summaryrefslogtreecommitdiff
path: root/compiler_and_linker/BackEnd/PowerPC/Scheduler/MachineSimulation750.c
diff options
context:
space:
mode:
authorAsh Wolf <ninji@wuffs.org>2023-01-26 11:30:47 +0000
committerAsh Wolf <ninji@wuffs.org>2023-01-26 11:30:47 +0000
commit094b96ca1df4a035b5f93c351f773306c0241f3f (patch)
tree95ce05e3ebe816c7ee7996206bb37ea17d8ca33c /compiler_and_linker/BackEnd/PowerPC/Scheduler/MachineSimulation750.c
parentfc0c4c0df7b583b55a08317cf1ef6a71d27c0440 (diff)
downloadMWCC-main.tar.gz
MWCC-main.zip
move lots of source files around to match their actual placement in the original treemain
Diffstat (limited to 'compiler_and_linker/BackEnd/PowerPC/Scheduler/MachineSimulation750.c')
-rw-r--r--compiler_and_linker/BackEnd/PowerPC/Scheduler/MachineSimulation750.c678
1 files changed, 678 insertions, 0 deletions
diff --git a/compiler_and_linker/BackEnd/PowerPC/Scheduler/MachineSimulation750.c b/compiler_and_linker/BackEnd/PowerPC/Scheduler/MachineSimulation750.c
new file mode 100644
index 0000000..d412df3
--- /dev/null
+++ b/compiler_and_linker/BackEnd/PowerPC/Scheduler/MachineSimulation750.c
@@ -0,0 +1,678 @@
+#include "compiler/Scheduler.h"
+#include "compiler/PCode.h"
+#include "compiler/PCodeInfo.h"
+
+// https://www.nxp.com/docs/en/reference-manual/MPC750UM.pdf
+
+typedef enum Stage {
+ BPU, // Branch Prediction Unit
+ IU1, // Integer Unit 1
+ IU2, // Integer Unit 2
+
+ LSU1, // Load/Store Unit
+ LSU2,
+
+ FPU1, // Floating Point Unit
+ FPU2,
+ FPU3,
+
+ SRU, // System Register Unit
+
+ NumStages
+} Stage;
+
+static struct {
+ // the instruction currently in this pipeline stage
+ PCode *instr;
+
+ // how many cycles are left for this instruction to finish
+ int remaining;
+} pipeline[NumStages];
+
+static PCode *iu1_completed_instruction;
+static PCode *iu2_completed_instruction;
+
+enum {
+ MaxEntries = 6
+};
+
+static struct {
+ // how many entries remain unused in the queue
+ unsigned int free;
+
+ // how many entries are currently used in the queue
+ unsigned int used;
+
+ // the index of the next instruction that will be retired
+ unsigned int nextToRetire;
+
+ // the index of the next free slot that will be used when an instruction is dispatched
+ unsigned int nextFreeSlot;
+
+ // circular array of entries in the completion queue
+ struct {
+ PCode *instr;
+ int completed;
+ } entries[MaxEntries];
+} completionbuffers;
+
+static struct {
+ // the initial stage for this instruction
+ Stage stage;
+
+ // the total amount of cycles required by this instruction
+ char latency;
+
+ // how long it takes to finish each stage
+ char cycles[3];
+
+ // does this instruction serialise?
+ char serializes;
+} instruction_timing[OPCODE_MAX] = {
+ BPU, 0, 0, 0, 0, 1, // PC_B
+ BPU, 0, 0, 0, 0, 1, // PC_BL
+ BPU, 0, 0, 0, 0, 1, // PC_BC
