From 9958475129a26624dbba18781bb9efa4cf5549a1 Mon Sep 17 00:00:00 2001 From: Tom Stellard Date: Mon, 24 Mar 2014 16:07:25 +0000 Subject: [PATCH] R600: Reorganize tablegen instruction definitions Each GPU family now has its own file. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@204615 91177308-0d34-0410-b5e6-96231b3b80d8 --- lib/Target/R600/AMDGPUInstructions.td | 3 + lib/Target/R600/CaymanInstructions.td | 221 +++++++ lib/Target/R600/EvergreenInstructions.td | 579 +++++++++++++++++ lib/Target/R600/R600Instructions.td | 783 +---------------------- lib/Target/R600/R700Instructions.td | 21 + 5 files changed, 826 insertions(+), 781 deletions(-) create mode 100644 lib/Target/R600/CaymanInstructions.td create mode 100644 lib/Target/R600/EvergreenInstructions.td create mode 100644 lib/Target/R600/R700Instructions.td diff --git a/lib/Target/R600/AMDGPUInstructions.td b/lib/Target/R600/AMDGPUInstructions.td index 4a9d92caf93..505fc81347b 100644 --- a/lib/Target/R600/AMDGPUInstructions.td +++ b/lib/Target/R600/AMDGPUInstructions.td @@ -415,6 +415,9 @@ class UMUL24Pattern : Pat < */ include "R600Instructions.td" +include "R700Instructions.td" +include "EvergreenInstructions.td" +include "CaymanInstructions.td" include "SIInstrInfo.td" diff --git a/lib/Target/R600/CaymanInstructions.td b/lib/Target/R600/CaymanInstructions.td new file mode 100644 index 00000000000..acd7bdecdcb --- /dev/null +++ b/lib/Target/R600/CaymanInstructions.td @@ -0,0 +1,221 @@ +//===-- CaymanInstructions.td - CM Instruction defs -------*- tablegen -*-===// +// +// The LLVM Compiler Infrastructure +// +// This file is distributed under the University of Illinois Open Source +// License. See LICENSE.TXT for details. +// +//===----------------------------------------------------------------------===// +// +// TableGen definitions for instructions which are available only on Cayman +// family GPUs. +// +//===----------------------------------------------------------------------===// + +def isCayman : Predicate<"Subtarget.hasCaymanISA()">; + +//===----------------------------------------------------------------------===// +// Cayman Instructions +//===----------------------------------------------------------------------===// + +let Predicates = [isCayman] in { + +def MULADD_INT24_cm : R600_3OP <0x08, "MULADD_INT24", + [(set i32:$dst, (add (mul I24:$src0, I24:$src1), i32:$src2))], VecALU +>; +def MUL_INT24_cm : R600_2OP <0x5B, "MUL_INT24", + [(set i32:$dst, (mul I24:$src0, I24:$src1))], VecALU +>; + +let isVector = 1 in { + +def RECIP_IEEE_cm : RECIP_IEEE_Common<0x86>; + +def MULLO_INT_cm : MULLO_INT_Common<0x8F>; +def MULHI_INT_cm : MULHI_INT_Common<0x90>; +def MULLO_UINT_cm : MULLO_UINT_Common<0x91>; +def MULHI_UINT_cm : MULHI_UINT_Common<0x92>; +def RECIPSQRT_CLAMPED_cm : RECIPSQRT_CLAMPED_Common<0x87>; +def EXP_IEEE_cm : EXP_IEEE_Common<0x81>; +def LOG_IEEE_cm : LOG_IEEE_Common<0x83>; +def RECIP_CLAMPED_cm : RECIP_CLAMPED_Common<0x84>; +def RECIPSQRT_IEEE_cm : RECIPSQRT_IEEE_Common<0x89>; +def SIN_cm : SIN_Common<0x8D>; +def COS_cm : COS_Common<0x8E>; +} // End isVector = 1 + +def : POW_Common ; + +defm DIV_cm : DIV_Common; + +// RECIP_UINT emulation for Cayman +// The multiplication scales from [0,1] to the unsigned integer range +def : Pat < + (AMDGPUurecip i32:$src0), + (FLT_TO_UINT_eg (MUL_IEEE (RECIP_IEEE_cm (UINT_TO_FLT_eg $src0)), + (MOV_IMM_I32 CONST.FP_UINT_MAX_PLUS_1))) +>; + + def CF_END_CM : CF_CLAUSE_EG<32, (ins), "CF_END"> { + let ADDR = 0; + let POP_COUNT = 0; + let COUNT = 0; + } + + +def : Pat<(fsqrt f32:$src), (MUL R600_Reg32:$src, (RECIPSQRT_CLAMPED_cm $src))>; + +class RAT_STORE_DWORD mask> : + CF_MEM_RAT_CACHELESS <0x14, 0, mask, + (ins rc:$rw_gpr, R600_TReg32_X:$index_gpr), + "STORE_DWORD $rw_gpr, $index_gpr", + [(global_store vt:$rw_gpr, i32:$index_gpr)]> { + let eop = 0; // This bit is not used on Cayman. +} + +def RAT_STORE_DWORD32 : RAT_STORE_DWORD ; +def RAT_STORE_DWORD64 : RAT_STORE_DWORD ; +def RAT_STORE_DWORD128 : RAT_STORE_DWORD ; + +class VTX_READ_cm buffer_id, dag outs, list pattern> + : VTX_WORD0_cm, VTX_READ { + + // Static fields + let VC_INST = 0; + let FETCH_TYPE = 2; + let FETCH_WHOLE_QUAD = 0; + let BUFFER_ID = buffer_id; + let SRC_REL = 0; + // XXX: We can infer this field based on the SRC_GPR. This would allow us + // to store vertex addresses in any channel, not just X. + let SRC_SEL_X = 0; + let SRC_SEL_Y = 0; + let STRUCTURED_READ = 0; + let LDS_REQ = 0; + let COALESCED_READ = 0; + + let Inst{31-0} = Word0; +} + +class VTX_READ_8_cm buffer_id, list pattern> + : VTX_READ_cm <"VTX_READ_8 $dst_gpr, $src_gpr", buffer_id, + (outs R600_TReg32_X:$dst_gpr), pattern> { + + let DST_SEL_X = 0; + let DST_SEL_Y = 7; // Masked + let DST_SEL_Z = 7; // Masked + let DST_SEL_W = 7; // Masked + let DATA_FORMAT = 1; // FMT_8 +} + +class VTX_READ_16_cm buffer_id, list pattern> + : VTX_READ_cm <"VTX_READ_16 $dst_gpr, $src_gpr", buffer_id, + (outs R600_TReg32_X:$dst_gpr), pattern> { + let DST_SEL_X = 0; + let DST_SEL_Y = 7; // Masked + let DST_SEL_Z = 7; // Masked + let DST_SEL_W = 7; // Masked + let DATA_FORMAT = 5; // FMT_16 + +} + +class VTX_READ_32_cm buffer_id, list pattern> + : VTX_READ_cm <"VTX_READ_32 $dst_gpr, $src_gpr", buffer_id, + (outs R600_TReg32_X:$dst_gpr), pattern> { + + let DST_SEL_X = 0; + let DST_SEL_Y = 7; // Masked + let DST_SEL_Z = 7; // Masked + let DST_SEL_W = 7; // Masked + let DATA_FORMAT = 0xD; // COLOR_32 + + // This is not really necessary, but there were some GPU hangs that appeared + // to be caused by ALU instructions in the next instruction group that wrote + // to the $src_gpr registers of the VTX_READ. + // e.g. + // %T3_X = VTX_READ_PARAM_32_eg %T2_X, 24 + // %T2_X = MOV %ZERO + //Adding this constraint prevents this from happening. + let Constraints = "$src_gpr.ptr = $dst_gpr"; +} + +class VTX_READ_64_cm buffer_id, list pattern> + : VTX_READ_cm <"VTX_READ_64 $dst_gpr, $src_gpr", buffer_id, + (outs R600_Reg64:$dst_gpr), pattern> { + + let DST_SEL_X = 0; + let DST_SEL_Y = 1; + let DST_SEL_Z = 7; + let DST_SEL_W = 7; + let DATA_FORMAT = 0x1D; // COLOR_32_32 +} + +class VTX_READ_128_cm buffer_id, list pattern> + : VTX_READ_cm <"VTX_READ_128 $dst_gpr.XYZW, $src_gpr", buffer_id, + (outs R600_Reg128:$dst_gpr), pattern> { + + let DST_SEL_X = 0; + let DST_SEL_Y = 1; + let DST_SEL_Z = 2; + let DST_SEL_W = 3; + let DATA_FORMAT = 0x22; // COLOR_32_32_32_32 + + // XXX: Need to force VTX_READ_128 instructions to write to the same register + // that holds its buffer address to avoid potential