI: [[LSHR16:%[0-9]+]]:_(s32) = G_LSHR [[UV5]], [[C2]](s32) + ; CI: [[LSHR17:%[0-9]+]]:_(s32) = G_LSHR [[UV5]], [[C3]](s32) + ; CI: [[COPY21:%[0-9]+]]:_(s32) = COPY [[UV5]](s32) + ; CI: [[COPY22:%[0-9]+]]:_(s32) = COPY [[LSHR15]](s32) + ; CI: [[COPY23:%[0-9]+]]:_(s32) = COPY [[LSHR16]](s32) + ; CI: [[COPY24:%[0-9]+]]:_(s32) = COPY [[LSHR17]](s32) + ; CI: [[BUILD_VECTOR5:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY21]](s32), [[COPY22]](s32), [[COPY23]](s32), [[COPY24]](s32) + ; CI: [[TRUNC5:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR5]](<4 x s32>) + ; CI: [[LSHR18:%[0-9]+]]:_(s32) = G_LSHR [[UV6]], [[C1]](s32) + ; CI: [[LSHR19:%[0-9]+]]:_(s32) = G_LSHR [[UV6]], [[C2]](s32) + ; CI: [[LSHR20:%[0-9]+]]:_(s32) = G_LSHR [[UV6]], [[C3]](s32) + ; CI: [[COPY25:%[0-9]+]]:_(s32) = COPY [[UV6]](s32) + ; CI: [[COPY26:%[0-9]+]]:_(s32) = COPY [[LSHR18]](s32) + ; CI: [[COPY27:%[0-9]+]]:_(s32) = COPY [[LSHR19]](s32) + ; CI: [[COPY28:%[0-9]+]]:_(s32) = COPY [[LSHR20]](s32) + ; CI: [[BUILD_VECTOR6:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY25]](s32), [[COPY26]](s32), [[COPY27]](s32), [[COPY28]](s32) + ; CI: [[TRUNC6:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR6]](<4 x s32>) + ; CI: [[LSHR21:%[0-9]+]]:_(s32) = G_LSHR [[UV7]], [[C1]](s32) + ; CI: [[LSHR22:%[0-9]+]]:_(s32) = G_LSHR [[UV7]], [[C2]](s32) + ; CI: [[LSHR23:%[0-9]+]]:_(s32) = G_LSHR [[UV7]], [[C3]](s32) + ; CI: [[COPY29:%[0-9]+]]:_(s32) = COPY [[UV7]](s32) + ; CI: [[COPY30:%[0-9]+]]:_(s32) = COPY [[LSHR21]](s32) + ; CI: [[COPY31:%[0-9]+]]:_(s32) = COPY [[LSHR22]](s32) + ; CI: [[COPY32:%[0-9]+]]:_(s32) = COPY [[LSHR23]](s32) + ; CI: [[BUILD_VECTOR7:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY29]](s32), [[COPY30]](s32), [[COPY31]](s32), [[COPY32]](s32) + ; CI: [[TRUNC7:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR7]](<4 x s32>) + ; CI: [[CONCAT_VECTORS:%[0-9]+]]:_(<32 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>), [[TRUNC2]](<4 x s8>), [[TRUNC3]](<4 x s8>), [[TRUNC4]](<4 x s8>), [[TRUNC5]](<4 x s8>), [[TRUNC6]](<4 x s8>), [[TRUNC7]](<4 x s8>) + ; CI: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[CONCAT_VECTORS]](<32 x s8>) ; VI-LABEL: name: test_load_flat_v32s8_align32 ; VI: [[COPY:%[0-9]+]]:_(p0) = COPY $vgpr0_vgpr1 ; VI: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p0) :: (load 16, align 32) ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 16 ; VI: [[PTR_ADD:%[0-9]+]]:_(p0) = G_PTR_ADD [[COPY]], [[C]](s64) ; VI: [[LOAD1:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[PTR_ADD]](p0) :: (load 16 + 16) - ; VI: [[CONCAT_VECTORS:%[0-9]+]]:_(<8 x s32>) = G_CONCAT_VECTORS [[LOAD]](<4 x s32>), [[LOAD1]](<4 x s32>) - ; VI: [[BITCAST:%[0-9]+]]:_(<32 x s8>) = G_BITCAST [[CONCAT_VECTORS]](<8 x s32>) - ; VI: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[BITCAST]](<32 x s8>) + ; VI: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32), [[UV2:%[0-9]+]]:_(s32), [[UV3:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[LOAD]](<4 x s32>) + ; VI: [[UV4:%[0-9]+]]:_(s32), [[UV5:%[0-9]+]]:_(s32), [[UV6:%[0-9]+]]:_(s32), [[UV7:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[LOAD1]](<4 x s32>) + ; VI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8 + ; VI: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C1]](s32) + ; VI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 16 + ; VI: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C2]](s32) + ; VI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 24 + ; VI: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C3]](s32) + ; VI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[UV]](s32) + ; VI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32) + ; VI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32) + ; VI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32) + ; VI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32) + ; VI: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>) + ; VI: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C1]](s32) + ; VI: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C2]](s32) + ; VI: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C3]](s32) + ; VI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[UV1]](s32) + ; VI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32) + ; VI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32) + ; VI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32) + ; VI: [[BUILD_VECTOR1:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY5]](s32), [[COPY6]](s32), [[COPY7]](s32), [[COPY8]](s32) + ; VI: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR1]](<4 x s32>) + ; VI: [[LSHR6:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C1]](s32) + ; VI: [[LSHR7:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C2]](s32) + ; VI: [[LSHR8:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C3]](s32) + ; VI: [[COPY9:%[0-9]+]]:_(s32) = COPY [[UV2]](s32) + ; VI: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LSHR6]](s32) + ; VI: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LSHR7]](s32) + ; VI: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LSHR8]](s32) + ; VI: [[BUILD_VECTOR2:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY9]](s32), [[COPY10]](s32), [[COPY11]](s32), [[COPY12]](s32) + ; VI: [[TRUNC2:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR2]](<4 x s32>) + ; VI: [[LSHR9:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C1]](s32) + ; VI: [[LSHR10:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C2]](s32) + ; VI: [[LSHR11:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C3]](s32) + ; VI: [[COPY13:%[0-9]+]]:_(s32) = COPY [[UV3]](s32) + ; VI: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LSHR9]](s32) + ; VI: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LSHR10]](s32) + ; VI: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LSHR11]](s32) + ; VI: [[BUILD_VECTOR3:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY13]](s32), [[COPY14]](s32), [[COPY15]](s32), [[COPY16]](s32) + ; VI: [[TRUNC3:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR3]](<4 x s32>) + ; VI: [[LSHR12:%[0-9]+]]:_(s32) = G_LSHR [[UV4]], [[C1]](s32) + ; VI: [[LSHR13:%[0-9]+]]:_(s32) = G_LSHR [[UV4]], [[C2]](s32) + ; VI: [[LSHR14:%[0-9]+]]:_(s32) = G_LSHR [[UV4]], [[C3]](s32) + ; VI: [[COPY17:%[0-9]+]]:_(s32) = COPY [[UV4]](s32) + ; VI: [[COPY18:%[0-9]+]]:_(s32) = COPY [[LSHR12]](s32) + ; VI: [[COPY19:%[0-9]+]]:_(s32) = COPY [[LSHR13]](s32) + ; VI: [[COPY20:%[0-9]+]]:_(s32) = COPY [[LSHR14]](s32) + ; VI: [[BUILD_VECTOR4:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY17]](s32), [[COPY18]](s32), [[COPY19]](s32), [[COPY20]](s32) + ; VI: [[TRUNC4:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR4]](<4 x s32>) + ; VI: [[LSHR15:%[0-9]+]]:_(s32) = G_LSHR [[UV5]], [[C1]](s32) + ; VI: [[LSHR16:%[0-9]+]]:_(s32) = G_LSHR [[UV5]], [[C2]](s32) + ; VI: [[LSHR17:%[0-9]+]]:_(s32) = G_LSHR [[UV5]], [[C3]](s32) + ; VI: [[COPY21:%[0-9]+]]:_(s32) = COPY [[UV5]](s32) + ; VI: [[COPY22:%[0-9]+]]:_(s32) = COPY [[LSHR15]](s32) + ; VI: [[COPY23:%[0-9]+]]:_(s32) = COPY [[LSHR16]](s32) + ; VI: [[COPY24:%[0-9]+]]:_(s32) = COPY [[LSHR17]](s32) + ; VI: [[BUILD_VECTOR5:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY21]](s32), [[COPY22]](s32), [[COPY23]](s32), [[COPY24]](s32) + ; VI: [[TRUNC5:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR5]](<4 x s32>) + ; VI: [[LSHR18:%[0-9]+]]:_(s32) = G_LSHR [[UV6]], [[C1]](s32) + ; VI: [[LSHR19:%[0-9]+]]:_(s32) = G_LSHR [[UV6]], [[C2]](s32) + ; VI: [[LSHR20:%[0-9]+]]:_(s32) = G_LSHR [[UV6]], [[C3]](s32) + ; VI: [[COPY25:%[0-9]+]]:_(s32) = COPY [[UV6]](s32) + ; VI: [[COPY26:%[0-9]+]]:_(s32) = COPY [[LSHR18]](s32) + ; VI: [[COPY27:%[0-9]+]]:_(s32) = COPY [[LSHR19]](s32) + ; VI: [[COPY28:%[0-9]+]]:_(s32) = COPY [[LSHR20]](s32) + ; VI: [[BUILD_VECTOR6:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY25]](s32), [[COPY26]](s32), [[COPY27]](s32), [[COPY28]](s32) + ; VI: [[TRUNC6:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR6]](<4 x s32>) + ; VI: [[LSHR21:%[0-9]+]]:_(s32) = G_LSHR [[UV7]], [[C1]](s32) + ; VI: [[LSHR22:%[0-9]+]]:_(s32) = G_LSHR [[UV7]], [[C2]](s32) + ; VI: [[LSHR23:%[0-9]+]]:_(s32) = G_LSHR [[UV7]], [[C3]](s32) + ; VI: [[COPY29:%[0-9]+]]:_(s32) = COPY [[UV7]](s32) + ; VI: [[COPY30:%[0-9]+]]:_(s32) = COPY [[LSHR21]](s32) + ; VI: [[COPY31:%[0-9]+]]:_(s32) = COPY [[LSHR22]](s32) + ; VI: [[COPY32:%[0-9]+]]:_(s32) = COPY [[LSHR23]](s32) + ; VI: [[BUILD_VECTOR7:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY29]](s32), [[COPY30]](s32), [[COPY31]](s32), [[COPY32]](s32) + ; VI: [[TRUNC7:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR7]](<4 x s32>) + ; VI: [[CONCAT_VECTORS:%[0-9]+]]:_(<32 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>), [[TRUNC2]](<4 x s8>), [[TRUNC3]](<4 x s8>), [[TRUNC4]](<4 x s8>), [[TRUNC5]](<4 x s8>), [[TRUNC6]](<4 x s8>), [[TRUNC7]](<4 x s8>) + ; VI: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[CONCAT_VECTORS]](<32 x s8>) ; GFX9-LABEL: name: test_load_flat_v32s8_align32 ; GFX9: [[COPY:%[0-9]+]]:_(p0) = COPY $vgpr0_vgpr1 ; GFX9: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p0) :: (load 16, align 32) ; GFX9: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 16 ; GFX9: [[PTR_ADD:%[0-9]+]]:_(p0) = G_PTR_ADD [[COPY]], [[C]](s64) ; GFX9: [[LOAD1:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[PTR_ADD]](p0) :: (load 16 + 16) - ; GFX9: [[CONCAT_VECTORS:%[0-9]+]]:_(<8 x s32>) = G_CONCAT_VECTORS [[LOAD]](<4 x s32>), [[LOAD1]](<4 x s32>) - ; GFX9: [[BITCAST:%[0-9]+]]:_(<32 x s8>) = G_BITCAST [[CONCAT_VECTORS]](<8 x s32>) - ; GFX9: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[BITCAST]](<32 x s8>) + ; GFX9: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32), [[UV2:%[0-9]+]]:_(s32), [[UV3:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[LOAD]](<4 x s32>) + ; GFX9: [[UV4:%[0-9]+]]:_(s32), [[UV5:%[0-9]+]]:_(s32), [[UV6:%[0-9]+]]:_(s32), [[UV7:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[LOAD1]](<4 x s32>) + ; GFX9: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8 + ; GFX9: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C1]](s32) + ; GFX9: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 16 + ; GFX9: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C2]](s32) + ; GFX9: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 24 + ; GFX9: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C3]](s32) + ; GFX9: [[COPY1:%[0-9]+]]:_(s32) = COPY [[UV]](s32) + ; GFX9: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32) + ; GFX9: [[BUILD_VECTOR_TRUNC:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY1]](s32), [[COPY2]](s32) + ; GFX9: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32) + ; GFX9: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32) + ; GFX9: [[BUILD_VECTOR_TRUNC1:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY3]](s32), [[COPY4]](s32) + ; GFX9: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC]](<2 x s16>), [[BUILD_VECTOR_TRUNC1]](<2 x s16>) + ; GFX9: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS]](<4 x s16>) + ; GFX9: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C1]](s32) + ; GFX9: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C2]](s32) + ; GFX9: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C3]](s32) + ; GFX9: [[COPY5:%[0-9]+]]:_(s32) = COPY [[UV1]](s32) + ; GFX9: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32) + ; GFX9: [[BUILD_VECTOR_TRUNC2:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY5]](s32), [[COPY6]](s32) + ; GFX9: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32) + ; GFX9: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32) + ; GFX9: [[BUILD_VECTOR_TRUNC3:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY7]](s32), [[COPY8]](s32) + ; GFX9: [[CONCAT_VECTORS1:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC2]](<2 x s16>), [[BUILD_VECTOR_TRUNC3]](<2 x s16>) + ; GFX9: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS1]](<4 x s16>) + ; GFX9: [[LSHR6:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C1]](s32) + ; GFX9: [[LSHR7:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C2]](s32) + ; GFX9: [[LSHR8:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C3]](s32) + ; GFX9: [[COPY9:%[0-9]+]]:_(s32) = COPY [[UV2]](s32) + ; GFX9: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LSHR6]](s32) + ; GFX9: [[BUILD_VECTOR_TRUNC4:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY9]](s32), [[COPY10]](s32) + ; GFX9: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LSHR7]](s32) + ; GFX9: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LSHR8]](s32) + ; GFX9: [[BUILD_VECTOR_TRUNC5:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY11]](s32), [[COPY12]](s32) + ; GFX9: [[CONCAT_VECTORS2:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC4]](<2 x s16>), [[BUILD_VECTOR_TRUNC5]](<2 x s16>) + ; GFX9: [[TRUNC2:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS2]](<4 x s16>) + ; GFX9: [[LSHR9:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C1]](s32) + ; GFX9: [[LSHR10:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C2]](s32) + ; GFX9: [[LSHR11:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C3]](s32) + ; GFX9: [[COPY13:%[0-9]+]]:_(s32) = COPY [[UV3]](s32) + ; GFX9: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LSHR9]](s32) + ; GFX9: [[BUILD_VECTOR_TRUNC6:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY13]](s32), [[COPY14]](s32) + ; GFX9: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LSHR10]](s32) + ; GFX9: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LSHR11]](s32) + ; GFX9: [[BUILD_VECTOR_TRUNC7:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY15]](s32), [[COPY16]](s32) + ; GFX9: [[CONCAT_VECTORS3:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC6]](<2 x s16>), [[BUILD_VECTOR_TRUNC7]](<2 x s16>) + ; GFX9: [[TRUNC3:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS3]](<4 x s16>) + ; GFX9: [[LSHR12:%[0-9]+]]:_(s32) = G_LSHR [[UV4]], [[C1]](s32) + ; GFX9: [[LSHR13:%[0-9]+]]:_(s32) = G_LSHR [[UV4]], [[C2]](s32) + ; GFX9: [[LSHR14:%[0-9]+]]:_(s32) = G_LSHR [[UV4]], [[C3]](s32) + ; GFX9: [[COPY17:%[0-9]+]]:_(s32) = COPY [[UV4]](s32) + ; GFX9: [[COPY18:%[0-9]+]]:_(s32) = COPY [[LSHR12]](s32) + ; GFX9: [[BUILD_VECTOR_TRUNC8:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY17]](s32), [[COPY18]](s32) + ; GFX9: [[COPY19:%[0-9]+]]:_(s32) = COPY [[LSHR13]](s32) + ; GFX9: [[COPY20:%[0-9]+]]:_(s32) = COPY [[LSHR14]](s32) + ; GFX9: [[BUILD_VECTOR_TRUNC9:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY19]](s32), [[COPY20]](s32) + ; GFX9: [[CONCAT_VECTORS4:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC8]](<2 x s16>), [[BUILD_VECTOR_TRUNC9]](<2 x s16>) + ; GFX9: [[TRUNC4:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS4]](<4 x s16>) + ; GFX9: [[LSHR15:%[0-9]+]]:_(s32) = G_LSHR [[UV5]], [[C1]](s32) + ; GFX9: [[LSHR16:%[0-9]+]]:_(s32) = G_LSHR [[UV5]], [[C2]](s32) + ; GFX9: [[LSHR17:%[0-9]+]]:_(s32) = G_LSHR [[UV5]], [[C3]](s32) + ; GFX9: [[COPY21:%[0-9]+]]:_(s32) = COPY [[UV5]](s32) + ; GFX9: [[COPY22:%[0-9]+]]:_(s32) = COPY [[LSHR15]](s32) + ; GFX9: [[BUILD_VECTOR_TRUNC10:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY21]](s32), [[COPY22]](s32) + ; GFX9: [[COPY23:%[0-9]+]]:_(s32) = COPY [[LSHR16]](s32) + ; GFX9: [[COPY24:%[0-9]+]]:_(s32) = COPY [[LSHR17]](s32) + ; GFX9: [[BUILD_VECTOR_TRUNC11:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY23]](s32), [[COPY24]](s32) + ; GFX9: [[CONCAT_VECTORS5:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC10]](<2 x s16>), [[BUILD_VECTOR_TRUNC11]](<2 x s16>) + ; GFX9: [[TRUNC5:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS5]](<4 x s16>) + ; GFX9: [[LSHR18:%[0-9]+]]:_(s32) = G_LSHR [[UV6]], [[C1]](s32) + ; GFX9: [[LSHR19:%[0-9]+]]:_(s32) = G_LSHR [[UV6]], [[C2]](s32) + ; GFX9: [[LSHR20:%[0-9]+]]:_(s32) = G_LSHR [[UV6]], [[C3]](s32) + ; GFX9: [[COPY25:%[0-9]+]]:_(s32) = COPY [[UV6]](s32) + ; GFX9: [[COPY26:%[0-9]+]]:_(s32) = COPY [[LSHR18]](s32) + ; GFX9: [[BUILD_VECTOR_TRUNC12:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY25]](s32), [[COPY26]](s32) + ; GFX9: [[COPY27:%[0-9]+]]:_(s32) = COPY [[LSHR19]](s32) + ; GFX9: [[COPY28:%[0-9]+]]:_(s32) = COPY [[LSHR20]](s32) + ; GFX9: [[BUILD_VECTOR_TRUNC13:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY27]](s32), [[COPY28]](s32) + ; GFX9: [[CONCAT_VECTORS6:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC12]](<2 x s16>), [[BUILD_VECTOR_TRUNC13]](<2 x s16>) + ; GFX9: [[TRUNC6:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS6]](<4 x s16>) + ; GFX9: [[LSHR21:%[0-9]+]]:_(s32) = G_LSHR [[UV7]], [[C1]](s32) + ; GFX9: [[LSHR22:%[0-9]+]]:_(s32) = G_LSHR [[UV7]], [[C2]](s32) + ; GFX9: [[LSHR23:%[0-9]+]]:_(s32) = G_LSHR [[UV7]], [[C3]](s32) + ; GFX9: [[COPY29:%[0-9]+]]:_(s32) = COPY [[UV7]](s32) + ; GFX9: [[COPY30:%[0-9]+]]:_(s32) = COPY [[LSHR21]](s32) + ; GFX9: [[BUILD_VECTOR_TRUNC14:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY29]](s32), [[COPY30]](s32) + ; GFX9: [[COPY31:%[0-9]+]]:_(s32) = COPY [[LSHR22]](s32) + ; GFX9: [[COPY32:%[0-9]+]]:_(s32) = COPY [[LSHR23]](s32) + ; GFX9: [[BUILD_VECTOR_TRUNC15:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY31]](s32), [[COPY32]](s32) + ; GFX9: [[CONCAT_VECTORS7:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC14]](<2 x s16>), [[BUILD_VECTOR_TRUNC15]](<2 x s16>) + ; GFX9: [[TRUNC7:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS7]](<4 x s16>) + ; GFX9: [[CONCAT_VECTORS8:%[0-9]+]]:_(<32 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>), [[TRUNC2]](<4 x s8>), [[TRUNC3]](<4 x s8>), [[TRUNC4]](<4 x s8>), [[TRUNC5]](<4 x s8>), [[TRUNC6]](<4 x s8>), [[TRUNC7]](<4 x s8>) + ; GFX9: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[CONCAT_VECTORS8]](<32 x s8>) ; CI-MESA-LABEL: name: test_load_flat_v32s8_align32 ; CI-MESA: [[COPY:%[0-9]+]]:_(p0) = COPY $vgpr0_vgpr1 ; CI-MESA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p0) :: (load 16, align 32) ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 16 ; CI-MESA: [[PTR_ADD:%[0-9]+]]:_(p0) = G_PTR_ADD [[COPY]], [[C]](s64) ; CI-MESA: [[LOAD1:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[PTR_ADD]](p0) :: (load 16 + 16) - ; CI-MESA: [[CONCAT_VECTORS:%[0-9]+]]:_(<8 x s32>) = G_CONCAT_VECTORS [[LOAD]](<4 x s32>), [[LOAD1]](<4 x s32>) - ; CI-MESA: [[BITCAST:%[0-9]+]]:_(<32 x s8>) = G_BITCAST [[CONCAT_VECTORS]](<8 x s32>) - ; CI-MESA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[BITCAST]](<32 x s8>) + ; CI-MESA: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32), [[UV2:%[0-9]+]]:_(s32), [[UV3:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[LOAD]](<4 x s32>) + ; CI-MESA: [[UV4:%[0-9]+]]:_(s32), [[UV5:%[0-9]+]]:_(s32), [[UV6:%[0-9]+]]:_(s32), [[UV7:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[LOAD1]](<4 x s32>) + ; CI-MESA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8 + ; CI-MESA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C1]](s32) + ; CI-MESA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 16 + ; CI-MESA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C2]](s32) + ; CI-MESA: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 24 + ; CI-MESA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C3]](s32) + ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[UV]](s32) + ; CI-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32) + ; CI-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32) + ; CI-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32) + ; CI-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32) + ; CI-MESA: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>) + ; CI-MESA: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C1]](s32) + ; CI-MESA: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C2]](s32) + ; CI-MESA: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C3]](s32) + ; CI-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[UV1]](s32) + ; CI-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32) + ; CI-MESA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32) + ; CI-MESA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32) + ; CI-MESA: [[BUILD_VECTOR1:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY5]](s32), [[COPY6]](s32), [[COPY7]](s32), [[COPY8]](s32) + ; CI-MESA: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR1]](<4 x s32>) + ; CI-MESA: [[LSHR6:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C1]](s32) + ; CI-MESA: [[LSHR7:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C2]](s32) + ; CI-MESA: [[LSHR8:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C3]](s32) + ; CI-MESA: [[COPY9:%[0-9]+]]:_(s32) = COPY [[UV2]](s32) + ; CI-MESA: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LSHR6]](s32) + ; CI-MESA: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LSHR7]](s32) + ; CI-MESA: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LSHR8]](s32) + ; CI-MESA: [[BUILD_VECTOR2:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY9]](s32), [[COPY10]](s32), [[COPY11]](s32), [[COPY12]](s32) + ; CI-MESA: [[TRUNC2:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR2]](<4 x s32>) + ; CI-MESA: [[LSHR9:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C1]](s32) + ; CI-MESA: [[LSHR10:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C2]](s32) + ; CI-MESA: [[LSHR11:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C3]](s32) + ; CI-MESA: [[COPY13:%[0-9]+]]:_(s32) = COPY [[UV3]](s32) + ; CI-MESA: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LSHR9]](s32) + ; CI-MESA: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LSHR10]](s32) + ; CI-MESA: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LSHR11]](s32) + ; CI-MESA: [[BUILD_VECTOR3:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY13]](s32), [[COPY14]](s32), [[COPY15]](s32), [[COPY16]](s32) + ; CI-MESA: [[TRUNC3:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR3]](<4 x s32>) + ; CI-MESA: [[LSHR12:%[0-9]+]]:_(s32) = G_LSHR [[UV4]], [[C1]](s32) + ; CI-MESA: [[LSHR13:%[0-9]+]]:_(s32) = G_LSHR [[UV4]], [[C2]](s32) + ; CI-MESA: [[LSHR14:%[0-9]+]]:_(s32) = G_LSHR [[UV4]], [[C3]](s32) + ; CI-MESA: [[COPY17:%[0-9]+]]:_(s32) = COPY [[UV4]](s32) + ; CI-MESA: [[COPY18:%[0-9]+]]:_(s32) = COPY [[LSHR12]](s32) + ; CI-MESA: [[COPY19:%[0-9]+]]:_(s32) = COPY [[LSHR13]](s32) + ; CI-MESA: [[COPY20:%[0-9]+]]:_(s32) = COPY [[LSHR14]](s32) + ; CI-MESA: [[BUILD_VECTOR4:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY17]](s32), [[COPY18]](s32), [[COPY19]](s32), [[COPY20]](s32) + ; CI-MESA: [[TRUNC4:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR4]](<4 x s32>) + ; CI-MESA: [[LSHR15:%[0-9]+]]:_(s32) = G_LSHR [[UV5]], [[C1]](s32) + ; CI-MESA: [[LSHR16:%[0-9]+]]:_(s32) = G_LSHR [[UV5]], [[C2]](s32) + ; CI-MESA: [[LSHR17:%[0-9]+]]:_(s32) = G_LSHR [[UV5]], [[C3]](s32) + ; CI-MESA: [[COPY21:%[0-9]+]]:_(s32) = COPY [[UV5]](s32) + ; CI-MESA: [[COPY22:%[0-9]+]]:_(s32) = COPY [[LSHR15]](s32) + ; CI-MESA: [[COPY23:%[0-9]+]]:_(s32) = COPY [[LSHR16]](s32) + ; CI-MESA: [[COPY24:%[0-9]+]]:_(s32) = COPY [[LSHR17]](s32) + ; CI-MESA: [[BUILD_VECTOR5:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY21]](s32), [[COPY22]](s32), [[COPY23]](s32), [[COPY24]](s32) + ; CI-MESA: [[TRUNC5:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR5]](<4 x s32>) + ; CI-MESA: [[LSHR18:%[0-9]+]]:_(s32) = G_LSHR [[UV6]], [[C1]](s32) + ; CI-MESA: [[LSHR19:%[0-9]+]]:_(s32) = G_LSHR [[UV6]], [[C2]](s32) + ; CI-MESA: [[LSHR20:%[0-9]+]]:_(s32) = G_LSHR [[UV6]], [[C3]](s32) + ; CI-MESA: [[COPY25:%[0-9]+]]:_(s32) = COPY [[UV6]](s32) + ; CI-MESA: [[COPY26:%[0-9]+]]:_(s32) = COPY [[LSHR18]](s32) + ; CI-MESA: [[COPY27:%[0-9]+]]:_(s32) = COPY [[LSHR19]](s32) + ; CI-MESA: [[COPY28:%[0-9]+]]:_(s32) = COPY [[LSHR20]](s32) + ; CI-MESA: [[BUILD_VECTOR6:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY25]](s32), [[COPY26]](s32), [[COPY27]](s32), [[COPY28]](s32) + ; CI-MESA: [[TRUNC6:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR6]](<4 x s32>) + ; CI-MESA: [[LSHR21:%[0-9]+]]:_(s32) = G_LSHR [[UV7]], [[C1]](s32) + ; CI-MESA: [[LSHR22:%[0-9]+]]:_(s32) = G_LSHR [[UV7]], [[C2]](s32) + ; CI-MESA: [[LSHR23:%[0-9]+]]:_(s32) = G_LSHR [[UV7]], [[C3]](s32) + ; CI-MESA: [[COPY29:%[0-9]+]]:_(s32) = COPY [[UV7]](s32) + ; CI-MESA: [[COPY30:%[0-9]+]]:_(s32) = COPY [[LSHR21]](s32) + ; CI-MESA: [[COPY31:%[0-9]+]]:_(s32) = COPY [[LSHR22]](s32) + ; CI-MESA: [[COPY32:%[0-9]+]]:_(s32) = COPY [[LSHR23]](s32) + ; CI-MESA: [[BUILD_VECTOR7:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY29]](s32), [[COPY30]](s32), [[COPY31]](s32), [[COPY32]](s32) + ; CI-MESA: [[TRUNC7:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR7]](<4 x s32>) + ; CI-MESA: [[CONCAT_VECTORS:%[0-9]+]]:_(<32 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>), [[TRUNC2]](<4 x s8>), [[TRUNC3]](<4 x s8>), [[TRUNC4]](<4 x s8>), [[TRUNC5]](<4 x s8>), [[TRUNC6]](<4 x s8>), [[TRUNC7]](<4 x s8>) + ; CI-MESA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[CONCAT_VECTORS]](<32 x s8>) ; GFX9-MESA-LABEL: name: test_load_flat_v32s8_align32 ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p0) = COPY $vgpr0_vgpr1 ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p0) :: (load 16, align 32) ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 16 ; GFX9-MESA: [[PTR_ADD:%[0-9]+]]:_(p0) = G_PTR_ADD [[COPY]], [[C]](s64) ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[PTR_ADD]](p0) :: (load 16 + 16) - ; GFX9-MESA: [[CONCAT_VECTORS:%[0-9]+]]:_(<8 x s32>) = G_CONCAT_VECTORS [[LOAD]](<4 x s32>), [[LOAD1]](<4 x s32>) - ; GFX9-MESA: [[BITCAST:%[0-9]+]]:_(<32 x s8>) = G_BITCAST [[CONCAT_VECTORS]](<8 x s32>) - ; GFX9-MESA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[BITCAST]](<32 