+ BPU, 0, 0, 0, 0, 1, // PC_BCLR
+ BPU, 0, 0, 0, 0, 1, // PC_BCCTR
+ BPU, 0, 0, 0, 0, 1, // PC_BT
+ BPU, 0, 0, 0, 0, 1, // PC_BTLR
+ BPU, 0, 0, 0, 0, 1, // PC_BTCTR
+ BPU, 0, 0, 0, 0, 1, // PC_BF
+ BPU, 0, 0, 0, 0, 1, // PC_BFLR
+ BPU, 0, 0, 0, 0, 1, // PC_BFCTR
+ BPU, 0, 0, 0, 0, 1, // PC_BDNZ
+ BPU, 0, 0, 0, 0, 1, // PC_BDNZT
+ BPU, 0, 0, 0, 0, 1, // PC_BDNZF
+ BPU, 0, 0, 0, 0, 1, // PC_BDZ
+ BPU, 0, 0, 0, 0, 1, // PC_BDZT
+ BPU, 0, 0, 0, 0, 1, // PC_BDZF
+ BPU, 0, 0, 0, 0, 1, // PC_BLR
+ BPU, 0, 0, 0, 0, 1, // PC_BCTR
+ BPU, 0, 0, 0, 0, 1, // PC_BCTRL
+ BPU, 0, 0, 0, 0, 1, // PC_BLRL
+ LSU1, 2, 1, 1, 0, 0, // PC_LBZ
+ LSU1, 2, 1, 1, 0, 0, // PC_LBZU
+ LSU1, 2, 1, 1, 0, 0, // PC_LBZX
+ LSU1, 2, 1, 1, 0, 0, // PC_LBZUX
+ LSU1, 2, 1, 1, 0, 0, // PC_LHZ
+ LSU1, 2, 1, 1, 0, 0, // PC_LHZU
+ LSU1, 2, 1, 1, 0, 0, // PC_LHZX
+ LSU1, 2, 1, 1, 0, 0, // PC_LHZUX
+ LSU1, 2, 1, 1, 0, 0, // PC_LHA
+ LSU1, 2, 1, 1, 0, 0, // PC_LHAU
+ LSU1, 2, 1, 1, 0, 0, // PC_LHAX
+ LSU1, 2, 1, 1, 0, 0, // PC_LHAUX
+ LSU1, 2, 1, 1, 0, 0, // PC_LHBRX
+ LSU1, 2, 1, 1, 0, 0, // PC_LWZ
+ LSU1, 2, 1, 1, 0, 0, // PC_LWZU
+ LSU1, 2, 1, 1, 0, 0, // PC_LWZX
+ LSU1, 2, 1, 1, 0, 0, // PC_LWZUX
+ LSU1, 2, 1, 1, 0, 0, // PC_LWBRX
+ LSU1, 2, 1, 1, 0, 0, // PC_LMW
+ LSU1, 2, 1, 1, 0, 0, // PC_STB
+ LSU1, 2, 1, 1, 0, 0, // PC_STBU
+ LSU1, 2, 1, 1, 0, 0, // PC_STBX
+ LSU1, 2, 1, 1, 0, 0, // PC_STBUX
+ LSU1, 2, 1, 1, 0, 0, // PC_STH
+ LSU1, 2, 1, 1, 0, 0, // PC_STHU
+ LSU1, 2, 1, 1, 0, 0, // PC_STHX
+ LSU1, 2, 1, 1, 0, 0, // PC_STHUX
+ LSU1, 2, 1, 1, 0, 0, // PC_STHBRX
+ LSU1, 2, 1, 1, 0, 0, // PC_STW
+ LSU1, 2, 1, 1, 0, 0, // PC_STWU
+ LSU1, 2, 1, 1, 0, 0, // PC_STWX
+ LSU1, 2, 1, 1, 0, 0, // PC_STWUX
+ LSU1, 2, 1, 1, 0, 0, // PC_STWBRX
+ LSU1, 2, 1, 1, 0, 0, // PC_STMW
+ LSU1, 3, 1, 2, 0, 0, // PC_DCBF
+ LSU1, 3, 1, 2, 0, 0, // PC_DCBST
+ LSU1, 2, 1, 1, 0, 0, // PC_DCBT
+ LSU1, 2, 1, 1, 0, 0, // PC_DCBTST
+ LSU1, 3, 1, 2, 0, 0, // PC_DCBZ
+ IU2, 1, 1, 0, 0, 0, // PC_ADD
+ IU2, 1, 1, 0, 0, 0, // PC_ADDC
+ IU2, 1, 1, 0, 0, 0, // PC_ADDE
+ IU2, 1, 1, 0, 0, 0, // PC_ADDI
+ IU2, 1, 1, 0, 0, 0, // PC_ADDIC
+ IU2, 1, 1, 0, 0, 0, // PC_ADDICR
+ IU2, 1, 1, 0, 0, 0, // PC_ADDIS
+ IU2, 1, 1, 0, 0, 0, // PC_ADDME
+ IU2, 1, 1, 0, 0, 0, // PC_ADDZE
+ IU1, 19, 19, 0, 0, 0, // PC_DIVW
+ IU1, 19, 19, 0, 0, 0, // PC_DIVWU
+ IU1, 5, 5, 0, 0, 0, // PC_MULHW
+ IU1, 6, 5, 0, 0, 0, // PC_MULHWU
+ IU1, 3, 3, 0, 0, 0, // PC_MULLI
+ IU1, 5, 5, 0, 0, 0, // PC_MULLW
+ IU2, 1, 1, 0, 0, 0, // PC_NEG
+ IU2, 1, 1, 0, 0, 0, // PC_SUBF
+ IU2, 1, 1, 0, 0, 0, // PC_SUBFC
+ IU2, 1, 1, 0, 0, 0, // PC_SUBFE
+ IU2, 1, 1, 0, 0, 0, // PC_SUBFIC
+ IU2, 1, 1, 0, 0, 0, // PC_SUBFME
+ IU2, 1, 1, 0, 0, 0, // PC_SUBFZE
+ IU2, 3, 1, 0, 0, 0, // PC_CMPI
+ IU2, 3, 1, 0, 0, 0, // PC_CMP
+ IU2, 3, 1, 0, 0, 0, // PC_CMPLI
+ IU2, 3, 1, 0, 0, 0, // PC_CMPL