hangs. We can't use + // the same constraint as VTX_READ_32_eg, because the $src_gpr.ptr and $dst + // registers are different sizes. +} + +//===----------------------------------------------------------------------===// +// VTX Read from parameter memory space +//===----------------------------------------------------------------------===// +def VTX_READ_PARAM_8_cm : VTX_READ_8_cm <0, + [(set i32:$dst_gpr, (load_param_exti8 ADDRVTX_READ:$src_gpr))] +>; + +def VTX_READ_PARAM_16_cm : VTX_READ_16_cm <0, + [(set i32:$dst_gpr, (load_param_exti16 ADDRVTX_READ:$src_gpr))] +>; + +def VTX_READ_PARAM_32_cm : VTX_READ_32_cm <0, + [(set i32:$dst_gpr, (load_param ADDRVTX_READ:$src_gpr))] +>; + +def VTX_READ_PARAM_64_cm : VTX_READ_64_cm <0, + [(set v2i32:$dst_gpr, (load_param ADDRVTX_READ:$src_gpr))] +>; + +def VTX_READ_PARAM_128_cm : VTX_READ_128_cm <0, + [(set v4i32:$dst_gpr, (load_param ADDRVTX_READ:$src_gpr))] +>; + +//===----------------------------------------------------------------------===// +// VTX Read from global memory space +//===----------------------------------------------------------------------===// + +// 8-bit reads +def VTX_READ_GLOBAL_8_cm : VTX_READ_8_cm <1, + [(set i32:$dst_gpr, (az_extloadi8_global ADDRVTX_READ:$src_gpr))] +>; + +def VTX_READ_GLOBAL_16_cm : VTX_READ_16_cm <1, + [(set i32:$dst_gpr, (az_extloadi16_global ADDRVTX_READ:$src_gpr))] +>; + +// 32-bit reads +def VTX_READ_GLOBAL_32_cm : VTX_READ_32_cm <1, + [(set i32:$dst_gpr, (global_load ADDRVTX_READ:$src_gpr))] +>; + +// 64-bit reads +def VTX_READ_GLOBAL_64_cm : VTX_READ_64_cm <1, + [(set v2i32:$dst_gpr, (global_load ADDRVTX_READ:$src_gpr))] +>; + +// 128-bit reads +def VTX_READ_GLOBAL_128_cm : VTX_READ_128_cm <1, + [(set v4i32:$dst_gpr, (global_load ADDRVTX_READ:$src_gpr))] +>; + +} // End isCayman + diff --git a/lib/Target/R600/EvergreenInstructions.td b/lib/Target/R600/EvergreenInstructions.td new file mode 100644 index 00000000000..384b98eebe8 --- /dev/null +++ b/lib/Target/R600/EvergreenInstructions.td @@ -0,0 +1,579 @@ +//===-- EvergreenInstructions.td - EG Instruction defs ----*- tablegen -*-===// +// +// The LLVM Compiler Infrastructure +// +// This file is distributed under the University of Illinois Open Source +// License. See LICENSE.TXT for details. +// +//===----------------------------------------------------------------------===// +// +// TableGen definitions for instructions which are: +// - Available to Evergreen and newer VLIW4/VLIW5 GPUs +// - Available only on Evergreen family GPUs. +// +//===----------------------------------------------------------------------===// + +def isEG : Predicate< + "Subtarget.getGeneration() >= AMDGPUSubtarget::EVERGREEN && " + "Subtarget.getGeneration() < AMDGPUSubtarget::SOUTHERN_ISLANDS && " + "!Subtarget.hasCaymanISA()" +>; + +def isEGorCayman : Predicate< + "Subtarget.getGeneration() == AMDGPUSubtarget::EVERGREEN ||" + "Subtarget.getGeneration() ==AMDGPUSubtarget::NORTHERN_ISLANDS" +>; + +//===----------------------------------------------------------------------===// +// Evergreen / Cayman store instructions +//===----------------------------------------------------------------------===// + +let Predicates = [isEGorCayman] in { + +class CF_MEM_RAT_CACHELESS rat_inst, bits<4> rat_id, bits<4> mask, dag ins, + string name, list pattern> + : EG_CF_RAT <0x57, rat_inst, rat_id, mask, (outs), ins, + "MEM_RAT_CACHELESS "#name, pattern>; + +class CF_MEM_RAT rat_inst, bits<4> rat_id, dag ins, string name, + list pattern> + : EG_CF_RAT <0x56, rat_inst, rat_id, 0xf /* mask */, (outs), ins, + "MEM_RAT "#name, pattern>; + +def RAT_MSKOR : CF_MEM_RAT <0x11, 0, + (ins R600_Reg128:$rw_gpr, R600_TReg32_X:$index_gpr), + "MSKOR $rw_gpr.XW, $index_gpr", + [(mskor_global v4i32:$rw_gpr, i32:$index_gpr)] +> { + let eop = 0; +} + +} // End let Predicates = [isEGorCayman] + +//===----------------------------------------------------------------------===// +// Evergreen Only instructions +//===----------------------------------------------------------------------===// + +let Predicates = [isEG] in { + +def RECIP_IEEE_eg : RECIP_IEEE_Common<0x86>; +defm DIV_eg : DIV_Common; + +def MULLO_INT_eg : MULLO_INT_Common<0x8F>; +def MULHI_INT_eg : MULHI_INT_Common<0x90>; +def MULLO_UINT_eg : MULLO_UINT_Common<0x91>; +def MULHI_UINT_eg : MULHI_UINT_Common<0x92>; +def RECIP_UINT_eg : RECIP_UINT_Common<0x94>; +def RECIPSQRT_CLAMPED_eg : RECIPSQRT_CLAMPED_Common<0x87>; +def EXP_IEEE_eg : EXP_IEEE_Common<0x81>; +def LOG_IEEE_eg : LOG_IEEE_Common<0x83>; +def RECIP_CLAMPED_eg : RECIP_CLAMPED_Common<0x84>; +def RECIPSQRT_IEEE_eg : RECIPSQRT_IEEE_Common<0x89>; +def SIN_eg : SIN_Common<0x8D>; +def COS_eg : COS_Common<0x8E>; + +def : POW_Common ; +def : Pat<(fsqrt f32:$src), (MUL $src, (RECIPSQRT_CLAMPED_eg $src))>; + +//===----------------------------------------------------------------------===// +// Memory read/write instructions +//===----------------------------------------------------------------------===// + +let usesCustomInserter = 1 in { + +// 32-bit store +def RAT_WRITE_CACHELESS_32_eg : CF_MEM_RAT_CACHELESS <0x2, 0, 0x1, + (ins R600_TReg32_X:$rw_gpr, R600_TReg32_X:$index_gpr, InstFlag:$eop), + "STORE_RAW $rw_gpr, $index_gpr, $eop", + [(global_store i32:$rw_gpr, i32:$index_gpr)] +>; + +// 64-bit store +def RAT_WRITE_CACHELESS_64_eg : CF_MEM_RAT_CACHELESS <0x2, 0, 0x3, + (ins R600_Reg64:$rw_gpr, R600_TReg32_X:$index_gpr, InstFlag:$eop), + "STORE_RAW $rw_gpr.XY, $index_gpr, $eop", + [(global_store v2i32:$rw_gpr, i32:$index_gpr)] +>; + +//128-bit store +def RAT_WRITE_CACHELESS_128_eg : CF_MEM_RAT_CACHELESS <0x2, 0, 0xf, + (ins R600_Reg128:$rw_gpr, R600_TReg32_X:$index_gpr, InstFlag:$eop), + "STORE_RAW $rw_gpr.XYZW, $index_gpr, $eop", + [(global_store v4i32:$rw_gpr, i32:$index_gpr)] +>; + +} // End usesCustomInserter = 1 + +class VTX_READ_eg buffer_id, dag outs, list pattern> + : VTX_WORD0_eg, VTX_READ { + + // Static fields + let VC_INST = 0; + let FETCH_TYPE = 2; + let FETCH_WHOLE_QUAD = 0; + let BUFFER_ID = buffer_id; + let SRC_REL = 0; + // XXX: We can infer this field based on the SRC_GPR. This would allow us + // to store vertex addresses in any channel, not just X. + let SRC_SEL_X = 0; + + let Inst{31-0} = Word0; +} + +class VTX_READ_8_eg buffer_id, list pattern> + : VTX_READ_eg <"VTX_READ_8 $dst_gpr, $src_gpr", buffer_id, + (outs R600_TReg32_X:$dst_gpr), pattern> { + + let MEGA_FETCH_COUNT = 1; + let DST_SEL_X = 0; + let DST_SEL_Y = 7; // Masked + let DST_SEL_Z = 7; // Masked + let DST_SEL_W = 7; // Masked + let