x s8>) + ; GFX9-MESA: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32), [[UV2:%[0-9]+]]:_(s32), [[UV3:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[LOAD]](<4 x s32>) + ; GFX9-MESA: [[UV4:%[0-9]+]]:_(s32), [[UV5:%[0-9]+]]:_(s32), [[UV6:%[0-9]+]]:_(s32), [[UV7:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[LOAD1]](<4 x s32>) + ; GFX9-MESA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8 + ; GFX9-MESA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C1]](s32) + ; GFX9-MESA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 16 + ; GFX9-MESA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C2]](s32) + ; GFX9-MESA: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 24 + ; GFX9-MESA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C3]](s32) + ; GFX9-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[UV]](s32) + ; GFX9-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY1]](s32), [[COPY2]](s32) + ; GFX9-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32) + ; GFX9-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC1:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY3]](s32), [[COPY4]](s32) + ; GFX9-MESA: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC]](<2 x s16>), [[BUILD_VECTOR_TRUNC1]](<2 x s16>) + ; GFX9-MESA: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS]](<4 x s16>) + ; GFX9-MESA: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C1]](s32) + ; GFX9-MESA: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C2]](s32) + ; GFX9-MESA: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C3]](s32) + ; GFX9-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[UV1]](s32) + ; GFX9-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC2:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY5]](s32), [[COPY6]](s32) + ; GFX9-MESA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32) + ; GFX9-MESA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC3:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY7]](s32), [[COPY8]](s32) + ; GFX9-MESA: [[CONCAT_VECTORS1:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC2]](<2 x s16>), [[BUILD_VECTOR_TRUNC3]](<2 x s16>) + ; GFX9-MESA: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS1]](<4 x s16>) + ; GFX9-MESA: [[LSHR6:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C1]](s32) + ; GFX9-MESA: [[LSHR7:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C2]](s32) + ; GFX9-MESA: [[LSHR8:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C3]](s32) + ; GFX9-MESA: [[COPY9:%[0-9]+]]:_(s32) = COPY [[UV2]](s32) + ; GFX9-MESA: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LSHR6]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC4:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY9]](s32), [[COPY10]](s32) + ; GFX9-MESA: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LSHR7]](s32) + ; GFX9-MESA: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LSHR8]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC5:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY11]](s32), [[COPY12]](s32) + ; GFX9-MESA: [[CONCAT_VECTORS2:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC4]](<2 x s16>), [[BUILD_VECTOR_TRUNC5]](<2 x s16>) + ; GFX9-MESA: [[TRUNC2:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS2]](<4 x s16>) + ; GFX9-MESA: [[LSHR9:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C1]](s32) + ; GFX9-MESA: [[LSHR10:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C2]](s32) + ; GFX9-MESA: [[LSHR11:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C3]](s32) + ; GFX9-MESA: [[COPY13:%[0-9]+]]:_(s32) = COPY [[UV3]](s32) + ; GFX9-MESA: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LSHR9]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC6:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY13]](s32), [[COPY14]](s32) + ; GFX9-MESA: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LSHR10]](s32) + ; GFX9-MESA: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LSHR11]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC7:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY15]](s32), [[COPY16]](s32) + ; GFX9-MESA: [[CONCAT_VECTORS3:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC6]](<2 x s16>), [[BUILD_VECTOR_TRUNC7]](<2 x s16>) + ; GFX9-MESA: [[TRUNC3:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS3]](<4 x s16>) + ; GFX9-MESA: [[LSHR12:%[0-9]+]]:_(s32) = G_LSHR [[UV4]], [[C1]](s32) + ; GFX9-MESA: [[LSHR13:%[0-9]+]]:_(s32) = G_LSHR [[UV4]], [[C2]](s32) + ; GFX9-MESA: [[LSHR14:%[0-9]+]]:_(s32) = G_LSHR [[UV4]], [[C3]](s32) + ; GFX9-MESA: [[COPY17:%[0-9]+]]:_(s32) = COPY [[UV4]](s32) + ; GFX9-MESA: [[COPY18:%[0-9]+]]:_(s32) = COPY [[LSHR12]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC8:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY17]](s32), [[COPY18]](s32) + ; GFX9-MESA: [[COPY19:%[0-9]+]]:_(s32) = COPY [[LSHR13]](s32) + ; GFX9-MESA: [[COPY20:%[0-9]+]]:_(s32) = COPY [[LSHR14]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC9:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY19]](s32), [[COPY20]](s32) + ; GFX9-MESA: [[CONCAT_VECTORS4:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC8]](<2 x s16>), [[BUILD_VECTOR_TRUNC9]](<2 x s16>) + ; GFX9-MESA: [[TRUNC4:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS4]](<4 x s16>) + ; GFX9-MESA: [[LSHR15:%[0-9]+]]:_(s32) = G_LSHR [[UV5]], [[C1]](s32) + ; GFX9-MESA: [[LSHR16:%[0-9]+]]:_(s32) = G_LSHR [[UV5]], [[C2]](s32) + ; GFX9-MESA: [[LSHR17:%[0-9]+]]:_(s32) = G_LSHR [[UV5]], [[C3]](s32) + ; GFX9-MESA: [[COPY21:%[0-9]+]]:_(s32) = COPY [[UV5]](s32) + ; GFX9-MESA: [[COPY22:%[0-9]+]]:_(s32) = COPY [[LSHR15]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC10:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY21]](s32), [[COPY22]](s32) + ; GFX9-MESA: [[COPY23:%[0-9]+]]:_(s32) = COPY [[LSHR16]](s32) + ; GFX9-MESA: [[COPY24:%[0-9]+]]:_(s32) = COPY [[LSHR17]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC11:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY23]](s32), [[COPY24]](s32) + ; GFX9-MESA: [[CONCAT_VECTORS5:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC10]](<2 x s16>), [[BUILD_VECTOR_TRUNC11]](<2 x s16>) + ; GFX9-MESA: [[TRUNC5:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS5]](<4 x s16>) + ; GFX9-MESA: [[LSHR18:%[0-9]+]]:_(s32) = G_LSHR [[UV6]], [[C1]](s32) + ; GFX9-MESA: [[LSHR19:%[0-9]+]]:_(s32) = G_LSHR [[UV6]], [[C2]](s32) + ; GFX9-MESA: [[LSHR20:%[0-9]+]]:_(s32) = G_LSHR [[UV6]], [[C3]](s32) + ; GFX9-MESA: [[COPY25:%[0-9]+]]:_(s32) = COPY [[UV6]](s32) + ; GFX9-MESA: [[COPY26:%[0-9]+]]:_(s32) = COPY [[LSHR18]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC12:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY25]](s32), [[COPY26]](s32) + ; GFX9-MESA: [[COPY27:%[0-9]+]]:_(s32) = COPY [[LSHR19]](s32) + ; GFX9-MESA: [[COPY28:%[0-9]+]]:_(s32) = COPY [[LSHR20]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC13:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY27]](s32), [[COPY28]](s32) + ; GFX9-MESA: [[CONCAT_VECTORS6:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC12]](<2 x s16>), [[BUILD_VECTOR_TRUNC13]](<2 x s16>) + ; GFX9-MESA: [[TRUNC6:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS6]](<4 x s16>) + ; GFX9-MESA: [[LSHR21:%[0-9]+]]:_(s32) = G_LSHR [[UV7]], [[C1]](s32) + ; GFX9-MESA: [[LSHR22:%[0-9]+]]:_(s32) = G_LSHR [[UV7]], [[C2]](s32) + ; GFX9-MESA: [[LSHR23:%[0-9]+]]:_(s32) = G_LSHR [[UV7]], [[C3]](s32) + ; GFX9-MESA: [[COPY29:%[0-9]+]]:_(s32) = COPY [[UV7]](s32) + ; GFX9-MESA: [[COPY30:%[0-9]+]]:_(s32) = COPY [[LSHR21]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC14:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY29]](s32), [[COPY30]](s32) + ; GFX9-MESA: [[COPY31:%[0-9]+]]:_(s32) = COPY [[LSHR22]](s32) + ; GFX9-MESA: [[COPY32:%[0-9]+]]:_(s32) = COPY [[LSHR23]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC15:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY31]](s32), [[COPY32]](s32) + ; GFX9-MESA: [[CONCAT_VECTORS7:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC14]](<2 x s16>), [[BUILD_VECTOR_TRUNC15]](<2 x s16>) + ; GFX9-MESA: [[TRUNC7:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS7]](<4 x s16>) + ; GFX9-MESA: [[CONCAT_VECTORS8:%[0-9]+]]:_(<32 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>), [[TRUNC2]](<4 x s8>), [[TRUNC3]](<4 x s8>), [[TRUNC4]](<4 x s8>), [[TRUNC5]](<4 x s8>), [[TRUNC6]](<4 x s8>), [[TRUNC7]](<4 x s8>) + ; GFX9-MESA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[CONCAT_VECTORS8]](<32 x s8>) %0:_(p0) = COPY $vgpr0_vgpr1 %1:_(<32 x s8>) = G_LOAD %0 :: (load 32, align 32, addrspace 0) $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY %1 Index: llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-load-global.mir =================================================================== --- llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-load-global.mir +++ llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-load-global.mir @@ -4801,33 +4801,173 @@ ; SI-LABEL: name: test_load_global_v8s8_align8 ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1 ; SI: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1) - ; SI: [[BITCAST:%[0-9]+]]:_(<8 x s8>) = G_BITCAST [[LOAD]](<2 x s32>) - ; SI: $vgpr0_vgpr1 = COPY [[BITCAST]](<8 x s8>) + ; SI: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[LOAD]](<2 x s32>) + ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8 + ; SI: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C]](s32) + ; SI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16 + ; SI: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C1]](s32) + ; SI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24 + ; SI: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C2]](s32) + ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[UV]](s32) + ; SI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32) + ; SI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32) + ; SI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32) + ; SI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32) + ; SI: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>) + ; SI: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C]](s32) + ; SI: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C1]](s32) + ; SI: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C2]](s32) + ; SI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[UV1]](s32) + ; SI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32) + ; SI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32) + ; SI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32) + ; SI: [[BUILD_VECTOR1:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY5]](s32), [[COPY6]](s32), [[COPY7]](s32), [[COPY8]](s32) + ; SI: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR1]](<4 x s32>) + ; SI: [[CONCAT_VECTORS:%[0-9]+]]:_(<8 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>) + ; SI: $vgpr0_vgpr1 = COPY [[CONCAT_VECTORS]](<8 x s8>) ; CI-HSA-LABEL: name: test_load_global_v8s8_align8 ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1 ; CI-HSA: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1) - ; CI-HSA: [[BITCAST:%[0-9]+]]:_(<8 x s8>) = G_BITCAST [[LOAD]](<2 x s32>) - ; CI-HSA: $vgpr0_vgpr1 = COPY [[BITCAST]](<8 x s8>) + ; CI-HSA: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[LOAD]](<2 x s32>) + ; CI-HSA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8 + ; CI-HSA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C]](s32) + ; CI-HSA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16 + ; CI-HSA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C1]](s32) + ; CI-HSA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24 + ; CI-HSA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C2]](s32) + ; CI-HSA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[UV]](s32) + ; CI-HSA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32) + ; CI-HSA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32) + ; CI-HSA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32) + ; CI-HSA: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32) + ; CI-HSA: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>) + ; CI-HSA: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C]](s32) + ; CI-HSA: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C1]](s32) + ; CI-HSA: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C2]](s32) + ; CI-HSA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[UV1]](s32) + ; CI-HSA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32) + ; CI-HSA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32) + ; CI-HSA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32) + ; CI-HSA: [[BUILD_VECTOR1:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY5]](s32), [[COPY6]](s32), [[COPY7]](s32), [[COPY8]](s32) + ; CI-HSA: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR1]](<4 x s32>) + ; CI-HSA: [[CONCAT_VECTORS:%[0-9]+]]:_(<8 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>) + ; CI-HSA: $vgpr0_vgpr1 = COPY [[CONCAT_VECTORS]](<8 x s8>) ; CI-MESA-LABEL: name: test_load_global_v8s8_align8 ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1 ; CI-MESA: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1) - ; CI-MESA: [[BITCAST:%[0-9]+]]:_(<8 x s8>) = G_BITCAST [[LOAD]](<2 x s32>) - ; CI-MESA: $vgpr0_vgpr1 = COPY [[BITCAST]](<8 x s8>) + ; CI-MESA: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[LOAD]](<2 x s32>) + ; CI-MESA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8 + ; CI-MESA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C]](s32) + ; CI-MESA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16 + ; CI-MESA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C1]](s32) + ; CI-MESA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24 + ; CI-MESA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C2]](s32) + ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[UV]](s32) + ; CI-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32) + ; CI-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32) + ; CI-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32) + ; CI-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32) + ; CI-MESA: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>) + ; CI-MESA: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C]](s32) + ; CI-MESA: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C1]](s32) + ; CI-MESA: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C2]](s32) + ; CI-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[UV1]](s32) + ; CI-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32) + ; CI-MESA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32) + ; CI-MESA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32) + ; CI-MESA: [[BUILD_VECTOR1:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY5]](s32), [[COPY6]](s32), [[COPY7]](s32), [[COPY8]](s32) + ; CI-MESA: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR1]](<4 x s32>) + ; CI-MESA: [[CONCAT_VECTORS:%[0-9]+]]:_(<8 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>) + ; CI-MESA: $vgpr0_vgpr1 = COPY [[CONCAT_VECTORS]](<8 x s8>) ; VI-LABEL: name: test_load_global_v8s8_align8 ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1 ; VI: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1) - ; VI: [[BITCAST:%[0-9]+]]:_(<8 x s8>) = G_BITCAST [[LOAD]](<2 x s32>) - ; VI: $vgpr0_vgpr1 = COPY [[BITCAST]](<8 x s8>) + ; VI: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[LOAD]](<2 x s32>) + ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8 + ; VI: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C]](s32) + ; VI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16 + ; VI: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C1]](s32) + ; VI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24 + ; VI: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C2]](s32) + ; VI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[UV]](s32) + ; VI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32) + ; VI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32) + ; VI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32) + ; VI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32) + ; VI: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>) + ; VI: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C]](s32) + ; VI: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C1]](s32) + ; VI: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C2]](s32) + ; VI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[UV1]](s32) + ; VI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32) + ; VI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32) + ; VI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32) + ; VI: [[BUILD_VECTOR1:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY5]](s32), [[COPY6]](s32), [[COPY7]](s32), [[COPY8]](s32) + ; VI: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR1]](<4 x s32>) + ; VI: [[CONCAT_VECTORS:%[0-9]+]]:_(<8 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>) + ; VI: $vgpr0_vgpr1 = COPY [[CONCAT_VECTORS]](<8 x s8>) ; GFX9-HSA-LABEL: name: test_load_global_v8s8_align8 ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1 ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1) - ; GFX9-HSA: [[BITCAST:%[0-9]+]]:_(<8 x s8>) = G_BITCAST [[LOAD]](<2 x s32>) - ; GFX9-HSA: $vgpr0_vgpr1 = COPY [[BITCAST]](<8 x s8>) + ; GFX9-HSA: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[LOAD]](<2 x s32>) + ; GFX9-HSA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8 + ; GFX9-HSA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C]](s32) + ; GFX9-HSA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16 + ; GFX9-HSA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C1]](s32) + ; GFX9-HSA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24 + ; GFX9-HSA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C2]](s32) + ; GFX9-HSA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[UV]](s32) + ; GFX9-HSA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32) + ; GFX9-HSA: [[BUILD_VECTOR_TRUNC:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY1]](s32), [[COPY2]](s32) + ; GFX9-HSA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32) + ; GFX9-HSA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32) + ; GFX9-HSA: [[BUILD_VECTOR_TRUNC1:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY3]](s32), [[COPY4]](s32) + ; GFX9-HSA: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC]](<2 x s16>), [[BUILD_VECTOR_TRUNC1]](<2 x s16>) + ; GFX9-HSA: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS]](<4 x s16>) + ; GFX9-HSA: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C]](s32) + ; GFX9-HSA: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C1]](s32) + ; GFX9-HSA: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C2]](s32) + ; GFX9-HSA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[UV1]](s32) + ; GFX9-HSA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32) + ; GFX9-HSA: [[BUILD_VECTOR_TRUNC2:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY5]](s32), [[COPY6]](s32) + ; GFX9-HSA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32) + ; GFX9-HSA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32) + ; GFX9-HSA: [[BUILD_VECTOR_TRUNC3:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY7]](s32), [[COPY8]](s32) + ; GFX9-HSA: [[CONCAT_VECTORS1:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC2]](<2 x s16>), [[BUILD_VECTOR_TRUNC3]](<2 x s16>) + ; GFX9-HSA: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS1]](<4 x s16>) + ; GFX9-HSA: [[CONCAT_VECTORS2:%[0-9]+]]:_(<8 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>) + ; GFX9-HSA: $vgpr0_vgpr1 = COPY [[CONCAT_VECTORS2]](<8 x s8>) ; GFX9-MESA-LABEL: name: test_load_global_v8s8_align8 ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1 ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p1) :: (load 8, addrspace 1) - ; GFX9-MESA: [[BITCAST:%[0-9]+]]:_(<8 x s8>) = G_BITCAST [[LOAD]](<2 x s32>) - ; GFX9-MESA: $vgpr0_vgpr1 = COPY [[BITCAST]](<8 x s8>) + ; GFX9-MESA: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[LOAD]](<2 x s32>) + ; GFX9-MESA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8 + ; GFX9-MESA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C]](s32) + ; GFX9-MESA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16 + ; GFX9-MESA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C1]](s32) + ; GFX9-MESA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24 + ; GFX9-MESA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C2]](s32) + ; GFX9-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[UV]](s32) + ; GFX9-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY1]](s32), [[COPY2]](s32) + ; GFX9-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32) + ; GFX9-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC1:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY3]](s32), [[COPY4]](s32) + ; GFX9-MESA: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC]](<2 x s16>), [[BUILD_VECTOR_TRUNC1]](<2 x s16>) + ; GFX9-MESA: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS]](<4 x s16>) + ; GFX9-MESA: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C]](s32) + ; GFX9-MESA: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C1]](s32) + ; GFX9-MESA: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C2]](s32) + ; GFX9-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[UV1]](s32) + ; GFX9-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC2:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY5]](s32), [[COPY6]](s32) + ; GFX9-MESA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32) + ; GFX9-MESA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC3:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY7]](s32), [[COPY8]](s32) + ; GFX9-MESA: [[CONCAT_VECTORS1:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC2]](<2 x s16>), [[BUILD_VECTOR_TRUNC3]](<2 x s16>) + ; GFX9-MESA: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS1]](<4 x s16>) + ; GFX9-MESA: [[CONCAT_VECTORS2:%[0-9]+]]:_(<8 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>) + ; GFX9-MESA: $vgpr0_vgpr1 = COPY [[CONCAT_VECTORS2]](<8 x s8>) %0:_(p1) = COPY $vgpr0_vgpr1 %1:_(<8 x s8>) = G_LOAD %0 :: (load 8, align 8, addrspace 1) $vgpr0_vgpr1 = COPY %1 @@ -4842,33 +4982,289 @@ ; SI-LABEL: name: test_load_global_v16s8_align16 ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1 ; SI: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1) - ; SI: [[BITCAST:%[0-9]+]]:_(<16 x s8>) = G_BITCAST [[LOAD]](<4 x s32>) - ; SI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](<16 x s8>) + ; SI: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32), [[UV2:%[0-9]+]]:_(s32), [[UV3:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[LOAD]](<4 x s32>) + ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8 + ; SI: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C]](s32) + ; SI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16 + ; SI: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C1]](s32) + ; SI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24 + ; SI: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C2]](s32) + ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[UV]](s32) + ; SI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32) + ; SI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32) + ; SI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32) + ; SI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32) + ; SI: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>) + ; SI: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C]](s32) + ; SI: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C1]](s32) + ; SI: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C2]](s32) + ; SI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[UV1]](s32) + ; SI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32) + ; SI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32) + ; SI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32) + ; SI: [[BUILD_VECTOR1:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY5]](s32), [[COPY6]](s32), [[COPY7]](s32), [[COPY8]](s32) + ; SI: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR1]](<4 x s32>) + ; SI: [[LSHR6:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C]](s32) + ; SI: [[LSHR7:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C1]](s32) + ; SI: [[LSHR8:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C2]](s32) + ; SI: [[COPY9:%[0-9]+]]:_(s32) = COPY [[UV2]](s32) + ; SI: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LSHR6]](s32) + ; SI: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LSHR7]](s32) + ; SI: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LSHR8]](s32) + ; SI: [[BUILD_VECTOR2:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY9]](s32), [[COPY10]](s32), [[COPY11]](s32), [[COPY12]](s32) + ; SI: [[TRUNC2:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR2]](<4 x s32>) + ; SI: [[LSHR9:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C]](s32) + ; SI: [[LSHR10:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C1]](s32) + ; SI: [[LSHR11:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C2]](s32) + ; SI: [[COPY13:%[0-9]+]]:_(s32) = COPY [[UV3]](s32) + ; SI: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LSHR9]](s32) + ; SI: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LSHR10]](s32) + ; SI: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LSHR11]](s32) + ; SI: [[BUILD_VECTOR3:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY13]](s32), [[COPY14]](s32), [[COPY15]](s32), [[COPY16]](s32) + ; SI: [[TRUNC3:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR3]](<4 x s32>) + ; SI: [[CONCAT_VECTORS:%[0-9]+]]:_(<16 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>), [[TRUNC2]](<4 x s8>), [[TRUNC3]](<4 x s8>) + ; SI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[CONCAT_VECTORS]](<16 x s8>) ; CI-HSA-LABEL: name: test_load_global_v16s8_align16 ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1 ; CI-HSA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1) - ; CI-HSA: [[BITCAST:%[0-9]+]]:_(<16 x s8>) = G_BITCAST [[LOAD]](<4 x s32>) - ; CI-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](<16 x s8>) + ; CI-HSA: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32), [[UV2:%[0-9]+]]:_(s32), [[UV3:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[LOAD]](<4 x s32>) + ; CI-HSA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8 + ; CI-HSA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C]](s32) + ; CI-HSA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16 + ; CI-HSA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C1]](s32) + ; CI-HSA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24 + ; CI-HSA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C2]](s32) + ; CI-HSA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[UV]](s32) + ; CI-HSA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32) + ; CI-HSA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32) + ; CI-HSA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32) + ; CI-HSA: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32) + ; CI-HSA: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>) + ; CI-HSA: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C]](s32) + ; CI-HSA: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C1]](s32) + ; CI-HSA: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C2]](s32) + ; CI-HSA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[UV1]](s32) + ; CI-HSA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32) + ; CI-HSA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32) + ; CI-HSA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32) + ; CI-HSA: [[BUILD_VECTOR1:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY5]](s32), [[COPY6]](s32), [[COPY7]](s32), [[COPY8]](s32) + ; CI-HSA: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR1]](<4 x s32>) + ; CI-HSA: [[LSHR6:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C]](s32) + ; CI-HSA: [[LSHR7:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C1]](s32) + ; CI-HSA: [[LSHR8:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C2]](s32) + ; CI-HSA: [[COPY9:%[0-9]+]]:_(s32) = COPY [[UV2]](s32) + ; CI-HSA: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LSHR6]](s32) + ; CI-HSA: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LSHR7]](s32) + ; CI-HSA: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LSHR8]](s32) + ; CI-HSA: [[BUILD_VECTOR2:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY9]](s32), [[COPY10]](s32), [[COPY11]](s32), [[COPY12]](s32) + ; CI-HSA: [[TRUNC2:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR2]](<4 x s32>) + ; CI-HSA: [[LSHR9:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C]](s32) + ; CI-HSA: [[LSHR10:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C1]](s32) + ; CI-HSA: [[LSHR11:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C2]](s32) + ; CI-HSA: [[COPY13:%[0-9]+]]:_(s32) = COPY [[UV3]](s32) + ; CI-HSA: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LSHR9]](s32) + ; CI-HSA: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LSHR10]](s32) + ; CI-HSA: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LSHR11]](s32) + ; CI-HSA: [[BUILD_VECTOR3:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY13]](s32), [[COPY14]](s32), [[COPY15]](s32), [[COPY16]](s32) + ; CI-HSA: [[TRUNC3:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR3]](<4 x s32>) + ; CI-HSA: [[CONCAT_VECTORS:%[0-9]+]]:_(<16 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>), [[TRUNC2]](<4 x s8>), [[TRUNC3]](<4 x s8>) + ; CI-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[CONCAT_VECTORS]](<16 x s8>) ; CI-MESA-LABEL: name: test_load_global_v16s8_align16 ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1 ; CI-MESA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1) - ; CI-MESA: [[BITCAST:%[0-9]+]]:_(<16 x s8>) = G_BITCAST [[LOAD]](<4 x s32>) - ; CI-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](<16 x s8>) + ; CI-MESA: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32), [[UV2:%[0-9]+]]:_(s32), [[UV3:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[LOAD]](<4 x s32>) + ; CI-MESA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8 + ; CI-MESA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C]](s32) + ; CI-MESA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16 + ; CI-MESA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C1]](s32) + ; CI-MESA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24 + ; CI-MESA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C2]](s32) + ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[UV]](s32) + ; CI-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32) + ; CI-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32) + ; CI-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32) + ; CI-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32) + ; CI-MESA: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>) + ; CI-MESA: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C]](s32) + ; CI-MESA: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C1]](s32) + ; CI-MESA: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C2]](s32) + ; CI-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[UV1]](s32) + ; CI-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32) + ; CI-MESA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32) + ; CI-MESA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32) + ; CI-MESA: [[BUILD_VECTOR1:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY5]](s32), [[COPY6]](s32), [[COPY7]](s32), [[COPY8]](s32) + ; CI-MESA: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR1]](<4 x s32>) + ; CI-MESA: [[LSHR6:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C]](s32) + ; CI-MESA: [[LSHR7:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C1]](s32) + ; CI-MESA: [[LSHR8:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C2]](s32) + ; CI-MESA: [[COPY9:%[0-9]+]]:_(s32) = COPY [[UV2]](s32) + ; CI-MESA: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LSHR6]](s32) + ; CI-MESA: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LSHR7]](s32) + ; CI-MESA: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LSHR8]](s32) + ; CI-MESA: [[BUILD_VECTOR2:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY9]](s32), [[COPY10]](s32), [[COPY11]](s32), [[COPY12]](s32) + ; CI-MESA: [[TRUNC2:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR2]](<4 x s32>) + ; CI-MESA: [[LSHR9:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C]](s32) + ; CI-MESA: [[LSHR10:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C1]](s32) + ; CI-MESA: [[LSHR11:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C2]](s32) + ; CI-MESA: [[COPY13:%[0-9]+]]:_(s32) = COPY [[UV3]](s32) + ; CI-MESA: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LSHR9]](s32) + ; CI-MESA: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LSHR10]](s32) + ; CI-MESA: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LSHR11]](s32) + ; CI-MESA: [[BUILD_VECTOR3:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY13]](s32), [[COPY14]](s32), [[COPY15]](s32), [[COPY16]](s32) + ; CI-MESA: [[TRUNC3:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR3]](<4 x s32>) + ; CI-MESA: [[CONCAT_VECTORS:%[0-9]+]]:_(<16 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>), [[TRUNC2]](<4 x s8>), [[TRUNC3]](<4 x s8>) + ; CI-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[CONCAT_VECTORS]](<16 x s8>) ; VI-LABEL: name: test_load_global_v16s8_align16 ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1 ; VI: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1) - ; VI: [[BITCAST:%[0-9]+]]:_(<16 x s8>) = G_BITCAST [[LOAD]](<4 x s32>) - ; VI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](<16 x s8>) + ; VI: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32), [[UV2:%[0-9]+]]:_(s32), [[UV3:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[LOAD]](<4 x s32>) + ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8 + ; VI: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C]](s32) + ; VI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16 + ; VI: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C1]](s32) + ; VI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24 + ; VI: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C2]](s32) + ; VI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[UV]](s32) + ; VI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32) + ; VI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32) + ; VI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32) + ; VI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32) + ; VI: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>) + ; VI: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C]](s32) + ; VI: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C1]](s32) + ; VI: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C2]](s32) + ; VI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[UV1]](s32) + ; VI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32) + ; VI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32) + ; VI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32) + ; VI: [[BUILD_VECTOR1:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY5]](s32), [[COPY6]](s32), [[COPY7]](s32), [[COPY8]](s32) + ; VI: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR1]](<4 x s32>) + ; VI: [[LSHR6:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C]](s32) + ; VI: [[LSHR7:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C1]](s32) + ; VI: [[LSHR8:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C2]](s32) + ; VI: [[COPY9:%[0-9]+]]:_(s32) = COPY [[UV2]](s32) + ; VI: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LSHR6]](s32) + ; VI: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LSHR7]](s32) + ; VI: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LSHR8]](s32) + ; VI: [[BUILD_VECTOR2:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY9]](s32), [[COPY10]](s32), [[COPY11]](s32), [[COPY12]](s32) + ; VI: [[TRUNC2:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR2]](<4 x s32>) + ; VI: [[LSHR9:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C]](s32) + ; VI: [[LSHR10:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C1]](s32) + ; VI: [[LSHR11:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C2]](s32) + ; VI: [[COPY13:%[0-9]+]]:_(s32) = COPY [[UV3]](s32) + ; VI: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LSHR9]](s32) + ; VI: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LSHR10]](s32) + ; VI: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LSHR11]](s32) + ; VI: [[BUILD_VECTOR3:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY13]](s32), [[COPY14]](s32), [[COPY15]](s32), [[COPY16]](s32) + ; VI: [[TRUNC3:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR3]](<4 x s32>) + ; VI: [[CONCAT_VECTORS:%[0-9]+]]:_(<16 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>), [[TRUNC2]](<4 x s8>), [[TRUNC3]](<4 x s8>) + ; VI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[CONCAT_VECTORS]](<16 x s8>) ; GFX9-HSA-LABEL: name: test_load_global_v16s8_align16 ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1 ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1) - ; GFX9-HSA: [[BITCAST:%[0-9]+]]:_(<16 x s8>) = G_BITCAST [[LOAD]](<4 x s32>) - ; GFX9-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](<16 x s8>) + ; GFX9-HSA: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32), [[UV2:%[0-9]+]]:_(s32), [[UV3:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[LOAD]](<4 x s32>) + ; GFX9-HSA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8 + ; GFX9-HSA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C]](s32) + ; GFX9-HSA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16 + ; GFX9-HSA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C1]](s32) + ; GFX9-HSA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24 + ; GFX9-HSA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C2]](s32) + ; GFX9-HSA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[UV]](s32) + ; GFX9-HSA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32) + ; GFX9-HSA: [[BUILD_VECTOR_TRUNC:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY1]](s32), [[COPY2]](s32) + ; GFX9-HSA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32) + ; GFX9-HSA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32) + ; GFX9-HSA: [[BUILD_VECTOR_TRUNC1:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY3]](s32), [[COPY4]](s32) + ; GFX9-HSA: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC]](<2 x s16>), [[BUILD_VECTOR_TRUNC1]](<2 x s16>) + ; GFX9-HSA: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS]](<4 x s16>) + ; GFX9-HSA: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C]](s32) + ; GFX9-HSA: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C1]](s32) + ; GFX9-HSA: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C2]](s32) + ; GFX9-HSA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[UV1]](s32) + ; GFX9-HSA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32) + ; GFX9-HSA: [[BUILD_VECTOR_TRUNC2:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY5]](s32), [[COPY6]](s32) + ; GFX9-HSA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32) + ; GFX9-HSA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32) + ; GFX9-HSA: [[BUILD_VECTOR_TRUNC3:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY7]](s32), [[COPY8]](s32) + ; GFX9-HSA: [[CONCAT_VECTORS1:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC2]](<2 x s16>), [[BUILD_VECTOR_TRUNC3]](<2 x s16>) + ; GFX9-HSA: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS1]](<4 x s16>) + ; GFX9-HSA: [[LSHR6:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C]](s32) + ; GFX9-HSA: [[LSHR7:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C1]](s32) + ; GFX9-HSA: [[LSHR8:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C2]](s32) + ; GFX9-HSA: [[COPY9:%[0-9]+]]:_(s32) = COPY [[UV2]](s32) + ; GFX9-HSA: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LSHR6]](s32) + ; GFX9-HSA: [[BUILD_VECTOR_TRUNC4:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY9]](s32), [[COPY10]](s32) + ; GFX9-HSA: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LSHR7]](s32) + ; GFX9-HSA: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LSHR8]](s32) + ; GFX9-HSA: [[BUILD_VECTOR_TRUNC5:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY11]](s32), [[COPY12]](s32) + ; GFX9-HSA: [[CONCAT_VECTORS2:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC4]](<2 x s16>), [[BUILD_VECTOR_TRUNC5]](<2 x s16>) + ; GFX9-HSA: [[TRUNC2:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS2]](<4 x s16>) + ; GFX9-HSA: [[LSHR9:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C]](s32) + ; GFX9-HSA: [[LSHR10:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C1]](s32) + ; GFX9-HSA: [[LSHR11:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C2]](s32) + ; GFX9-HSA: [[COPY13:%[0-9]+]]:_(s32) = COPY [[UV3]](s32) + ; GFX9-HSA: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LSHR9]](s32) + ; GFX9-HSA: [[BUILD_VECTOR_TRUNC6:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY13]](s32), [[COPY14]](s32) + ; GFX9-HSA: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LSHR10]](s32) + ; GFX9-HSA: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LSHR11]](s32) + ; GFX9-HSA: [[BUILD_VECTOR_TRUNC7:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY15]](s32), [[COPY16]](s32) + ; GFX9-HSA: [[CONCAT_VECTORS3:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC6]](<2 x s16>), [[BUILD_VECTOR_TRUNC7]](<2 x s16>) + ; GFX9-HSA: [[TRUNC3:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS3]](<4 x s16>) + ; GFX9-HSA: [[CONCAT_VECTORS4:%[0-9]+]]:_(<16 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>), [[TRUNC2]](<4 x s8>), [[TRUNC3]](<4 x s8>) + ; GFX9-HSA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[CONCAT_VECTORS4]](<16 x s8>) ; GFX9-MESA-LABEL: name: test_load_global_v16s8_align16 ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1 ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<4 x s32>) = G_LOAD [[COPY]](p1) :: (load 16, addrspace 1) - ; GFX9-MESA: [[BITCAST:%[0-9]+]]:_(<16 x s8>) = G_BITCAST [[LOAD]](<4 x s32>) - ; GFX9-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](<16 x s8>) + ; GFX9-MESA: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32), [[UV2:%[0-9]+]]:_(s32), [[UV3:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[LOAD]](<4 x s32>) + ; GFX9-MESA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8 + ; GFX9-MESA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C]](s32) + ; GFX9-MESA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16 + ; GFX9-MESA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C1]](s32) + ; GFX9-MESA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24 + ; GFX9-MESA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C2]](s32) + ; GFX9-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[UV]](s32) + ; GFX9-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY1]](s32), [[COPY2]](s32) + ; GFX9-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32) + ; GFX9-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC1:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY3]](s32), [[COPY4]](s32) + ; GFX9-MESA: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC]](<2 x s16>), [[BUILD_VECTOR_TRUNC1]](<2 x s16>) + ; GFX9-MESA: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS]](<4 x s16>) + ; GFX9-MESA: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C]](s32) + ; GFX9-MESA: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C1]](s32) + ; GFX9-MESA: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C2]](s32) + ; GFX9-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[UV1]](s32) + ; GFX9-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC2:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY5]](s32), [[COPY6]](s32) + ; GFX9-MESA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32) + ; GFX9-MESA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC3:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY7]](s32), [[COPY8]](s32) + ; GFX9-MESA: [[CONCAT_VECTORS1:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC2]](<2 x s16>), [[BUILD_VECTOR_TRUNC3]](<2 x s16>) + ; GFX9-MESA: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS1]](<4 x s16>) + ; GFX9-MESA: [[LSHR6:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C]](s32) + ; GFX9-MESA: [[LSHR7:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C1]](s32) + ; GFX9-MESA: [[LSHR8:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C2]](s32) + ; GFX9-MESA: [[COPY9:%[0-9]+]]:_(s32) = COPY [[UV2]](s32) + ; GFX9-MESA: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LSHR6]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC4:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY9]](s32), [[COPY10]](s32) + ; GFX9-MESA: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LSHR7]](s32) + ; GFX9-MESA: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LSHR8]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC5:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY11]](s32), [[COPY12]](s32) + ; GFX9-MESA: [[CONCAT_VECTORS2:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC4]](<2 x s16>), [[BUILD_VECTOR_TRUNC5]](<2 x s16>) + ; GFX9-MESA: [[TRUNC2:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS2]](<4 x s16>) + ; GFX9-MESA: [[LSHR9:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C]](s32) + ; GFX9-MESA: [[LSHR10:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C1]](s32) + ; GFX9-MESA: [[LSHR11:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C2]](s32) + ; GFX9-MESA: [[COPY13:%[0-9]+]]:_(s32) = COPY [[UV3]](s32) + ; GFX9-MESA: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LSHR9]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC6:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY13]](s32), [[COPY14]](s32) + ; GFX9-MESA: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LSHR10]](s32) + ; GFX9-MESA: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LSHR11]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC7:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY15]](s32), [[COPY16]](s32) + ; GFX9-MESA: [[CONCAT_VECTORS3:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC6]](<2 x s16>), [[BUILD_VECTOR_TRUNC7]](<2 x s16>) + ; GFX9-MESA: [[TRUNC3:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS3]](<4 x s16>) + ; GFX9-MESA: [[CONCAT_VECTORS4:%[0-9]+]]:_(<16 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>), [[TRUNC2]](<4 x s8>), [[TRUNC3]](<4 x s8>) + ; GFX9-MESA: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[CONCAT_VECTORS4]](<16 x s8>) %0:_(p1) = COPY $vgpr0_vgpr1 %1:_(<16 x s8>) = G_LOAD %0 :: (load 16, align 16, addrspace 1) $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %1 @@ -4883,33 +5279,521 @@ ; SI-LABEL: name: test_load_global_v32s8_align32 ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1 ; SI: [[LOAD:%[0-9]+]]:_(<8 x s32>) = G_LOAD [[COPY]](p1) :: (load 32, addrspace 1) - ; SI: [[BITCAST:%[0-9]+]]:_(<32 x s8>) = G_BITCAST [[LOAD]](<8 x s32>) - ; SI: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[BITCAST]](<32 x s8>) + ; SI: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32), [[UV2:%[0-9]+]]:_(s32), [[UV3:%[0-9]+]]:_(s32), [[UV4:%[0-9]+]]:_(s32), [[UV5:%[0-9]+]]:_(s32), [[UV6:%[0-9]+]]:_(s32), [[UV7:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[LOAD]](<8 x s32>) + ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8 + ; SI: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C]](s32) + ; SI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16 + ; SI: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C1]](s32) + ; SI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24 + ; SI: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C2]](s32) + ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[UV]](s32) + ; SI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32) + ; SI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32) + ; SI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32) + ; SI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32) + ; SI: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>) + ; SI: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C]](s32) + ; SI: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C1]](s32) + ; SI: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C2]](s32) + ; SI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[UV1]](s32) + ; SI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32) + ; SI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32) + ; SI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32) + ; SI: [[BUILD_VECTOR1:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY5]](s32), [[COPY6]](s32), [[COPY7]](s32), [[COPY8]](s32) + ; SI: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR1]](<4 x s32>) + ; SI: [[LSHR6:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C]](s32) + ; SI: [[LSHR7:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C1]](s32) + ; SI: [[LSHR8:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C2]](s32) + ; SI: [[COPY9:%[0-9]+]]:_(s32) = COPY [[UV2]](s32) + ; SI: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LSHR6]](s32) + ; SI: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LSHR7]](s32) + ; SI: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LSHR8]](s32) + ; SI: [[BUILD_VECTOR2:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY9]](s32), [[COPY10]](s32), [[COPY11]](s32), [[COPY12]](s32) + ; SI: [[TRUNC2:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR2]](<4 x s32>) + ; SI: [[LSHR9:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C]](s32) + ; SI: [[LSHR10:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C1]](s32) + ; SI: [[LSHR11:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C2]](s32) + ; SI: [[COPY13:%[0-9]+]]:_(s32) = COPY [[UV3]](s32) + ; SI: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LSHR9]](s32) + ; SI: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LSHR10]](s32) + ; SI: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LSHR11]](s32) + ; SI: [[BUILD_VECTOR3:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY13]](s32), [[COPY14]](s32), [[COPY15]](s32), [[COPY16]](s32) + ; SI: [[TRUNC3:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR3]](<4 x s32>) + ; SI: [[LSHR12:%[0-9]+]]:_(s32) = G_LSHR [[UV4]], [[C]](s32) + ; SI: [[LSHR13:%[0-9]+]]:_(s32) = G_LSHR [[UV4]], [[C1]](s32) + ; SI: [[LSHR14:%[0-9]+]]:_(s32) = G_LSHR [[UV4]], [[C2]](s32) + ; SI: [[COPY17:%[0-9]+]]:_(s32) = COPY [[UV4]](s32) + ; SI: [[COPY18:%[0-9]+]]:_(s32) = COPY [[LSHR12]](s32) + ; SI: [[COPY19:%[0-9]+]]:_(s32) = COPY [[LSHR13]](s32) + ; SI: [[COPY20:%[0-9]+]]:_(s32) = COPY [[LSHR14]](s32) + ; SI: [[BUILD_VECTOR4:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY17]](s32), [[COPY18]](s32), [[COPY19]](s32), [[COPY20]](s32) + ; SI: [[TRUNC4:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR4]](<4 x s32>) + ; SI: [[LSHR15:%[0-9]+]]:_(s32) = G_LSHR [[UV5]], [[C]](s32) + ; SI: [[LSHR16:%[0-9]+]]:_(s32) = G_LSHR [[UV5]], [[C1]](s32) + ; SI: [[LSHR17:%[0-9]+]]:_(s32) = G_LSHR [[UV5]], [[C2]](s32) + ; SI: [[COPY21:%[0-9]+]]:_(s32) = COPY [[UV5]](s32) + ; SI: [[COPY22:%[0-9]+]]:_(s32) = COPY [[LSHR15]](s32) + ; SI: [[COPY23:%[0-9]+]]:_(s32) = COPY [[LSHR16]](s32) + ; SI: [[COPY24:%[0-9]+]]:_(s32) = COPY [[LSHR17]](s32) + ; SI: [[BUILD_VECTOR5:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY21]](s32), [[COPY22]](s32), [[COPY23]](s32), [[COPY24]](s32) + ; SI: [[TRUNC5:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR5]](<4 x s32>) + ; SI: [[LSHR18:%[0-9]+]]:_(s32) = G_LSHR [[UV6]], [[C]](s32) + ; SI: [[LSHR19:%[0-9]+]]:_(s32) = G_LSHR [[UV6]], [[C1]](s32) + ; SI: [[LSHR20:%[0-9]+]]:_(s32) = G_LSHR [[UV6]], [[C2]](s32) + ; SI: [[COPY25:%[0-9]+]]:_(s32) = COPY [[UV6]](s32) + ; SI: [[COPY26:%[0-9]+]]:_(s32) = COPY [[LSHR18]](s32) + ; SI: [[COPY27:%[0-9]+]]:_(s32) = COPY [[LSHR19]](s32) + ; SI: [[COPY28:%[0-9]+]]:_(s32) = COPY [[LSHR20]](s32) + ; SI: [[BUILD_VECTOR6:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY25]](s32), [[COPY26]](s32), [[COPY27]](s32), [[COPY28]](s32) + ; SI: [[TRUNC6:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR6]](<4 x s32>) + ; SI: [[LSHR21:%[0-9]+]]:_(s32) = G_LSHR [[UV7]], [[C]](s32) + ; SI: [[LSHR22:%[0-9]+]]:_(s32) = G_LSHR [[UV7]], [[C1]](s32) + ; SI: [[LSHR23:%[0-9]+]]:_(s32) = G_LSHR [[UV7]], [[C2]](s32) + ; SI: [[COPY29:%[0-9]+]]:_(s32) = COPY [[UV7]](s32) + ; SI: [[COPY30:%[0-9]+]]:_(s32) = COPY [[LSHR21]](s32) + ; SI: [[COPY31:%[0-9]+]]:_(s32) = COPY [[LSHR22]](s32) + ; SI: [[COPY32:%[0-9]+]]:_(s32) = COPY [[LSHR23]](s32) + ; SI: [[BUILD_VECTOR7:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY29]](s32), [[COPY30]](s32), [[COPY31]](s32), [[COPY32]](s32) + ; SI: [[TRUNC7:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR7]](<4 x s32>) + ; SI: [[CONCAT_VECTORS:%[0-9]+]]:_(<32 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>), [[TRUNC2]](<4 x s8>), [[TRUNC3]](<4 x s8>), [[TRUNC4]](<4 x s8>), [[TRUNC5]](<4 x s8>), [[TRUNC6]](<4 x s8>), [[TRUNC7]](<4 x s8>) + ; SI: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[CONCAT_VECTORS]](<32 x s8>) ; CI-HSA-LABEL: name: test_load_global_v32s8_align32 ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1 ; CI-HSA: [[LOAD:%[0-9]+]]:_(<8 x s32>) = G_LOAD [[COPY]](p1) :: (load 32, addrspace 1) - ; CI-HSA: [[BITCAST:%[0-9]+]]:_(<32 x s8>) = G_BITCAST [[LOAD]](<8 x s32>) - ; CI-HSA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[BITCAST]](<32 x s8>) + ; CI-HSA: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32), [[UV2:%[0-9]+]]:_(s32), [[UV3:%[0-9]+]]:_(s32), [[UV4:%[0-9]+]]:_(s32), [[UV5:%[0-9]+]]:_(s32), [[UV6:%[0-9]+]]:_(s32), [[UV7:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[LOAD]](<8 x s32>) + ; CI-HSA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8 + ; CI-HSA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C]](s32) + ; CI-HSA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16 + ; CI-HSA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C1]](s32) + ; CI-HSA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24 + ; CI-HSA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C2]](s32) + ; CI-HSA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[UV]](s32) + ; CI-HSA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32) + ; CI-HSA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32) + ; CI-HSA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32) + ; CI-HSA: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32) + ; CI-HSA: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>) + ; CI-HSA: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C]](s32) + ; CI-HSA: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C1]](s32) + ; CI-HSA: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C2]](s32) + ; CI-HSA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[UV1]](s32) + ; CI-HSA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32) + ; CI-HSA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32) + ; CI-HSA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32) + ; CI-HSA: [[BUILD_VECTOR1:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY5]](s32), [[COPY6]](s32), [[COPY7]](s32), [[COPY8]](s32) + ; CI-HSA: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR1]](<4 x s32>) + ; CI-HSA: [[LSHR6:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C]](s32) + ; CI-HSA: [[LSHR7:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C1]](s32) + ; CI-HSA: [[LSHR8:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C2]](s32) + ; CI-HSA: [[COPY9:%[0-9]+]]:_(s32) = COPY [[UV2]](s32) + ; CI-HSA: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LSHR6]](s32) + ; CI-HSA: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LSHR7]](s32) + ; CI-HSA: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LSHR8]](s32) + ; CI-HSA: [[BUILD_VECTOR2:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY9]](s32), [[COPY10]](s32), [[COPY11]](s32), [[COPY12]](s32) + ; CI-HSA: [[TRUNC2:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR2]](<4 x s32>) + ; CI-HSA: [[LSHR9:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C]](s32) + ; CI-HSA: [[LSHR10:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C1]](s32) + ; CI-HSA: [[LSHR11:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C2]](s32) + ; CI-HSA: [[COPY13:%[0-9]+]]:_(s32) = COPY [[UV3]](s32) + ; CI-HSA: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LSHR9]](s32) + ; CI-HSA: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LSHR10]](s32) + ; CI-HSA: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LSHR11]](s32) + ; CI-HSA: [[BUILD_VECTOR3:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY13]](s32), [[COPY14]](s32), [[COPY15]](s32), [[COPY16]](s32) + ; CI-HSA: [[TRUNC3:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR3]](<4 x s32>) + ; CI-HSA: [[LSHR12:%[0-9]+]]:_(s32) = G_LSHR [[UV4]], [[C]](s32) + ; CI-HSA: [[LSHR13:%[0-9]+]]:_(s32) = G_LSHR [[UV4]], [[C1]](s32) + ; CI-HSA: [[LSHR14:%[0-9]+]]:_(s32) = G_LSHR [[UV4]], [[C2]](s32) + ; CI-HSA: [[COPY17:%[0-9]+]]:_(s32) = COPY [[UV4]](s32) + ; CI-HSA: [[COPY18:%[0-9]+]]:_(s32) = COPY [[LSHR12]](s32) + ; CI-HSA: [[COPY19:%[0-9]+]]:_(s32) = COPY [[LSHR13]](s32) + ; CI-HSA: [[COPY20:%[0-9]+]]:_(s32) = COPY [[LSHR14]](s32) + ; CI-HSA: [[BUILD_VECTOR4:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY17]](s32), [[COPY18]](s32), [[COPY19]](s32), [[COPY20]](s32) + ; CI-HSA: [[TRUNC4:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR4]](<4 x s32>) + ; CI-HSA: [[LSHR15:%[0-9]+]]:_(s32) = G_LSHR [[UV5]], [[C]](s32) + ; CI-HSA: [[LSHR16:%[0-9]+]]:_(s32) = G_LSHR [[UV5]], [[C1]](s32) + ; CI-HSA: [[LSHR17:%[0-9]+]]:_(s32) = G_LSHR [[UV5]], [[C2]](s32) + ; CI-HSA: [[COPY21:%[0-9]+]]:_(s32) = COPY [[UV5]](s32) + ; CI-HSA: [[COPY22:%[0-9]+]]:_(s32) = COPY [[LSHR15]](s32) + ; CI-HSA: [[COPY23:%[0-9]+]]:_(s32) = COPY [[LSHR16]](s32) + ; CI-HSA: [[COPY24:%[0-9]+]]:_(s32) = COPY [[LSHR17]](s32) + ; CI-HSA: [[BUILD_VECTOR5:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY21]](s32), [[COPY22]](s32), [[COPY23]](s32), [[COPY24]](s32) + ; CI-HSA: [[TRUNC5:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR5]](<4 x s32>) + ; CI-HSA: [[LSHR18:%[0-9]+]]:_(s32) = G_LSHR [[UV6]], [[C]](s32) + ; CI-HSA: [[LSHR19:%[0-9]+]]:_(s32) = G_LSHR [[UV6]], [[C1]](s32) + ; CI-HSA: [[LSHR20:%[0-9]+]]:_(s32) = G_LSHR [[UV6]], [[C2]](s32) + ; CI-HSA: [[COPY25:%[0-9]+]]:_(s32) = COPY [[UV6]](s32) + ; CI-HSA: [[COPY26:%[0-9]+]]:_(s32) = COPY [[LSHR18]](s32) + ; CI-HSA: [[COPY27:%[0-9]+]]:_(s32) = COPY [[LSHR19]](s32) + ; CI-HSA: [[COPY28:%[0-9]+]]:_(s32) = COPY [[LSHR20]](s32) + ; CI-HSA: [[BUILD_VECTOR6:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY25]](s32), [[COPY26]](s32), [[COPY27]](s32), [[COPY28]](s32) + ; CI-HSA: [[TRUNC6:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR6]](<4 x s32>) + ; CI-HSA: [[LSHR21:%[0-9]+]]:_(s32) = G_LSHR [[UV7]], [[C]](s32) + ; CI-HSA: [[LSHR22:%[0-9]+]]:_(s32) = G_LSHR [[UV7]], [[C1]](s32) + ; CI-HSA: [[LSHR23:%[0-9]+]]:_(s32) = G_LSHR [[UV7]], [[C2]](s32) + ; CI-HSA: [[COPY29:%[0-9]+]]:_(s32) = COPY [[UV7]](s32) + ; CI-HSA: [[COPY30:%[0-9]+]]:_(s32) = COPY [[LSHR21]](s32) + ; CI-HSA: [[COPY31:%[0-9]+]]:_(s32) = COPY [[LSHR22]](s32) + ; CI-HSA: [[COPY32:%[0-9]+]]:_(s32) = COPY [[LSHR23]](s32) + ; CI-HSA: [[BUILD_VECTOR7:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY29]](s32), [[COPY30]](s32), [[COPY31]](s32), [[COPY32]](s32) + ; CI-HSA: [[TRUNC7:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR7]](<4 x s32>) + ; CI-HSA: [[CONCAT_VECTORS:%[0-9]+]]:_(<32 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>), [[TRUNC2]](<4 x s8>), [[TRUNC3]](<4 x s8>), [[TRUNC4]](<4 x s8>), [[TRUNC5]](<4 x s8>), [[TRUNC6]](<4 x s8>), [[TRUNC7]](<4 x s8>) + ; CI-HSA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[CONCAT_VECTORS]](<32 x s8>) ; CI-MESA-LABEL: name: test_load_global_v32s8_align32 ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1 ; CI-MESA: [[LOAD:%[0-9]+]]:_(<8 x s32>) = G_LOAD [[COPY]](p1) :: (load 32, addrspace 1) - ; CI-MESA: [[BITCAST:%[0-9]+]]:_(<32 x s8>) = G_BITCAST [[LOAD]](<8 x s32>) - ; CI-MESA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[BITCAST]](<32 x s8>) + ; CI-MESA: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32), [[UV2:%[0-9]+]]:_(s32), [[UV3:%[0-9]+]]:_(s32), [[UV4:%[0-9]+]]:_(s32), [[UV5:%[0-9]+]]:_(s32), [[UV6:%[0-9]+]]:_(s32), [[UV7:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[LOAD]](<8 x s32>) + ; CI-MESA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8 + ; CI-MESA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C]](s32) + ; CI-MESA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16 + ; CI-MESA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C1]](s32) + ; CI-MESA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24 + ; CI-MESA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C2]](s32) + ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[UV]](s32) + ; CI-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32) + ; CI-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32) + ; CI-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32) + ; CI-MESA: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32) + ; CI-MESA: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>) + ; CI-MESA: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C]](s32) + ; CI-MESA: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C1]](s32) + ; CI-MESA: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C2]](s32) + ; CI-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[UV1]](s32) + ; CI-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32) + ; CI-MESA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32) + ; CI-MESA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32) + ; CI-MESA: [[BUILD_VECTOR1:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY5]](s32), [[COPY6]](s32), [[COPY7]](s32), [[COPY8]](s32) + ; CI-MESA: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR1]](<4 x s32>) + ; CI-MESA: [[LSHR6:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C]](s32) + ; CI-MESA: [[LSHR7:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C1]](s32) + ; CI-MESA: [[LSHR8:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C2]](s32) + ; CI-MESA: [[COPY9:%[0-9]+]]:_(s32) = COPY [[UV2]](s32) + ; CI-MESA: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LSHR6]](s32) + ; CI-MESA: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LSHR7]](s32) + ; CI-MESA: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LSHR8]](s32) + ; CI-MESA: [[BUILD_VECTOR2:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY9]](s32), [[COPY10]](s32), [[COPY11]](s32), [[COPY12]](s32) + ; CI-MESA: [[TRUNC2:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR2]](<4 x s32>) + ; CI-MESA: [[LSHR9:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C]](s32) + ; CI-MESA: [[LSHR10:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C1]](s32) + ; CI-MESA: [[LSHR11:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C2]](s32) + ; CI-MESA: [[COPY13:%[0-9]+]]:_(s32) = COPY [[UV3]](s32) + ; CI-MESA: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LSHR9]](s32) + ; CI-MESA: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LSHR10]](s32) + ; CI-MESA: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LSHR11]](s32) + ; CI-MESA: [[BUILD_VECTOR3:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY13]](s32), [[COPY14]](s32), [[COPY15]](s32), [[COPY16]](s32) + ; CI-MESA: [[TRUNC3:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR3]](<4 x s32>) + ; CI-MESA: [[LSHR12:%[0-9]+]]:_(s32) = G_LSHR [[UV4]], [[C]](s32) + ; CI-MESA: [[LSHR13:%[0-9]+]]:_(s32) = G_LSHR [[UV4]], [[C1]](s32) + ; CI-MESA: [[LSHR14:%[0-9]+]]:_(s32) = G_LSHR [[UV4]], [[C2]](s32) + ; CI-MESA: [[COPY17:%[0-9]+]]:_(s32) = COPY [[UV4]](s32) + ; CI-MESA: [[COPY18:%[0-9]+]]:_(s32) = COPY [[LSHR12]](s32) + ; CI-MESA: [[COPY19:%[0-9]+]]:_(s32) = COPY [[LSHR13]](s32) + ; CI-MESA: [[COPY20:%[0-9]+]]:_(s32) = COPY [[LSHR14]](s32) + ; CI-MESA: [[BUILD_VECTOR4:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY17]](s32), [[COPY18]](s32), [[COPY19]](s32), [[COPY20]](s32) + ; CI-MESA: [[TRUNC4:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR4]](<4 x s32>) + ; CI-MESA: [[LSHR15:%[0-9]+]]:_(s32) = G_LSHR [[UV5]], [[C]](s32) + ; CI-MESA: [[LSHR16:%[0-9]+]]:_(s32) = G_LSHR [[UV5]], [[C1]](s32) + ; CI-MESA: [[LSHR17:%[0-9]+]]:_(s32) = G_LSHR [[UV5]], [[C2]](s32) + ; CI-MESA: [[COPY21:%[0-9]+]]:_(s32) = COPY [[UV5]](s32) + ; CI-MESA: [[COPY22:%[0-9]+]]:_(s32) = COPY [[LSHR15]](s32) + ; CI-MESA: [[COPY23:%[0-9]+]]:_(s32) = COPY [[LSHR16]](s32) + ; CI-MESA: [[COPY24:%[0-9]+]]:_(s32) = COPY [[LSHR17]](s32) + ; CI-MESA: [[BUILD_VECTOR5:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY21]](s32), [[COPY22]](s32), [[COPY23]](s32), [[COPY24]](s32) + ; CI-MESA: [[TRUNC5:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR5]](<4 x s32>) + ; CI-MESA: [[LSHR18:%[0-9]+]]:_(s32) = G_LSHR [[UV6]], [[C]](s32) + ; CI-MESA: [[LSHR19:%[0-9]+]]:_(s32) = G_LSHR [[UV6]], [[C1]](s32) + ; CI-MESA: [[LSHR20:%[0-9]+]]:_(s32) = G_LSHR [[UV6]], [[C2]](s32) + ; CI-MESA: [[COPY25:%[0-9]+]]:_(s32) = COPY [[UV6]](s32) + ; CI-MESA: [[COPY26:%[0-9]+]]:_(s32) = COPY [[LSHR18]](s32) + ; CI-MESA: [[COPY27:%[0-9]+]]:_(s32) = COPY [[LSHR19]](s32) + ; CI-MESA: [[COPY28:%[0-9]+]]:_(s32) = COPY [[LSHR20]](s32) + ; CI-MESA: [[BUILD_VECTOR6:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY25]](s32), [[COPY26]](s32), [[COPY27]](s32), [[COPY28]](s32) + ; CI-MESA: [[TRUNC6:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR6]](<4 x s32>) + ; CI-MESA: [[LSHR21:%[0-9]+]]:_(s32) = G_LSHR [[UV7]], [[C]](s32) + ; CI-MESA: [[LSHR22:%[0-9]+]]:_(s32) = G_LSHR [[UV7]], [[C1]](s32) + ; CI-MESA: [[LSHR23:%[0-9]+]]:_(s32) = G_LSHR [[UV7]], [[C2]](s32) + ; CI-MESA: [[COPY29:%[0-9]+]]:_(s32) = COPY [[UV7]](s32) + ; CI-MESA: [[COPY30:%[0-9]+]]:_(s32) = COPY [[LSHR21]](s32) + ; CI-MESA: [[COPY31:%[0-9]+]]:_(s32) = COPY [[LSHR22]](s32) + ; CI-MESA: [[COPY32:%[0-9]+]]:_(s32) = COPY [[LSHR23]](s32) + ; CI-MESA: [[BUILD_VECTOR7:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY29]](s32), [[COPY30]](s32), [[COPY31]](s32), [[COPY32]](s32) + ; CI-MESA: [[TRUNC7:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR7]](<4 x s32>) + ; CI-MESA: [[CONCAT_VECTORS:%[0-9]+]]:_(<32 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>), [[TRUNC2]](<4 x s8>), [[TRUNC3]](<4 x s8>), [[TRUNC4]](<4 x s8>), [[TRUNC5]](<4 x s8>), [[TRUNC6]](<4 x s8>), [[TRUNC7]](<4 x s8>) + ; CI-MESA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[CONCAT_VECTORS]](<32 x s8>) ; VI-LABEL: name: test_load_global_v32s8_align32 ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1 ; VI: [[LOAD:%[0-9]+]]:_(<8 x s32>) = G_LOAD [[COPY]](p1) :: (load 32, addrspace 1) - ; VI: [[BITCAST:%[0-9]+]]:_(<32 x s8>) = G_BITCAST [[LOAD]](<8 x s32>) - ; VI: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[BITCAST]](<32 x s8>) + ; VI: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32), [[UV2:%[0-9]+]]:_(s32), [[UV3:%[0-9]+]]:_(s32), [[UV4:%[0-9]+]]:_(s32), [[UV5:%[0-9]+]]:_(s32), [[UV6:%[0-9]+]]:_(s32), [[UV7:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[LOAD]](<8 x s32>) + ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8 + ; VI: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C]](s32) + ; VI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16 + ; VI: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C1]](s32) + ; VI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24 + ; VI: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C2]](s32) + ; VI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[UV]](s32) + ; VI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32) + ; VI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32) + ; VI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32) + ; VI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32) + ; VI: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>) + ; VI: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C]](s32) + ; VI: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C1]](s32) + ; VI: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C2]](s32) + ; VI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[UV1]](s32) + ; VI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32) + ; VI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32) + ; VI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32) + ; VI: [[BUILD_VECTOR1:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY5]](s32), [[COPY6]](s32), [[COPY7]](s32), [[COPY8]](s32) + ; VI: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR1]](<4 x s32>) + ; VI: [[LSHR6:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C]](s32) + ; VI: [[LSHR7:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C1]](s32) + ; VI: [[LSHR8:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C2]](s32) + ; VI: [[COPY9:%[0-9]+]]:_(s32) = COPY [[UV2]](s32) + ; VI: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LSHR6]](s32) + ; VI: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LSHR7]](s32) + ; VI: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LSHR8]](s32) + ; VI: [[BUILD_VECTOR2:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY9]](s32), [[COPY10]](s32), [[COPY11]](s32), [[COPY12]](s32) + ; VI: [[TRUNC2:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR2]](<4 x s32>) + ; VI: [[LSHR9:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C]](s32) + ; VI: [[LSHR10:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C1]](s32) + ; VI: [[LSHR11:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C2]](s32) + ; VI: [[COPY13:%[0-9]+]]:_(s32) = COPY [[UV3]](s32) + ; VI: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LSHR9]](s32) + ; VI: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LSHR10]](s32) + ; VI: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LSHR11]](s32) + ; VI: [[BUILD_VECTOR3:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY13]](s32), [[COPY14]](s32), [[COPY15]](s32), [[COPY16]](s32) + ; VI: [[TRUNC3:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR3]](<4 x s32>) + ; VI: [[LSHR12:%[0-9]+]]:_(s32) = G_LSHR [[UV4]], [[C]](s32) + ; VI: [[LSHR13:%[0-9]+]]:_(s32) = G_LSHR [[UV4]], [[C1]](s32) + ; VI: [[LSHR14:%[0-9]+]]:_(s32) = G_LSHR [[UV4]], [[C2]](s32) + ; VI: [[COPY17:%[0-9]+]]:_(s32) = COPY [[UV4]](s32) + ; VI: [[COPY18:%[0-9]+]]:_(s32) = COPY [[LSHR12]](s32) + ; VI: [[COPY19:%[0-9]+]]:_(s32) = COPY [[LSHR13]](s32) + ; VI: [[COPY20:%[0-9]+]]:_(s32) = COPY [[LSHR14]](s32) + ; VI: [[BUILD_VECTOR4:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY17]](s32), [[COPY18]](s32), [[COPY19]](s32), [[COPY20]](s32) + ; VI: [[TRUNC4:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR4]](<4 x s32>) + ; VI: [[LSHR15:%[0-9]+]]:_(s32) = G_LSHR [[UV5]], [[C]](s32) + ; VI: [[LSHR16:%[0-9]+]]:_(s32) = G_LSHR [[UV5]], [[C1]](s32) + ; VI: [[LSHR17:%[0-9]+]]:_(s32) = G_LSHR [[UV5]], [[C2]](s32) + ; VI: [[COPY21:%[0-9]+]]:_(s32) = COPY [[UV5]](s32) + ; VI: [[COPY22:%[0-9]+]]:_(s32) = COPY [[LSHR15]](s32) + ; VI: [[COPY23:%[0-9]+]]:_(s32) = COPY [[LSHR16]](s32) + ; VI: [[COPY24:%[0-9]+]]:_(s32) = COPY [[LSHR17]](s32) + ; VI: [[BUILD_VECTOR5:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY21]](s32), [[COPY22]](s32), [[COPY23]](s32), [[COPY24]](s32) + ; VI: [[TRUNC5:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR5]](<4 x s32>) + ; VI: [[LSHR18:%[0-9]+]]:_(s32) = G_LSHR [[UV6]], [[C]](s32) + ; VI: [[LSHR19:%[0-9]+]]:_(s32) = G_LSHR [[UV6]], [[C1]](s32) + ; VI: [[LSHR20:%[0-9]+]]:_(s32) = G_LSHR [[UV6]], [[C2]](s32) + ; VI: [[COPY25:%[0-9]+]]:_(s32) = COPY [[UV6]](s32) + ; VI: [[COPY26:%[0-9]+]]:_(s32) = COPY [[LSHR18]](s32) + ; VI: [[COPY27:%[0-9]+]]:_(s32) = COPY [[LSHR19]](s32) + ; VI: [[COPY28:%[0-9]+]]:_(s32) = COPY [[LSHR20]](s32) + ; VI: [[BUILD_VECTOR6:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY25]](s32), [[COPY26]](s32), [[COPY27]](s32), [[COPY28]](s32) + ; VI: [[TRUNC6:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR6]](<4 x s32>) + ; VI: [[LSHR21:%[0-9]+]]:_(s32) = G_LSHR [[UV7]], [[C]](s32) + ; VI: [[LSHR22:%[0-9]+]]:_(s32) = G_LSHR [[UV7]], [[C1]](s32) + ; VI: [[LSHR23:%[0-9]+]]:_(s32) = G_LSHR [[UV7]], [[C2]](s32) + ; VI: [[COPY29:%[0-9]+]]:_(s32) = COPY [[UV7]](s32) + ; VI: [[COPY30:%[0-9]+]]:_(s32) = COPY [[LSHR21]](s32) + ; VI: [[COPY31:%[0-9]+]]:_(s32) = COPY [[LSHR22]](s32) + ; VI: [[COPY32:%[0-9]+]]:_(s32) = COPY [[LSHR23]](s32) + ; VI: [[BUILD_VECTOR7:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY29]](s32), [[COPY30]](s32), [[COPY31]](s32), [[COPY32]](s32) + ; VI: [[TRUNC7:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR7]](<4 x s32>) + ; VI: [[CONCAT_VECTORS:%[0-9]+]]:_(<32 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>), [[TRUNC2]](<4 x s8>), [[TRUNC3]](<4 x s8>), [[TRUNC4]](<4 x s8>), [[TRUNC5]](<4 x s8>), [[TRUNC6]](<4 x s8>), [[TRUNC7]](<4 x s8>) + ; VI: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[CONCAT_VECTORS]](<32 x s8>) ; GFX9-HSA-LABEL: name: test_load_global_v32s8_align32 ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1 ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(<8 x s32>) = G_LOAD [[COPY]](p1) :: (load 32, addrspace 1) - ; GFX9-HSA: [[BITCAST:%[0-9]+]]:_(<32 x s8>) = G_BITCAST [[LOAD]](<8 x s32>) - ; GFX9-HSA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[BITCAST]](<32 x s8>) + ; GFX9-HSA: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32), [[UV2:%[0-9]+]]:_(s32), [[UV3:%[0-9]+]]:_(s32), [[UV4:%[0-9]+]]:_(s32), [[UV5:%[0-9]+]]:_(s32), [[UV6:%[0-9]+]]:_(s32), [[UV7:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[LOAD]](<8 x s32>) + ; GFX9-HSA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8 + ; GFX9-HSA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C]](s32) + ; GFX9-HSA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16 + ; GFX9-HSA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C1]](s32) + ; GFX9-HSA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24 + ; GFX9-HSA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C2]](s32) + ; GFX9-HSA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[UV]](s32) + ; GFX9-HSA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32) + ; GFX9-HSA: [[BUILD_VECTOR_TRUNC:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY1]](s32), [[COPY2]](s32) + ; GFX9-HSA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32) + ; GFX9-HSA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32) + ; GFX9-HSA: [[BUILD_VECTOR_TRUNC1:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY3]](s32), [[COPY4]](s32) + ; GFX9-HSA: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC]](<2 x s16>), [[BUILD_VECTOR_TRUNC1]](<2 x s16>) + ; GFX9-HSA: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS]](<4 x s16>) + ; GFX9-HSA: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C]](s32) + ; GFX9-HSA: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C1]](s32) + ; GFX9-HSA: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C2]](s32) + ; GFX9-HSA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[UV1]](s32) + ; GFX9-HSA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32) + ; GFX9-HSA: [[BUILD_VECTOR_TRUNC2:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY5]](s32), [[COPY6]](s32) + ; GFX9-HSA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32) + ; GFX9-HSA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32) + ; GFX9-HSA: [[BUILD_VECTOR_TRUNC3:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY7]](s32), [[COPY8]](s32) + ; GFX9-HSA: [[CONCAT_VECTORS1:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC2]](<2 x s16>), [[BUILD_VECTOR_TRUNC3]](<2 x s16>) + ; GFX9-HSA: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS1]](<4 x s16>) + ; GFX9-HSA: [[LSHR6:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C]](s32) + ; GFX9-HSA: [[LSHR7:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C1]](s32) + ; GFX9-HSA: [[LSHR8:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C2]](s32) + ; GFX9-HSA: [[COPY9:%[0-9]+]]:_(s32) = COPY [[UV2]](s32) + ; GFX9-HSA: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LSHR6]](s32) + ; GFX9-HSA: [[BUILD_VECTOR_TRUNC4:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY9]](s32), [[COPY10]](s32) + ; GFX9-HSA: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LSHR7]](s32) + ; GFX9-HSA: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LSHR8]](s32) + ; GFX9-HSA: [[BUILD_VECTOR_TRUNC5:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY11]](s32), [[COPY12]](s32) + ; GFX9-HSA: [[CONCAT_VECTORS2:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC4]](<2 