+ IU2, 1, 1, 0, 0, 0, // PC_ANDI
+ IU2, 1, 1, 0, 0, 0, // PC_ANDIS
+ IU2, 1, 1, 0, 0, 0, // PC_ORI
+ IU2, 1, 1, 0, 0, 0, // PC_ORIS
+ IU2, 1, 1, 0, 0, 0, // PC_XORI
+ IU2, 1, 1, 0, 0, 0, // PC_XORIS
+ IU2, 1, 1, 0, 0, 0, // PC_AND
+ IU2, 1, 1, 0, 0, 0, // PC_OR
+ IU2, 1, 1, 0, 0, 0, // PC_XOR
+ IU2, 1, 1, 0, 0, 0, // PC_NAND
+ IU2, 1, 1, 0, 0, 0, // PC_NOR
+ IU2, 1, 1, 0, 0, 0, // PC_EQV
+ IU2, 1, 1, 0, 0, 0, // PC_ANDC
+ IU2, 1, 1, 0, 0, 0, // PC_ORC
+ IU2, 1, 1, 0, 0, 0, // PC_EXTSB
+ IU2, 1, 1, 0, 0, 0, // PC_EXTSH
+ IU2, 1, 1, 0, 0, 0, // PC_CNTLZW
+ IU2, 1, 1, 0, 0, 0, // PC_RLWINM
+ IU2, 1, 1, 0, 0, 0, // PC_RLWNM
+ IU2, 1, 1, 0, 0, 0, // PC_RLWIMI
+ IU2, 1, 1, 0, 0, 0, // PC_SLW
+ IU2, 1, 1, 0, 0, 0, // PC_SRW
+ IU2, 1, 1, 0, 0, 0, // PC_SRAWI
+ IU2, 1, 1, 0, 0, 0, // PC_SRAW
+ SRU, 1, 1, 0, 0, 1, // PC_CRAND
+ SRU, 1, 1, 0, 0, 1, // PC_CRANDC
+ SRU, 1, 1, 0, 0, 1, // PC_CREQV
+ SRU, 1, 1, 0, 0, 1, // PC_CRNAND
+ SRU, 1, 1, 0, 0, 1, // PC_CRNOR
+ SRU, 1, 1, 0, 0, 1, // PC_CROR
+ SRU, 1, 1, 0, 0, 1, // PC_CRORC
+ SRU, 1, 1, 0, 0, 1, // PC_CRXOR
+ SRU, 1, 1, 0, 0, 1, // PC_MCRF
+ SRU, 2, 2, 0, 0, 1, // PC_MTXER
+ SRU, 2, 2, 0, 0, 1, // PC_MTCTR
+ SRU, 2, 2, 0, 0, 1, // PC_MTLR
+ SRU, 1, 1, 0, 0, 1, // PC_MTCRF
+ SRU, 1, 1, 0, 0, 0, // PC_MTMSR
+ SRU, 1, 1, 0, 0, 1, // PC_MTSPR
+ SRU, 1, 1, 0, 0, 1, // PC_MFMSR
+ SRU, 1, 1, 0, 0, 1, // PC_MFSPR
+ SRU, 1, 1, 0, 0, 1, // PC_MFXER
+ SRU, 1, 1, 0, 0, 1, // PC_MFCTR
+ SRU, 1, 1, 0, 0, 1, // PC_MFLR
+ SRU, 1, 1, 0, 0, 1, // PC_MFCR
+ FPU1, 3, 1, 1, 1, 0, // PC_MFFS
+ FPU1, 3, 1, 1, 1, 0, // PC_MTFSF
+ SRU, 1, 1, 0, 0, 1, // PC_EIEIO
+ SRU, 2, 2, 0, 0, 1, // PC_ISYNC
+ SRU, 3, 3, 0, 0, 1, // PC_SYNC
+ SRU, 1, 1, 0, 0, 1, // PC_RFI
+ IU2, 1, 1, 0, 0, 0, // PC_LI
+ IU2, 1, 1, 0, 0, 0, // PC_LIS
+ IU2, 1, 1, 0, 0, 0, // PC_MR
+ IU2, 1, 1, 0, 0, 0, // PC_NOP
+ IU2, 1, 1, 0, 0, 0, // PC_NOT
+ LSU1, 2, 1, 1, 0, 0, // PC_LFS
+ LSU1, 2, 1, 1, 0, 0, // PC_LFSU
+ LSU1, 2, 1, 1, 0, 0, // PC_LFSX
+ LSU1, 2, 1, 1, 0, 0, // PC_LFSUX
+ LSU1, 2, 1, 1, 0, 0, // PC_LFD
+ LSU1, 2, 1, 1, 0, 0, // PC_LFDU
+ LSU1, 2, 1, 1, 0, 0, // PC_LFDX
+ LSU1, 2, 1, 1, 0, 0, // PC_LFDUX
+ LSU1, 2, 1, 1, 0, 0, // PC_STFS
+ LSU1, 2, 1, 1, 0, 0, // PC_STFSU
+ LSU1, 2, 1, 1, 0, 0, // PC_STFSX
+ LSU1, 2, 1, 1, 0, 0, // PC_STFSUX
+ LSU1, 2, 1, 1, 0, 0, // PC_STFD
+ LSU1, 2, 1, 1, 0, 0, // PC_STFDU
+ LSU1, 2, 1, 1, 0, 0, // PC_STFDX
+ LSU1, 2, 1, 1, 0, 0, // PC_STFDUX
+ FPU1, 3, 1, 1, 1, 0, // PC_FMR
+ FPU1, 3, 1, 1, 1, 0, // PC_FABS
+ FPU1, 3, 1, 1, 1, 0, // PC_FNEG
+ FPU1, 3, 1, 1, 1, 0, // PC_FNABS
+ FPU1, 3, 1, 1, 1, 0, // PC_FADD
+ FPU1, 3, 1, 1, 1, 0, // PC_FADDS
+ FPU1, 3, 1, 1, 1, 0, // PC_FSUB
+ FPU1, 3, 1, 1, 1, 0, // PC_FSUBS
+ FPU1, 4, 2, 1, 1, 0, // PC_FMUL
+ FPU1, 3, 1, 1, 1, 0, // PC_FMULS
+ FPU1, 31, 31, 0, 0, 0, // PC_FDIV