DATA_FORMAT = 1; // FMT_8 +} + +class VTX_READ_16_eg buffer_id, list pattern> + : VTX_READ_eg <"VTX_READ_16 $dst_gpr, $src_gpr", buffer_id, + (outs R600_TReg32_X:$dst_gpr), pattern> { + let MEGA_FETCH_COUNT = 2; + let DST_SEL_X = 0; + let DST_SEL_Y = 7; // Masked + let DST_SEL_Z = 7; // Masked + let DST_SEL_W = 7; // Masked + let DATA_FORMAT = 5; // FMT_16 + +} + +class VTX_READ_32_eg buffer_id, list pattern> + : VTX_READ_eg <"VTX_READ_32 $dst_gpr, $src_gpr", buffer_id, + (outs R600_TReg32_X:$dst_gpr), pattern> { + + let MEGA_FETCH_COUNT = 4; + let DST_SEL_X = 0; + let DST_SEL_Y = 7; // Masked + let DST_SEL_Z = 7; // Masked + let DST_SEL_W = 7; // Masked + let DATA_FORMAT = 0xD; // COLOR_32 + + // This is not really necessary, but there were some GPU hangs that appeared + // to be caused by ALU instructions in the next instruction group that wrote + // to the $src_gpr registers of the VTX_READ. + // e.g. + // %T3_X = VTX_READ_PARAM_32_eg %T2_X, 24 + // %T2_X = MOV %ZERO + //Adding this constraint prevents this from happening. + let Constraints = "$src_gpr.ptr = $dst_gpr"; +} + +class VTX_READ_64_eg buffer_id, list pattern> + : VTX_READ_eg <"VTX_READ_64 $dst_gpr.XY, $src_gpr", buffer_id, + (outs R600_Reg64:$dst_gpr), pattern> { + + let MEGA_FETCH_COUNT = 8; + let DST_SEL_X = 0; + let DST_SEL_Y = 1; + let DST_SEL_Z = 7; + let DST_SEL_W = 7; + let DATA_FORMAT = 0x1D; // COLOR_32_32 +} + +class VTX_READ_128_eg buffer_id, list pattern> + : VTX_READ_eg <"VTX_READ_128 $dst_gpr.XYZW, $src_gpr", buffer_id, + (outs R600_Reg128:$dst_gpr), pattern> { + + let MEGA_FETCH_COUNT = 16; + let DST_SEL_X = 0; + let DST_SEL_Y = 1; + let DST_SEL_Z = 2; + let DST_SEL_W = 3; + let DATA_FORMAT = 0x22; // COLOR_32_32_32_32 + + // XXX: Need to force VTX_READ_128 instructions to write to the same register + // that holds its buffer address to avoid potential hangs. We can't use + // the same constraint as VTX_READ_32_eg, because the $src_gpr.ptr and $dst + // registers are different sizes. +} + +//===----------------------------------------------------------------------===// +// VTX Read from parameter memory space +//===----------------------------------------------------------------------===// + +def VTX_READ_PARAM_8_eg : VTX_READ_8_eg <0, + [(set i32:$dst_gpr, (load_param_exti8 ADDRVTX_READ:$src_gpr))] +>; + +def VTX_READ_PARAM_16_eg : VTX_READ_16_eg <0, + [(set i32:$dst_gpr, (load_param_exti16 ADDRVTX_READ:$src_gpr))] +>; + +def VTX_READ_PARAM_32_eg : VTX_READ_32_eg <0, + [(set i32:$dst_gpr, (load_param ADDRVTX_READ:$src_gpr))] +>; + +def VTX_READ_PARAM_64_eg : VTX_READ_64_eg <0, + [(set v2i32:$dst_gpr, (load_param ADDRVTX_READ:$src_gpr))] +>; + +def VTX_READ_PARAM_128_eg : VTX_READ_128_eg <0, + [(set v4i32:$dst_gpr, (load_param ADDRVTX_READ:$src_gpr))] +>; + +//===----------------------------------------------------------------------===// +// VTX Read from global memory space +//===----------------------------------------------------------------------===// + +// 8-bit reads +def VTX_READ_GLOBAL_8_eg : VTX_READ_8_eg <1, + [(set i32:$dst_gpr, (az_extloadi8_global ADDRVTX_READ:$src_gpr))] +>; + +def VTX_READ_GLOBAL_16_eg : VTX_READ_16_eg <1, + [(set i32:$dst_gpr, (az_extloadi16_global ADDRVTX_READ:$src_gpr))] +>; + +// 32-bit reads +def VTX_READ_GLOBAL_32_eg : VTX_READ_32_eg <1, + [(set i32:$dst_gpr, (global_load ADDRVTX_READ:$src_gpr))] +>; + +// 64-bit reads +def VTX_READ_GLOBAL_64_eg : VTX_READ_64_eg <1, + [(set v2i32:$dst_gpr, (global_load ADDRVTX_READ:$src_gpr))] +>; + +// 128-bit reads +def VTX_READ_GLOBAL_128_eg : VTX_READ_128_eg <1, + [(set v4i32:$dst_gpr, (global_load ADDRVTX_READ:$src_gpr))] +>; + +} // End Predicates = [isEG] + +//===----------------------------------------------------------------------===// +// Evergreen / Cayman Instructions +//===----------------------------------------------------------------------===// + +let Predicates = [isEGorCayman] in { + +// BFE_UINT - bit_extract, an optimization for mask and shift +// Src0 = Input +// Src1 = Offset +// Src2 = Width +// +// bit_extract = (Input << (32 - Offset - Width)) >> (32 - Width) +// +// Example Usage: +// (Offset, Width) +// +// (0, 8) = (Input << 24) >> 24 = (Input & 0xff) >> 0 +// (8, 8) = (Input << 16) >> 24 = (Input & 0xffff) >> 8 +// (16, 8) = (Input << 8) >> 24 = (Input & 0xffffff) >> 16 +// (24, 8) = (Input << 0) >> 24 = (Input & 0xffffffff) >> 24 +def BFE_UINT_eg : R600_3OP <0x4, "BFE_UINT", + [(set i32:$dst, (AMDGPUbfe_u32 i32:$src0, i32:$src1, i32:$src2))], + VecALU +>; + +def BFE_INT_eg : R600_3OP <0x4, "BFE_INT", + [(set i32:$dst, (AMDGPUbfe_i32 i32:$src0, i32:$src1, i32:$src2))], + VecALU +>; + +// XXX: This pattern is broken, disabling for now. See comment in +// AMDGPUInstructions.td for more info. +// def : BFEPattern ; + +def BFI_INT_eg : R600_3OP <0x06, "BFI_INT", [], VecALU>; +defm : BFIPatterns ; + +def MULADD_UINT24_eg : R600_3OP <0x10, "MULADD_UINT24", + [(set i32:$dst, (add (mul U24:$src0, U24:$src1), i32:$src2))], VecALU +>; +def BIT_ALIGN_INT_eg : R600_3OP <0xC, "BIT_ALIGN_INT", [], VecALU>; +def : ROTRPattern ; +def MULADD_eg : MULADD_Common<0x14>; +def MULADD_IEEE_eg : MULADD_IEEE_Common<0x18>; +def ASHR_eg : ASHR_Common<0x15>; +def LSHR_eg : LSHR_Common<0x16>; +def LSHL_eg : LSHL_Common<0x17>; +def CNDE_eg : CNDE_Common<0x19>; +def CNDGT_eg : CNDGT_Common<0x1A>; +def CNDGE_eg : CNDGE_Common<0x1B>; +def MUL_LIT_eg : MUL_LIT_Common<0x1F>; +def LOG_CLAMPED_eg : LOG_CLAMPED_Common<0x82>; +def MUL_UINT24_eg : R600_2OP <0xB5, "MUL_UINT24", + [(set i32:$dst, (mul U24:$src0, U24:$src1))], VecALU +>; +def DOT4_eg : DOT4_Common<0xBE>; +defm CUBE_eg : CUBE_Common<0xC0>; + +let hasSideEffects = 1 in { + def MOVA_INT_eg : R600_1OP <0xCC, "MOVA_INT", [], VecALU>; +} + +def TGSI_LIT_Z_eg : TGSI_LIT_Z_Common; + +def FLT_TO_INT_eg : FLT_TO_INT_Common<0x50> { + let Pattern = []; + let Itinerary = AnyALU; +} + +def INT_TO_FLT_eg : INT_TO_FLT_Common<0x9B>; + +def FLT_TO_UINT_eg : FLT_TO_UINT_Common<0x9A> { + let Pattern = []; +} + +def UINT_TO_FLT_eg : UINT_TO_FLT_Common<0x9C>; + +def GROUP_BARRIER : InstR600 < + (outs), (ins), " GROUP_BARRIER", [(int_AMDGPU_barrier_local)], AnyALU>, + R600ALU_Word0, + R600ALU_Word1_OP2 <0x54> { + + let dst = 0; + let dst_rel = 0; + let src0 = 0; + let src0_rel = 0; + let src0_neg = 0; + let src0_abs = 0; + let src1 = 0; + let src1_rel = 0; + let src1_neg = 0; + let src1_abs = 0; + let write = 0; + let omod = 0; + let clamp = 0; + let last = 1; + let bank_swizzle = 0; + let pred_sel = 0; + let update_exec_mask = 0; + let update_pred = 0; + + let Inst{31-0} = Word0; + let Inst{63-32} = Word1; + + let ALUInst = 1; +} + +//===----------------------------------------------------------------------===// +// LDS Instructions +//===----------------------------------------------------------------------===// +class R600_LDS op, dag outs, dag ins, string asm, + list pattern = []> : + + InstR600 , + R600_ALU_LDS_Word0, + R600LDS_Word1 { + + bits<6> offset = 0; + let lds_op = op; + + let Word1{27} = offset{0}; + let Word1{12} = offset{1}; + let Word1{28} = offset{2}; + let Word1{31} = offset{3}; + let Word0{12} = offset{4}; + let Word0{25} = offset{5}; + + + let Inst{31-0} = Word0; + let Inst{63-32} = Word1; + + let ALUInst = 1; + let HasNativeOperands = 1; + let UseNamedOperandTable = 1; +} + +class R600_LDS_1A lds_op, string name, list pattern> : R600_LDS < + lds_op, + (outs R600_Reg32:$dst), + (ins R600_Reg32:$src0, REL:$src0_rel, SEL:$src0_sel, + LAST:$last, R600_Pred:$pred_sel, + BANK_SWIZZLE:$bank_swizzle), + " "#name#" $last OQAP, $src0$src0_rel $pred_sel", + pattern + > { + + let src1 = 0; + let src1_rel = 0; + let src2 = 0; + let src2_rel = 0; + + let usesCustomInserter = 1; + let LDS_1A = 1; + let DisableEncoding = "$dst"; +} + +class R600_LDS_1A1D lds_op, dag outs, string name, list pattern, + string dst =""> : + R600_LDS < + lds_op, outs, + (ins R600_Reg32:$src0, REL:$src0_rel, SEL:$src0_sel, + R600_Reg32:$src1, REL:$src1_rel, SEL:$src1_sel, + LAST:$last, R600_Pred:$pred_sel, + BANK_SWIZZLE:$bank_swizzle), + " "#name#" $last "#dst#"$src0$src0_rel, $src1$src1_rel, $pred_sel", + pattern + > { + + field string BaseOp; + + let src2 = 0; + let src2_rel = 0; + let LDS_1A1D = 1; +} + +class R600_LDS_1A1D_NORET lds_op, string name, list pattern> : + R600_LDS_1A1D { + let BaseOp = name; +} + +class R600_LDS_1A1D_RET lds_op, string name, list pattern> : + R600_LDS_1A1D { + + let BaseOp = name; + let usesCustomInserter = 1; + let DisableEncoding = "$dst"; +} + +class R600_LDS_1A2D lds_op, string name, list pattern> : + R600_LDS < + lds_op, + (outs), + (ins R600_Reg32:$src0, REL:$src0_rel, SEL:$src0_sel, + R600_Reg32:$src1, REL:$src1_rel, SEL:$src1_sel, + R600_Reg32:$src2, REL:$src2_rel, SEL:$src2_sel, + LAST:$last, R600_Pred:$pred_sel, BANK_SWIZZLE:$bank_swizzle), + " "#name# "$last $src0$src0_rel, $src1$src1_rel, $src2$src2_rel, $pred_sel", + pattern> { + let LDS_1A2D = 1; +} + +def LDS_ADD : R600_LDS_1A1D_NORET <0x0, "LDS_ADD", [] >; +def LDS_SUB : R600_LDS_1A1D_NORET <0x1, "LDS_SUB", [] >; +def LDS_WRITE : R600_LDS_1A1D_NORET <0xD, "LDS_WRITE", + [(local_store (i32 R600_Reg32:$src1), R600_Reg32:$src0)] +>; +def LDS_BYTE_WRITE : R600_LDS_1A1D_NORET<0x12, "LDS_BYTE_WRITE", + [(truncstorei8_local i32:$src1, i32:$src0)] +>; +def LDS_SHORT_WRITE : R600_LDS_1A1D_NORET<0x13, "LDS_SHORT_WRITE", + [(truncstorei16_local i32:$src1, i32:$src0)] +>; +def LDS_ADD_RET : R600_LDS_1A1D_RET <0x20, "LDS_ADD", + [(set i32:$dst, (atomic_load_add_local i32:$src0, i32:$src1))] +>; +def LDS_SUB_RET : R600_LDS_1A1D_RET <0x21, "LDS_SUB", + [(set i32:$dst, (atomic_load_sub_local i32:$src0, i32:$src1))] +>; +def LDS_READ_RET : R600_LDS_1A <0x32, "LDS_READ_RET", + [(set (i32 R600_Reg32:$dst), (local_load R600_Reg32:$src0))] +>; +def LDS_BYTE_READ_RET : R600_LDS_1A <0x36, "LDS_BYTE_READ_RET", + [(set i32:$dst, (sextloadi8_local i32:$src0))] +>; +def LDS_UBYTE_READ_RET : R600_LDS_1A <0x37, "LDS_UBYTE_READ_RET", + [(set i32:$dst, (az_extloadi8_local i32:$src0))] +>; +def LDS_SHORT_READ_RET : R600_LDS_1A <0x38, "LDS_SHORT_READ_RET", + [(set i32:$dst, (sextloadi16_local i32:$src0))] +>; +def LDS_USHORT_READ_RET : R600_LDS_1A <0x39, "LDS_USHORT_READ_RET", + [(set i32:$dst, (az_extloadi16_local i32:$src0))] +>; + +// TRUNC is used for the FLT_TO_INT instructions to work around a +// perceived problem where the rounding modes are applied differently +// depending on the instruction and the slot they are in. +// See: +// https://bugs.freedesktop.org/show_bug.cgi?id=50232 +// Mesa commit: a1a0974401c467cb86ef818f22df67c21774a38c +// +// XXX: Lowering SELECT_CC will sometimes generate fp_to_[su]int nodes, +// which do not need to be truncated since the fp values are 0.0f or 1.0f. +// We should look into handling these cases separately. +def : Pat<(fp_to_sint f32:$src0), (FLT_TO_INT_eg (TRUNC $src0))>; + +def : Pat<(fp_to_uint f32:$src0), (FLT_TO_UINT_eg (TRUNC $src0))>; + +// SHA-256 Patterns +def : SHA256MaPattern ; + +def : FROUNDPat ; + +def EG_ExportSwz : ExportSwzInst { + let Word1{19-16} = 0; // BURST_COUNT + let Word1{20} = 0; // VALID_PIXEL_MODE + let Word1{21} = eop; + let Word1{29-22} = inst; + let Word1{30} = 0; // MARK + let Word1{31} = 1; // BARRIER +} +defm : ExportPattern; + +def EG_ExportBuf : ExportBufInst { + let Word1{19-16} = 0; // BURST_COUNT + let Word1{20} = 0; // VALID_PIXEL_MODE + let Word1{21} = eop; + let Word1{29-22} = inst; + let Word1{30} = 0; // MARK + let Word1{31} = 1; // BARRIER +} +defm : SteamOutputExportPattern; + +def CF_TC_EG : CF_CLAUSE_EG<1, (ins i32imm:$ADDR, i32imm:$COUNT), + "TEX $COUNT @$ADDR"> { + let POP_COUNT = 0; +} +def CF_VC_EG : CF_CLAUSE_EG<2, (ins i32imm:$ADDR, i32imm:$COUNT), + "VTX $COUNT @$ADDR"> { + let POP_COUNT = 0; +} +def WHILE_LOOP_EG : CF_CLAUSE_EG<6, (ins i32imm:$ADDR), + "LOOP_START_DX10 @$ADDR"> { + let POP_COUNT = 0; + let COUNT = 0; +} +def END_LOOP_EG : CF_CLAUSE_EG<5, (ins i32imm:$ADDR), "END_LOOP @$ADDR"> { + let POP_COUNT = 0; + let COUNT = 0; +} +def LOOP_BREAK_EG : CF_CLAUSE_EG<9, (ins i32imm:$ADDR), + "LOOP_BREAK @$ADDR"> { + let POP_COUNT = 0; + let COUNT = 0; +} +def CF_CONTINUE_EG : CF_CLAUSE_EG<8, (ins i32imm:$ADDR), + "CONTINUE @$ADDR"> { + let POP_COUNT = 0; + let COUNT = 0; +} +def CF_JUMP_EG : CF_CLAUSE_EG<10, (ins i32imm:$ADDR, i32imm:$POP_COUNT), + "JUMP @$ADDR POP:$POP_COUNT"> { + let COUNT = 0; +} +def CF_PUSH_EG : CF_CLAUSE_EG<11, (ins i32imm:$ADDR, i32imm:$POP_COUNT), + "PUSH @$ADDR POP:$POP_COUNT"> { + let COUNT = 0; +} +def CF_ELSE_EG : CF_CLAUSE_EG<13, (ins i32imm:$ADDR, i32imm:$POP_COUNT), + "ELSE @$ADDR POP:$POP_COUNT"> { + let COUNT = 0; +} +def CF_CALL_FS_EG : CF_CLAUSE_EG<19, (ins), "CALL_FS"> { + let ADDR = 0; + let COUNT = 0; + let POP_COUNT = 0; +} +def POP_EG : CF_CLAUSE_EG<14, (ins i32imm:$ADDR, i32imm:$POP_COUNT), + "POP @$ADDR POP:$POP_COUNT"> { + let COUNT = 0; +} +def CF_END_EG : CF_CLAUSE_EG<0, (ins), "CF_END"> { + let COUNT = 0; + let POP_COUNT = 0; + let ADDR = 0; + let END_OF_PROGRAM = 1; +} + +} // End Predicates = [isEGorCayman] diff --git