diff options
Diffstat (limited to 'lib/Target/NVPTX/NVPTXInstrInfo.td')
| -rw-r--r-- | lib/Target/NVPTX/NVPTXInstrInfo.td | 50 |
1 files changed, 46 insertions, 4 deletions
diff --git a/lib/Target/NVPTX/NVPTXInstrInfo.td b/lib/Target/NVPTX/NVPTXInstrInfo.td index e94250b38d..725d6fc91c 100644 --- a/lib/Target/NVPTX/NVPTXInstrInfo.td +++ b/lib/Target/NVPTX/NVPTXInstrInfo.td @@ -158,6 +158,7 @@ def do_SQRTF32_APPROX : Predicate<"!usePrecSqrtF32()">; def do_SQRTF32_RN : Predicate<"usePrecSqrtF32()">; def hasHWROT32 : Predicate<"Subtarget.hasHWROT32()">; +def noHWROT32 : Predicate<"!Subtarget.hasHWROT32()">; def true : Predicate<"1">; @@ -1085,6 +1086,43 @@ multiclass RSHIFT_FORMAT<string OpcStr, SDNode OpNode> { defm SRA : RSHIFT_FORMAT<"shr.s", sra>; defm SRL : RSHIFT_FORMAT<"shr.u", srl>; +// +// Rotate: use ptx shf instruction if available. +// + +// 32 bit r2 = rotl r1, n +// => +// r2 = shf.l r1, r1, n +def ROTL32imm_hw : NVPTXInst<(outs Int32Regs:$dst), + (ins Int32Regs:$src, i32imm:$amt), + "shf.l.wrap.b32 \t$dst, $src, $src, $amt;", + [(set Int32Regs:$dst, (rotl Int32Regs:$src, (i32 imm:$amt)))]>, + Requires<[hasHWROT32]> ; + +def ROTL32reg_hw : NVPTXInst<(outs Int32Regs:$dst), + (ins Int32Regs:$src, Int32Regs:$amt), + "shf.l.wrap.b32 \t$dst, $src, $src, $amt;", + [(set Int32Regs:$dst, (rotl Int32Regs:$src, Int32Regs:$amt))]>, + Requires<[hasHWROT32]>; + +// 32 bit r2 = rotr r1, n +// => +// r2 = shf.r r1, r1, n +def ROTR32imm_hw : NVPTXInst<(outs Int32Regs:$dst), + (ins Int32Regs:$src, i32imm:$amt), + "shf.r.wrap.b32 \t$dst, $src, $src, $amt;", + [(set Int32Regs:$dst, (rotr Int32Regs:$src, (i32 imm:$amt)))]>, + Requires<[hasHWROT32]>; + +def ROTR32reg_hw : NVPTXInst<(outs Int32Regs:$dst), + (ins Int32Regs:$src, Int32Regs:$amt), + "shf.r.wrap.b32 \t$dst, $src, $src, $amt;", + [(set Int32Regs:$dst, (rotr Int32Regs:$src, Int32Regs:$amt))]>, + Requires<[hasHWROT32]>; + +// +// Rotate: if ptx shf instruction is not available, then use shift+add +// // 32bit def ROT32imm_sw : NVPTXInst<(outs Int32Regs:$dst), (ins Int32Regs:$src, i32imm:$amt1, i32imm:$amt2), @@ -1102,9 +1140,11 @@ def SUB_FRM_32 : SDNodeXForm<imm, [{ }]>; def : Pat<(rotl Int32Regs:$src, (i32 imm:$amt)), - (ROT32imm_sw Int32Regs:$src, imm:$amt, (SUB_FRM_32 node:$amt))>; + (ROT32imm_sw Int32Regs:$src, imm:$amt, (SUB_FRM_32 node:$amt))>, + Requires<[noHWROT32]>; def : Pat<(rotr Int32Regs:$src, (i32 imm:$amt)), - (ROT32imm_sw Int32Regs:$src, (SUB_FRM_32 node:$amt), imm:$amt)>; + (ROT32imm_sw Int32Regs:$src, (SUB_FRM_32 node:$amt), imm:$amt)>, + Requires<[noHWROT32]>; def ROTL32reg_sw : NVPTXInst<(outs Int32Regs:$dst), (ins Int32Regs:$src, Int32Regs:$amt), @@ -1117,7 +1157,8 @@ def ROTL32reg_sw : NVPTXInst<(outs Int32Regs:$dst), (ins Int32Regs:$src, !strconcat("shr.b32 \t%rhs, $src, %amt2;\n\t", !strconcat("add.u32 \t$dst, %lhs, %rhs;\n\t", !strconcat("}}", ""))))))))), - [(set Int32Regs:$dst, (rotl Int32Regs:$src, Int32Regs:$amt))]>; + [(set Int32Regs:$dst, (rotl Int32Regs:$src, Int32Regs:$amt))]>, + Requires<[noHWROT32]>; def ROTR32reg_sw : NVPTXInst<(outs Int32Regs:$dst), (ins Int32Regs:$src, Int32Regs:$amt), @@ -1130,7 +1171,8 @@ def ROTR32reg_sw : NVPTXInst<(outs Int32Regs:$dst), (ins Int32Regs:$src, !strconcat("shl.b32 \t%rhs, $src, %amt2;\n\t", !strconcat("add.u32 \t$dst, %lhs, %rhs;\n\t", !strconcat("}}", ""))))))))), - [(set Int32Regs:$dst, (rotr Int32Regs:$src, Int32Regs:$amt))]>; + [(set Int32Regs:$dst, (rotr Int32Regs:$src, Int32Regs:$amt))]>, + Requires<[noHWROT32]>; // 64bit def ROT64imm_sw : NVPTXInst<(outs Int64Regs:$dst), (ins Int64Regs:$src, |