+ FPU1, 17, 17, 0, 0, 0, // PC_FDIVS
+ FPU1, 4, 2, 1, 1, 0, // PC_FMADD
+ FPU1, 3, 1, 1, 1, 0, // PC_FMADDS
+ FPU1, 4, 2, 1, 1, 0, // PC_FMSUB
+ FPU1, 3, 1, 1, 1, 0, // PC_FMSUBS
+ FPU1, 4, 2, 1, 1, 0, // PC_FNMADD
+ FPU1, 3, 1, 1, 1, 0, // PC_FNMADDS
+ FPU1, 4, 2, 1, 1, 0, // PC_FNMSUB
+ FPU1, 3, 1, 1, 1, 0, // PC_FNMSUBS
+ FPU1, 10, 10, 0, 0, 0, // PC_FRES
+ FPU1, 3, 1, 1, 1, 0, // PC_FRSQRTE
+ FPU1, 3, 1, 1, 1, 0, // PC_FSEL
+ FPU1, 3, 1, 1, 1, 0, // PC_FRSP
+ FPU1, 3, 1, 1, 1, 0, // PC_FCTIW
+ FPU1, 3, 1, 1, 1, 0, // PC_FCTIWZ
+ FPU1, 3, 1, 1, 1, 0, // PC_FCMPU
+ FPU1, 3, 1, 1, 1, 0, // PC_FCMPO
+ LSU1, 1, 1, 0, 0, 0, // PC_LWARX
+ LSU1, 1, 1, 0, 0, 0, // PC_LSWI
+ LSU1, 1, 1, 0, 0, 0, // PC_LSWX
+ LSU1, 1, 1, 0, 0, 0, // PC_STFIWX
+ LSU1, 1, 1, 0, 0, 0, // PC_STSWI
+ LSU1, 1, 1, 0, 0, 0, // PC_STSWX
+ LSU1, 1, 1, 0, 0, 0, // PC_STWCX
+ IU1, 1, 1, 0, 0, 1, // PC_ECIWX
+ IU1, 1, 1, 0, 0, 1, // PC_ECOWX
+ IU1, 1, 1, 0, 0, 0, // PC_DCBI
+ IU1, 1, 1, 0, 0, 0, // PC_ICBI
+ IU1, 1, 1, 0, 0, 0, // PC_MCRFS
+ IU1, 1, 1, 0, 0, 0, // PC_MCRXR
+ IU1, 1, 1, 0, 0, 0, // PC_MFTB
+ IU1, 1, 1, 0, 0, 0, // PC_MFSR
+ IU1, 1, 1, 0, 0, 0, // PC_MTSR
+ IU1, 1, 1, 0, 0, 0, // PC_MFSRIN
+ IU1, 1, 1, 0, 0, 0, // PC_MTSRIN
+ IU1, 1, 1, 0, 0, 0, // PC_MTFSB0
+ IU1, 1, 1, 0, 0, 0, // PC_MTFSB1
+ IU1, 1, 1, 0, 0, 0, // PC_MTFSFI
+ IU1, 1, 1, 0, 0, 1, // PC_SC
+ FPU1, 1, 1, 0, 0, 0, // PC_FSQRT
+ FPU1, 1, 1, 0, 0, 0, // PC_FSQRTS
+ IU1, 1, 1, 0, 0, 0, // PC_TLBIA
+ IU1, 1, 1, 0, 0, 0, // PC_TLBIE
+ IU1, 1, 1, 0, 0, 0, // PC_TLBLD
+ IU1, 1, 1, 0, 0, 0, // PC_TLBLI
+ IU1, 1, 1, 0, 0, 0, // PC_TLBSYNC
+ IU1, 1, 1, 0, 0, 1, // PC_TW
+ IU1, 1, 1, 0, 0, 1, // PC_TRAP
+ IU1, 1, 1, 0, 0, 1, // PC_TWI
+ IU1, 1, 1, 0, 0, 1, // PC_OPWORD
+ IU1, 1, 1, 0, 0, 0, // PC_MFROM
+ IU1, 1, 1, 0, 0, 1, // PC_DSA
+ IU1, 1, 1, 0, 0, 1, // PC_ESA
+ IU1, 0, 0, 0, 0, 0, // PC_DCCCI
+ IU1, 0, 0, 0, 0, 0, // PC_DCREAD
+ IU1, 0, 0, 0, 0, 0, // PC_ICBT
+ IU1, 0, 0, 0, 0, 0, // PC_ICCCI
+ IU1, 0, 0, 0, 0, 0, // PC_ICREAD
+ IU1, 0, 0, 0, 0, 0, // PC_RFCI
+ IU1, 0, 0, 0, 0, 0, // PC_TLBRE
+ IU1, 0, 0, 0, 0, 0, // PC_TLBSX
+ IU1, 0, 0, 0, 0, 0, // PC_TLBWE
+ IU1, 0, 0, 0, 0, 0, // PC_WRTEE
+ IU1, 0, 0, 0, 0, 0, // PC_WRTEEI
+ IU1, 0, 0, 0, 0, 0, // PC_MFDCR
+ IU1, 0, 0, 0, 0, 0, // PC_MTDCR
+ IU1, 0, 0, 0, 0, 0, // PC_DCBA
+ BPU, 0, 0, 0, 0, 0, // PC_DSS
+ BPU, 0, 0, 0, 0, 0, // PC_DSSALL
+ BPU, 0, 0, 0, 0, 0, // PC_DST
+ BPU, 0, 0, 0, 0, 0, // PC_DSTT
+ BPU, 0, 0, 0, 0, 0, // PC_DSTST
+ BPU, 0, 0, 0, 0, 0, // PC_DSTSTT
+ BPU, 0, 0, 0, 0, 0, // PC_LVEBX
+ BPU, 0, 0, 0, 0, 0, // PC_LVEHX
+ BPU, 0, 0, 0, 0, 0, // PC_LVEWX
+ BPU, 0, 0, 0, 0, 0, // PC_LVSL
+ BPU, 0, 0, 0, 0, 0, // PC_LVSR
+ BPU, 0, 0, 0, 0, 0, // PC_LVX
+ BPU, 0, 0, 0, 0, 0, // PC_LVXL
+ BPU, 0, 0, 0, 0, 0, // PC_STVEBX
+ BPU, 0, 0, 0, 0, 0, // PC_STVEHX