a/lib/Target/R600/R600Instructions.td b/lib/Target/R600/R600Instructions.td index ae3d8747a4d..d2075c0577b 100644 --- a/lib/Target/R600/R600Instructions.td +++ b/lib/Target/R600/R600Instructions.td @@ -7,7 +7,8 @@ // //===----------------------------------------------------------------------===// // -// R600 Tablegen instruction definitions +// TableGen definitions for instructions which are available on R600 family +// GPUs. // //===----------------------------------------------------------------------===// @@ -335,17 +336,6 @@ def load_param_exti8 : LoadParamFrag; def load_param_exti16 : LoadParamFrag; def isR600 : Predicate<"Subtarget.getGeneration() <= AMDGPUSubtarget::R700">; -def isR700 : Predicate<"Subtarget.getGeneration() == AMDGPUSubtarget::R700">; -def isEG : Predicate< - "Subtarget.getGeneration() >= AMDGPUSubtarget::EVERGREEN && " - "Subtarget.getGeneration() < AMDGPUSubtarget::SOUTHERN_ISLANDS && " - "!Subtarget.hasCaymanISA()">; - -def isCayman : Predicate<"Subtarget.hasCaymanISA()">; -def isEGorCayman : Predicate<"Subtarget.getGeneration() == " - "AMDGPUSubtarget::EVERGREEN" - "|| Subtarget.getGeneration() ==" - "AMDGPUSubtarget::NORTHERN_ISLANDS">; def isR600toCayman : Predicate< "Subtarget.getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS">; @@ -1267,570 +1257,6 @@ let Predicates = [isR600] in { } -//===----------------------------------------------------------------------===// -// R700 Only instructions -//===----------------------------------------------------------------------===// - -let Predicates = [isR700] in { - def SIN_r700 : SIN_Common<0x6E>; - def COS_r700 : COS_Common<0x6F>; -} - -//===----------------------------------------------------------------------===// -// Evergreen / Cayman store instructions -//===----------------------------------------------------------------------===// - -let Predicates = [isEGorCayman] in { - -class CF_MEM_RAT_CACHELESS rat_inst, bits<4> rat_id, bits<4> mask, dag ins, - string name, list pattern> - : EG_CF_RAT <0x57, rat_inst, rat_id, mask, (outs), ins, - "MEM_RAT_CACHELESS "#name, pattern>; - -class CF_MEM_RAT rat_inst, bits<4> rat_id, dag ins, string name, - list pattern> - : EG_CF_RAT <0x56, rat_inst, rat_id, 0xf /* mask */, (outs), ins, - "MEM_RAT "#name, pattern>; - -def RAT_MSKOR : CF_MEM_RAT <0x11, 0, - (ins R600_Reg128:$rw_gpr, R600_TReg32_X:$index_gpr), - "MSKOR $rw_gpr.XW, $index_gpr", - [(mskor_global v4i32:$rw_gpr, i32:$index_gpr)] -> { - let eop = 0; -} - -} // End Predicates = [isEGorCayman] - - -//===----------------------------------------------------------------------===// -// Evergreen Only instructions -//===----------------------------------------------------------------------===// - -let Predicates = [isEG] in { - -def RECIP_IEEE_eg : RECIP_IEEE_Common<0x86>; -defm DIV_eg : DIV_Common; - -def MULLO_INT_eg : MULLO_INT_Common<0x8F>; -def MULHI_INT_eg : MULHI_INT_Common<0x90>; -def MULLO_UINT_eg : MULLO_UINT_Common<0x91>; -def MULHI_UINT_eg : MULHI_UINT_Common<0x92>; -def RECIP_UINT_eg : RECIP_UINT_Common<0x94>; -def RECIPSQRT_CLAMPED_eg : RECIPSQRT_CLAMPED_Common<0x87>; -def EXP_IEEE_eg : EXP_IEEE_Common<0x81>; -def LOG_IEEE_eg : LOG_IEEE_Common<0x83>; -def RECIP_CLAMPED_eg : RECIP_CLAMPED_Common<0x84>; -def RECIPSQRT_IEEE_eg : RECIPSQRT_IEEE_Common<0x89>; -def SIN_eg : SIN_Common<0x8D>; -def COS_eg : COS_Common<0x8E>; - -def : POW_Common ; -def : Pat<(fsqrt f32:$src), (MUL $src, (RECIPSQRT_CLAMPED_eg $src))>; - -//===----------------------------------------------------------------------===// -// Memory read/write instructions -//===----------------------------------------------------------------------===// - -let usesCustomInserter = 1 in { - -// 32-bit store -def RAT_WRITE_CACHELESS_32_eg : CF_MEM_RAT_CACHELESS <0x2, 0, 0x1, - (ins R600_TReg32_X:$rw_gpr, R600_TReg32_X:$index_gpr, InstFlag:$eop), - "STORE_RAW $rw_gpr, $index_gpr, $eop", - [(global_store i32:$rw_gpr, i32:$index_gpr)] ->; - -// 64-bit store -def RAT_WRITE_CACHELESS_64_eg : CF_MEM_RAT_CACHELESS <0x2, 0, 0x3, - (ins R600_Reg64:$rw_gpr, R600_TReg32_X:$index_gpr, InstFlag:$eop), - "STORE_RAW $rw_gpr.XY, $index_gpr, $eop", - [(global_store v2i32:$rw_gpr, i32:$index_gpr)] ->; - -//128-bit store -def RAT_WRITE_CACHELESS_128_eg : CF_MEM_RAT_CACHELESS <0x2, 0, 0xf, - (ins R600_Reg128:$rw_gpr, R600_TReg32_X:$index_gpr, InstFlag:$eop), - "STORE_RAW $rw_gpr.XYZW, $index_gpr, $eop", - [(global_store v4i32:$rw_gpr, i32:$index_gpr)] ->; - -} // End usesCustomInserter = 1 - -class VTX_READ_eg buffer_id, dag outs, list pattern> - : VTX_WORD0_eg, VTX_READ { - - // Static fields - let VC_INST = 0; - let FETCH_TYPE = 2; - let FETCH_WHOLE_QUAD = 0; - let BUFFER_ID = buffer_id; - let SRC_REL = 0; - // XXX: We can infer this field based on the SRC_GPR. This would allow us - // to store vertex addresses in any channel, not just X. - let SRC_SEL_X = 0; - - let Inst{31-0} = Word0; -} - -class VTX_READ_8_eg buffer_id, list pattern> - : VTX_READ_eg <"VTX_READ_8 $dst_gpr, $src_gpr", buffer_id, - (outs R600_TReg32_X:$dst_gpr), pattern> { - - let MEGA_FETCH_COUNT = 1; - let DST_SEL_X = 0; - let DST_SEL_Y = 7; // Masked - let DST_SEL_Z = 7; // Masked - let DST_SEL_W = 7; // Masked - let DATA_FORMAT = 1; // FMT_8 -} - -class VTX_READ_16_eg buffer_id, list pattern> - : VTX_READ_eg <"VTX_READ_16 $dst_gpr, $src_gpr", buffer_id, - (outs R600_TReg32_X:$dst_gpr), pattern> { - let MEGA_FETCH_COUNT = 2; - let DST_SEL_X = 0; - let DST_SEL_Y = 7; // Masked - let DST_SEL_Z = 7; // Masked - let DST_SEL_W = 7; // Masked - let DATA_FORMAT = 5; // FMT_16 - -} - -class VTX_READ_32_eg buffer_id, list pattern> - : VTX_READ_eg <"VTX_READ_32 $dst_gpr, $src_gpr", buffer_id, - (outs R600_TReg32_X:$dst_gpr), pattern> { - - let MEGA_FETCH_COUNT = 4; - let DST_SEL_X = 0; - let DST_SEL_Y = 7; // Masked - let DST_SEL_Z = 7; // Masked - let DST_SEL_W = 7; // Masked - let DATA_FORMAT = 0xD; // COLOR_32 - - // This is not really necessary, but there were some GPU hangs that appeared - // to be caused by ALU instructions in the next instruction group that wrote - // to the $src_gpr registers of the VTX_READ. - // e.g. - // %T3_X = VTX_READ_PARAM_32_eg %T2_X, 24 - // %T2_X = MOV %ZERO - //Adding this constraint prevents this from happening. - let Constraints = "$src_gpr.ptr = $dst_gpr"; -} - -class VTX_READ_64_eg buffer_id, list pattern> - : VTX_READ_eg <"VTX_READ_64 $dst_gpr.XY, $src_gpr", buffer_id, - (outs R600_Reg64:$dst_gpr), pattern> { - - let MEGA_FETCH_COUNT = 8; - let DST_SEL_X = 0; - let DST_SEL_Y = 1; - let DST_SEL_Z = 7; - let DST_SEL_W = 7; - let DATA_FORMAT = 0x1D; // COLOR_32_32 -} - -class VTX_READ_128_eg buffer_id, list pattern> - : VTX_READ_eg <"VTX_READ_128 $dst_gpr.XYZW, $src_gpr", buffer_id, - (outs R600_Reg128:$dst_gpr), pattern> { - - let MEGA_FETCH_COUNT = 16; - let DST_SEL_X = 0; - let DST_SEL_Y = 1; - let DST_SEL_Z = 2; - let DST_SEL_W = 3; - let DATA_FORMAT = 0x22; // COLOR_32_32_32_32 - - // XXX: Need to force VTX_READ_128 instructions to write to the same register - // that holds its buffer address to avoid potential hangs. We can't use - // the same constraint as VTX_READ_32_eg, because the $src_gpr.ptr and $dst - // registers are different sizes. -} - -//===----------------------------------------------------------------------===// -// VTX Read from parameter memory space -//===----------------------------------------------------------------------===// - -def VTX_READ_PARAM_8_eg : VTX_READ_8_eg <0, - [(set i32:$dst_gpr, (load_param_exti8 ADDRVTX_READ:$src_gpr))] ->; - -def VTX_READ_PARAM_16_eg : VTX_READ_16_eg <0, - [(set i32:$dst_gpr, (load_param_exti16 ADDRVTX_READ:$src_gpr))] ->; - -def VTX_READ_PARAM_32_eg : VTX_READ_32_eg <0, - [(set i32:$dst_gpr, (load_param ADDRVTX_READ:$src_gpr))] ->; - -def VTX_READ_PARAM_64_eg : VTX_READ_64_eg <0, - [(set v2i32:$dst_gpr, (load_param ADDRVTX_READ:$src_gpr))] ->; - -def VTX_READ_PARAM_128_eg : VTX_READ_128_eg <0, - [(set v4i32:$dst_gpr, (load_param ADDRVTX_READ:$src_gpr))] ->; - -//===----------------------------------------------------------------------===// -// VTX Read from global memory space -//===----------------------------------------------------------------------===// - -// 8-bit reads -def VTX_READ_GLOBAL_8_eg : VTX_READ_8_eg <1, - [(set i32:$dst_gpr, (az_extloadi8_global ADDRVTX_READ:$src_gpr))] ->; - -def VTX_READ_GLOBAL_16_eg : VTX_READ_16_eg <1, - [(set i32:$dst_gpr, (az_extloadi16_global ADDRVTX_READ:$src_gpr))] ->; - -// 32-bit reads -def VTX_READ_GLOBAL_32_eg : VTX_READ_32_eg <1, - [(set i32:$dst_gpr, (global_load ADDRVTX_READ:$src_gpr))] ->; - -// 64-bit reads -def VTX_READ_GLOBAL_64_eg : VTX_READ_64_eg <1, - [(set v2i32:$dst_gpr, (global_load ADDRVTX_READ:$src_gpr))] ->; - -// 128-bit reads -def VTX_READ_GLOBAL_128_eg : VTX_READ_128_eg <1, - [(set v4i32:$dst_gpr, (global_load ADDRVTX_READ:$src_gpr))] ->; - -} // End Predicates = [isEG] - -//===----------------------------------------------------------------------===// -// Evergreen / Cayman Instructions -//===----------------------------------------------------------------------===// - -let Predicates = [isEGorCayman] in { - - // BFE_UINT - bit_extract, an optimization for mask and shift - // Src0 = Input - // Src1 = Offset - // Src2 = Width - // - // bit_extract = (Input << (32 - Offset - Width)) >> (32 - Width) - // - // Example Usage: - // (Offset, Width) - // - // (0, 8) = (Input << 24) >> 24 = (Input & 0xff) >> 0 - // (8, 8) = (Input << 16) >> 24 = (Input & 0xffff) >> 8 - // (16, 8) = (Input << 8) >> 24 = (Input & 0xffffff) >> 16 - // (24, 8) = (Input << 0) >> 24 = (Input & 0xffffffff) >> 24 - def BFE_UINT_eg : R600_3OP <0x4, "BFE_UINT", - [(set i32:$dst, (AMDGPUbfe_u32 i32:$src0, i32:$src1, i32:$src2))], - VecALU - >; - - def BFE_INT_eg : R600_3OP <0x4, "BFE_INT", - [(set i32:$dst, (AMDGPUbfe_i32 i32:$src0, i32:$src1, i32:$src2))], - VecALU - >; - -// XXX: This pattern is broken, disabling for now. See comment in -// AMDGPUInstructions.td for more info. -// def : BFEPattern ; - - def BFI_INT_eg : R600_3OP <0x06, "BFI_INT", [], VecALU>; - defm : BFIPatterns ; - - def MULADD_UINT24_eg : R600_3OP <0x10, "MULADD_UINT24", - [(set i32:$dst, (add (mul U24:$src0, U24:$src1), i32:$src2))], VecALU - >; - def BIT_ALIGN_INT_eg : R600_3OP <0xC, "BIT_ALIGN_INT", [], VecALU>; - def : ROTRPattern ; - - def MULADD_eg : MULADD_Common<0x14>; - def MULADD_IEEE_eg : MULADD_IEEE_Common<0x18>; - def ASHR_eg : ASHR_Common<0x15>; - def LSHR_eg : LSHR_Common<0x16>; - def LSHL_eg : LSHL_Common<0x17>; - def CNDE_eg : CNDE_Common<0x19>; - def CNDGT_eg : CNDGT_Common<0x1A>; - def CNDGE_eg : CNDGE_Common<0x1B>; - def MUL_LIT_eg : MUL_LIT_Common<0x1F>; - def LOG_CLAMPED_eg : LOG_CLAMPED_Common<0x82>; - def MUL_UINT24_eg : R600_2OP <0xB5, "MUL_UINT24", - [(set i32:$dst, (mul U24:$src0, U24:$src1))], VecALU - >; - def DOT4_eg : DOT4_Common<0xBE>; - defm CUBE_eg : CUBE_Common<0xC0>; - -let hasSideEffects = 1 in { - def MOVA_INT_eg : R600_1OP <0xCC, "MOVA_INT", [], VecALU>; -} - - def TGSI_LIT_Z_eg : TGSI_LIT_Z_Common; - - def FLT_TO_INT_eg : FLT_TO_INT_Common<0x50> { - let Pattern = []; - let Itinerary = AnyALU; - } - - def INT_TO_FLT_eg : INT_TO_FLT_Common<0x9B>; - - def FLT_TO_UINT_eg : FLT_TO_UINT_Common<0x9A> { - let Pattern = []; - } - - def UINT_TO_FLT_eg : UINT_TO_FLT_Common<0x9C>; - -def GROUP_BARRIER : InstR600 < - (outs), (ins), " GROUP_BARRIER", [(int_AMDGPU_barrier_local)], AnyALU>, - R600ALU_Word0, - R600ALU_Word1_OP2 <0x54> { - - let dst = 0; - let dst_rel = 0; - let src0 = 0; - let src0_rel = 0; - let src0_neg = 0; - let src0_abs = 0; - let src1 = 0; - let src1_rel = 0; - let src1_neg = 0; - let src1_abs = 0; - let write = 0; - let omod = 0; - let clamp = 0; - let last = 1; - let bank_swizzle = 0; - let pred_sel = 0; - let update_exec_mask = 0; - let update_pred = 0; - - let Inst{31-0} = Word0; - let Inst{63-32} = Word1; - - let ALUInst = 1; -} - -//===----------------------------------------------------------------------===// -// LDS Instructions -//===----------------------------------------------------------------------===// -class R600_LDS op, dag outs, dag ins, string asm, - list pattern = []> : - - InstR600 , - R600_ALU_LDS_Word0, - R600LDS_Word1 { - - bits<6> offset = 0; - let lds_op = op; - - let Word1{27} = offset{0}; - let Word1{12} = offset{1}; - let Word1{28} = offset{2}; - let Word1{31} = offset{3}; - let Word0{12} = offset{4}; - let Word0{25} = offset{5}; - - - let Inst{31-0} = Word0; - let Inst{63-32} = Word1; - - let ALUInst = 1; - let HasNativeOperands = 1; - let UseNamedOperandTable = 1; -} - -class R600_LDS_1A lds_op, string name, list pattern> : R600_LDS < - lds_op, - (outs R600_Reg32:$dst), - (ins R600_Reg32:$src0, REL:$src0_rel, SEL:$src0_sel, - LAST:$last, R600_Pred:$pred_sel, - BANK_SWIZZLE:$bank_swizzle), - " "#name#" $last OQAP, $src0$src0_rel $pred_sel", - pattern - > { - - let src1 = 0; - let src1_rel = 0; - let src2 = 0; - let src2_rel = 0; - - let usesCustomInserter = 1; - let LDS_1A = 1; - let DisableEncoding = "$dst"; -} - -class R600_LDS_1A1D lds_op, dag outs, string name, list pattern, - string dst =""> : - R600_LDS < - lds_op, outs, - (ins R600_Reg32:$src0, REL:$src0_rel, SEL:$src0_sel, - R600_Reg32:$src1, REL:$src1_rel, SEL:$src1_sel, - LAST:$last, R600_Pred:$pred_sel, - BANK_SWIZZLE:$bank_swizzle), - " "#name#" $last "#dst#"$src0$src0_rel, $src1$src1_rel, $pred_sel", - pattern - > { - - field string BaseOp; - - let src2 = 0; - let src2_rel = 0; - let LDS_1A1D = 1; -} - -class R600_LDS_1A1D_NORET lds_op, string name, list pattern> : - R600_LDS_1A1D { - let BaseOp = name; -} - -class R600_LDS_1A1D_RET lds_op, string name, list pattern> : - R600_LDS_1A1D { - - let BaseOp = name; - let usesCustomInserter = 1; - let DisableEncoding = "$dst"; -} - -class R600_LDS_1A2D lds_op, string name, list pattern> : - R600_LDS < - lds_op, - (outs), - (ins R600_Reg32:$src0, REL:$src0_rel, SEL:$src0_sel, - R600_Reg32:$src1, REL:$src1_rel, SEL:$src1_sel, - R600_Reg32:$src2, REL:$src2_rel, SEL:$src2_sel, - LAST:$last, R600_Pred:$pred_sel, BANK_SWIZZLE:$bank_swizzle), - " "#name# "$last $src0$src0_rel, $src1$src1_rel, $src2$src2_rel, $pred_sel", - pattern> { - let LDS_1A2D = 1; -} - -def LDS_ADD : R600_LDS_1A1D_NORET <0x0, "LDS_ADD", [] >; -def LDS_SUB : R600_LDS_1A1D_NORET <0x1, "LDS_SUB", [] >; -def LDS_WRITE : R600_LDS_1A1D_NORET <0xD, "LDS_WRITE", - [(local_store (i32 R600_Reg32:$src1), R600_Reg32:$src0)] ->; -def LDS_BYTE_WRITE : R600_LDS_1A1D_NORET<0x12, "LDS_BYTE_WRITE", - [(truncstorei8_local i32:$src1, i32:$src0)] ->; -def LDS_SHORT_WRITE : R600_LDS_1A1D_NORET<0x13, "LDS_SHORT_WRITE", - [(truncstorei16_local i32:$src1, i32:$src0)] ->; -def LDS_ADD_RET : R600_LDS_1A1D_RET <0x20, "LDS_ADD", - [(set i32:$dst, (atomic_load_add_local i32:$src0, i32:$src1))] ->; -def LDS_SUB_RET : R600_LDS_1A1D_RET <0x21, "LDS_SUB", - [(set i32:$dst, (atomic_load_sub_local i32:$src0, i32:$src1))] ->; -def LDS_READ_RET : R600_LDS_1A <0x32, "LDS_READ_RET", - [(set (i32 R600_Reg32:$dst), (local_load R600_Reg32:$src0))] ->; -def LDS_BYTE_READ_RET : R600_LDS_1A <0x36, "LDS_BYTE_READ_RET", - [(set i32:$dst, (sextloadi8_local i32:$src0))] ->; -def LDS_UBYTE_READ_RET : R600_LDS_1A <0x37, "LDS_UBYTE_READ_RET", - [(set i32:$dst, (az_extloadi8_local i32:$src0))] ->; -def LDS_SHORT_READ_RET : R600_LDS_1A <0x38, "LDS_SHORT_READ_RET", - [(set i32:$dst, (sextloadi16_local i32:$src0))] ->; -def LDS_USHORT_READ_RET : R600_LDS_1A <0x39, "LDS_USHORT_READ_RET", - [(set i32:$dst, (az_extloadi16_local i32:$src0))] ->; - - // TRUNC is used for the FLT_TO_INT instructions to work around a - // perceived problem where the rounding modes are applied differently - // depending on the instruction and the slot they are in. - // See: - // https://bugs.freedesktop.org/show_bug.cgi?id=50232 - // Mesa commit: a1a0974401c467cb86ef818f22df67c21774a38c - // - // XXX: Lowering SELECT_CC will sometimes generate fp_to_[su]int nodes, - // which do not need to be truncated since the fp values are 0.0f or 1.0f. - // We should look into handling these cases separately. - def : Pat<(fp_to_sint f32:$src0), (FLT_TO_INT_eg (TRUNC $src0))>; - - def : Pat<(fp_to_uint f32:$src0), (FLT_TO_UINT_eg (TRUNC $src0))>; - - // SHA-256 Patterns - def : SHA256MaPattern ; - - def : FROUNDPat ; - - def EG_ExportSwz : ExportSwzInst { - let Word1{19-16} = 0; // BURST_COUNT - let Word1{20} = 0; // VALID_PIXEL_MODE - let Word1{21} = eop; - let Word1{29-22} = inst; - let Word1{30} = 0; // MARK - let Word1{31} = 1; // BARRIER - } - defm : ExportPattern; - - def EG_ExportBuf : ExportBufInst { - let Word1{19-16} = 0; // BURST_COUNT - let Word1{20} = 0; // VALID_PIXEL_MODE - let Word1{21} = eop; - let Word1{29-22} = inst; - let Word1{30} = 0; // MARK - let Word1{31} = 1; // BARRIER - } - defm : SteamOutputExportPattern; - - def CF_TC_EG : CF_CLAUSE_EG<1, (ins i32imm:$ADDR, i32imm:$COUNT), - "TEX $COUNT @$ADDR"> { - let POP_COUNT = 0; - } - def CF_VC_EG : CF_CLAUSE_EG<2, (ins i32imm:$ADDR, i32imm:$COUNT), - "VTX $COUNT @$ADDR"> { - let POP_COUNT = 0; - } - def WHILE_LOOP_EG : CF_CLAUSE_EG<6, (ins i32imm:$ADDR), - "LOOP_START_DX10 @$ADDR"> { - let POP_COUNT = 0; - let COUNT = 0; - } - def END_LOOP_EG : CF_CLAUSE_EG<5, (ins i32imm:$ADDR), "END_LOOP @$ADDR"> { - let POP_COUNT = 0; - let COUNT = 0; - } - def LOOP_BREAK_EG : CF_CLAUSE_EG<9, (ins i32imm:$ADDR), - "LOOP_BREAK @$ADDR"> { - let POP_COUNT = 0; - let COUNT = 0; - } - def CF_CONTINUE_EG : CF_CLAUSE_EG<8, (ins i32imm:$ADDR), - "CONTINUE @$ADDR"> { - let POP_COUNT = 0; - let COUNT = 0; - } - def CF_JUMP_EG : CF_CLAUSE_EG<10, (ins i32imm:$ADDR, i32imm:$POP_COUNT), - "JUMP @$ADDR POP:$POP_COUNT"> { - let COUNT = 0; - } - def CF_PUSH_EG : CF_CLAUSE_EG<11, (ins i32imm:$ADDR, i32imm:$POP_COUNT), - "PUSH @$ADDR POP:$POP_COUNT"> { - let COUNT = 0; - } - def CF_ELSE_EG : CF_CLAUSE_EG<13, (ins i32imm:$ADDR, i32imm:$POP_COUNT), - "ELSE @$ADDR POP:$POP_COUNT"> { - let COUNT = 0; - } - def CF_CALL_FS_EG : CF_CLAUSE_EG<19, (ins), "CALL_FS"> { - let ADDR = 0; - let COUNT = 0; - let POP_COUNT = 0; - } - def POP_EG : CF_CLAUSE_EG<14, (ins i32imm:$ADDR, i32imm:$POP_COUNT), - "POP @$ADDR POP:$POP_COUNT"> { - let COUNT = 0; - } - def CF_END_EG : CF_CLAUSE_EG<0, (ins), "CF_END"> { - let COUNT = 0; - let POP_COUNT = 0; - let ADDR = 0; - let END_OF_PROGRAM = 1; - } - -} // End Predicates = [isEGorCayman] //===----------------------------------------------------------------------===// // Regist loads and stores - for indirect addressing @@ -1838,216 +1264,11 @@ def LDS_USHORT_READ_RET : R600_LDS_1A <0x39, "LDS_USHORT_READ_RET", defm R600_ : RegisterLoadStore ; -//===----------------------------------------------------------------------===// -// Cayman Instructions -//===----------------------------------------------------------------------===// - -let Predicates = [isCayman] in { - -def MULADD_INT24_cm : R600_3OP <0x08, "MULADD_INT24", - [(set i32:$dst, (add (mul I24:$src0, I24:$src1), i32:$src2))], VecALU ->; -def MUL_INT24_cm : R600_2OP <0x5B, "MUL_INT24", - [(set i32:$dst, (mul I24:$src0, I24:$src1))], VecALU ->; - -let isVector = 1 in { - -def RECIP_IEEE_cm : RECIP_IEEE_Common<0x86>; - -def MULLO_INT_cm : MULLO_INT_Common<0x8F>; -def MULHI_INT_cm : MULHI_INT_Common<0x90>; -def MULLO_UINT_cm : MULLO_UINT_Common<0x91>; -def MULHI_UINT_cm : MULHI_UINT_Common<0x92>; -def RECIPSQRT_CLAMPED_cm : RECIPSQRT_CLAMPED_Common<0x87>; -def EXP_IEEE_cm : EXP_IEEE_Common<0x81>; -def LOG_IEEE_cm : LOG_IEEE_Common<0x83>; -def RECIP_CLAMPED_cm : RECIP_CLAMPED_Common<0x84>; -def RECIPSQRT_IEEE_cm : RECIPSQRT_IEEE_Common<0x89>; -def SIN_cm : SIN_Common<0x8D>; -def