Skip to content

Commit 2456349

Browse files
michaelmaitlandmichaelmaitland
committed
[RISCV][VLOPT] Add support for 11.10 div instructions
1 parent 1c94388 commit 2456349

File tree

2 files changed

+168
-1
lines changed

2 files changed

+168
-1
lines changed

‎llvm/lib/Target/RISCV/RISCVVLOptimizer.cpp‎

Lines changed: 8 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -546,7 +546,14 @@ static bool isSupportedInstr(const MachineInstr &MI) {
546546
case RISCV::VMULHSU_VV:
547547
case RISCV::VMULHSU_VX:
548548
// 11.11. Vector Integer Divide Instructions
549-
// FIXME: Add support for 11.11 instructions
549+
case RISCV::VDIVU_VV:
550+
case RISCV::VDIVU_VX:
551+
case RISCV::VDIV_VV:
552+
case RISCV::VDIV_VX:
553+
case RISCV::VREMU_VV:
554+
case RISCV::VREMU_VX:
555+
case RISCV::VREM_VV:
556+
case RISCV::VREM_VX:
550557
// 11.12. Vector Widening Integer Multiply Instructions
551558
// FIXME: Add support for 11.12 instructions
552559
// 11.13. Vector Single-Width Integer Multiply-Add Instructions

‎llvm/test/CodeGen/RISCV/rvv/vl-opt-instrs.ll‎

Lines changed: 160 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -804,6 +804,166 @@ define <vscale x 4 x i32> @vmulhsu_vx(<vscale x 4 x i32> %a, i32 %b, iXLen %vl)
804804
ret <vscale x 4 x i32> %2
805805
}
806806