+ BPU, 0, 0, 0, 0, 0, // PC_STVEWX
+ BPU, 0, 0, 0, 0, 0, // PC_STVX
+ BPU, 0, 0, 0, 0, 0, // PC_STVXL
+ BPU, 0, 0, 0, 0, 0, // PC_MFVSCR
+ BPU, 0, 0, 0, 0, 0, // PC_MTVSCR
+ BPU, 0, 0, 0, 0, 0, // PC_VADDCUW
+ BPU, 0, 0, 0, 0, 0, // PC_VADDFP
+ BPU, 0, 0, 0, 0, 0, // PC_VADDSBS
+ BPU, 0, 0, 0, 0, 0, // PC_VADDSHS
+ BPU, 0, 0, 0, 0, 0, // PC_VADDSWS
+ BPU, 0, 0, 0, 0, 0, // PC_VADDUBM
+ BPU, 0, 0, 0, 0, 0, // PC_VADDUBS
+ BPU, 0, 0, 0, 0, 0, // PC_VADDUHM
+ BPU, 0, 0, 0, 0, 0, // PC_VADDUHS
+ BPU, 0, 0, 0, 0, 0, // PC_VADDUWM
+ BPU, 0, 0, 0, 0, 0, // PC_VADDUWS
+ BPU, 0, 0, 0, 0, 0, // PC_VAND
+ BPU, 0, 0, 0, 0, 0, // PC_VANDC
+ BPU, 0, 0, 0, 0, 0, // PC_VAVGSB
+ BPU, 0, 0, 0, 0, 0, // PC_VAVGSH
+ BPU, 0, 0, 0, 0, 0, // PC_VAVGSW
+ BPU, 0, 0, 0, 0, 0, // PC_VAVGUB
+ BPU, 0, 0, 0, 0, 0, // PC_VAVGUH
+ BPU, 0, 0, 0, 0, 0, // PC_VAVGUW
+ BPU, 0, 0, 0, 0, 0, // PC_VCFSX
+ BPU, 0, 0, 0, 0, 0, // PC_VCFUX
+ BPU, 0, 0, 0, 0, 0, // PC_VCMPBFP
+ BPU, 0, 0, 0, 0, 0, // PC_VCMPEQFP
+ BPU, 0, 0, 0, 0, 0, // PC_VCMPEQUB
+ BPU, 0, 0, 0, 0, 0, // PC_VCMPEQUH
+ BPU, 0, 0, 0, 0, 0, // PC_VCMPEQUW
+ BPU, 0, 0, 0, 0, 0, // PC_VCMPGEFP
+ BPU, 0, 0, 0, 0, 0, // PC_VCMPGTFP
+ BPU, 0, 0, 0, 0, 0, // PC_VCMPGTSB
+ BPU, 0, 0, 0, 0, 0, // PC_VCMPGTSH
+ BPU, 0, 0, 0, 0, 0, // PC_VCMPGTSW
+ BPU, 0, 0, 0, 0, 0, // PC_VCMPGTUB
+ BPU, 0, 0, 0, 0, 0, // PC_VCMPGTUH
+ BPU, 0, 0, 0, 0, 0, // PC_VCMPGTUW
+ BPU, 0, 0, 0, 0, 0, // PC_VCTSXS
+ BPU, 0, 0, 0, 0, 0, // PC_VCTUXS
+ BPU, 0, 0, 0, 0, 0, // PC_VEXPTEFP
+ BPU, 0, 0, 0, 0, 0, // PC_VLOGEFP
+ BPU, 0, 0, 0, 0, 0, // PC_VMAXFP
+ BPU, 0, 0, 0, 0, 0, // PC_VMAXSB
+ BPU, 0, 0, 0, 0, 0, // PC_VMAXSH
+ BPU, 0, 0, 0, 0, 0, // PC_VMAXSW
+ BPU, 0, 0, 0, 0, 0, // PC_VMAXUB
+ BPU, 0, 0, 0, 0, 0, // PC_VMAXUH
+ BPU, 0, 0, 0, 0, 0, // PC_VMAXUW
+ BPU, 0, 0, 0, 0, 0, // PC_VMINFP
+ BPU, 0, 0, 0, 0, 0, // PC_VMINSB
+ BPU, 0, 0, 0, 0, 0, // PC_VMINSH
+ BPU, 0, 0, 0, 0, 0, // PC_VMINSW
+ BPU, 0, 0, 0, 0, 0, // PC_VMINUB
+ BPU, 0, 0, 0, 0, 0, // PC_VMINUH
+ BPU, 0, 0, 0, 0, 0, // PC_VMINUW
+ BPU, 0, 0, 0, 0, 0, // PC_VMRGHB
+ BPU, 0, 0, 0, 0, 0, // PC_VMRGHH
+ BPU, 0, 0, 0, 0, 0, // PC_VMRGHW
+ BPU, 0, 0, 0, 0, 0, // PC_VMRGLB
+ BPU, 0, 0, 0, 0, 0, // PC_VMRGLH
+ BPU, 0, 0, 0, 0, 0, // PC_VMRGLW
+ BPU, 0, 0, 0, 0, 0, // PC_VMULESB
+ BPU, 0, 0, 0, 0, 0, // PC_VMULESH
+ BPU, 0, 0, 0, 0, 0, // PC_VMULEUB
+ BPU, 0, 0, 0, 0, 0, // PC_VMULEUH
+ BPU, 0, 0, 0, 0, 0, // PC_VMULOSB
+ BPU, 0, 0, 0, 0, 0, // PC_VMULOSH
+ BPU, 0, 0, 0, 0, 0, // PC_VMULOUB
+ BPU, 0, 0, 0, 0, 0, // PC_VMULOUH
+ BPU, 0, 0, 0, 0, 0, // PC_VNOR
+ BPU, 0, 0, 0, 0, 0, // PC_VOR
+ BPU, 0, 0, 0, 0, 0, // PC_VPKPX
+ BPU, 0, 0, 0, 0, 0, // PC_VPKSHSS
+ BPU, 0, 0, 0, 0, 0, // PC_VPKSHUS
+ BPU, 0, 0, 0, 0, 0, // PC_VPKSWSS
+ BPU, 0, 0, 0, 0, 0, // PC_VPKSWUS
+ BPU, 0, 0, 0, 0, 0, // PC_VPKUHUM