COS_cm : COS_Common<0x8E>; -} // End isVector = 1 - -def : POW_Common ; - -defm DIV_cm : DIV_Common; - -// RECIP_UINT emulation for Cayman -// The multiplication scales from [0,1] to the unsigned integer range -def : Pat < - (AMDGPUurecip i32:$src0), - (FLT_TO_UINT_eg (MUL_IEEE (RECIP_IEEE_cm (UINT_TO_FLT_eg $src0)), - (MOV_IMM_I32 CONST.FP_UINT_MAX_PLUS_1))) ->; - - def CF_END_CM : CF_CLAUSE_EG<32, (ins), "CF_END"> { - let ADDR = 0; - let POP_COUNT = 0; - let COUNT = 0; - } - - -def : Pat<(fsqrt f32:$src), (MUL R600_Reg32:$src, (RECIPSQRT_CLAMPED_cm $src))>; - -class RAT_STORE_DWORD mask> : - CF_MEM_RAT_CACHELESS <0x14, 0, mask, - (ins rc:$rw_gpr, R600_TReg32_X:$index_gpr), - "STORE_DWORD $rw_gpr, $index_gpr", - [(global_store vt:$rw_gpr, i32:$index_gpr)]> { - let eop = 0; // This bit is not used on Cayman. -} - -def RAT_STORE_DWORD32 : RAT_STORE_DWORD ; -def RAT_STORE_DWORD64 : RAT_STORE_DWORD ; -def RAT_STORE_DWORD128 : RAT_STORE_DWORD ; - -class VTX_READ_cm buffer_id, dag outs, list pattern> - : VTX_WORD0_cm, VTX_READ { - - // Static fields - let VC_INST = 0; - let FETCH_TYPE = 2; - let FETCH_WHOLE_QUAD = 0; - let BUFFER_ID = buffer_id; - let SRC_REL = 0; - // XXX: We can infer this field based on the SRC_GPR. This would allow us - // to store vertex addresses in any channel, not just X. - let SRC_SEL_X = 0; - let SRC_SEL_Y = 0; - let STRUCTURED_READ = 0; - let LDS_REQ = 0; - let COALESCED_READ = 0; - - let Inst{31-0} = Word0; -} - -class VTX_READ_8_cm buffer_id, list pattern> - : VTX_READ_cm <"VTX_READ_8 $dst_gpr, $src_gpr", buffer_id, - (outs R600_TReg32_X:$dst_gpr), pattern> { - - let DST_SEL_X = 0; - let DST_SEL_Y = 7; // Masked - let DST_SEL_Z = 7; // Masked - let DST_SEL_W = 7; // Masked - let DATA_FORMAT = 1; // FMT_8 -} - -class VTX_READ_16_cm buffer_id, list pattern> - : VTX_READ_cm <"VTX_READ_16 $dst_gpr, $src_gpr", buffer_id, - (outs R600_TReg32_X:$dst_gpr), pattern> { - let DST_SEL_X = 0; - let DST_SEL_Y = 7; // Masked - let DST_SEL_Z = 7; // Masked - let DST_SEL_W = 7; // Masked - let DATA_FORMAT = 5; // FMT_16 - -} - -class VTX_READ_32_cm buffer_id, list pattern> - : VTX_READ_cm <"VTX_READ_32 $dst_gpr, $src_gpr", buffer_id, - (outs R600_TReg32_X:$dst_gpr), pattern> { - - let DST_SEL_X = 0; - let DST_SEL_Y = 7; // Masked - let DST_SEL_Z = 7; // Masked - let DST_SEL_W = 7; // Masked - let DATA_FORMAT = 0xD; // COLOR_32 - - // This is not really necessary, but there were some GPU hangs that appeared - // to be caused by ALU instructions in the next instruction group that wrote - // to the $src_gpr registers of the VTX_READ. - // e.g. - // %T3_X = VTX_READ_PARAM_32_eg %T2_X, 24 - // %T2_X = MOV %ZERO - //Adding this constraint prevents this from happening. - let Constraints = "$src_gpr.ptr = $dst_gpr"; -} - -class VTX_READ_64_cm buffer_id, list pattern> - : VTX_READ_cm <"VTX_READ_64 $dst_gpr, $src_gpr", buffer_id, - (outs R600_Reg64:$dst_gpr), pattern> { - - let DST_SEL_X = 0; - let DST_SEL_Y = 1; - let DST_SEL_Z = 7; - let DST_SEL_W = 7; - let DATA_FORMAT = 0x1D; // COLOR_32_32 -} - -class VTX_READ_128_cm buffer_id, list pattern> - : VTX_READ_cm <"VTX_READ_128 $dst_gpr.XYZW, $src_gpr", buffer_id, - (outs R600_Reg128:$dst_gpr), pattern> { - - let DST_SEL_X = 0; - let DST_SEL_Y = 1; - let DST_SEL_Z = 2; - let DST_SEL_W = 3; - let DATA_FORMAT = 0x22; // COLOR_32_32_32_32 - - // XXX: Need to force VTX_READ_128 instructions to write to the same register - // that holds its buffer address to avoid potential hangs. We can't use - // the same constraint as VTX_READ_32_eg, because the $src_gpr.ptr and $dst - // registers are different sizes. -} - -//===----------------------------------------------------------------------===// -// VTX Read from parameter memory space -//===----------------------------------------------------------------------===// -def VTX_READ_PARAM_8_cm : VTX_READ_8_cm <0, - [(set i32:$dst_gpr, (load_param_exti8 ADDRVTX_READ:$src_gpr))] ->; - -def VTX_READ_PARAM_16_cm : VTX_READ_16_cm <0, - [(set i32:$dst_gpr, (load_param_exti16 ADDRVTX_READ:$src_gpr))] ->; - -def VTX_READ_PARAM_32_cm : VTX_READ_32_cm <0, - [(set i32:$dst_gpr, (load_param ADDRVTX_READ:$src_gpr))] ->; - -def VTX_READ_PARAM_64_cm : VTX_READ_64_cm <0, - [(set v2i32:$dst_gpr, (load_param ADDRVTX_READ:$src_gpr))] ->; - -def VTX_READ_PARAM_128_cm : VTX_READ_128_cm <0, - [(set v4i32:$dst_gpr, (load_param ADDRVTX_READ:$src_gpr))] ->; - -//===----------------------------------------------------------------------===// -// VTX Read from global memory space -//===----------------------------------------------------------------------===// - -// 8-bit reads -def VTX_READ_GLOBAL_8_cm : VTX_READ_8_cm <1, - [(set i32:$dst_gpr, (az_extloadi8_global ADDRVTX_READ:$src_gpr))] ->; - -def VTX_READ_GLOBAL_16_cm : VTX_READ_16_cm <1, - [(set i32:$dst_gpr, (az_extloadi16_global ADDRVTX_READ:$src_gpr))] ->; - -// 32-bit reads -def VTX_READ_GLOBAL_32_cm : VTX_READ_32_cm <1, - [(set i32:$dst_gpr, (global_load ADDRVTX_READ:$src_gpr))] ->; - -// 64-bit reads -def VTX_READ_GLOBAL_64_cm : VTX_READ_64_cm <1, - [(set v2i32:$dst_gpr, (global_load ADDRVTX_READ:$src_gpr))] ->; - -// 128-bit reads -def VTX_READ_GLOBAL_128_cm : VTX_READ_128_cm <1, - [(set v4i32:$dst_gpr, (global_load ADDRVTX_READ:$src_gpr))] ->; - -} // End isCayman //===----------------------------------------------------------------------===// // Branch Instructions //===----------------------------------------------------------------------===// - def IF_PREDICATE_SET : ILFormat<(outs), (ins GPRI32:$src), "IF_PREDICATE_SET $src", []>; diff --git a/lib/Target/R600/R700Instructions.td b/lib/Target/R600/R700Instructions.td new file mode 100644 index 00000000000..9aad85da4e9 --- /dev/null +++ b/lib/Target/R600/R700Instructions.td @@ -0,0 +1,21 @@ +//===-- R700Instructions.td - R700 Instruction defs -------*- tablegen -*-===// +// +// The LLVM Compiler Infrastructure +// +// This file is distributed under the University of Illinois Open Source +// License. See LICENSE.TXT for details. +// +//===----------------------------------------------------------------------===// +// +// TableGen definitions for instructions which are: +// - Available to R700 and newer VLIW4/VLIW5 GPUs +// - Available only on R700 family GPUs. +// +//===----------------------------------------------------------------------===// + +def isR700 : Predicate<"Subtarget.getGeneration() == AMDGPUSubtarget::R700">; + +let Predicates = [isR700] in { + def SIN_r700 : SIN_Common<0x6E>; + def COS_r700 : COS_Common<0x6F>; +}