diff options
Diffstat (limited to 'lib/Target/R600/R600RegisterInfo.td')
-rw-r--r-- | lib/Target/R600/R600RegisterInfo.td | 48 |
1 files changed, 45 insertions, 3 deletions
diff --git a/lib/Target/R600/R600RegisterInfo.td b/lib/Target/R600/R600RegisterInfo.td index 68bcd20..cc667d9 100644 --- a/lib/Target/R600/R600RegisterInfo.td +++ b/lib/Target/R600/R600RegisterInfo.td @@ -18,18 +18,28 @@ class R600RegWithChan <string name, bits<9> sel, string chan> : class R600Reg_128<string n, list<Register> subregs, bits<16> encoding> : RegisterWithSubRegs<n, subregs> { + field bits<2> chan_encoding = 0; let Namespace = "AMDGPU"; let SubRegIndices = [sub0, sub1, sub2, sub3]; - let HWEncoding = encoding; + let HWEncoding{8-0} = encoding{8-0}; + let HWEncoding{10-9} = chan_encoding; } class R600Reg_64<string n, list<Register> subregs, bits<16> encoding> : RegisterWithSubRegs<n, subregs> { + field bits<2> chan_encoding = 0; let Namespace = "AMDGPU"; let SubRegIndices = [sub0, sub1]; let HWEncoding = encoding; + let HWEncoding{8-0} = encoding{8-0}; + let HWEncoding{10-9} = chan_encoding; } +class R600Reg_64Vertical<int lo, int hi, string chan> : R600Reg_64 < + "V"#lo#hi#"_"#chan, + [!cast<Register>("T"#lo#"_"#chan), !cast<Register>("T"#hi#"_"#chan)], + lo +>; foreach Index = 0-127 in { foreach Chan = [ "X", "Y", "Z", "W" ] in { @@ -54,6 +64,24 @@ foreach Index = 0-127 in { Index>; } +foreach Chan = [ "X", "Y", "Z", "W"] in { + + let chan_encoding = !if(!eq(Chan, "X"), 0, + !if(!eq(Chan, "Y"), 1, + !if(!eq(Chan, "Z"), 2, + !if(!eq(Chan, "W"), 3, 0)))) in { + def V0123_#Chan : R600Reg_128 <"V0123_"#Chan, + [!cast<Register>("T0_"#Chan), + !cast<Register>("T1_"#Chan), + !cast<Register>("T2_"#Chan), + !cast<Register>("T3_"#Chan)], + 0>; + def V01_#Chan : R600Reg_64Vertical<0, 1, Chan>; + def V23_#Chan : R600Reg_64Vertical<2, 3, Chan>; + } +} + + // KCACHE_BANK0 foreach Index = 159-128 in { foreach Chan = [ "X", "Y", "Z", "W" ] in { @@ -130,8 +158,14 @@ def ALU_PARAM : R600Reg<"Param", 0>; let isAllocatable = 0 in { -// XXX: Only use the X channel, until we support wider stack widths -def R600_Addr : RegisterClass <"AMDGPU", [i32], 127, (add (sequence "Addr%u_X", 0, 127))>; +def R600_Addr : RegisterClass <"AMDGPU", [i32], 32, (add (sequence "Addr%u_X", 0, 127))>; + +// We only use Addr_[YZW] for vertical vectors. +// FIXME if we add more vertical vector registers we will need to ad more +// registers to these classes. +def R600_Addr_Y : RegisterClass <"AMDGPU", [i32], 32, (add Addr0_Y)>; +def R600_Addr_Z : RegisterClass <"AMDGPU", [i32], 32, (add Addr0_Z)>; +def R600_Addr_W : RegisterClass <"AMDGPU", [i32], 32, (add Addr0_W)>; def R600_LDS_SRC_REG : RegisterClass<"AMDGPU", [i32], 32, (add OQA, OQB, OQAP, OQBP, LDS_DIRECT_A, LDS_DIRECT_B)>; @@ -206,5 +240,13 @@ def R600_Reg128 : RegisterClass<"AMDGPU", [v4f32, v4i32], 128, let CopyCost = -1; } +def R600_Reg128Vertical : RegisterClass<"AMDGPU", [v4f32, v4i32], 128, + (add V0123_W, V0123_Z, V0123_Y, V0123_X) +>; + def R600_Reg64 : RegisterClass<"AMDGPU", [v2f32, v2i32], 64, (add (sequence "T%u_XY", 0, 63))>; + +def R600_Reg64Vertical : RegisterClass<"AMDGPU", [v2f32, v2i32], 64, + (add V01_X, V01_Y, V01_Z, V01_W, + V23_X, V23_Y, V23_Z, V23_W)>; |