+ BPU, 0, 0, 0, 0, 0, // PC_VPKUHUS
+ BPU, 0, 0, 0, 0, 0, // PC_VPKUWUM
+ BPU, 0, 0, 0, 0, 0, // PC_VPKUWUS
+ BPU, 0, 0, 0, 0, 0, // PC_VREFP
+ BPU, 0, 0, 0, 0, 0, // PC_VRFIM
+ BPU, 0, 0, 0, 0, 0, // PC_VRFIN
+ BPU, 0, 0, 0, 0, 0, // PC_VRFIP
+ BPU, 0, 0, 0, 0, 0, // PC_VRFIZ
+ BPU, 0, 0, 0, 0, 0, // PC_VRLB
+ BPU, 0, 0, 0, 0, 0, // PC_VRLH
+ BPU, 0, 0, 0, 0, 0, // PC_VRLW
+ BPU, 0, 0, 0, 0, 0, // PC_VRSQRTEFP
+ BPU, 0, 0, 0, 0, 0, // PC_VSL
+ BPU, 0, 0, 0, 0, 0, // PC_VSLB
+ BPU, 0, 0, 0, 0, 0, // PC_VSLH
+ BPU, 0, 0, 0, 0, 0, // PC_VSLO
+ BPU, 0, 0, 0, 0, 0, // PC_VSLW
+ BPU, 0, 0, 0, 0, 0, // PC_VSPLTB
+ BPU, 0, 0, 0, 0, 0, // PC_VSPLTH
+ BPU, 0, 0, 0, 0, 0, // PC_VSPLTW
+ BPU, 0, 0, 0, 0, 0, // PC_VSPLTISB
+ BPU, 0, 0, 0, 0, 0, // PC_VSPLTISH
+ BPU, 0, 0, 0, 0, 0, // PC_VSPLTISW
+ BPU, 0, 0, 0, 0, 0, // PC_VSR
+ BPU, 0, 0, 0, 0, 0, // PC_VSRAB
+ BPU, 0, 0, 0, 0, 0, // PC_VSRAH
+ BPU, 0, 0, 0, 0, 0, // PC_VSRAW
+ BPU, 0, 0, 0, 0, 0, // PC_VSRB
+ BPU, 0, 0, 0, 0, 0, // PC_VSRH
+ BPU, 0, 0, 0, 0, 0, // PC_VSRO
+ BPU, 0, 0, 0, 0, 0, // PC_VSRW
+ BPU, 0, 0, 0, 0, 0, // PC_VSUBCUW
+ BPU, 0, 0, 0, 0, 0, // PC_VSUBFP
+ BPU, 0, 0, 0, 0, 0, // PC_VSUBSBS
+ BPU, 0, 0, 0, 0, 0, // PC_VSUBSHS
+ BPU, 0, 0, 0, 0, 0, // PC_VSUBSWS
+ BPU, 0, 0, 0, 0, 0, // PC_VSUBUBM
+ BPU, 0, 0, 0, 0, 0, // PC_VSUBUBS
+ BPU, 0, 0, 0, 0, 0, // PC_VSUBUHM
+ BPU, 0, 0, 0, 0, 0, // PC_VSUBUHS
+ BPU, 0, 0, 0, 0, 0, // PC_VSUBUWM
+ BPU, 0, 0, 0, 0, 0, // PC_VSUBUWS
+ BPU, 0, 0, 0, 0, 0, // PC_VSUMSWS
+ BPU, 0, 0, 0, 0, 0, // PC_VSUM2SWS
+ BPU, 0, 0, 0, 0, 0, // PC_VSUM4SBS
+ BPU, 0, 0, 0, 0, 0, // PC_VSUM4SHS
+ BPU, 0, 0, 0, 0, 0, // PC_VSUM4UBS
+ BPU, 0, 0, 0, 0, 0, // PC_VUPKHPX
+ BPU, 0, 0, 0, 0, 0, // PC_VUPKHSB
+ BPU, 0, 0, 0, 0, 0, // PC_VUPKHSH
+ BPU, 0, 0, 0, 0, 0, // PC_VUPKLPX
+ BPU, 0, 0, 0, 0, 0, // PC_VUPKLSB
+ BPU, 0, 0, 0, 0, 0, // PC_VUPKLSH
+ BPU, 0, 0, 0, 0, 0, // PC_VXOR
+ BPU, 0, 0, 0, 0, 0, // PC_VMADDFP
+ BPU, 0, 0, 0, 0, 0, // PC_VMHADDSHS
+ BPU, 0, 0, 0, 0, 0, // PC_VMHRADDSHS
+ BPU, 0, 0, 0, 0, 0, // PC_VMLADDUHM
+ BPU, 0, 0, 0, 0, 0, // PC_VMSUMMBM
+ BPU, 0, 0, 0, 0, 0, // PC_VMSUMSHM
+ BPU, 0, 0, 0, 0, 0, // PC_VMSUMSHS
+ BPU, 0, 0, 0, 0, 0, // PC_VMSUMUBM
+ BPU, 0, 0, 0, 0, 0, // PC_VMSUMUHM
+ BPU, 0, 0, 0, 0, 0, // PC_VMSUMUHS
+ BPU, 0, 0, 0, 0, 0, // PC_VNMSUBFP
+ BPU, 0, 0, 0, 0, 0, // PC_VPERM
+ BPU, 0, 0, 0, 0, 0, // PC_VSEL
+ BPU, 0, 0, 0, 0, 0, // PC_VSLDOI
+ BPU, 0, 0, 0, 0, 0, // PC_VMR
+ BPU, 0, 0, 0, 0, 0, // PC_VMRP
+ BPU, 0, 0, 0, 0, 0, // PC_SLE
+ BPU, 0, 0, 0, 0, 0, // PC_SLEQ
+ BPU, 0, 0, 0, 0, 0, // PC_SLIQ
+ BPU, 0, 0, 0, 0, 0, // PC_SLLIQ
+ BPU, 0, 0, 0, 0, 0, // PC_SLLQ
+ BPU, 0, 0, 0, 0, 0, // PC_SLQ
+ BPU, 0, 0, 0, 0, 0, // PC_SRAIQ
+ BPU, 0, 0, 0, 0, 0, // PC_SRAQ
+ BPU, 0, 0, 0, 0, 0, // PC_SRE
+ BPU, 0, 0, 0, 0, 0, // PC_SREA
+ BPU, 0, 0, 0, 0, 0, // PC_SREQ
+ BPU, 0, 0, 0, 0, 0, // PC_SRIQ
+ BPU, 0, 0, 0, 0, 0, // PC_SRLIQ
+ BPU, 0, 0, 0, 0, 0, // PC_SRLQ
+ BPU, 0, 0, 0, 0, 0, // PC_SRQ
+ BPU, 0, 0, 0, 0, 0, // PC_MASKG
+ BPU, 0, 0, 0, 0, 0, // PC_MASKIR
+ BPU, 0, 0, 0, 0, 0, // PC_LSCBX
+ BPU, 0, 0, 0, 0, 0, // PC_DIV
+ BPU, 0, 0, 0, 0, 0, // PC_DIVS
+ BPU, 0, 0, 0, 0, 0, // PC_DOZ
+ BPU, 0, 0, 0, 0, 0, // PC_MUL
+ BPU, 0, 0, 0, 0, 0, // PC_NABS
+ BPU, 0, 0, 0, 0, 0, // PC_ABS
+ BPU, 0, 0, 0, 0, 0, // PC_CLCS
+ BPU, 0, 0, 0, 0, 0, // PC_DOZI
+ BPU, 0, 0, 0, 0, 0, // PC_RLMI
+ BPU, 0, 0, 0, 0, 0, // PC_RRIB
+};
+
+static void advance(int firstStage, int oldStage, int newStage) {
+ PCode *instr = pipeline[oldStage].instr;
+ int cycles = instruction_timing[instr->op].cycles[newStage - firstStage];
+ pipeline[newStage].instr = instr;
+ pipeline[newStage].remaining = cycles;
+ pipeline[oldStage].instr = NULL;
+}
+
+static void assign_completion_buffer(PCode *instr) {
+ completionbuffers.used++;
+ completionbuffers.free--;
+ completionbuffers.entries[completionbuffers.nextFreeSlot].instr = instr;
+ completionbuffers.entries[completionbuffers.nextFreeSlot].completed = 0;
+ completionbuffers.nextFreeSlot = (completionbuffers.nextFreeSlot + 1) % MaxEntries;
+}
+
+static void complete_instruction(int stage) {
+ PCode *instr = pipeline[stage].instr;
+ int buf = 0;
+ while (buf < MaxEntries && completionbuffers.entries[buf].instr != instr)
+ buf++;
+
+ completionbuffers.entries[buf].completed = 1;
+ pipeline[stage].instr = NULL;
+
+ if (stage == IU1)
+ iu1_completed_instruction = instr;
+ else if (stage == IU2)
+ iu2_completed_instruction = instr;
+}
+
+static void retire_instruction(void) {
+ completionbuffers.entries[completionbuffers.nextToRetire].instr = NULL;
+ completionbuffers.used--;
+ completionbuffers.free++;
+ completionbuffers.nextToRetire = (completionbuffers.nextToRetire + 1) % MaxEntries;
+}
+
+static int latency(PCode *instr) {
+ int cycles = instruction_timing[instr->op].latency;
+ if (PCODE_FLAG_SET_F(instr) & fRecordBit)
+ cycles += 2;
+ if (instr->op == PC_LMW || instr->op == PC_STMW)
+ cycles += instr->argCount - 2;
+ return cycles;
+}
+
+static void initialize(void) {
+ int stage;
+ int i;
+
+ for (stage = 0; stage < NumStages; stage++)
+ pipeline[stage].instr = NULL;
+
+ completionbuffers.free = MaxEntries;
+ completionbuffers.used = 0;
+ completionbuffers.nextToRetire = 0;
+ completionbuffers.nextFreeSlot = 0;
+ for (i = 0; i < MaxEntries; i++)
+ completionbuffers.entries[i].instr = NULL;
+
+ iu1_completed_instruction = NULL;
+ iu2_completed_instruction = NULL;
+}
+
+static int can_issue(PCode *instr) {
+ int stage;
+
+ if (completionbuffers.free == 0)
+ return 0;
+
+ stage = instruction_timing[instr->op].stage;
+
+ if (stage == IU2) {
+ PCode *check;
+ int isClear1 = !pipeline[IU1].instr;
+ int isClear2 = !pipeline[IU2].instr;
+ if (!isClear1 && !isClear2)
+ return 0;
+ if (isClear1 && isClear2)
+ return 1;
+
+ if (isClear1)
+ check = pipeline[IU2].instr;
+ else
+ check = pipeline[IU1].instr;
+
+ if (is_dependent(instr, check, RegClass_GPR))
+ return 0;
+ if (is_dependent(instr, iu1_completed_instruction, RegClass_GPR))
+ return 0;
+ if (is_dependent(instr, iu2_completed_instruction, RegClass_GPR))
+ return 0;
+ } else {
+ if (pipeline[stage].instr)
+ return 0;
+ }
+
+ if ((instr->flags & fIsWrite) && pipeline[LSU2].instr && (pipeline[LSU2].instr->flags & fIsWrite))
+ return 0;
+
+ return 1;
+}
+
+static void issue(PCode *instr) {
+ int stage = instruction_timing[instr->op].stage;
+ int cycles = instruction_timing[instr->op].cycles[0];
+ assign_completion_buffer(instr);
+ if (stage == IU2 && !pipeline[IU1].instr)
+ stage = IU1;
+ pipeline[stage].instr = instr;
+ pipeline[stage].remaining = cycles;
+}
+
+static void advance_clock(void) {
+ int stage;
+
+ iu1_completed_instruction = NULL;
+ iu2_completed_instruction = NULL;
+
+ for (stage = 0; stage < NumStages; stage++) {
+ if (pipeline[stage].instr && pipeline[stage].remaining)
+ --pipeline[stage].remaining;
+ }
+
+ if (completionbuffers.used && completionbuffers.entries[completionbuffers.nextToRetire].completed) {
+ retire_instruction();
+ if (completionbuffers.used && completionbuffers.entries[completionbuffers.nextToRetire].completed) {
+ retire_instruction();
+ }
+ }
+
+ if (pipeline[IU1].instr && pipeline[IU1].remaining == 0)
+ complete_instruction(IU1);
+ if (pipeline[LSU2].instr && pipeline[LSU2].remaining == 0)
+ complete_instruction(LSU2);
+ if (pipeline[FPU3].instr && pipeline[FPU3].remaining == 0)
+ complete_instruction(FPU3);
+ if (pipeline[SRU].instr && pipeline[SRU].remaining == 0)
+ complete_instruction(SRU);
+ if (pipeline[BPU].instr && pipeline[BPU].remaining == 0)
+ complete_instruction(BPU);
+ if (pipeline[IU2].instr && pipeline[IU2].remaining == 0)
+ complete_instruction(IU2);
+
+ if (
+ pipeline[FPU1].instr &&
+ pipeline[FPU1].remaining == 0 &&
+ (pipeline[FPU1].instr->op == PC_FDIV || pipeline[FPU1].instr->op == PC_FDIVS)
+ )
+ complete_instruction(FPU1);
+
+ if (pipeline[FPU2].instr && pipeline[FPU2].remaining == 0 && !pipeline[FPU3].instr)
+ advance(FPU1, FPU2, FPU3);
+ if (pipeline[FPU1].instr && pipeline[FPU1].remaining == 0 && !pipeline[FPU2].instr)
+ advance(FPU1, FPU1, FPU2);
+
+ if (pipeline[LSU1].instr && pipeline[LSU1].remaining == 0 && !pipeline[LSU2].instr)
+ advance(LSU1, LSU1, LSU2);
+}
+
+static int serializes(PCode *instr) {
+ return instruction_timing[instr->op].serializes;
+}
+
+MachineInfo machine750 = {
+ 2,
+ 1,
+ 0,
+ &latency,
+ &initialize,
+ &can_issue,
+ &issue,
+ &advance_clock,
+ &serializes,
+ &default_uses_vpermute_unit
+};