807+
define <vscale x 4 x i32> @vdivu_vv(<vscale x 4 x i32> %a, <vscale x 4 x i32> %b, iXLen %vl) {
808+
; NOVLOPT-LABEL: vdivu_vv:
809+
; NOVLOPT: # %bb.0:
810+
; NOVLOPT-NEXT: vsetvli a1, zero, e32, m2, ta, ma
811+
; NOVLOPT-NEXT: vdivu.vv v8, v8, v10
812+
; NOVLOPT-NEXT: vsetvli zero, a0, e32, m2, ta, ma
813+
; NOVLOPT-NEXT: vmul.vv v8, v8, v10
814+
; NOVLOPT-NEXT: ret
815+
;
816+
; VLOPT-LABEL: vdivu_vv:
817+
; VLOPT: # %bb.0:
818+
; VLOPT-NEXT: vsetvli zero, a0, e32, m2, ta, ma
819+
; VLOPT-NEXT: vdivu.vv v8, v8, v10
820+
; VLOPT-NEXT: vmul.vv v8, v8, v10
821+
; VLOPT-NEXT: ret
822+
%1 = call <vscale x 4 x i32> @llvm.riscv.vdivu.nxv4i32.nxv4i32(<vscale x 4 x i32> poison, <vscale x 4 x i32> %a, <vscale x 4 x i32> %b, iXLen -1)
823+
%2 = call <vscale x 4 x i32> @llvm.riscv.vmul.nxv4i32.nxv4i32(<vscale x 4 x i32> poison, <vscale x 4 x i32> %1, <vscale x 4 x i32> %b, iXLen %vl)
824+
ret <vscale x 4 x i32> %2
825+
}
826+
827+
define <vscale x 4 x i32> @vdivu_vx(<vscale x 4 x i32> %a, i32 %b, iXLen %vl) {
828+
; NOVLOPT-LABEL: vdivu_vx:
829+
; NOVLOPT: # %bb.0:
830+
; NOVLOPT-NEXT: vsetvli a2, zero, e32, m2, ta, ma
831+
; NOVLOPT-NEXT: vdivu.vx v10, v8, a0
832+
; NOVLOPT-NEXT: vsetvli zero, a1, e32, m2, ta, ma
833+
; NOVLOPT-NEXT: vmul.vv v8, v10, v8
834+
; NOVLOPT-NEXT: ret
835+
;
836+
; VLOPT-LABEL: vdivu_vx:
837+
; VLOPT: # %bb.0:
838+
; VLOPT-NEXT: vsetvli zero, a1, e32, m2, ta, ma
839+
; VLOPT-NEXT: vdivu.vx v10, v8, a0
840+
; VLOPT-NEXT: vmul.vv v8, v10, v8
841+
; VLOPT-NEXT: ret
842+
%1 = call <vscale x 4 x i32> @llvm.riscv.vdivu.nxv4i32.nxv4i32(<vscale x 4 x i32> poison, <vscale x 4 x i32> %a, i32 %b, iXLen -1)
843+
%2 = call <vscale x 4 x i32> @llvm.riscv.vmul.nxv4i32.nxv4i32(<vscale x 4 x i32> poison, <vscale x 4 x i32> %1, <vscale x 4 x i32> %a, iXLen %vl)
844+
ret <vscale x 4 x i32> %2
845+
}
846+
847+
define <vscale x 4 x i32> @vdiv_vv(<vscale x 4 x i32> %a, <vscale x 4 x i32> %b, iXLen %vl) {
848+
; NOVLOPT-LABEL: vdiv_vv:
849+
; NOVLOPT: # %bb.0:
850+
; NOVLOPT-NEXT: vsetvli a1, zero, e32, m2, ta, ma
851+
; NOVLOPT-NEXT: vdiv.vv v8, v8, v10
852+
; NOVLOPT-NEXT: vsetvli zero, a0, e32, m2, ta, ma
853+
; NOVLOPT-NEXT: vmul.vv v8, v8, v10
854+
; NOVLOPT-NEXT: ret
855+
;
856+
; VLOPT-LABEL: vdiv_vv:
857+
; VLOPT: # %bb.0:
858+
; VLOPT-NEXT: vsetvli zero, a0, e32, m2, ta, ma
859+
; VLOPT-NEXT: vdiv.vv v8, v8, v10
860+
; VLOPT-NEXT: vmul.vv v8, v8, v10
861+
; VLOPT-NEXT: ret
862+
%1 = call <vscale x 4 x i32> @llvm.riscv.vdiv.nxv4i32.nxv4i32(<vscale x 4 x i32> poison, <vscale x 4 x i32> %a, <vscale x 4 x i32> %b, iXLen -1)
863+
%2 = call <vscale x 4 x i32> @llvm.riscv.vmul.nxv4i32.nxv4i32(<vscale x 4 x i32> poison, <vscale x 4 x i32> %1, <vscale x 4 x i32> %b, iXLen %vl)
864+
ret <vscale x 4 x i32> %2
865+
}
866+
867+
define <vscale x 4 x i32> @vdiv_vx(<vscale x 4 x i32> %a, i32 %b, iXLen %vl) {
868+
; NOVLOPT-LABEL: vdiv_vx:
869+
; NOVLOPT: # %bb.0:
870+
; NOVLOPT-NEXT: vsetvli a2, zero, e32, m2, ta, ma
871+
; NOVLOPT-NEXT: vdiv.vx v10, v8, a0
872+
; NOVLOPT-NEXT: vsetvli zero, a1, e32, m2, ta, ma
873+
; NOVLOPT-NEXT: vmul.vv v8, v10, v8
874+
; NOVLOPT-NEXT: ret
875+
;
876+
; VLOPT-LABEL: vdiv_vx:
877+
; VLOPT: # %bb.0:
878+
; VLOPT-NEXT: vsetvli zero, a1, e32, m2, ta, ma
879+
; VLOPT-NEXT: vdiv.vx v10, v8, a0
880+
; VLOPT-NEXT: vmul.vv v8, v10, v8
881+
; VLOPT-NEXT: ret
882+
%1 = call <vscale x 4 x i32> @llvm.riscv.vdiv.nxv4i32.nxv4i32(<vscale x 4 x i32> poison, <vscale x 4 x i32> %a, i32 %b, iXLen -1)
883+
%2 = call <vscale x 4 x i32> @llvm.riscv.vmul.nxv4i32.nxv4i32(<vscale x 4 x i32> poison, <vscale x 4 x i32> %1, <vscale x 4 x i32> %a, iXLen %vl)
884+
ret <vscale x 4 x i32> %2
885+
}
886+
887+
define <vscale x 4 x i32> @vremu_vv(<vscale x 4 x i32> %a, <vscale x 4 x i32> %b, iXLen %vl) {