x s16>), [[BUILD_VECTOR_TRUNC5]](<2 x s16>) + ; GFX9-HSA: [[TRUNC2:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS2]](<4 x s16>) + ; GFX9-HSA: [[LSHR9:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C]](s32) + ; GFX9-HSA: [[LSHR10:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C1]](s32) + ; GFX9-HSA: [[LSHR11:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C2]](s32) + ; GFX9-HSA: [[COPY13:%[0-9]+]]:_(s32) = COPY [[UV3]](s32) + ; GFX9-HSA: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LSHR9]](s32) + ; GFX9-HSA: [[BUILD_VECTOR_TRUNC6:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY13]](s32), [[COPY14]](s32) + ; GFX9-HSA: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LSHR10]](s32) + ; GFX9-HSA: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LSHR11]](s32) + ; GFX9-HSA: [[BUILD_VECTOR_TRUNC7:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY15]](s32), [[COPY16]](s32) + ; GFX9-HSA: [[CONCAT_VECTORS3:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC6]](<2 x s16>), [[BUILD_VECTOR_TRUNC7]](<2 x s16>) + ; GFX9-HSA: [[TRUNC3:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS3]](<4 x s16>) + ; GFX9-HSA: [[LSHR12:%[0-9]+]]:_(s32) = G_LSHR [[UV4]], [[C]](s32) + ; GFX9-HSA: [[LSHR13:%[0-9]+]]:_(s32) = G_LSHR [[UV4]], [[C1]](s32) + ; GFX9-HSA: [[LSHR14:%[0-9]+]]:_(s32) = G_LSHR [[UV4]], [[C2]](s32) + ; GFX9-HSA: [[COPY17:%[0-9]+]]:_(s32) = COPY [[UV4]](s32) + ; GFX9-HSA: [[COPY18:%[0-9]+]]:_(s32) = COPY [[LSHR12]](s32) + ; GFX9-HSA: [[BUILD_VECTOR_TRUNC8:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY17]](s32), [[COPY18]](s32) + ; GFX9-HSA: [[COPY19:%[0-9]+]]:_(s32) = COPY [[LSHR13]](s32) + ; GFX9-HSA: [[COPY20:%[0-9]+]]:_(s32) = COPY [[LSHR14]](s32) + ; GFX9-HSA: [[BUILD_VECTOR_TRUNC9:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY19]](s32), [[COPY20]](s32) + ; GFX9-HSA: [[CONCAT_VECTORS4:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC8]](<2 x s16>), [[BUILD_VECTOR_TRUNC9]](<2 x s16>) + ; GFX9-HSA: [[TRUNC4:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS4]](<4 x s16>) + ; GFX9-HSA: [[LSHR15:%[0-9]+]]:_(s32) = G_LSHR [[UV5]], [[C]](s32) + ; GFX9-HSA: [[LSHR16:%[0-9]+]]:_(s32) = G_LSHR [[UV5]], [[C1]](s32) + ; GFX9-HSA: [[LSHR17:%[0-9]+]]:_(s32) = G_LSHR [[UV5]], [[C2]](s32) + ; GFX9-HSA: [[COPY21:%[0-9]+]]:_(s32) = COPY [[UV5]](s32) + ; GFX9-HSA: [[COPY22:%[0-9]+]]:_(s32) = COPY [[LSHR15]](s32) + ; GFX9-HSA: [[BUILD_VECTOR_TRUNC10:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY21]](s32), [[COPY22]](s32) + ; GFX9-HSA: [[COPY23:%[0-9]+]]:_(s32) = COPY [[LSHR16]](s32) + ; GFX9-HSA: [[COPY24:%[0-9]+]]:_(s32) = COPY [[LSHR17]](s32) + ; GFX9-HSA: [[BUILD_VECTOR_TRUNC11:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY23]](s32), [[COPY24]](s32) + ; GFX9-HSA: [[CONCAT_VECTORS5:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC10]](<2 x s16>), [[BUILD_VECTOR_TRUNC11]](<2 x s16>) + ; GFX9-HSA: [[TRUNC5:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS5]](<4 x s16>) + ; GFX9-HSA: [[LSHR18:%[0-9]+]]:_(s32) = G_LSHR [[UV6]], [[C]](s32) + ; GFX9-HSA: [[LSHR19:%[0-9]+]]:_(s32) = G_LSHR [[UV6]], [[C1]](s32) + ; GFX9-HSA: [[LSHR20:%[0-9]+]]:_(s32) = G_LSHR [[UV6]], [[C2]](s32) + ; GFX9-HSA: [[COPY25:%[0-9]+]]:_(s32) = COPY [[UV6]](s32) + ; GFX9-HSA: [[COPY26:%[0-9]+]]:_(s32) = COPY [[LSHR18]](s32) + ; GFX9-HSA: [[BUILD_VECTOR_TRUNC12:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY25]](s32), [[COPY26]](s32) + ; GFX9-HSA: [[COPY27:%[0-9]+]]:_(s32) = COPY [[LSHR19]](s32) + ; GFX9-HSA: [[COPY28:%[0-9]+]]:_(s32) = COPY [[LSHR20]](s32) + ; GFX9-HSA: [[BUILD_VECTOR_TRUNC13:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY27]](s32), [[COPY28]](s32) + ; GFX9-HSA: [[CONCAT_VECTORS6:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC12]](<2 x s16>), [[BUILD_VECTOR_TRUNC13]](<2 x s16>) + ; GFX9-HSA: [[TRUNC6:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS6]](<4 x s16>) + ; GFX9-HSA: [[LSHR21:%[0-9]+]]:_(s32) = G_LSHR [[UV7]], [[C]](s32) + ; GFX9-HSA: [[LSHR22:%[0-9]+]]:_(s32) = G_LSHR [[UV7]], [[C1]](s32) + ; GFX9-HSA: [[LSHR23:%[0-9]+]]:_(s32) = G_LSHR [[UV7]], [[C2]](s32) + ; GFX9-HSA: [[COPY29:%[0-9]+]]:_(s32) = COPY [[UV7]](s32) + ; GFX9-HSA: [[COPY30:%[0-9]+]]:_(s32) = COPY [[LSHR21]](s32) + ; GFX9-HSA: [[BUILD_VECTOR_TRUNC14:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY29]](s32), [[COPY30]](s32) + ; GFX9-HSA: [[COPY31:%[0-9]+]]:_(s32) = COPY [[LSHR22]](s32) + ; GFX9-HSA: [[COPY32:%[0-9]+]]:_(s32) = COPY [[LSHR23]](s32) + ; GFX9-HSA: [[BUILD_VECTOR_TRUNC15:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY31]](s32), [[COPY32]](s32) + ; GFX9-HSA: [[CONCAT_VECTORS7:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC14]](<2 x s16>), [[BUILD_VECTOR_TRUNC15]](<2 x s16>) + ; GFX9-HSA: [[TRUNC7:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS7]](<4 x s16>) + ; GFX9-HSA: [[CONCAT_VECTORS8:%[0-9]+]]:_(<32 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>), [[TRUNC2]](<4 x s8>), [[TRUNC3]](<4 x s8>), [[TRUNC4]](<4 x s8>), [[TRUNC5]](<4 x s8>), [[TRUNC6]](<4 x s8>), [[TRUNC7]](<4 x s8>) + ; GFX9-HSA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[CONCAT_VECTORS8]](<32 x s8>) ; GFX9-MESA-LABEL: name: test_load_global_v32s8_align32 ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1 ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(<8 x s32>) = G_LOAD [[COPY]](p1) :: (load 32, addrspace 1) - ; GFX9-MESA: [[BITCAST:%[0-9]+]]:_(<32 x s8>) = G_BITCAST [[LOAD]](<8 x s32>) - ; GFX9-MESA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[BITCAST]](<32 x s8>) + ; GFX9-MESA: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32), [[UV2:%[0-9]+]]:_(s32), [[UV3:%[0-9]+]]:_(s32), [[UV4:%[0-9]+]]:_(s32), [[UV5:%[0-9]+]]:_(s32), [[UV6:%[0-9]+]]:_(s32), [[UV7:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[LOAD]](<8 x s32>) + ; GFX9-MESA: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8 + ; GFX9-MESA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C]](s32) + ; GFX9-MESA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16 + ; GFX9-MESA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C1]](s32) + ; GFX9-MESA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24 + ; GFX9-MESA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C2]](s32) + ; GFX9-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[UV]](s32) + ; GFX9-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY1]](s32), [[COPY2]](s32) + ; GFX9-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32) + ; GFX9-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC1:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY3]](s32), [[COPY4]](s32) + ; GFX9-MESA: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC]](<2 x s16>), [[BUILD_VECTOR_TRUNC1]](<2 x s16>) + ; GFX9-MESA: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS]](<4 x s16>) + ; GFX9-MESA: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C]](s32) + ; GFX9-MESA: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C1]](s32) + ; GFX9-MESA: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C2]](s32) + ; GFX9-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[UV1]](s32) + ; GFX9-MESA: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC2:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY5]](s32), [[COPY6]](s32) + ; GFX9-MESA: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32) + ; GFX9-MESA: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC3:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY7]](s32), [[COPY8]](s32) + ; GFX9-MESA: [[CONCAT_VECTORS1:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC2]](<2 x s16>), [[BUILD_VECTOR_TRUNC3]](<2 x s16>) + ; GFX9-MESA: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS1]](<4 x s16>) + ; GFX9-MESA: [[LSHR6:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C]](s32) + ; GFX9-MESA: [[LSHR7:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C1]](s32) + ; GFX9-MESA: [[LSHR8:%[0-9]+]]:_(s32) = G_LSHR [[UV2]], [[C2]](s32) + ; GFX9-MESA: [[COPY9:%[0-9]+]]:_(s32) = COPY [[UV2]](s32) + ; GFX9-MESA: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LSHR6]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC4:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY9]](s32), [[COPY10]](s32) + ; GFX9-MESA: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LSHR7]](s32) + ; GFX9-MESA: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LSHR8]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC5:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY11]](s32), [[COPY12]](s32) + ; GFX9-MESA: [[CONCAT_VECTORS2:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC4]](<2 x s16>), [[BUILD_VECTOR_TRUNC5]](<2 x s16>) + ; GFX9-MESA: [[TRUNC2:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS2]](<4 x s16>) + ; GFX9-MESA: [[LSHR9:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C]](s32) + ; GFX9-MESA: [[LSHR10:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C1]](s32) + ; GFX9-MESA: [[LSHR11:%[0-9]+]]:_(s32) = G_LSHR [[UV3]], [[C2]](s32) + ; GFX9-MESA: [[COPY13:%[0-9]+]]:_(s32) = COPY [[UV3]](s32) + ; GFX9-MESA: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LSHR9]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC6:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY13]](s32), [[COPY14]](s32) + ; GFX9-MESA: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LSHR10]](s32) + ; GFX9-MESA: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LSHR11]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC7:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY15]](s32), [[COPY16]](s32) + ; GFX9-MESA: [[CONCAT_VECTORS3:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC6]](<2 x s16>), [[BUILD_VECTOR_TRUNC7]](<2 x s16>) + ; GFX9-MESA: [[TRUNC3:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS3]](<4 x s16>) + ; GFX9-MESA: [[LSHR12:%[0-9]+]]:_(s32) = G_LSHR [[UV4]], [[C]](s32) + ; GFX9-MESA: [[LSHR13:%[0-9]+]]:_(s32) = G_LSHR [[UV4]], [[C1]](s32) + ; GFX9-MESA: [[LSHR14:%[0-9]+]]:_(s32) = G_LSHR [[UV4]], [[C2]](s32) + ; GFX9-MESA: [[COPY17:%[0-9]+]]:_(s32) = COPY [[UV4]](s32) + ; GFX9-MESA: [[COPY18:%[0-9]+]]:_(s32) = COPY [[LSHR12]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC8:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY17]](s32), [[COPY18]](s32) + ; GFX9-MESA: [[COPY19:%[0-9]+]]:_(s32) = COPY [[LSHR13]](s32) + ; GFX9-MESA: [[COPY20:%[0-9]+]]:_(s32) = COPY [[LSHR14]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC9:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY19]](s32), [[COPY20]](s32) + ; GFX9-MESA: [[CONCAT_VECTORS4:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC8]](<2 x s16>), [[BUILD_VECTOR_TRUNC9]](<2 x s16>) + ; GFX9-MESA: [[TRUNC4:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS4]](<4 x s16>) + ; GFX9-MESA: [[LSHR15:%[0-9]+]]:_(s32) = G_LSHR [[UV5]], [[C]](s32) + ; GFX9-MESA: [[LSHR16:%[0-9]+]]:_(s32) = G_LSHR [[UV5]], [[C1]](s32) + ; GFX9-MESA: [[LSHR17:%[0-9]+]]:_(s32) = G_LSHR [[UV5]], [[C2]](s32) + ; GFX9-MESA: [[COPY21:%[0-9]+]]:_(s32) = COPY [[UV5]](s32) + ; GFX9-MESA: [[COPY22:%[0-9]+]]:_(s32) = COPY [[LSHR15]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC10:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY21]](s32), [[COPY22]](s32) + ; GFX9-MESA: [[COPY23:%[0-9]+]]:_(s32) = COPY [[LSHR16]](s32) + ; GFX9-MESA: [[COPY24:%[0-9]+]]:_(s32) = COPY [[LSHR17]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC11:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY23]](s32), [[COPY24]](s32) + ; GFX9-MESA: [[CONCAT_VECTORS5:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC10]](<2 x s16>), [[BUILD_VECTOR_TRUNC11]](<2 x s16>) + ; GFX9-MESA: [[TRUNC5:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS5]](<4 x s16>) + ; GFX9-MESA: [[LSHR18:%[0-9]+]]:_(s32) = G_LSHR [[UV6]], [[C]](s32) + ; GFX9-MESA: [[LSHR19:%[0-9]+]]:_(s32) = G_LSHR [[UV6]], [[C1]](s32) + ; GFX9-MESA: [[LSHR20:%[0-9]+]]:_(s32) = G_LSHR [[UV6]], [[C2]](s32) + ; GFX9-MESA: [[COPY25:%[0-9]+]]:_(s32) = COPY [[UV6]](s32) + ; GFX9-MESA: [[COPY26:%[0-9]+]]:_(s32) = COPY [[LSHR18]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC12:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY25]](s32), [[COPY26]](s32) + ; GFX9-MESA: [[COPY27:%[0-9]+]]:_(s32) = COPY [[LSHR19]](s32) + ; GFX9-MESA: [[COPY28:%[0-9]+]]:_(s32) = COPY [[LSHR20]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC13:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY27]](s32), [[COPY28]](s32) + ; GFX9-MESA: [[CONCAT_VECTORS6:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC12]](<2 x s16>), [[BUILD_VECTOR_TRUNC13]](<2 x s16>) + ; GFX9-MESA: [[TRUNC6:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS6]](<4 x s16>) + ; GFX9-MESA: [[LSHR21:%[0-9]+]]:_(s32) = G_LSHR [[UV7]], [[C]](s32) + ; GFX9-MESA: [[LSHR22:%[0-9]+]]:_(s32) = G_LSHR [[UV7]], [[C1]](s32) + ; GFX9-MESA: [[LSHR23:%[0-9]+]]:_(s32) = G_LSHR [[UV7]], [[C2]](s32) + ; GFX9-MESA: [[COPY29:%[0-9]+]]:_(s32) = COPY [[UV7]](s32) + ; GFX9-MESA: [[COPY30:%[0-9]+]]:_(s32) = COPY [[LSHR21]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC14:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY29]](s32), [[COPY30]](s32) + ; GFX9-MESA: [[COPY31:%[0-9]+]]:_(s32) = COPY [[LSHR22]](s32) + ; GFX9-MESA: [[COPY32:%[0-9]+]]:_(s32) = COPY [[LSHR23]](s32) + ; GFX9-MESA: [[BUILD_VECTOR_TRUNC15:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY31]](s32), [[COPY32]](s32) + ; GFX9-MESA: [[CONCAT_VECTORS7:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC14]](<2 x s16>), [[BUILD_VECTOR_TRUNC15]](<2 x s16>) + ; GFX9-MESA: [[TRUNC7:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS7]](<4 x s16>) + ; GFX9-MESA: [[CONCAT_VECTORS8:%[0-9]+]]:_(<32 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>), [[TRUNC2]](<4 x s8>), [[TRUNC3]](<4 x s8>), [[TRUNC4]](<4 x s8>), [[TRUNC5]](<4 x s8>), [[TRUNC6]](<4 x s8>), [[TRUNC7]](<4 x s8>) + ; GFX9-MESA: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[CONCAT_VECTORS8]](<32 x s8>) %0:_(p1) = COPY $vgpr0_vgpr1 %1:_(<32 x s8>) = G_LOAD %0 :: (load 32, align 32, addrspace 1) $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY %1 Index: llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-load-local.mir =================================================================== --- llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-load-local.mir +++ llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-load-local.mir @@ -5654,28 +5654,142 @@ ; SI-LABEL: name: test_load_local_v8s8_align8 ; SI: [[COPY:%[0-9]+]]:_(p3) = COPY $vgpr0 ; SI: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p3) :: (load 8, addrspace 3) - ; SI: [[BITCAST:%[0-9]+]]:_(<8 x s8>) = G_BITCAST [[LOAD]](<2 x s32>) - ; SI: $vgpr0_vgpr1 = COPY [[BITCAST]](<8 x s8>) + ; SI: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[LOAD]](<2 x s32>) + ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8 + ; SI: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C]](s32) + ; SI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16 + ; SI: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C1]](s32) + ; SI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24 + ; SI: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C2]](s32) + ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[UV]](s32) + ; SI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32) + ; SI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32) + ; SI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32) + ; SI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32) + ; SI: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>) + ; SI: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C]](s32) + ; SI: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C1]](s32) + ; SI: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C2]](s32) + ; SI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[UV1]](s32) + ; SI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32) + ; SI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32) + ; SI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32) + ; SI: [[BUILD_VECTOR1:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY5]](s32), [[COPY6]](s32), [[COPY7]](s32), [[COPY8]](s32) + ; SI: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR1]](<4 x s32>) + ; SI: [[CONCAT_VECTORS:%[0-9]+]]:_(<8 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>) + ; SI: $vgpr0_vgpr1 = COPY [[CONCAT_VECTORS]](<8 x s8>) ; CI-LABEL: name: test_load_local_v8s8_align8 ; CI: [[COPY:%[0-9]+]]:_(p3) = COPY $vgpr0 ; CI: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p3) :: (load 8, addrspace 3) - ; CI: [[BITCAST:%[0-9]+]]:_(<8 x s8>) = G_BITCAST [[LOAD]](<2 x s32>) - ; CI: $vgpr0_vgpr1 = COPY [[BITCAST]](<8 x s8>) + ; CI: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[LOAD]](<2 x s32>) + ; CI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8 + ; CI: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C]](s32) + ; CI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16 + ; CI: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C1]](s32) + ; CI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24 + ; CI: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C2]](s32) + ; CI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[UV]](s32) + ; CI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32) + ; CI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32) + ; CI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32) + ; CI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32) + ; CI: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>) + ; CI: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C]](s32) + ; CI: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C1]](s32) + ; CI: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C2]](s32) + ; CI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[UV1]](s32) + ; CI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32) + ; CI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32) + ; CI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32) + ; CI: [[BUILD_VECTOR1:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY5]](s32), [[COPY6]](s32), [[COPY7]](s32), [[COPY8]](s32) + ; CI: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR1]](<4 x s32>) + ; CI: [[CONCAT_VECTORS:%[0-9]+]]:_(<8 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>) + ; CI: $vgpr0_vgpr1 = COPY [[CONCAT_VECTORS]](<8 x s8>) ; CI-DS128-LABEL: name: test_load_local_v8s8_align8 ; CI-DS128: [[COPY:%[0-9]+]]:_(p3) = COPY $vgpr0 ; CI-DS128: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p3) :: (load 8, addrspace 3) - ; CI-DS128: [[BITCAST:%[0-9]+]]:_(<8 x s8>) = G_BITCAST [[LOAD]](<2 x s32>) - ; CI-DS128: $vgpr0_vgpr1 = COPY [[BITCAST]](<8 x s8>) + ; CI-DS128: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[LOAD]](<2 x s32>) + ; CI-DS128: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8 + ; CI-DS128: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C]](s32) + ; CI-DS128: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16 + ; CI-DS128: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C1]](s32) + ; CI-DS128: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24 + ; CI-DS128: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C2]](s32) + ; CI-DS128: [[COPY1:%[0-9]+]]:_(s32) = COPY [[UV]](s32) + ; CI-DS128: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32) + ; CI-DS128: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32) + ; CI-DS128: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32) + ; CI-DS128: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32) + ; CI-DS128: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>) + ; CI-DS128: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C]](s32) + ; CI-DS128: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C1]](s32) + ; CI-DS128: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C2]](s32) + ; CI-DS128: [[COPY5:%[0-9]+]]:_(s32) = COPY [[UV1]](s32) + ; CI-DS128: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32) + ; CI-DS128: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32) + ; CI-DS128: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32) + ; CI-DS128: [[BUILD_VECTOR1:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY5]](s32), [[COPY6]](s32), [[COPY7]](s32), [[COPY8]](s32) + ; CI-DS128: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR1]](<4 x s32>) + ; CI-DS128: [[CONCAT_VECTORS:%[0-9]+]]:_(<8 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>) + ; CI-DS128: $vgpr0_vgpr1 = COPY [[CONCAT_VECTORS]](<8 x s8>) ; VI-LABEL: name: test_load_local_v8s8_align8 ; VI: [[COPY:%[0-9]+]]:_(p3) = COPY $vgpr0 ; VI: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p3) :: (load 8, addrspace 3) - ; VI: [[BITCAST:%[0-9]+]]:_(<8 x s8>) = G_BITCAST [[LOAD]](<2 x s32>) - ; VI: $vgpr0_vgpr1 = COPY [[BITCAST]](<8 x s8>) + ; VI: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[LOAD]](<2 x s32>) + ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8 + ; VI: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C]](s32) + ; VI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16 + ; VI: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C1]](s32) + ; VI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24 + ; VI: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C2]](s32) + ; VI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[UV]](s32) + ; VI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32) + ; VI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32) + ; VI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32) + ; VI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32) + ; VI: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>) + ; VI: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C]](s32) + ; VI: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C1]](s32) + ; VI: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C2]](s32) + ; VI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[UV1]](s32) + ; VI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32) + ; VI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32) + ; VI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32) + ; VI: [[BUILD_VECTOR1:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY5]](s32), [[COPY6]](s32), [[COPY7]](s32), [[COPY8]](s32) + ; VI: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR1]](<4 x s32>) + ; VI: [[CONCAT_VECTORS:%[0-9]+]]:_(<8 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>) + ; VI: $vgpr0_vgpr1 = COPY [[CONCAT_VECTORS]](<8 x s8>) ; GFX9-LABEL: name: test_load_local_v8s8_align8 ; GFX9: [[COPY:%[0-9]+]]:_(p3) = COPY $vgpr0 ; GFX9: [[LOAD:%[0-9]+]]:_(<2 x s32>) = G_LOAD [[COPY]](p3) :: (load 8, addrspace 3) - ; GFX9: [[BITCAST:%[0-9]+]]:_(<8 x s8>) = G_BITCAST [[LOAD]](<2 x s32>) - ; GFX9: $vgpr0_vgpr1 = COPY [[BITCAST]](<8 x s8>) + ; GFX9: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[LOAD]](<2 x s32>) + ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8 + ; GFX9: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C]](s32) + ; GFX9: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16 + ; GFX9: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C1]](s32) + ; GFX9: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 24 + ; GFX9: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[UV]], [[C2]](s32) + ; GFX9: [[COPY1:%[0-9]+]]:_(s32) = COPY [[UV]](s32) + ; GFX9: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32) + ; GFX9: [[BUILD_VECTOR_TRUNC:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY1]](s32), [[COPY2]](s32) + ; GFX9: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32) + ; GFX9: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32) + ; GFX9: [[BUILD_VECTOR_TRUNC1:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY3]](s32), [[COPY4]](s32) + ; GFX9: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC]](<2 x s16>), [[BUILD_VECTOR_TRUNC1]](<2 x s16>) + ; GFX9: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS]](<4 x s16>) + ; GFX9: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C]](s32) + ; GFX9: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C1]](s32) + ; GFX9: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[UV1]], [[C2]](s32) + ; GFX9: [[COPY5:%[0-9]+]]:_(s32) = COPY [[UV1]](s32) + ; GFX9: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32) + ; GFX9: [[BUILD_VECTOR_TRUNC2:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY5]](s32), [[COPY6]](s32) + ; GFX9: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32) + ; GFX9: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32) + ; GFX9: [[BUILD_VECTOR_TRUNC3:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY7]](s32), [[COPY8]](s32) + ; GFX9: [[CONCAT_VECTORS1:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC2]](<2 x s16>), [[BUILD_VECTOR_TRUNC3]](<2 x s16>) + ; GFX9: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS1]](<4 x s16>) + ; GFX9: [[CONCAT_VECTORS2:%[0-9]+]]:_(<8 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>) + ; GFX9: $vgpr0_vgpr1 = COPY [[CONCAT_VECTORS2]](<8 x s8>) %0:_(p3) = COPY $vgpr0 %1:_(<8 x s8>) = G_LOAD %0 :: (load 8, align 8, addrspace 3) $vgpr0_vgpr1 = COPY %1 @@ -5690,315 +5804,334 @@ ; SI-LABEL: name: test_load_local_v16s8_align16 ; SI: [[COPY:%[0-9]+]]:_(p3) = COPY $vgpr0 ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p3) :: (load 1, addrspace 3) - ; SI: [[TRUNC:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD]](s32) ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1 ; SI: [[PTR_ADD:%[0-9]+]]:_(p3) = G_PTR_ADD [[COPY]], [[C]](s32) ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p3) :: (load 1 + 1, addrspace 3) - ; SI: [[TRUNC1:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD1]](s32) ; SI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2 ; SI: [[PTR_ADD1:%[0-9]+]]:_(p3) = G_PTR_ADD [[COPY]], [[C1]](s32) ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p3) :: (load 1 + 2, addrspace 3) - ; SI: [[TRUNC2:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD2]](s32) ; SI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3 ; SI: [[PTR_ADD2:%[0-9]+]]:_(p3) = G_PTR_ADD [[COPY]], [[C2]](s32) ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p3) :: (load 1 + 3, addrspace 3) - ; SI: [[TRUNC3:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD3]](s32) - ; SI: [[MV:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC]](s8), [[TRUNC1]](s8), [[TRUNC2]](s8), [[TRUNC3]](s8) ; SI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 4 ; SI: [[PTR_ADD3:%[0-9]+]]:_(p3) = G_PTR_ADD [[COPY]], [[C3]](s32) ; SI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p3) :: (load 1 + 4, addrspace 3) - ; SI: [[TRUNC4:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD4]](s32) ; SI: [[PTR_ADD4:%[0-9]+]]:_(p3) = G_PTR_ADD [[PTR_ADD3]], [[C]](s32) ; SI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p3) :: (load 1 + 5, addrspace 3) - ; SI: [[TRUNC5:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD5]](s32) ; SI: [[PTR_ADD5:%[0-9]+]]:_(p3) = G_PTR_ADD [[PTR_ADD3]], [[C1]](s32) ; SI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p3) :: (load 1 + 6, addrspace 3) - ; SI: [[TRUNC6:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD6]](s32) ; SI: [[PTR_ADD6:%[0-9]+]]:_(p3) = G_PTR_ADD [[PTR_ADD3]], [[C2]](s32) ; SI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD6]](p3) :: (load 1 + 7, addrspace 3) - ; SI: [[TRUNC7:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD7]](s32) - ; SI: [[MV1:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC4]](s8), [[TRUNC5]](s8), [[TRUNC6]](s8), [[TRUNC7]](s8) - ; SI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[MV]](s32), [[MV1]](s32) ; SI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 8 ; SI: [[PTR_ADD7:%[0-9]+]]:_(p3) = G_PTR_ADD [[COPY]], [[C4]](s32) ; SI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD7]](p3) :: (load 1 + 8, addrspace 3) - ; SI: [[TRUNC8:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD8]](s32) ; SI: [[PTR_ADD8:%[0-9]+]]:_(p3) = G_PTR_ADD [[PTR_ADD7]], [[C]](s32) ; SI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD8]](p3) :: (load 1 + 9, addrspace 3) - ; SI: [[TRUNC9:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD9]](s32) ; SI: [[PTR_ADD9:%[0-9]+]]:_(p3) = G_PTR_ADD [[PTR_ADD7]], [[C1]](s32) ; SI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD9]](p3) :: (load 1 + 10, addrspace 3) - ; SI: [[TRUNC10:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD10]](s32) ; SI: [[PTR_ADD10:%[0-9]+]]:_(p3) = G_PTR_ADD [[PTR_ADD7]], [[C2]](s32) ; SI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD10]](p3) :: (load 1 + 11, addrspace 3) - ; SI: [[TRUNC11:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD11]](s32) - ; SI: [[MV2:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC8]](s8), [[TRUNC9]](s8), [[TRUNC10]](s8), [[TRUNC11]](s8) ; SI: [[PTR_ADD11:%[0-9]+]]:_(p3) = G_PTR_ADD [[PTR_ADD7]], [[C3]](s32) ; SI: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD11]](p3) :: (load 1 + 12, addrspace 3) - ; SI: [[TRUNC12:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD12]](s32) ; SI: [[PTR_ADD12:%[0-9]+]]:_(p3) = G_PTR_ADD [[PTR_ADD11]], [[C]](s32) ; SI: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD12]](p3) :: (load 1 + 13, addrspace 3) - ; SI: [[TRUNC13:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD13]](s32) ; SI: [[PTR_ADD13:%[0-9]+]]:_(p3) = G_PTR_ADD [[PTR_ADD11]], [[C1]](s32) ; SI: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD13]](p3) :: (load 1 + 14, addrspace 3) - ; SI: [[TRUNC14:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD14]](s32) ; SI: [[PTR_ADD14:%[0-9]+]]:_(p3) = G_PTR_ADD [[PTR_ADD11]], [[C2]](s32) ; SI: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD14]](p3) :: (load 1 + 15, addrspace 3) - ; SI: [[TRUNC15:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD15]](s32) - ; SI: [[MV3:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC12]](s8), [[TRUNC13]](s8), [[TRUNC14]](s8), [[TRUNC15]](s8) - ; SI: [[BUILD_VECTOR1:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[MV2]](s32), [[MV3]](s32) - ; SI: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s32>) = G_CONCAT_VECTORS [[BUILD_VECTOR]](<2 x s32>), [[BUILD_VECTOR1]](<2 x s32>) - ; SI: [[BITCAST:%[0-9]+]]:_(<16 x s8>) = G_BITCAST [[CONCAT_VECTORS]](<4 x s32>) - ; SI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](<16 x s8>) + ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32) + ; SI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32) + ; SI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32) + ; SI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32) + ; SI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32) + ; SI: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>) + ; SI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD4]](s32) + ; SI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32) + ; SI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD6]](s32) + ; SI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32) + ; SI: [[BUILD_VECTOR1:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY5]](s32), [[COPY6]](s32), [[COPY7]](s32), [[COPY8]](s32) + ; SI: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR1]](<4 x s32>) + ; SI: [[COPY9:%[0-9]+]]:_(s32) = COPY [[LOAD8]](s32) + ; SI: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LOAD9]](s32) + ; SI: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LOAD10]](s32) + ; SI: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LOAD11]](s32) + ; SI: [[BUILD_VECTOR2:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY9]](s32), [[COPY10]](s32), [[COPY11]](s32), [[COPY12]](s32) + ; SI: [[TRUNC2:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR2]](<4 x s32>) + ; SI: [[COPY13:%[0-9]+]]:_(s32) = COPY [[LOAD12]](s32) + ; SI: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LOAD13]](s32) + ; SI: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LOAD14]](s32) + ; SI: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LOAD15]](s32) + ; SI: [[BUILD_VECTOR3:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY13]](s32), [[COPY14]](s32), [[COPY15]](s32), [[COPY16]](s32) + ; SI: [[TRUNC3:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR3]](<4 x s32>) + ; SI: [[CONCAT_VECTORS:%[0-9]+]]:_(<16 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>), [[TRUNC2]](<4 x s8>), [[TRUNC3]](<4 x s8>) + ; SI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[CONCAT_VECTORS]](<16 x s8>) ; CI-LABEL: name: test_load_local_v16s8_align16 ; CI: [[COPY:%[0-9]+]]:_(p3) = COPY $vgpr0 ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p3) :: (load 1, addrspace 3) - ; CI: [[TRUNC:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD]](s32) ; CI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1 ; CI: [[PTR_ADD:%[0-9]+]]:_(p3) = G_PTR_ADD [[COPY]], [[C]](s32) ; CI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p3) :: (load 1 + 1, addrspace 3) - ; CI: [[TRUNC1:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD1]](s32) ; CI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2 ; CI: [[PTR_ADD1:%[0-9]+]]:_(p3) = G_PTR_ADD [[COPY]], [[C1]](s32) ; CI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p3) :: (load 1 + 2, addrspace 3) - ; CI: [[TRUNC2:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD2]](s32) ; CI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3 ; CI: [[PTR_ADD2:%[0-9]+]]:_(p3) = G_PTR_ADD [[COPY]], [[C2]](s32) ; CI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p3) :: (load 1 + 3, addrspace 3) - ; CI: [[TRUNC3:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD3]](s32) - ; CI: [[MV:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC]](s8), [[TRUNC1]](s8), [[TRUNC2]](s8), [[TRUNC3]](s8) ; CI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 4 ; CI: [[PTR_ADD3:%[0-9]+]]:_(p3) = G_PTR_ADD [[COPY]], [[C3]](s32) ; CI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p3) :: (load 1 + 4, addrspace 3) - ; CI: [[TRUNC4:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD4]](s32) ; CI: [[PTR_ADD4:%[0-9]+]]:_(p3) = G_PTR_ADD [[PTR_ADD3]], [[C]](s32) ; CI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p3) :: (load 1 + 5, addrspace 3) - ; CI: [[TRUNC5:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD5]](s32) ; CI: [[PTR_ADD5:%[0-9]+]]:_(p3) = G_PTR_ADD [[PTR_ADD3]], [[C1]](s32) ; CI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p3) :: (load 1 + 6, addrspace 3) - ; CI: [[TRUNC6:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD6]](s32) ; CI: [[PTR_ADD6:%[0-9]+]]:_(p3) = G_PTR_ADD [[PTR_ADD3]], [[C2]](s32) ; CI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD6]](p3) :: (load 1 + 7, addrspace 3) - ; CI: [[TRUNC7:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD7]](s32) - ; CI: [[MV1:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC4]](s8), [[TRUNC5]](s8), [[TRUNC6]](s8), [[TRUNC7]](s8) - ; CI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[MV]](s32), [[MV1]](s32) ; CI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 8 ; CI: [[PTR_ADD7:%[0-9]+]]:_(p3) = G_PTR_ADD [[COPY]], [[C4]](s32) ; CI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD7]](p3) :: (load 1 + 8, addrspace 3) - ; CI: [[TRUNC8:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD8]](s32) ; CI: [[PTR_ADD8:%[0-9]+]]:_(p3) = G_PTR_ADD [[PTR_ADD7]], [[C]](s32) ; CI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD8]](p3) :: (load 1 + 9, addrspace 3) - ; CI: [[TRUNC9:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD9]](s32) ; CI: [[PTR_ADD9:%[0-9]+]]:_(p3) = G_PTR_ADD [[PTR_ADD7]], [[C1]](s32) ; CI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD9]](p3) :: (load 1 + 10, addrspace 3) - ; CI: [[TRUNC10:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD10]](s32) ; CI: [[PTR_ADD10:%[0-9]+]]:_(p3) = G_PTR_ADD [[PTR_ADD7]], [[C2]](s32) ; CI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD10]](p3) :: (load 1 + 11, addrspace 3) - ; CI: [[TRUNC11:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD11]](s32) - ; CI: [[MV2:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC8]](s8), [[TRUNC9]](s8), [[TRUNC10]](s8), [[TRUNC11]](s8) ; CI: [[PTR_ADD11:%[0-9]+]]:_(p3) = G_PTR_ADD [[PTR_ADD7]], [[C3]](s32) ; CI: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD11]](p3) :: (load 1 + 12, addrspace 3) - ; CI: [[TRUNC12:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD12]](s32) ; CI: [[PTR_ADD12:%[0-9]+]]:_(p3) = G_PTR_ADD [[PTR_ADD11]], [[C]](s32) ; CI: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD12]](p3) :: (load 1 + 13, addrspace 3) - ; CI: [[TRUNC13:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD13]](s32) ; CI: [[PTR_ADD13:%[0-9]+]]:_(p3) = G_PTR_ADD [[PTR_ADD11]], [[C1]](s32) ; CI: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD13]](p3) :: (load 1 + 14, addrspace 3) - ; CI: [[TRUNC14:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD14]](s32) ; CI: [[PTR_ADD14:%[0-9]+]]:_(p3) = G_PTR_ADD [[PTR_ADD11]], [[C2]](s32) ; CI: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD14]](p3) :: (load 1 + 15, addrspace 3) - ; CI: [[TRUNC15:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD15]](s32) - ; CI: [[MV3:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC12]](s8), [[TRUNC13]](s8), [[TRUNC14]](s8), [[TRUNC15]](s8) - ; CI: [[BUILD_VECTOR1:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[MV2]](s32), [[MV3]](s32) - ; CI: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s32>) = G_CONCAT_VECTORS [[BUILD_VECTOR]](<2 x s32>), [[BUILD_VECTOR1]](<2 x s32>) - ; CI: [[BITCAST:%[0-9]+]]:_(<16 x s8>) = G_BITCAST [[CONCAT_VECTORS]](<4 x s32>) - ; CI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](<16 x s8>) + ; CI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32) + ; CI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32) + ; CI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32) + ; CI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32) + ; CI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32) + ; CI: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>) + ; CI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD4]](s32) + ; CI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32) + ; CI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD6]](s32) + ; CI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32) + ; CI: [[BUILD_VECTOR1:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY5]](s32), [[COPY6]](s32), [[COPY7]](s32), [[COPY8]](s32) + ; CI: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR1]](<4 x s32>) + ; CI: [[COPY9:%[0-9]+]]:_(s32) = COPY [[LOAD8]](s32) + ; CI: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LOAD9]](s32) + ; CI: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LOAD10]](s32) + ; CI: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LOAD11]](s32) + ; CI: [[BUILD_VECTOR2:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY9]](s32), [[COPY10]](s32), [[COPY11]](s32), [[COPY12]](s32) + ; CI: [[TRUNC2:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR2]](<4 x s32>) + ; CI: [[COPY13:%[0-9]+]]:_(s32) = COPY [[LOAD12]](s32) + ; CI: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LOAD13]](s32) + ; CI: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LOAD14]](s32) + ; CI: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LOAD15]](s32) + ; CI: [[BUILD_VECTOR3:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY13]](s32), [[COPY14]](s32), [[COPY15]](s32), [[COPY16]](s32) + ; CI: [[TRUNC3:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR3]](<4 x s32>) + ; CI: [[CONCAT_VECTORS:%[0-9]+]]:_(<16 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>), [[TRUNC2]](<4 x s8>), [[TRUNC3]](<4 x s8>) + ; CI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[CONCAT_VECTORS]](<16 x s8>) ; CI-DS128-LABEL: name: test_load_local_v16s8_align16 ; CI-DS128: [[COPY:%[0-9]+]]:_(p3) = COPY $vgpr0 ; CI-DS128: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p3) :: (load 1, addrspace 3) - ; CI-DS128: [[TRUNC:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD]](s32) ; CI-DS128: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1 ; CI-DS128: [[PTR_ADD:%[0-9]+]]:_(p3) = G_PTR_ADD [[COPY]], [[C]](s32) ; CI-DS128: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p3) :: (load 1 + 1, addrspace 3) - ; CI-DS128: [[TRUNC1:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD1]](s32) ; CI-DS128: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2 ; CI-DS128: [[PTR_ADD1:%[0-9]+]]:_(p3) = G_PTR_ADD [[COPY]], [[C1]](s32) ; CI-DS128: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p3) :: (load 1 + 2, addrspace 3) - ; CI-DS128: [[TRUNC2:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD2]](s32) ; CI-DS128: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3 ; CI-DS128: [[PTR_ADD2:%[0-9]+]]:_(p3) = G_PTR_ADD [[COPY]], [[C2]](s32) ; CI-DS128: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p3) :: (load 1 + 3, addrspace 3) - ; CI-DS128: [[TRUNC3:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD3]](s32) - ; CI-DS128: [[MV:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC]](s8), [[TRUNC1]](s8), [[TRUNC2]](s8), [[TRUNC3]](s8) ; CI-DS128: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 4 ; CI-DS128: [[PTR_ADD3:%[0-9]+]]:_(p3) = G_PTR_ADD [[COPY]], [[C3]](s32) ; CI-DS128: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p3) :: (load 1 + 4, addrspace 3) - ; CI-DS128: [[TRUNC4:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD4]](s32) ; CI-DS128: [[PTR_ADD4:%[0-9]+]]:_(p3) = G_PTR_ADD [[PTR_ADD3]], [[C]](s32) ; CI-DS128: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p3) :: (load 1 + 5, addrspace 3) - ; CI-DS128: [[TRUNC5:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD5]](s32) ; CI-DS128: [[PTR_ADD5:%[0-9]+]]:_(p3) = G_PTR_ADD [[PTR_ADD3]], [[C1]](s32) ; CI-DS128: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p3) :: (load 1 + 6, addrspace 3) - ; CI-DS128: [[TRUNC6:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD6]](s32) ; CI-DS128: [[PTR_ADD6:%[0-9]+]]:_(p3) = G_PTR_ADD [[PTR_ADD3]], [[C2]](s32) ; CI-DS128: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD6]](p3) :: (load 1 + 7, addrspace 3) - ; CI-DS128: [[TRUNC7:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD7]](s32) - ; CI-DS128: [[MV1:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC4]](s8), [[TRUNC5]](s8), [[TRUNC6]](s8), [[TRUNC7]](s8) ; CI-DS128: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 8 ; CI-DS128: [[PTR_ADD7:%[0-9]+]]:_(p3) = G_PTR_ADD [[COPY]], [[C4]](s32) ; CI-DS128: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD7]](p3) :: (load 1 + 8, addrspace 3) - ; CI-DS128: [[TRUNC8:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD8]](s32) ; CI-DS128: [[PTR_ADD8:%[0-9]+]]:_(p3) = G_PTR_ADD [[PTR_ADD7]], [[C]](s32) ; CI-DS128: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD8]](p3) :: (load 1 + 9, addrspace 3) - ; CI-DS128: [[TRUNC9:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD9]](s32) ; CI-DS128: [[PTR_ADD9:%[0-9]+]]:_(p3) = G_PTR_ADD [[PTR_ADD7]], [[C1]](s32) ; CI-DS128: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD9]](p3) :: (load 1 + 10, addrspace 3) - ; CI-DS128: [[TRUNC10:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD10]](s32) ; CI-DS128: [[PTR_ADD10:%[0-9]+]]:_(p3) = G_PTR_ADD [[PTR_ADD7]], [[C2]](s32) ; CI-DS128: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD10]](p3) :: (load 1 + 11, addrspace 3) - ; CI-DS128: [[TRUNC11:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD11]](s32) - ; CI-DS128: [[MV2:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC8]](s8), [[TRUNC9]](s8), [[TRUNC10]](s8), [[TRUNC11]](s8) ; CI-DS128: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 12 ; CI-DS128: [[PTR_ADD11:%[0-9]+]]:_(p3) = G_PTR_ADD [[COPY]], [[C5]](s32) ; CI-DS128: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD11]](p3) :: (load 1 + 12, addrspace 3) - ; CI-DS128: [[TRUNC12:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD12]](s32) ; CI-DS128: [[PTR_ADD12:%[0-9]+]]:_(p3) = G_PTR_ADD [[PTR_ADD11]], [[C]](s32) ; CI-DS128: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD12]](p3) :: (load 1 + 13, addrspace 3) - ; CI-DS128: [[TRUNC13:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD13]](s32) ; CI-DS128: [[PTR_ADD13:%[0-9]+]]:_(p3) = G_PTR_ADD [[PTR_ADD11]], [[C1]](s32) ; CI-DS128: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD13]](p3) :: (load 1 + 14, addrspace 3) - ; CI-DS128: [[TRUNC14:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD14]](s32) ; CI-DS128: [[PTR_ADD14:%[0-9]+]]:_(p3) = G_PTR_ADD [[PTR_ADD11]], [[C2]](s32) ; CI-DS128: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD14]](p3) :: (load 1 + 15, addrspace 3) - ; CI-DS128: [[TRUNC15:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD15]](s32) - ; CI-DS128: [[MV3:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC12]](s8), [[TRUNC13]](s8), [[TRUNC14]](s8), [[TRUNC15]](s8) - ; CI-DS128: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[MV]](s32), [[MV1]](s32), [[MV2]](s32), [[MV3]](s32) - ; CI-DS128: [[BITCAST:%[0-9]+]]:_(<16 x s8>) = G_BITCAST [[BUILD_VECTOR]](<4 x s32>) - ; CI-DS128: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](<16 x s8>) + ; CI-DS128: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32) + ; CI-DS128: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32) + ; CI-DS128: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32) + ; CI-DS128: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32) + ; CI-DS128: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32) + ; CI-DS128: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>) + ; CI-DS128: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD4]](s32) + ; CI-DS128: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32) + ; CI-DS128: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD6]](s32) + ; CI-DS128: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32) + ; CI-DS128: [[BUILD_VECTOR1:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY5]](s32), [[COPY6]](s32), [[COPY7]](s32), [[COPY8]](s32) + ; CI-DS128: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR1]](<4 x s32>) + ; CI-DS128: [[COPY9:%[0-9]+]]:_(s32) = COPY [[LOAD8]](s32) + ; CI-DS128: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LOAD9]](s32) + ; CI-DS128: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LOAD10]](s32) + ; CI-DS128: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LOAD11]](s32) + ; CI-DS128: [[BUILD_VECTOR2:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY9]](s32), [[COPY10]](s32), [[COPY11]](s32), [[COPY12]](s32) + ; CI-DS128: [[TRUNC2:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR2]](<4 x s32>) + ; CI-DS128: [[COPY13:%[0-9]+]]:_(s32) = COPY [[LOAD12]](s32) + ; CI-DS128: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LOAD13]](s32) + ; CI-DS128: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LOAD14]](s32) + ; CI-DS128: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LOAD15]](s32) + ; CI-DS128: [[BUILD_VECTOR3:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY13]](s32), [[COPY14]](s32), [[COPY15]](s32), [[COPY16]](s32) + ; CI-DS128: [[TRUNC3:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR3]](<4 x s32>) + ; CI-DS128: [[CONCAT_VECTORS:%[0-9]+]]:_(<16 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>), [[TRUNC2]](<4 x s8>), [[TRUNC3]](<4 x s8>) + ; CI-DS128: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[CONCAT_VECTORS]](<16 x s8>) ; VI-LABEL: name: test_load_local_v16s8_align16 ; VI: [[COPY:%[0-9]+]]:_(p3) = COPY $vgpr0 ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p3) :: (load 1, addrspace 3) - ; VI: [[TRUNC:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD]](s32) ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1 ; VI: [[PTR_ADD:%[0-9]+]]:_(p3) = G_PTR_ADD [[COPY]], [[C]](s32) ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p3) :: (load 1 + 1, addrspace 3) - ; VI: [[TRUNC1:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD1]](s32) ; VI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2 ; VI: [[PTR_ADD1:%[0-9]+]]:_(p3) = G_PTR_ADD [[COPY]], [[C1]](s32) ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p3) :: (load 1 + 2, addrspace 3) - ; VI: [[TRUNC2:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD2]](s32) ; VI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3 ; VI: [[PTR_ADD2:%[0-9]+]]:_(p3) = G_PTR_ADD [[COPY]], [[C2]](s32) ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p3) :: (load 1 + 3, addrspace 3) - ; VI: [[TRUNC3:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD3]](s32) - ; VI: [[MV:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC]](s8), [[TRUNC1]](s8), [[TRUNC2]](s8), [[TRUNC3]](s8) ; VI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 4 ; VI: [[PTR_ADD3:%[0-9]+]]:_(p3) = G_PTR_ADD [[COPY]], [[C3]](s32) ; VI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p3) :: (load 1 + 4, addrspace 3) - ; VI: [[TRUNC4:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD4]](s32) ; VI: [[PTR_ADD4:%[0-9]+]]:_(p3) = G_PTR_ADD [[PTR_ADD3]], [[C]](s32) ; VI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p3) :: (load 1 + 5, addrspace 3) - ; VI: [[TRUNC5:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD5]](s32) ; VI: [[PTR_ADD5:%[0-9]+]]:_(p3) = G_PTR_ADD [[PTR_ADD3]], [[C1]](s32) ; VI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p3) :: (load 1 + 6, addrspace 3) - ; VI: [[TRUNC6:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD6]](s32) ; VI: [[PTR_ADD6:%[0-9]+]]:_(p3) = G_PTR_ADD [[PTR_ADD3]], [[C2]](s32) ; VI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD6]](p3) :: (load 1 + 7, addrspace 3) - ; VI: [[TRUNC7:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD7]](s32) - ; VI: [[MV1:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC4]](s8), [[TRUNC5]](s8), [[TRUNC6]](s8), [[TRUNC7]](s8) ; VI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 8 ; VI: [[PTR_ADD7:%[0-9]+]]:_(p3) = G_PTR_ADD [[COPY]], [[C4]](s32) ; VI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD7]](p3) :: (load 1 + 8, addrspace 3) - ; VI: [[TRUNC8:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD8]](s32) ; VI: [[PTR_ADD8:%[0-9]+]]:_(p3) = G_PTR_ADD [[PTR_ADD7]], [[C]](s32) ; VI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD8]](p3) :: (load 1 + 9, addrspace 3) - ; VI: [[TRUNC9:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD9]](s32) ; VI: [[PTR_ADD9:%[0-9]+]]:_(p3) = G_PTR_ADD [[PTR_ADD7]], [[C1]](s32) ; VI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD9]](p3) :: (load 1 + 10, addrspace 3) - ; VI: [[TRUNC10:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD10]](s32) ; VI: [[PTR_ADD10:%[0-9]+]]:_(p3) = G_PTR_ADD [[PTR_ADD7]], [[C2]](s32) ; VI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD10]](p3) :: (load 1 + 11, addrspace 3) - ; VI: [[TRUNC11:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD11]](s32) - ; VI: [[MV2:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC8]](s8), [[TRUNC9]](s8), [[TRUNC10]](s8), [[TRUNC11]](s8) ; VI: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 12 ; VI: [[PTR_ADD11:%[0-9]+]]:_(p3) = G_PTR_ADD [[COPY]], [[C5]](s32) ; VI: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD11]](p3) :: (load 1 + 12, addrspace 3) - ; VI: [[TRUNC12:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD12]](s32) ; VI: [[PTR_ADD12:%[0-9]+]]:_(p3) = G_PTR_ADD [[PTR_ADD11]], [[C]](s32) ; VI: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD12]](p3) :: (load 1 + 13, addrspace 3) - ; VI: [[TRUNC13:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD13]](s32) ; VI: [[PTR_ADD13:%[0-9]+]]:_(p3) = G_PTR_ADD [[PTR_ADD11]], [[C1]](s32) ; VI: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD13]](p3) :: (load 1 + 14, addrspace 3) - ; VI: [[TRUNC14:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD14]](s32) ; VI: [[PTR_ADD14:%[0-9]+]]:_(p3) = G_PTR_ADD [[PTR_ADD11]], [[C2]](s32) ; VI: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD14]](p3) :: (load 1 + 15, addrspace 3) - ; VI: [[TRUNC15:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD15]](s32) - ; VI: [[MV3:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC12]](s8), [[TRUNC13]](s8), [[TRUNC14]](s8), [[TRUNC15]](s8) - ; VI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[MV]](s32), [[MV1]](s32), [[MV2]](s32), [[MV3]](s32) - ; VI: [[BITCAST:%[0-9]+]]:_(<16 x s8>) = G_BITCAST [[BUILD_VECTOR]](<4 x s32>) - ; VI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](<16 x s8>) + ; VI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32) + ; VI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32) + ; VI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32) + ; VI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32) + ; VI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32) + ; VI: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>) + ; VI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD4]](s32) + ; VI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32) + ; VI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD6]](s32) + ; VI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32) + ; VI: [[BUILD_VECTOR1:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY5]](s32), [[COPY6]](s32), [[COPY7]](s32), [[COPY8]](s32) + ; VI: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR1]](<4 x s32>) + ; VI: [[COPY9:%[0-9]+]]:_(s32) = COPY [[LOAD8]](s32) + ; VI: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LOAD9]](s32) + ; VI: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LOAD10]](s32) + ; VI: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LOAD11]](s32) + ; VI: [[BUILD_VECTOR2:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY9]](s32), [[COPY10]](s32), [[COPY11]](s32), [[COPY12]](s32) + ; VI: [[TRUNC2:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR2]](<4 x s32>) + ; VI: [[COPY13:%[0-9]+]]:_(s32) = COPY [[LOAD12]](s32) + ; VI: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LOAD13]](s32) + ; VI: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LOAD14]](s32) + ; VI: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LOAD15]](s32) + ; VI: [[BUILD_VECTOR3:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY13]](s32), [[COPY14]](s32), [[COPY15]](s32), [[COPY16]](s32) + ; VI: [[TRUNC3:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR3]](<4 x s32>) + ; VI: [[CONCAT_VECTORS:%[0-9]+]]:_(<16 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>), [[TRUNC2]](<4 x s8>), [[TRUNC3]](<4 x s8>) + ; VI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[CONCAT_VECTORS]](<16 x s8>) ; GFX9-LABEL: name: test_load_local_v16s8_align16 ; GFX9: [[COPY:%[0-9]+]]:_(p3) = COPY $vgpr0 ; GFX9: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p3) :: (load 1, addrspace 3) - ; GFX9: [[TRUNC:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD]](s32) ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1 ; GFX9: [[PTR_ADD:%[0-9]+]]:_(p3) = G_PTR_ADD [[COPY]], [[C]](s32) ; GFX9: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p3) :: (load 1 + 1, addrspace 3) - ; GFX9: [[TRUNC1:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD1]](s32) ; GFX9: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2 ; GFX9: [[PTR_ADD1:%[0-9]+]]:_(p3) = G_PTR_ADD [[COPY]], [[C1]](s32) ; GFX9: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p3) :: (load 1 + 2, addrspace 3) - ; GFX9: [[TRUNC2:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD2]](s32) ; GFX9: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3 ; GFX9: [[PTR_ADD2:%[0-9]+]]:_(p3) = G_PTR_ADD [[COPY]], [[C2]](s32) ; GFX9: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p3) :: (load 1 + 3, addrspace 3) - ; GFX9: [[TRUNC3:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD3]](s32) - ; GFX9: [[MV:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC]](s8), [[TRUNC1]](s8), [[TRUNC2]](s8), [[TRUNC3]](s8) ; GFX9: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 4 ; GFX9: [[PTR_ADD3:%[0-9]+]]:_(p3) = G_PTR_ADD [[COPY]], [[C3]](s32) ; GFX9: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p3) :: (load 1 + 4, addrspace 3) - ; GFX9: [[TRUNC4:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD4]](s32) ; GFX9: [[PTR_ADD4:%[0-9]+]]:_(p3) = G_PTR_ADD [[PTR_ADD3]], [[C]](s32) ; GFX9: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p3) :: (load 1 + 5, addrspace 3) - ; GFX9: [[TRUNC5:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD5]](s32) ; GFX9: [[PTR_ADD5:%[0-9]+]]:_(p3) = G_PTR_ADD [[PTR_ADD3]], [[C1]](s32) ; GFX9: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p3) :: (load 1 + 6, addrspace 3) - ; GFX9: [[TRUNC6:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD6]](s32) ; GFX9: [[PTR_ADD6:%[0-9]+]]:_(p3) = G_PTR_ADD [[PTR_ADD3]], [[C2]](s32) ; GFX9: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD6]](p3) :: (load 1 + 7, addrspace 3) - ; GFX9: [[TRUNC7:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD7]](s32) - ; GFX9: [[MV1:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC4]](s8), [[TRUNC5]](s8), [[TRUNC6]](s8), [[TRUNC7]](s8) ; GFX9: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 8 ; GFX9: [[PTR_ADD7:%[0-9]+]]:_(p3) = G_PTR_ADD [[COPY]], [[C4]](s32) ; GFX9: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD7]](p3) :: (load 1 + 8, addrspace 3) - ; GFX9: [[TRUNC8:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD8]](s32) ; GFX9: [[PTR_ADD8:%[0-9]+]]:_(p3) = G_PTR_ADD [[PTR_ADD7]], [[C]](s32) ; GFX9: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD8]](p3) :: (load 1 + 9, addrspace 3) - ; GFX9: [[TRUNC9:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD9]](s32) ; GFX9: [[PTR_ADD9:%[0-9]+]]:_(p3) = G_PTR_ADD [[PTR_ADD7]], [[C1]](s32) ; GFX9: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD9]](p3) :: (load 1 + 10, addrspace 3) - ; GFX9: [[TRUNC10:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD10]](s32) ; GFX9: [[PTR_ADD10:%[0-9]+]]:_(p3) = G_PTR_ADD [[PTR_ADD7]], [[C2]](s32) ; GFX9: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD10]](p3) :: (load 1 + 11, addrspace 3) - ; GFX9: [[TRUNC11:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD11]](s32) - ; GFX9: [[MV2:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC8]](s8), [[TRUNC9]](s8), [[TRUNC10]](s8), [[TRUNC11]](s8) ; GFX9: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 12 ; GFX9: [[PTR_ADD11:%[0-9]+]]:_(p3) = G_PTR_ADD [[COPY]], [[C5]](s32) ; GFX9: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD11]](p3) :: (load 1 + 12, addrspace 3) - ; GFX9: [[TRUNC12:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD12]](s32) ; GFX9: [[PTR_ADD12:%[0-9]+]]:_(p3) = G_PTR_ADD [[PTR_ADD11]], [[C]](s32) ; GFX9: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD12]](p3) :: (load 1 + 13, addrspace 3) - ; GFX9: [[TRUNC13:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD13]](s32) ; GFX9: [[PTR_ADD13:%[0-9]+]]:_(p3) = G_PTR_ADD [[PTR_ADD11]], [[C1]](s32) ; GFX9: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD13]](p3) :: (load 1 + 14, addrspace 3) - ; GFX9: [[TRUNC14:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD14]](s32) ; GFX9: [[PTR_ADD14:%[0-9]+]]:_(p3) = G_PTR_ADD [[PTR_ADD11]], [[C2]](s32) ; GFX9: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD14]](p3) :: (load 1 + 15, addrspace 3) - ; GFX9: [[TRUNC15:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD15]](s32) - ; GFX9: [[MV3:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC12]](s8), [[TRUNC13]](s8), [[TRUNC14]](s8), [[TRUNC15]](s8) - ; GFX9: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[MV]](s32), [[MV1]](s32), [[MV2]](s32), [[MV3]](s32) - ; GFX9: [[BITCAST:%[0-9]+]]:_(<16 x s8>) = G_BITCAST [[BUILD_VECTOR]](<4 x s32>) - ; GFX9: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](<16 x s8>) + ; GFX9: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32) + ; GFX9: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32) + ; GFX9: [[BUILD_VECTOR_TRUNC:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY1]](s32), [[COPY2]](s32) + ; GFX9: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32) + ; GFX9: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32) + ; GFX9: [[BUILD_VECTOR_TRUNC1:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY3]](s32), [[COPY4]](s32) + ; GFX9: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC]](<2 x s16>), [[BUILD_VECTOR_TRUNC1]](<2 x s16>) + ; GFX9: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS]](<4 x s16>) + ; GFX9: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD4]](s32) + ; GFX9: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32) + ; GFX9: [[BUILD_VECTOR_TRUNC2:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY5]](s32), [[COPY6]](s32) + ; GFX9: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD6]](s32) + ; GFX9: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32) + ; GFX9: [[BUILD_VECTOR_TRUNC3:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY7]](s32), [[COPY8]](s32) + ; GFX9: [[CONCAT_VECTORS1:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC2]](<2 x s16>), [[BUILD_VECTOR_TRUNC3]](<2 x s16>) + ; GFX9: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS1]](<4 x s16>) + ; GFX9: [[COPY9:%[0-9]+]]:_(s32) = COPY [[LOAD8]](s32) + ; GFX9: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LOAD9]](s32) + ; GFX9: [[BUILD_VECTOR_TRUNC4:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY9]](s32), [[COPY10]](s32) + ; GFX9: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LOAD10]](s32) + ; GFX9: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LOAD11]](s32) + ; GFX9: [[BUILD_VECTOR_TRUNC5:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY11]](s32), [[COPY12]](s32) + ; GFX9: [[CONCAT_VECTORS2:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC4]](<2 x s16>), [[BUILD_VECTOR_TRUNC5]](<2 x s16>) + ; GFX9: [[TRUNC2:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS2]](<4 x s16>) + ; GFX9: [[COPY13:%[0-9]+]]:_(s32) = COPY [[LOAD12]](s32) + ; GFX9: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LOAD13]](s32) + ; GFX9: [[BUILD_VECTOR_TRUNC6:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY13]](s32), [[COPY14]](s32) + ; GFX9: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LOAD14]](s32) + ; GFX9: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LOAD15]](s32) + ; GFX9: [[BUILD_VECTOR_TRUNC7:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY15]](s32), [[COPY16]](s32) + ; GFX9: [[CONCAT_VECTORS3:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC6]](<2 x s16>), [[BUILD_VECTOR_TRUNC7]](<2 x s16>) + ; GFX9: [[TRUNC3:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS3]](<4 x s16>) + ; GFX9: [[CONCAT_VECTORS4:%[0-9]+]]:_(<16 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>), [[TRUNC2]](<4 x s8>), [[TRUNC3]](<4 x s8>) + ; GFX9: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[CONCAT_VECTORS4]](<16 x s8>) %0:_(p3) = COPY $vgpr0 %1:_(<16 x s8>) = G_LOAD %0 :: (load 16, align 1, addrspace 3) $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %1 Index: llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-load-private.mir =================================================================== --- llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-load-private.mir +++ llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-load-private.mir @@ -4431,36 +4431,120 @@ ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4 ; SI: [[PTR_ADD:%[0-9]+]]:_(p5) = G_PTR_ADD [[COPY]], [[C]](s32) ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p5) :: (load 4 + 4, addrspace 5) - ; SI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[LOAD]](s32), [[LOAD1]](s32) - ; SI: [[BITCAST:%[0-9]+]]:_(<8 x s8>) = G_BITCAST [[BUILD_VECTOR]](<2 x s32>) - ; SI: $vgpr0_vgpr1 = COPY [[BITCAST]](<8 x s8>) + ; SI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8 + ; SI: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C1]](s32) + ; SI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 16 + ; SI: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C2]](s32) + ; SI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 24 + ; SI: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C3]](s32) + ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32) + ; SI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32) + ; SI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32) + ; SI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32) + ; SI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32) + ; SI: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>) + ; SI: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[LOAD1]], [[C1]](s32) + ; SI: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[LOAD1]], [[C2]](s32) + ; SI: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[LOAD1]], [[C3]](s32) + ; SI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32) + ; SI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32) + ; SI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32) + ; SI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32) + ; SI: [[BUILD_VECTOR1:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY5]](s32), [[COPY6]](s32), [[COPY7]](s32), [[COPY8]](s32) + ; SI: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR1]](<4 x s32>) + ; SI: [[CONCAT_VECTORS:%[0-9]+]]:_(<8 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>) + ; SI: $vgpr0_vgpr1 = COPY [[CONCAT_VECTORS]](<8 x s8>) ; CI-LABEL: name: test_load_private_v8s8_align8 ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0 ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, align 8, addrspace 5) ; CI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4 ; CI: [[PTR_ADD:%[0-9]+]]:_(p5) = G_PTR_ADD [[COPY]], [[C]](s32) ; CI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p5) :: (load 4 + 4, addrspace 5) - ; CI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[LOAD]](s32), [[LOAD1]](s32) - ; CI: [[BITCAST:%[0-9]+]]:_(<8 x s8>) = G_BITCAST [[BUILD_VECTOR]](<2 x s32>) - ; CI: $vgpr0_vgpr1 = COPY [[BITCAST]](<8 x s8>) + ; CI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8 + ; CI: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C1]](s32) + ; CI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 16 + ; CI: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C2]](s32) + ; CI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 24 + ; CI: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C3]](s32) + ; CI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32) + ; CI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32) + ; CI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32) + ; CI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32) + ; CI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32) + ; CI: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>) + ; CI: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[LOAD1]], [[C1]](s32) + ; CI: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[LOAD1]], [[C2]](s32) + ; CI: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[LOAD1]], [[C3]](s32) + ; CI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32) + ; CI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32) + ; CI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32) + ; CI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32) + ; CI: [[BUILD_VECTOR1:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY5]](s32), [[COPY6]](s32), [[COPY7]](s32), [[COPY8]](s32) + ; CI: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR1]](<4 x s32>) + ; CI: [[CONCAT_VECTORS:%[0-9]+]]:_(<8 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>) + ; CI: $vgpr0_vgpr1 = COPY [[CONCAT_VECTORS]](<8 x s8>) ; VI-LABEL: name: test_load_private_v8s8_align8 ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0 ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, align 8, addrspace 5) ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4 ; VI: [[PTR_ADD:%[0-9]+]]:_(p5) = G_PTR_ADD [[COPY]], [[C]](s32) ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p5) :: (load 4 + 4, addrspace 5) - ; VI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[LOAD]](s32), [[LOAD1]](s32) - ; VI: [[BITCAST:%[0-9]+]]:_(<8 x s8>) = G_BITCAST [[BUILD_VECTOR]](<2 x s32>) - ; VI: $vgpr0_vgpr1 = COPY [[BITCAST]](<8 x s8>) + ; VI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8 + ; VI: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C1]](s32) + ; VI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 16 + ; VI: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C2]](s32) + ; VI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 24 + ; VI: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C3]](s32) + ; VI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32) + ; VI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32) + ; VI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32) + ; VI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32) + ; VI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32) + ; VI: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>) + ; VI: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[LOAD1]], [[C1]](s32) + ; VI: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[LOAD1]], [[C2]](s32) + ; VI: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[LOAD1]], [[C3]](s32) + ; VI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32) + ; VI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32) + ; VI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32) + ; VI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32) + ; VI: [[BUILD_VECTOR1:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY5]](s32), [[COPY6]](s32), [[COPY7]](s32), [[COPY8]](s32) + ; VI: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR1]](<4 x s32>) + ; VI: [[CONCAT_VECTORS:%[0-9]+]]:_(<8 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>) + ; VI: $vgpr0_vgpr1 = COPY [[CONCAT_VECTORS]](<8 x s8>) ; GFX9-LABEL: name: test_load_private_v8s8_align8 ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0 ; GFX9: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 4, align 8, addrspace 5) ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 4 ; GFX9: [[PTR_ADD:%[0-9]+]]:_(p5) = G_PTR_ADD [[COPY]], [[C]](s32) ; GFX9: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p5) :: (load 4 + 4, addrspace 5) - ; GFX9: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[LOAD]](s32), [[LOAD1]](s32) - ; GFX9: [[BITCAST:%[0-9]+]]:_(<8 x s8>) = G_BITCAST [[BUILD_VECTOR]](<2 x s32>) - ; GFX9: $vgpr0_vgpr1 = COPY [[BITCAST]](<8 x s8>) + ; GFX9: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8 + ; GFX9: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C1]](s32) + ; GFX9: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 16 + ; GFX9: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C2]](s32) + ; GFX9: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 24 + ; GFX9: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C3]](s32) + ; GFX9: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32) + ; GFX9: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32) + ; GFX9: [[BUILD_VECTOR_TRUNC:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY1]](s32), [[COPY2]](s32) + ; GFX9: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LSHR1]](s32) + ; GFX9: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LSHR2]](s32) + ; GFX9: [[BUILD_VECTOR_TRUNC1:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY3]](s32), [[COPY4]](s32) + ; GFX9: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC]](<2 x s16>), [[BUILD_VECTOR_TRUNC1]](<2 x s16>) + ; GFX9: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS]](<4 x s16>) + ; GFX9: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[LOAD1]], [[C1]](s32) + ; GFX9: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[LOAD1]], [[C2]](s32) + ; GFX9: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[LOAD1]], [[C3]](s32) + ; GFX9: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32) + ; GFX9: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LSHR3]](s32) + ; GFX9: [[BUILD_VECTOR_TRUNC2:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY5]](s32), [[COPY6]](s32) + ; GFX9: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LSHR4]](s32) + ; GFX9: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LSHR5]](s32) + ; GFX9: [[BUILD_VECTOR_TRUNC3:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY7]](s32), [[COPY8]](s32) + ; GFX9: [[CONCAT_VECTORS1:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC2]](<2 x s16>), [[BUILD_VECTOR_TRUNC3]](<2 x s16>) + ; GFX9: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS1]](<4 x s16>) + ; GFX9: [[CONCAT_VECTORS2:%[0-9]+]]:_(<8 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>) + ; GFX9: $vgpr0_vgpr1 = COPY [[CONCAT_VECTORS2]](<8 x s8>) %0:_(p5) = COPY $vgpr0 %1:_(<8 x s8>) = G_LOAD %0 :: (load 8, align 8, addrspace 5) $vgpr0_vgpr1 = COPY %1 @@ -4475,251 +4559,271 @@ ; SI-LABEL: name: test_load_private_v16s8_align16 ; SI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0 ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 56) - ; SI: [[TRUNC:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD]](s32) ; SI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1 ; SI: [[PTR_ADD:%[0-9]+]]:_(p5) = G_PTR_ADD [[COPY]], [[C]](s32) ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p5) :: (load 1 + 1, addrspace 56) - ; SI: [[TRUNC1:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD1]](s32) ; SI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2 ; SI: [[PTR_ADD1:%[0-9]+]]:_(p5) = G_PTR_ADD [[COPY]], [[C1]](s32) ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p5) :: (load 1 + 2, addrspace 56) - ; SI: [[TRUNC2:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD2]](s32) ; SI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3 ; SI: [[PTR_ADD2:%[0-9]+]]:_(p5) = G_PTR_ADD [[COPY]], [[C2]](s32) ; SI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p5) :: (load 1 + 3, addrspace 56) - ; SI: [[TRUNC3:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD3]](s32) - ; SI: [[MV:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC]](s8), [[TRUNC1]](s8), [[TRUNC2]](s8), [[TRUNC3]](s8) ; SI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 4 ; SI: [[PTR_ADD3:%[0-9]+]]:_(p5) = G_PTR_ADD [[COPY]], [[C3]](s32) ; SI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p5) :: (load 1 + 4, addrspace 56) - ; SI: [[TRUNC4:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD4]](s32) ; SI: [[PTR_ADD4:%[0-9]+]]:_(p5) = G_PTR_ADD [[PTR_ADD3]], [[C]](s32) ; SI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p5) :: (load 1 + 5, addrspace 56) - ; SI: [[TRUNC5:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD5]](s32) ; SI: [[PTR_ADD5:%[0-9]+]]:_(p5) = G_PTR_ADD [[PTR_ADD3]], [[C1]](s32) ; SI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p5) :: (load 1 + 6, addrspace 56) - ; SI: [[TRUNC6:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD6]](s32) ; SI: [[PTR_ADD6:%[0-9]+]]:_(p5) = G_PTR_ADD [[PTR_ADD3]], [[C2]](s32) ; SI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD6]](p5) :: (load 1 + 7, addrspace 56) - ; SI: [[TRUNC7:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD7]](s32) - ; SI: [[MV1:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC4]](s8), [[TRUNC5]](s8), [[TRUNC6]](s8), [[TRUNC7]](s8) ; SI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 8 ; SI: [[PTR_ADD7:%[0-9]+]]:_(p5) = G_PTR_ADD [[COPY]], [[C4]](s32) ; SI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD7]](p5) :: (load 1 + 8, addrspace 56) - ; SI: [[TRUNC8:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD8]](s32) ; SI: [[PTR_ADD8:%[0-9]+]]:_(p5) = G_PTR_ADD [[PTR_ADD7]], [[C]](s32) ; SI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD8]](p5) :: (load 1 + 9, addrspace 56) - ; SI: [[TRUNC9:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD9]](s32) ; SI: [[PTR_ADD9:%[0-9]+]]:_(p5) = G_PTR_ADD [[PTR_ADD7]], [[C1]](s32) ; SI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD9]](p5) :: (load 1 + 10, addrspace 56) - ; SI: [[TRUNC10:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD10]](s32) ; SI: [[PTR_ADD10:%[0-9]+]]:_(p5) = G_PTR_ADD [[PTR_ADD7]], [[C2]](s32) ; SI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD10]](p5) :: (load 1 + 11, addrspace 56) - ; SI: [[TRUNC11:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD11]](s32) - ; SI: [[MV2:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC8]](s8), [[TRUNC9]](s8), [[TRUNC10]](s8), [[TRUNC11]](s8) ; SI: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 12 ; SI: [[PTR_ADD11:%[0-9]+]]:_(p5) = G_PTR_ADD [[COPY]], [[C5]](s32) ; SI: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD11]](p5) :: (load 1 + 12, addrspace 56) - ; SI: [[TRUNC12:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD12]](s32) ; SI: [[PTR_ADD12:%[0-9]+]]:_(p5) = G_PTR_ADD [[PTR_ADD11]], [[C]](s32) ; SI: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD12]](p5) :: (load 1 + 13, addrspace 56) - ; SI: [[TRUNC13:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD13]](s32) ; SI: [[PTR_ADD13:%[0-9]+]]:_(p5) = G_PTR_ADD [[PTR_ADD11]], [[C1]](s32) ; SI: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD13]](p5) :: (load 1 + 14, addrspace 56) - ; SI: [[TRUNC14:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD14]](s32) ; SI: [[PTR_ADD14:%[0-9]+]]:_(p5) = G_PTR_ADD [[PTR_ADD11]], [[C2]](s32) ; SI: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD14]](p5) :: (load 1 + 15, addrspace 56) - ; SI: [[TRUNC15:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD15]](s32) - ; SI: [[MV3:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC12]](s8), [[TRUNC13]](s8), [[TRUNC14]](s8), [[TRUNC15]](s8) - ; SI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[MV]](s32), [[MV1]](s32), [[MV2]](s32), [[MV3]](s32) - ; SI: [[BITCAST:%[0-9]+]]:_(<16 x s8>) = G_BITCAST [[BUILD_VECTOR]](<4 x s32>) - ; SI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](<16 x s8>) + ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32) + ; SI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32) + ; SI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32) + ; SI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32) + ; SI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32) + ; SI: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>) + ; SI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD4]](s32) + ; SI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32) + ; SI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD6]](s32) + ; SI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32) + ; SI: [[BUILD_VECTOR1:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY5]](s32), [[COPY6]](s32), [[COPY7]](s32), [[COPY8]](s32) + ; SI: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR1]](<4 x s32>) + ; SI: [[COPY9:%[0-9]+]]:_(s32) = COPY [[LOAD8]](s32) + ; SI: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LOAD9]](s32) + ; SI: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LOAD10]](s32) + ; SI: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LOAD11]](s32) + ; SI: [[BUILD_VECTOR2:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY9]](s32), [[COPY10]](s32), [[COPY11]](s32), [[COPY12]](s32) + ; SI: [[TRUNC2:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR2]](<4 x s32>) + ; SI: [[COPY13:%[0-9]+]]:_(s32) = COPY [[LOAD12]](s32) + ; SI: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LOAD13]](s32) + ; SI: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LOAD14]](s32) + ; SI: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LOAD15]](s32) + ; SI: [[BUILD_VECTOR3:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY13]](s32), [[COPY14]](s32), [[COPY15]](s32), [[COPY16]](s32) + ; SI: [[TRUNC3:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR3]](<4 x s32>) + ; SI: [[CONCAT_VECTORS:%[0-9]+]]:_(<16 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>), [[TRUNC2]](<4 x s8>), [[TRUNC3]](<4 x s8>) + ; SI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[CONCAT_VECTORS]](<16 x s8>) ; CI-LABEL: name: test_load_private_v16s8_align16 ; CI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0 ; CI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 56) - ; CI: [[TRUNC:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD]](s32) ; CI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1 ; CI: [[PTR_ADD:%[0-9]+]]:_(p5) = G_PTR_ADD [[COPY]], [[C]](s32) ; CI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p5) :: (load 1 + 1, addrspace 56) - ; CI: [[TRUNC1:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD1]](s32) ; CI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2 ; CI: [[PTR_ADD1:%[0-9]+]]:_(p5) = G_PTR_ADD [[COPY]], [[C1]](s32) ; CI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p5) :: (load 1 + 2, addrspace 56) - ; CI: [[TRUNC2:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD2]](s32) ; CI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3 ; CI: [[PTR_ADD2:%[0-9]+]]:_(p5) = G_PTR_ADD [[COPY]], [[C2]](s32) ; CI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p5) :: (load 1 + 3, addrspace 56) - ; CI: [[TRUNC3:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD3]](s32) - ; CI: [[MV:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC]](s8), [[TRUNC1]](s8), [[TRUNC2]](s8), [[TRUNC3]](s8) ; CI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 4 ; CI: [[PTR_ADD3:%[0-9]+]]:_(p5) = G_PTR_ADD [[COPY]], [[C3]](s32) ; CI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p5) :: (load 1 + 4, addrspace 56) - ; CI: [[TRUNC4:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD4]](s32) ; CI: [[PTR_ADD4:%[0-9]+]]:_(p5) = G_PTR_ADD [[PTR_ADD3]], [[C]](s32) ; CI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p5) :: (load 1 + 5, addrspace 56) - ; CI: [[TRUNC5:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD5]](s32) ; CI: [[PTR_ADD5:%[0-9]+]]:_(p5) = G_PTR_ADD [[PTR_ADD3]], [[C1]](s32) ; CI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p5) :: (load 1 + 6, addrspace 56) - ; CI: [[TRUNC6:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD6]](s32) ; CI: [[PTR_ADD6:%[0-9]+]]:_(p5) = G_PTR_ADD [[PTR_ADD3]], [[C2]](s32) ; CI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD6]](p5) :: (load 1 + 7, addrspace 56) - ; CI: [[TRUNC7:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD7]](s32) - ; CI: [[MV1:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC4]](s8), [[TRUNC5]](s8), [[TRUNC6]](s8), [[TRUNC7]](s8) ; CI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 8 ; CI: [[PTR_ADD7:%[0-9]+]]:_(p5) = G_PTR_ADD [[COPY]], [[C4]](s32) ; CI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD7]](p5) :: (load 1 + 8, addrspace 56) - ; CI: [[TRUNC8:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD8]](s32) ; CI: [[PTR_ADD8:%[0-9]+]]:_(p5) = G_PTR_ADD [[PTR_ADD7]], [[C]](s32) ; CI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD8]](p5) :: (load 1 + 9, addrspace 56) - ; CI: [[TRUNC9:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD9]](s32) ; CI: [[PTR_ADD9:%[0-9]+]]:_(p5) = G_PTR_ADD [[PTR_ADD7]], [[C1]](s32) ; CI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD9]](p5) :: (load 1 + 10, addrspace 56) - ; CI: [[TRUNC10:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD10]](s32) ; CI: [[PTR_ADD10:%[0-9]+]]:_(p5) = G_PTR_ADD [[PTR_ADD7]], [[C2]](s32) ; CI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD10]](p5) :: (load 1 + 11, addrspace 56) - ; CI: [[TRUNC11:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD11]](s32) - ; CI: [[MV2:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC8]](s8), [[TRUNC9]](s8), [[TRUNC10]](s8), [[TRUNC11]](s8) ; CI: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 12 ; CI: [[PTR_ADD11:%[0-9]+]]:_(p5) = G_PTR_ADD [[COPY]], [[C5]](s32) ; CI: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD11]](p5) :: (load 1 + 12, addrspace 56) - ; CI: [[TRUNC12:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD12]](s32) ; CI: [[PTR_ADD12:%[0-9]+]]:_(p5) = G_PTR_ADD [[PTR_ADD11]], [[C]](s32) ; CI: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD12]](p5) :: (load 1 + 13, addrspace 56) - ; CI: [[TRUNC13:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD13]](s32) ; CI: [[PTR_ADD13:%[0-9]+]]:_(p5) = G_PTR_ADD [[PTR_ADD11]], [[C1]](s32) ; CI: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD13]](p5) :: (load 1 + 14, addrspace 56) - ; CI: [[TRUNC14:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD14]](s32) ; CI: [[PTR_ADD14:%[0-9]+]]:_(p5) = G_PTR_ADD [[PTR_ADD11]], [[C2]](s32) ; CI: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD14]](p5) :: (load 1 + 15, addrspace 56) - ; CI: [[TRUNC15:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD15]](s32) - ; CI: [[MV3:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC12]](s8), [[TRUNC13]](s8), [[TRUNC14]](s8), [[TRUNC15]](s8) - ; CI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[MV]](s32), [[MV1]](s32), [[MV2]](s32), [[MV3]](s32) - ; CI: [[BITCAST:%[0-9]+]]:_(<16 x s8>) = G_BITCAST [[BUILD_VECTOR]](<4 x s32>) - ; CI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](<16 x s8>) + ; CI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32) + ; CI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32) + ; CI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32) + ; CI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32) + ; CI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32) + ; CI: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>) + ; CI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD4]](s32) + ; CI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32) + ; CI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD6]](s32) + ; CI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32) + ; CI: [[BUILD_VECTOR1:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY5]](s32), [[COPY6]](s32), [[COPY7]](s32), [[COPY8]](s32) + ; CI: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR1]](<4 x s32>) + ; CI: [[COPY9:%[0-9]+]]:_(s32) = COPY [[LOAD8]](s32) + ; CI: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LOAD9]](s32) + ; CI: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LOAD10]](s32) + ; CI: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LOAD11]](s32) + ; CI: [[BUILD_VECTOR2:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY9]](s32), [[COPY10]](s32), [[COPY11]](s32), [[COPY12]](s32) + ; CI: [[TRUNC2:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR2]](<4 x s32>) + ; CI: [[COPY13:%[0-9]+]]:_(s32) = COPY [[LOAD12]](s32) + ; CI: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LOAD13]](s32) + ; CI: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LOAD14]](s32) + ; CI: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LOAD15]](s32) + ; CI: [[BUILD_VECTOR3:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY13]](s32), [[COPY14]](s32), [[COPY15]](s32), [[COPY16]](s32) + ; CI: [[TRUNC3:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR3]](<4 x s32>) + ; CI: [[CONCAT_VECTORS:%[0-9]+]]:_(<16 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>), [[TRUNC2]](<4 x s8>), [[TRUNC3]](<4 x s8>) + ; CI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[CONCAT_VECTORS]](<16 x s8>) ; VI-LABEL: name: test_load_private_v16s8_align16 ; VI: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0 ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 56) - ; VI: [[TRUNC:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD]](s32) ; VI: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1 ; VI: [[PTR_ADD:%[0-9]+]]:_(p5) = G_PTR_ADD [[COPY]], [[C]](s32) ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p5) :: (load 1 + 1, addrspace 56) - ; VI: [[TRUNC1:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD1]](s32) ; VI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2 ; VI: [[PTR_ADD1:%[0-9]+]]:_(p5) = G_PTR_ADD [[COPY]], [[C1]](s32) ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p5) :: (load 1 + 2, addrspace 56) - ; VI: [[TRUNC2:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD2]](s32) ; VI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3 ; VI: [[PTR_ADD2:%[0-9]+]]:_(p5) = G_PTR_ADD [[COPY]], [[C2]](s32) ; VI: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p5) :: (load 1 + 3, addrspace 56) - ; VI: [[TRUNC3:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD3]](s32) - ; VI: [[MV:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC]](s8), [[TRUNC1]](s8), [[TRUNC2]](s8), [[TRUNC3]](s8) ; VI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 4 ; VI: [[PTR_ADD3:%[0-9]+]]:_(p5) = G_PTR_ADD [[COPY]], [[C3]](s32) ; VI: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p5) :: (load 1 + 4, addrspace 56) - ; VI: [[TRUNC4:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD4]](s32) ; VI: [[PTR_ADD4:%[0-9]+]]:_(p5) = G_PTR_ADD [[PTR_ADD3]], [[C]](s32) ; VI: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p5) :: (load 1 + 5, addrspace 56) - ; VI: [[TRUNC5:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD5]](s32) ; VI: [[PTR_ADD5:%[0-9]+]]:_(p5) = G_PTR_ADD [[PTR_ADD3]], [[C1]](s32) ; VI: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p5) :: (load 1 + 6, addrspace 56) - ; VI: [[TRUNC6:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD6]](s32) ; VI: [[PTR_ADD6:%[0-9]+]]:_(p5) = G_PTR_ADD [[PTR_ADD3]], [[C2]](s32) ; VI: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD6]](p5) :: (load 1 + 7, addrspace 56) - ; VI: [[TRUNC7:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD7]](s32) - ; VI: [[MV1:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC4]](s8), [[TRUNC5]](s8), [[TRUNC6]](s8), [[TRUNC7]](s8) ; VI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 8 ; VI: [[PTR_ADD7:%[0-9]+]]:_(p5) = G_PTR_ADD [[COPY]], [[C4]](s32) ; VI: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD7]](p5) :: (load 1 + 8, addrspace 56) - ; VI: [[TRUNC8:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD8]](s32) ; VI: [[PTR_ADD8:%[0-9]+]]:_(p5) = G_PTR_ADD [[PTR_ADD7]], [[C]](s32) ; VI: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD8]](p5) :: (load 1 + 9, addrspace 56) - ; VI: [[TRUNC9:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD9]](s32) ; VI: [[PTR_ADD9:%[0-9]+]]:_(p5) = G_PTR_ADD [[PTR_ADD7]], [[C1]](s32) ; VI: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD9]](p5) :: (load 1 + 10, addrspace 56) - ; VI: [[TRUNC10:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD10]](s32) ; VI: [[PTR_ADD10:%[0-9]+]]:_(p5) = G_PTR_ADD [[PTR_ADD7]], [[C2]](s32) ; VI: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD10]](p5) :: (load 1 + 11, addrspace 56) - ; VI: [[TRUNC11:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD11]](s32) - ; VI: [[MV2:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC8]](s8), [[TRUNC9]](s8), [[TRUNC10]](s8), [[TRUNC11]](s8) ; VI: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 12 ; VI: [[PTR_ADD11:%[0-9]+]]:_(p5) = G_PTR_ADD [[COPY]], [[C5]](s32) ; VI: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD11]](p5) :: (load 1 + 12, addrspace 56) - ; VI: [[TRUNC12:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD12]](s32) ; VI: [[PTR_ADD12:%[0-9]+]]:_(p5) = G_PTR_ADD [[PTR_ADD11]], [[C]](s32) ; VI: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD12]](p5) :: (load 1 + 13, addrspace 56) - ; VI: [[TRUNC13:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD13]](s32) ; VI: [[PTR_ADD13:%[0-9]+]]:_(p5) = G_PTR_ADD [[PTR_ADD11]], [[C1]](s32) ; VI: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD13]](p5) :: (load 1 + 14, addrspace 56) - ; VI: [[TRUNC14:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD14]](s32) ; VI: [[PTR_ADD14:%[0-9]+]]:_(p5) = G_PTR_ADD [[PTR_ADD11]], [[C2]](s32) ; VI: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD14]](p5) :: (load 1 + 15, addrspace 56) - ; VI: [[TRUNC15:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD15]](s32) - ; VI: [[MV3:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC12]](s8), [[TRUNC13]](s8), [[TRUNC14]](s8), [[TRUNC15]](s8) - ; VI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[MV]](s32), [[MV1]](s32), [[MV2]](s32), [[MV3]](s32) - ; VI: [[BITCAST:%[0-9]+]]:_(<16 x s8>) = G_BITCAST [[BUILD_VECTOR]](<4 x s32>) - ; VI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](<16 x s8>) + ; VI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32) + ; VI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32) + ; VI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32) + ; VI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32) + ; VI: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY1]](s32), [[COPY2]](s32), [[COPY3]](s32), [[COPY4]](s32) + ; VI: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR]](<4 x s32>) + ; VI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD4]](s32) + ; VI: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32) + ; VI: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD6]](s32) + ; VI: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32) + ; VI: [[BUILD_VECTOR1:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY5]](s32), [[COPY6]](s32), [[COPY7]](s32), [[COPY8]](s32) + ; VI: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR1]](<4 x s32>) + ; VI: [[COPY9:%[0-9]+]]:_(s32) = COPY [[LOAD8]](s32) + ; VI: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LOAD9]](s32) + ; VI: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LOAD10]](s32) + ; VI: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LOAD11]](s32) + ; VI: [[BUILD_VECTOR2:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY9]](s32), [[COPY10]](s32), [[COPY11]](s32), [[COPY12]](s32) + ; VI: [[TRUNC2:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR2]](<4 x s32>) + ; VI: [[COPY13:%[0-9]+]]:_(s32) = COPY [[LOAD12]](s32) + ; VI: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LOAD13]](s32) + ; VI: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LOAD14]](s32) + ; VI: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LOAD15]](s32) + ; VI: [[BUILD_VECTOR3:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[COPY13]](s32), [[COPY14]](s32), [[COPY15]](s32), [[COPY16]](s32) + ; VI: [[TRUNC3:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[BUILD_VECTOR3]](<4 x s32>) + ; VI: [[CONCAT_VECTORS:%[0-9]+]]:_(<16 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>), [[TRUNC2]](<4 x s8>), [[TRUNC3]](<4 x s8>) + ; VI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[CONCAT_VECTORS]](<16 x s8>) ; GFX9-LABEL: name: test_load_private_v16s8_align16 ; GFX9: [[COPY:%[0-9]+]]:_(p5) = COPY $vgpr0 ; GFX9: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p5) :: (load 1, addrspace 56) - ; GFX9: [[TRUNC:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD]](s32) ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 1 ; GFX9: [[PTR_ADD:%[0-9]+]]:_(p5) = G_PTR_ADD [[COPY]], [[C]](s32) ; GFX9: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p5) :: (load 1 + 1, addrspace 56) - ; GFX9: [[TRUNC1:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD1]](s32) ; GFX9: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 2 ; GFX9: [[PTR_ADD1:%[0-9]+]]:_(p5) = G_PTR_ADD [[COPY]], [[C1]](s32) ; GFX9: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p5) :: (load 1 + 2, addrspace 56) - ; GFX9: [[TRUNC2:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD2]](s32) ; GFX9: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 3 ; GFX9: [[PTR_ADD2:%[0-9]+]]:_(p5) = G_PTR_ADD [[COPY]], [[C2]](s32) ; GFX9: [[LOAD3:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD2]](p5) :: (load 1 + 3, addrspace 56) - ; GFX9: [[TRUNC3:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD3]](s32) - ; GFX9: [[MV:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC]](s8), [[TRUNC1]](s8), [[TRUNC2]](s8), [[TRUNC3]](s8) ; GFX9: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 4 ; GFX9: [[PTR_ADD3:%[0-9]+]]:_(p5) = G_PTR_ADD [[COPY]], [[C3]](s32) ; GFX9: [[LOAD4:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD3]](p5) :: (load 1 + 4, addrspace 56) - ; GFX9: [[TRUNC4:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD4]](s32) ; GFX9: [[PTR_ADD4:%[0-9]+]]:_(p5) = G_PTR_ADD [[PTR_ADD3]], [[C]](s32) ; GFX9: [[LOAD5:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD4]](p5) :: (load 1 + 5, addrspace 56) - ; GFX9: [[TRUNC5:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD5]](s32) ; GFX9: [[PTR_ADD5:%[0-9]+]]:_(p5) = G_PTR_ADD [[PTR_ADD3]], [[C1]](s32) ; GFX9: [[LOAD6:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD5]](p5) :: (load 1 + 6, addrspace 56) - ; GFX9: [[TRUNC6:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD6]](s32) ; GFX9: [[PTR_ADD6:%[0-9]+]]:_(p5) = G_PTR_ADD [[PTR_ADD3]], [[C2]](s32) ; GFX9: [[LOAD7:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD6]](p5) :: (load 1 + 7, addrspace 56) - ; GFX9: [[TRUNC7:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD7]](s32) - ; GFX9: [[MV1:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC4]](s8), [[TRUNC5]](s8), [[TRUNC6]](s8), [[TRUNC7]](s8) ; GFX9: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 8 ; GFX9: [[PTR_ADD7:%[0-9]+]]:_(p5) = G_PTR_ADD [[COPY]], [[C4]](s32) ; GFX9: [[LOAD8:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD7]](p5) :: (load 1 + 8, addrspace 56) - ; GFX9: [[TRUNC8:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD8]](s32) ; GFX9: [[PTR_ADD8:%[0-9]+]]:_(p5) = G_PTR_ADD [[PTR_ADD7]], [[C]](s32) ; GFX9: [[LOAD9:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD8]](p5) :: (load 1 + 9, addrspace 56) - ; GFX9: [[TRUNC9:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD9]](s32) ; GFX9: [[PTR_ADD9:%[0-9]+]]:_(p5) = G_PTR_ADD [[PTR_ADD7]], [[C1]](s32) ; GFX9: [[LOAD10:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD9]](p5) :: (load 1 + 10, addrspace 56) - ; GFX9: [[TRUNC10:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD10]](s32) ; GFX9: [[PTR_ADD10:%[0-9]+]]:_(p5) = G_PTR_ADD [[PTR_ADD7]], [[C2]](s32) ; GFX9: [[LOAD11:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD10]](p5) :: (load 1 + 11, addrspace 56) - ; GFX9: [[TRUNC11:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD11]](s32) - ; GFX9: [[MV2:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC8]](s8), [[TRUNC9]](s8), [[TRUNC10]](s8), [[TRUNC11]](s8) ; GFX9: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 12 ; GFX9: [[PTR_ADD11:%[0-9]+]]:_(p5) = G_PTR_ADD [[COPY]], [[C5]](s32) ; GFX9: [[LOAD12:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD11]](p5) :: (load 1 + 12, addrspace 56) - ; GFX9: [[TRUNC12:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD12]](s32) ; GFX9: [[PTR_ADD12:%[0-9]+]]:_(p5) = G_PTR_ADD [[PTR_ADD11]], [[C]](s32) ; GFX9: [[LOAD13:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD12]](p5) :: (load 1 + 13, addrspace 56) - ; GFX9: [[TRUNC13:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD13]](s32) ; GFX9: [[PTR_ADD13:%[0-9]+]]:_(p5) = G_PTR_ADD [[PTR_ADD11]], [[C1]](s32) ; GFX9: [[LOAD14:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD13]](p5) :: (load 1 + 14, addrspace 56) - ; GFX9: [[TRUNC14:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD14]](s32) ; GFX9: [[PTR_ADD14:%[0-9]+]]:_(p5) = G_PTR_ADD [[PTR_ADD11]], [[C2]](s32) ; GFX9: [[LOAD15:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD14]](p5) :: (load 1 + 15, addrspace 56) - ; GFX9: [[TRUNC15:%[0-9]+]]:_(s8) = G_TRUNC [[LOAD15]](s32) - ; GFX9: [[MV3:%[0-9]+]]:_(s32) = G_MERGE_VALUES [[TRUNC12]](s8), [[TRUNC13]](s8), [[TRUNC14]](s8), [[TRUNC15]](s8) - ; GFX9: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s32>) = G_BUILD_VECTOR [[MV]](s32), [[MV1]](s32), [[MV2]](s32), [[MV3]](s32) - ; GFX9: [[BITCAST:%[0-9]+]]:_(<16 x s8>) = G_BITCAST [[BUILD_VECTOR]](<4 x s32>) - ; GFX9: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BITCAST]](<16 x s8>) + ; GFX9: [[COPY1:%[0-9]+]]:_(s32) = COPY [[LOAD]](s32) + ; GFX9: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32) + ; GFX9: [[BUILD_VECTOR_TRUNC:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY1]](s32), [[COPY2]](s32) + ; GFX9: [[COPY3:%[0-9]+]]:_(s32) = COPY [[LOAD2]](s32) + ; GFX9: [[COPY4:%[0-9]+]]:_(s32) = COPY [[LOAD3]](s32) + ; GFX9: [[BUILD_VECTOR_TRUNC1:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY3]](s32), [[COPY4]](s32) + ; GFX9: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC]](<2 x s16>), [[BUILD_VECTOR_TRUNC1]](<2 x s16>) + ; GFX9: [[TRUNC:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS]](<4 x s16>) + ; GFX9: [[COPY5:%[0-9]+]]:_(s32) = COPY [[LOAD4]](s32) + ; GFX9: [[COPY6:%[0-9]+]]:_(s32) = COPY [[LOAD5]](s32) + ; GFX9: [[BUILD_VECTOR_TRUNC2:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY5]](s32), [[COPY6]](s32) + ; GFX9: [[COPY7:%[0-9]+]]:_(s32) = COPY [[LOAD6]](s32) + ; GFX9: [[COPY8:%[0-9]+]]:_(s32) = COPY [[LOAD7]](s32) + ; GFX9: [[BUILD_VECTOR_TRUNC3:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY7]](s32), [[COPY8]](s32) + ; GFX9: [[CONCAT_VECTORS1:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC2]](<2 x s16>), [[BUILD_VECTOR_TRUNC3]](<2 x s16>) + ; GFX9: [[TRUNC1:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS1]](<4 x s16>) + ; GFX9: [[COPY9:%[0-9]+]]:_(s32) = COPY [[LOAD8]](s32) + ; GFX9: [[COPY10:%[0-9]+]]:_(s32) = COPY [[LOAD9]](s32) + ; GFX9: [[BUILD_VECTOR_TRUNC4:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY9]](s32), [[COPY10]](s32) + ; GFX9: [[COPY11:%[0-9]+]]:_(s32) = COPY [[LOAD10]](s32) + ; GFX9: [[COPY12:%[0-9]+]]:_(s32) = COPY [[LOAD11]](s32) + ; GFX9: [[BUILD_VECTOR_TRUNC5:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY11]](s32), [[COPY12]](s32) + ; GFX9: [[CONCAT_VECTORS2:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC4]](<2 x s16>), [[BUILD_VECTOR_TRUNC5]](<2 x s16>) + ; GFX9: [[TRUNC2:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS2]](<4 x s16>) + ; GFX9: [[COPY13:%[0-9]+]]:_(s32) = COPY [[LOAD12]](s32) + ; GFX9: [[COPY14:%[0-9]+]]:_(s32) = COPY [[LOAD13]](s32) + ; GFX9: [[BUILD_VECTOR_TRUNC6:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY13]](s32), [[COPY14]](s32) + ; GFX9: [[COPY15:%[0-9]+]]:_(s32) = COPY [[LOAD14]](s32) + ; GFX9: [[COPY16:%[0-9]+]]:_(s32) = COPY [[LOAD15]](s32) + ; GFX9: [[BUILD_VECTOR_TRUNC7:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[COPY15]](s32), [[COPY16]](s32) + ; GFX9: [[CONCAT_VECTORS3:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC6]](<2 x s16>), [[BUILD_VECTOR_TRUNC7]](<2 x s16>) + ; GFX9: [[TRUNC3:%[0-9]+]]:_(<4 x s8>) = G_TRUNC [[CONCAT_VECTORS3]](<4 x s16>) + ; GFX9: [[CONCAT_VECTORS4:%[0-9]+]]:_(<16 x s8>) = G_CONCAT_VECTORS [[TRUNC]](<4 x s8>), [[TRUNC1]](<4 x s8>), [[TRUNC2]](<4 x s8>), [[TRUNC3]](<4 x s8>) + ; GFX9: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[CONCAT_VECTORS4]](<16 x s8>) %0:_(p5) = COPY $vgpr0 %1:_(<16 x s8>) = G_LOAD %0 :: (load 16, align 1, addrspace 56) $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %1