Skip to content

Commit c2c4db8

Browse files
michaelmaitlandmichaelmaitland
authored
[RISCV][VLOPT] Add support for 11.11 div instructions (#112201)
This adds support for these instructions and also tests getOperandInfo for these instructions as well.
1 parent 82e89c0 commit c2c4db8

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
@@ -554,7 +554,14 @@ static bool isSupportedInstr(const MachineInstr &MI) {
554554
case RISCV::VMULHSU_VV:
555555
case RISCV::VMULHSU_VX:
556556
// 11.11. Vector Integer Divide Instructions
557-
// FIXME: Add support for 11.11 instructions
557+
case RISCV::VDIVU_VV:
558+
case RISCV::VDIVU_VX:
559+
case RISCV::VDIV_VV:
560+
case RISCV::VDIV_VX:
561+
case RISCV::VREMU_VV:
562+
case RISCV::VREMU_VX:
563+
case RISCV::VREM_VV:
564+
case RISCV::VREM_VX:
558565
// 11.12. Vector Widening Integer Multiply Instructions
559566
// FIXME: Add support for 11.12 instructions
560567
// 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
@@ -962,6 +962,166 @@ define <vscale x 4 x i32> @vmulhsu_vx(<vscale x 4 x i32> %a, i32 %b, iXLen %vl)
962962
ret <vscale x 4 x i32> %2
963963
}
964964

965+
define <vscale x 4 x i32> @vdivu_vv(<vscale x 4 x i32> %a, <vscale x 4 x i32> %b, iXLen %vl) {
966+
; NOVLOPT-LABEL: vdivu_vv:
967+
; NOVLOPT: # %bb.0:
968+
; NOVLOPT-NEXT: vsetvli a1, zero, e32, m2, ta, ma
969+
; NOVLOPT-NEXT: vdivu.vv v8, v8, v10
970+
; NOVLOPT-NEXT: vsetvli zero, a0, e32, m2, ta, ma
971+
; NOVLOPT-NEXT: vmul.vv v8, v8, v10
972+
; NOVLOPT-NEXT: ret
973+
;
974+
; VLOPT-LABEL: vdivu_vv:
975+
; VLOPT: # %bb.0:
976+
; VLOPT-NEXT: vsetvli zero, a0, e32, m2, ta, ma
977+
; VLOPT-NEXT: vdivu.vv v8, v8, v10
978+
; VLOPT-NEXT: vmul.vv v8, v8, v10
979+
; VLOPT-NEXT: ret
980+
%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)
981+
%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)
982+
ret <vscale x 4 x i32> %2
983+
}
984+
985+
define <vscale x 4 x i32> @vdivu_vx(<vscale x 4 x i32> %a, i32 %b, iXLen %vl) {
986+
; NOVLOPT-LABEL: vdivu_vx:
987+
; NOVLOPT: # %bb.0:
988+
; NOVLOPT-NEXT: vsetvli a2, zero, e32, m2, ta, ma
989+
; NOVLOPT-NEXT: vdivu.vx v10, v8, a0
990+
; NOVLOPT-NEXT: vsetvli zero, a1, e32, m2, ta, ma
991+
; NOVLOPT-NEXT: vmul.vv v8, v10, v8
992+
; NOVLOPT-NEXT: ret
993+
;
994+
; VLOPT-LABEL: vdivu_vx:
995+
; VLOPT: # %bb.0:
996+
; VLOPT-NEXT: vsetvli zero, a1, e32, m2, ta, ma
997+
; VLOPT-NEXT: vdivu.vx v10, v8, a0
998+
; VLOPT-NEXT: vmul.vv v8, v10, v8
999+
; VLOPT-NEXT: ret
1000+
%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)
1001+
%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)
1002+
ret <vscale x 4 x i32> %2
1003+
}
1004+
1005+
define <vscale x 4 x i32> @vdiv_vv(<vscale x 4 x i32> %a, <vscale x 4 x i32> %b, iXLen %vl) {
1006+
; NOVLOPT-LABEL: vdiv_vv:
1007+
; NOVLOPT: # %bb.0:
1008+
; NOVLOPT-NEXT: vsetvli a1, zero, e32, m2, ta, ma
1009+
; NOVLOPT-NEXT: vdiv.vv v8, v8, v10
1010+
; NOVLOPT-NEXT: vsetvli zero, a0, e32, m2, ta, ma
1011+
; NOVLOPT-NEXT: vmul.vv v8, v8, v10
1012+
; NOVLOPT-NEXT: ret
1013+
;
1014+
; VLOPT-LABEL: vdiv_vv:
1015+
; VLOPT: # %bb.0:
1016+
; VLOPT-NEXT: vsetvli zero, a0, e32, m2, ta, ma
1017+
; VLOPT-NEXT: vdiv.vv v8, v8, v10
1018+
; VLOPT-NEXT: vmul.vv v8, v8, v10
1019+
; VLOPT-NEXT: ret
1020+
%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)
1021+
%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)
1022+
ret <vscale x 4 x i32> %2
1023+
}
1024+
1025+
define <vscale x 4 x i32> @vdiv_vx(<vscale x 4 x i32> %a, i32 %b, iXLen %vl) {
1026+
; NOVLOPT-LABEL: vdiv_vx:
1027+
; NOVLOPT: # %bb.0:
1028+
; NOVLOPT-NEXT: vsetvli a2, zero, e32, m2, ta, ma
1029+
; NOVLOPT-NEXT: vdiv.vx v10, v8, a0
1030+
; NOVLOPT-NEXT: vsetvli zero, a1, e32, m2, ta, ma
1031+
; NOVLOPT-NEXT: vmul.vv v8, v10, v8
1032+
; NOVLOPT-NEXT: ret
1033+
;
1034+
; VLOPT-LABEL: vdiv_vx:
1035+
; VLOPT: # %bb.0:
1036+
; VLOPT-NEXT: vsetvli zero, a1, e32, m2, ta, ma
1037+
; VLOPT-NEXT: vdiv.vx v10, v8, a0
1038+
; VLOPT-NEXT: vmul.vv v8, v10, v8
1039+
; VLOPT-NEXT: ret
1040+
%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)
1041+
%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)
1042+
ret <vscale x 4 x i32> %2
1043+
}
1044+
1045+
define <vscale x 4 x i32> @vremu_vv(<vscale x 4 x i32> %a, <vscale x 4 x i32> %b, iXLen %vl) {
1046+
; NOVLOPT-LABEL: vremu_vv:
1047+
; NOVLOPT: # %bb.0:
1048+
; NOVLOPT-NEXT: vsetvli a1, zero, e32, m2, ta, ma
1049+
; NOVLOPT-NEXT: vremu.vv v8, v8, v10
1050+
; NOVLOPT-NEXT: vsetvli zero, a0, e32, m2, ta, ma
1051+
; NOVLOPT-NEXT: vmul.vv v8, v8, v10
1052+
; NOVLOPT-NEXT: ret
1053+
;
1054+
; VLOPT-LABEL: vremu_vv:
1055+
; VLOPT: # %bb.0:
1056+
; VLOPT-NEXT: vsetvli zero, a0, e32, m2, ta, ma
1057+
; VLOPT-NEXT: vremu.vv v8, v8, v10
1058+
; VLOPT-NEXT: vmul.vv v8, v8, v10
1059+
; VLOPT-NEXT: ret
1060+
%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)
1061+
%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)
1062+
ret <vscale x 4 x i32> %2
1063+
}
1064+
1065+
define <vscale x 4 x i32> @vremu_vx(<vscale x 4 x i32> %a, i32 %b, iXLen %vl) {
1066+
; NOVLOPT-LABEL: vremu_vx:
1067+
; NOVLOPT: # %bb.0:
1068+
; NOVLOPT-NEXT: vsetvli a2, zero, e32, m2, ta, ma
1069+
; NOVLOPT-NEXT: vremu.vx v10, v8, a0
1070+
; NOVLOPT-NEXT: vsetvli zero, a1, e32, m2, ta, ma
1071+
; NOVLOPT-NEXT: vmul.vv v8, v10, v8
1072+
; NOVLOPT-NEXT: ret
1073+
;
1074+
; VLOPT-LABEL: vremu_vx:
1075+
; VLOPT: # %bb.0:
1076+
; VLOPT-NEXT: vsetvli zero, a1, e32, m2, ta, ma
1077+
; VLOPT-NEXT: vremu.vx v10, v8, a0
1078+
; VLOPT-NEXT: vmul.vv v8, v10, v8
1079+
; VLOPT-NEXT: ret
1080+
%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)
1081+
%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)
1082+
ret <vscale x 4 x i32> %2
1083+
}
1084+
1085+
define <vscale x 4 x i32> @vrem_vv(<vscale x 4 x i32> %a, <vscale x 4 x i32> %b, iXLen %vl) {
1086+
; NOVLOPT-LABEL: vrem_vv:
1087+
; NOVLOPT: # %bb.0:
1088+
; NOVLOPT-NEXT: vsetvli a1, zero, e32, m2, ta, ma
1089+
; NOVLOPT-NEXT: vrem.vv v8, v8, v10
1090+
; NOVLOPT-NEXT: vsetvli zero, a0, e32, m2, ta, ma
1091+
; NOVLOPT-NEXT: vmul.vv v8, v8, v10
1092+
; NOVLOPT-NEXT: ret
1093+
;
1094+
; VLOPT-LABEL: vrem_vv:
1095+
; VLOPT: # %bb.0:
1096+
; VLOPT-NEXT: vsetvli zero, a0, e32, m2, ta, ma
1097+
; VLOPT-NEXT: vrem.vv v8, v8, v10
1098+
; VLOPT-NEXT: vmul.vv v8, v8, v10
1099+
; VLOPT-NEXT: ret
1100+
%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)
1101+
%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)
1102+
ret <vscale x 4 x i32> %2
1103+
}
1104+
1105+
define <vscale x 4 x i32> @vrem_vx(<vscale x 4 x i32> %a, i32 %b, iXLen %vl) {
1106+
; NOVLOPT-LABEL: vrem_vx:
1107+
; NOVLOPT: # %bb.0:
1108+
; NOVLOPT-NEXT: vsetvli a2, zero, e32, m2, ta, ma
1109+
; NOVLOPT-NEXT: vrem.vx v10, v8, a0
1110+
; NOVLOPT-NEXT: vsetvli zero, a1, e32, m2, ta, ma
1111+
; NOVLOPT-NEXT: vmul.vv v8, v10, v8
1112+
; NOVLOPT-NEXT: ret
1113+
;
1114+
; VLOPT-LABEL: vrem_vx:
1115+
; VLOPT: # %bb.0:
1116+
; VLOPT-NEXT: vsetvli zero, a1, e32, m2, ta, ma
1117+
; VLOPT-NEXT: vrem.vx v10, v8, a0
1118+
; VLOPT-NEXT: vmul.vv v8, v10, v8
1119+
; VLOPT-NEXT: ret
1120+
%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)
1121+
%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)
1122+
ret <vscale x 4 x i32> %2
1123+
}
1124+
9651125
define <vscale x 4 x i32> @vwmacc_vx(<vscale x 4 x i16> %a, i16 %b, iXLen %vl) {
9661126
; NOVLOPT-LABEL: vwmacc_vx:
9671127
; NOVLOPT: # %bb.0:

0 commit comments

Comments
 (0)