888+
; NOVLOPT-LABEL: vremu_vv:
889+
; NOVLOPT: # %bb.0:
890+
; NOVLOPT-NEXT: vsetvli a1, zero, e32, m2, ta, ma
891+
; NOVLOPT-NEXT: vremu.vv v8, v8, v10
892+
; NOVLOPT-NEXT: vsetvli zero, a0, e32, m2, ta, ma
893+
; NOVLOPT-NEXT: vmul.vv v8, v8, v10
894+
; NOVLOPT-NEXT: ret
895+
;
896+
; VLOPT-LABEL: vremu_vv:
897+
; VLOPT: # %bb.0:
898+
; VLOPT-NEXT: vsetvli zero, a0, e32, m2, ta, ma
899+
; VLOPT-NEXT: vremu.vv v8, v8, v10
900+
; VLOPT-NEXT: vmul.vv v8, v8, v10
901+
; VLOPT-NEXT: ret
902+
%1 = call <vscale x 4 x i32> @llvm.riscv.vremu.nxv4i32.nxv4i32(<vscale x 4 x i32> poison, <vscale x 4 x i32> %a, <vscale x 4 x i32> %b, iXLen -1)
903+
%2 = call <vscale x 4 x i32> @llvm.riscv.vmul.nxv4i32.nxv4i32(<vscale x 4 x i32> poison, <vscale x 4 x i32> %1, <vscale x 4 x i32> %b, iXLen %vl)
904+
ret <vscale x 4 x i32> %2
905+
}
906+
907+
define <vscale x 4 x i32> @vremu_vx(<vscale x 4 x i32> %a, i32 %b, iXLen %vl) {
908+
; NOVLOPT-LABEL: vremu_vx:
909+
; NOVLOPT: # %bb.0:
910+
; NOVLOPT-NEXT: vsetvli a2, zero, e32, m2, ta, ma
911+
; NOVLOPT-NEXT: vremu.vx v10, v8, a0
912+
; NOVLOPT-NEXT: vsetvli zero, a1, e32, m2, ta, ma
913+
; NOVLOPT-NEXT: vmul.vv v8, v10, v8
914+
; NOVLOPT-NEXT: ret
915+
;
916+
; VLOPT-LABEL: vremu_vx:
917+
; VLOPT: # %bb.0:
918+
; VLOPT-NEXT: vsetvli zero, a1, e32, m2, ta, ma
919+
; VLOPT-NEXT: vremu.vx v10, v8, a0
920+
; VLOPT-NEXT: vmul.vv v8, v10, v8
921+
; VLOPT-NEXT: ret
922+
%1 = call <vscale x 4 x i32> @llvm.riscv.vremu.nxv4i32.nxv4i32(<vscale x 4 x i32> poison, <vscale x 4 x i32> %a, i32 %b, iXLen -1)
923+
%2 = call <vscale x 4 x i32> @llvm.riscv.vmul.nxv4i32.nxv4i32(<vscale x 4 x i32> poison, <vscale x 4 x i32> %1, <vscale x 4 x i32> %a, iXLen %vl)
924+
ret <vscale x 4 x i32> %2
925+
}
926+
927+
define <vscale x 4 x i32> @vrem_vv(<vscale x 4 x i32> %a, <vscale x 4 x i32> %b, iXLen %vl) {
928+
; NOVLOPT-LABEL: vrem_vv:
929+
; NOVLOPT: # %bb.0:
930+
; NOVLOPT-NEXT: vsetvli a1, zero, e32, m2, ta, ma
931+
; NOVLOPT-NEXT: vrem.vv v8, v8, v10
932+
; NOVLOPT-NEXT: vsetvli zero, a0, e32, m2, ta, ma
933+
; NOVLOPT-NEXT: vmul.vv v8, v8, v10
934+
; NOVLOPT-NEXT: ret
935+
;
936+
; VLOPT-LABEL: vrem_vv:
937+
; VLOPT: # %bb.0:
938+
; VLOPT-NEXT: vsetvli zero, a0, e32, m2, ta, ma
939+
; VLOPT-NEXT: vrem.vv v8, v8, v10
940+
; VLOPT-NEXT: vmul.vv v8, v8, v10
941+
; VLOPT-NEXT: ret
942+
%1 = call <vscale x 4 x i32> @llvm.riscv.vrem.nxv4i32.nxv4i32(<vscale x 4 x i32> poison, <vscale x 4 x i32> %a, <vscale x 4 x i32> %b, iXLen -1)
943+
%2 = call <vscale x 4 x i32> @llvm.riscv.vmul.nxv4i32.nxv4i32(<vscale x 4 x i32> poison, <vscale x 4 x i32> %1, <vscale x 4 x i32> %b, iXLen %vl)
944+
ret <vscale x 4 x i32> %2
945+
}
946+
947+
define <vscale x 4 x i32> @vrem_vx(<vscale x 4 x i32> %a, i32 %b, iXLen %vl) {
948+
; NOVLOPT-LABEL: vrem_vx:
949+
; NOVLOPT: # %bb.0:
950+
; NOVLOPT-NEXT: vsetvli a2, zero, e32, m2, ta, ma
951+
; NOVLOPT-NEXT: vrem.vx v10, v8, a0
952+
; NOVLOPT-NEXT: vsetvli zero, a1, e32, m2, ta, ma
953+
; NOVLOPT-NEXT: vmul.vv v8, v10, v8
954+
; NOVLOPT-NEXT: ret
955+
;
956+
; VLOPT-LABEL: vrem_vx:
957+
; VLOPT: # %bb.0:
958+
; VLOPT-NEXT: vsetvli zero, a1, e32, m2, ta, ma
959+
; VLOPT-NEXT: vrem.vx v10, v8, a0
960+
; VLOPT-NEXT: vmul.vv v8, v10, v8
961+
; VLOPT-NEXT: ret
962+
%1 = call <vscale x 4 x i32> @llvm.riscv.vrem.nxv4i32.nxv4i32(<vscale x 4 x i32> poison, <vscale x 4 x i32> %a, i32 %b, iXLen -1)
963+
%2 = call <vscale x 4 x i32> @llvm.riscv.vmul.nxv4i32.nxv4i32(<vscale x 4 x i32> poison, <vscale x 4 x i32> %1, <vscale x 4 x i32> %a, iXLen %vl)
964+
ret <vscale x 4 x i32> %2
965+
}
966+
807967
define <vscale x 4 x i32> @vwmacc_vx(<vscale x 4 x i16> %a, i16 %b, iXLen %vl) {
808968
; NOVLOPT-LABEL: vwmacc_vx:
809969
; NOVLOPT: # %bb.0:

0 commit comments

Comments
 (0)