/* * QuickThreads -- Threads-building toolkit. * Copyright (c) 1993 by David Keppel * * Permission to use, copy, modify and distribute this software and * its documentation for any purpose and without fee is hereby * granted, provided that the above copyright notice and this notice * appear in all copies. This software is provided as a * proof-of-concept and for demonstration purposes; there is no * representation about the suitability of this software for any * purpose. */ .file "ksr1.s" .def .debug; .endef .align 128 .globl qt_blocki .globl qt_blocki$TXT .globl qt_block .globl qt_block$TXT .globl qt_start$TXT .globl qt_start .globl qt_abort$TXT .globl qt_abort .globl qt_vstart .globl qt_vstart$TXT # # KSR convention: on procedure calls, load both the procedure address # and a pointer to a constant block. The address of function `f' is # `f$TXT', and the constant block address is `f'. The constant block # has several reserved values: # # 8 bytes fpu register save mask # 4 bytes ipu register save mask # 4 bytes ceu register save mask # f: f$TXT # ... whatever you want ... (not quite...read on) # # Note, by the way, that a pointer to a function is passed as a # pointer to the constant area, and the constant area has the text # address. # # # Procedures that do not return structures prefix their code with # # proc$TXT: # finop; cxnop # finop; cxnop # # # Calls to those procedures branch to a 16 byte offset (4 instrs) in # to the procedure to skip those instructions. # # Procedures that return structures use a different code prefix: # # proc$TXT: # finop; beq.qt %rc, %rc, 24 # return value entry # finop; cxnop # finop; movi8 0, %rc # no return value entry # # # Calls that want the returned structure branch directly to the # procedure address. Callers that don't want (or aren't expecting) a # return value branche 16 bytes in to the procedure, which will zero # %rc, telling the called procedure not to return a structure. # # # On entry: # %i2 -- control block of helper function to run # (dereference to get helper) # %i3 -- a1 # %i4 -- a2 # %i5 -- sp of new to run # .data .half 0x0, 0x0, 0x7ffff000, 0x7fff8000 qt_blocki: qt_abort: .word qt_blocki$TXT .word qt_restore$TXT .text qt_abort$TXT: qt_blocki$TXT: finop ; cxnop # entry prefix finop ; cxnop # entry prefix add8.ntr 75,%i31,%i31 ; movi8 512,%c5 # ICR; stk adjust finop ; ssub8.ntr 0,%sp,%c5,%sp finop ; st8 %fp,504(%sp) # Save caller's fp finop ; st8 %cp,496(%sp) # Save caller's cp finop ; ld8 8(%c10),%c5 # ld qt_restore$TXT finop ; st8 %c14,0(%sp) # Save special ret addr finop ; mov8_8 %c10, %cp # Our cp finop ; sadd8.ntr 0,%sp,%c5,%fp # Our frame ptr finop ; st8 %c5,8(%sp) # st qt_restore$TXT # # CEU registers %c15-%c24, %c26-%c30 (%c14 we restore later) # finop ; st8 %c15,456(%sp) finop ; st8 %c16,448(%sp) finop ; st8 %c17,440(%sp) finop ; st8 %c18,432(%sp) finop ; st8 %c19,424(%sp) finop ; st8 %c20,416(%sp) finop ; st8 %c21,408(%sp) finop ; st8 %c22,400(%sp) finop ; st8 %c23,392(%sp) finop ; st8 %c24,384(%sp) # # %c25 is the Enclosing Frame Pointer (EFP) -- since C doesn't # use nested procedures, we ignore it (leaving a gap, though) # finop ; st8 %c26,368(%sp) finop ; st8 %c27,360(%sp) finop ; st8 %c28,352(%sp) finop ; st8 %c29,344(%sp) finop ; st8 %c30,336(%sp) # # IPU registers %i12-%i30 # finop ; st8 %i12,328(%sp) finop ; st8 %i13,320(%sp) finop ; st8 %i14,312(%sp) finop ; st8 %i15,304(%sp) # (gap to get alignment for st64) # -- Doesn't work on version 1.1.3 of the OS # finop ; st64 %i16,256(%sp) finop ; st8 %i16,256(%sp) finop ; st8 %i17,248(%sp) finop ; st8 %i18,240(%sp) finop ; st8 %i19,232(%sp) finop ; st8 %i20,224(%sp) finop ; st8 %i21,216(%sp) finop ; st8 %i22,208(%sp) finop ; st8 %i23,200(%sp) finop ; st8 %i24,192(%sp) finop ; st8 %i25,184(%sp) finop ; st8 %i26,176(%sp) finop ; st8 %i27,168(%sp) finop ; st8 %i28,160(%sp) finop ; st8 %i29,152(%sp) finop ; st8 %i30,144(%sp) # # FPU already saved, or saving not necessary # # # Switch to the stack passed in as fourth argument to the block # routine (%i5) and call the helper routine passed in as the first # argument (%i2). Note that the address of the helper's constant # block is passed in, so we must derefence it to get the helper's text # address. # finop ; movb8_8 %i2,%c10 # helper's ConstBlock finop ; cxnop # Delay slot, fill w/ finop ; cxnop # .. 2 st8 from above finop ; ld8 0(%c10),%c4 # load addr of helper finop ; movb8_8 %sp, %i2 # 1st arg to helper # is this stack; other # args remain in regs finop ; movb8_8 %i5,%sp # switch stacks finop ; jsr %c14,16(%c4) # call helper movi8 3, %i0 ; movi8 0,%c8 # nargs brain dmg finop ; cxnop finop ; cxnop # # Here is where behavior differs for threads being restored and threads # being started. Blocked threads have a pointer to qt_restore$TXT on # the top of their stacks; manufactured stacks have a pointer to qt_start$TXT # on the top of their stacks. With this setup, starting threads # skip the (unecessary) restore operations. # # We jump to an offset of 16 to either (1) skip past the two noop pairs # at the start of qt_start$TXT, or (2) skip past the two noop pairs # after qt_restore$TXT. # finop ; ld8 8(%sp),%c4 finop ; cxnop finop ; cxnop finop ; jmp 16(%c4) qt_restore$TXT: finop ; cxnop finop ; cxnop # # Point of Restore: # # The helper funtion will return here. Any result it has placed in # a return register (most likely %i0) will not get overwritten below # and will consequently be the return value of the blocking routine. # # # CEU registers %c15-%c24, %c26-%c30 (%c14 we restore later) # finop ; ld8 456(%sp),%c15 finop ; ld8 448(%sp),%c16 finop ; ld8 440(%sp),%c17 finop ; ld8 432(%sp),%c18 finop ; ld8 424(%sp),%c19 finop ; ld8 416(%sp),%c20 finop ; ld8 408(%sp),%c21 finop ; ld8 400(%sp),%c22 finop ; ld8 392(%sp),%c23 finop ; ld8 384(%sp),%c24 # # %c25 is the Enclosing Frame Pointer (EFP) -- since C doesn't # use nested procedures, we ignore it (leaving a gap, though) # finop ; ld8 368(%sp),%c26 finop ; ld8 360(%sp),%c27 finop ; ld8 352(%sp),%c28 finop ; ld8 344(%sp),%c29 finop ; ld8 336(%sp),%c30 # # IPU registers %i12-%i30 # finop ; ld8 328(%sp),%i12 finop ; ld8 320(%sp),%i13 finop ; ld8 312(%sp),%i14 finop ; ld8 304(%sp),%i15 # (gap to get alignment for ld64) # -- Doesn't work on version 1.1.3 of the OS # finop ; ld64 256(%sp),%i16 finop ; ld8 256(%sp),%i16 finop ; ld8 248(%sp),%i17 finop ; ld8 240(%sp),%i18 finop ; ld8 232(%sp),%i19 finop ; ld8 224(%sp),%i20 finop ; ld8 216(%sp),%i21 finop ; ld8 208(%sp),%i22 finop ; ld8 200(%sp),%i23 finop ; ld8 192(%sp),%i24 finop ; ld8 184(%sp),%i25 finop ; ld8 176(%sp),%i26 finop ; ld8 168(%sp),%i27 finop ; ld8 160(%sp),%i28 finop ; ld8 152(%sp),%i29 finop ; ld8 144(%sp),%i30 # # FPU registers don't need to be loaded, or will be loaded by an # enclosing scope (e.g., if this is called by qt_block). # # # Load the special registers. We don't load the stack ptr because # the new stack is passed in as an argument, we don't load the EFP # because we don't use it, and we load the return address specially # off the top of the stack. # finop ; ld8 0(%sp),%c14 # return addr finop ; ld8 496(%sp),%cp finop ; ld8 504(%sp),%fp finop ; jmp 32(%c14) # jump back to thread finop ; movi8 512,%c5 # stack adjust finop ; sadd8.ntr 0,%sp,%c5,%sp .data .half 0x0, 0x0, 0x7ffff000, 0x7fff8000 qt_block: .word qt_block$TXT .word qt_error .word qt_error$TXT .word qt_blocki # # Handle saving and restoring the FPU regs, relying on qt_blocki # to save and restore the remaining registers. # .text qt_block$TXT: finop ; cxnop # entry prefix finop ; cxnop # entry prefix add8.ntr 29,%i31,%i31 ; movi8 512,%c5 # ICR; stk adjust finop ; ssub8.ntr 0,%sp,%c5,%sp finop ; st8 %fp,504(%sp) # Save caller's fp finop ; st8 %cp,496(%sp) # Save caller's cp finop ; st8 %c14,488(%sp) # store ret addr finop ; sadd8.ntr 0,%sp,%c5,%fp # Our frame ptr finop ; mov8_8 %c10, %cp # Our cp # # Store 8 registers at once...destination must be a multiple of 64 # finop ; st64 %f16,384(%sp) finop ; st64 %f24,320(%sp) finop ; st64 %f32,256(%sp) finop ; st64 %f40,192(%sp) finop ; st64 %f48,128(%sp) finop ; st64 %f56,64(%sp) # # Call the integer blocking routine, passing the arguments passed to us # finop ; ld8 24(%cp), %c10 finop ; cxnop finop ; jsr %c14, qt_blocki$TXT finop ; cxnop finop ; cxnop movi8 4,%i0 ; movi8 0,%c8 # nargs brain dmg # # Load 8 registers at once...source must be a multiple of 64 # finop ; ld64 64(%sp),%f56 finop ; ld64 128(%sp),%f48 finop ; ld64 192(%sp),%f40 finop ; ld64 256(%sp),%f32 finop ; ld64 320(%sp),%f24 finop ; ld64 384(%sp),%f16 finop ; ld8 488(%sp),%c14 finop ; ld8 496(%sp),%cp finop ; ld8 504(%sp),%fp finop ; jmp 32(%c14) # jump back to thread finop ; movi8 512,%c5 # stack adjust finop ; sadd8.ntr 0,%sp,%c5,%sp .data .half 0x0, 0x0, 0x7ffff000, 0x7fff8000 qt_start: .word qt_start$TXT # # A new thread is set up to "appear" as if it were executing code at # the beginning of qt_start and then it called a blocking routine # (qt_blocki). So when a new thread starts to run, it gets unblocked # by the code above and "returns" to `qt_start$TXT' in the # restore step of the switch. Blocked threads jump to 16(qt_restore$TXT), # and starting threads jump to 16(qt_start$TXT). # .text qt_start$TXT: finop ; cxnop # finop ; cxnop # finop ; ld8 40(%sp),%c10 # `only' constant block finop ; ld8 32(%sp),%i4 # `userf' arg. finop ; ld8 24(%sp),%i3 # `t' arg. finop ; ld8 0(%c10),%c4 # `only' text location finop ; ld8 16(%sp),%i2 # `u' arg. finop ; cxnop finop ; jsr %c14,16(%c4) # call `only' # # Pop the frame used to store the thread's initial data # finop ; sadd8.ntr 0,%sp,128,%sp finop ; cxnop movi8 2,%i0 ; movi8 0,%c8 # nargs brain dmg # # If we ever return, it's an error. # finop ; jmp qt_error$TXT finop ; cxnop finop ; cxnop movi8 0,%i0 ; movi8 0,%c8 # nargs brain dmg # # This stuff is broken # .data .half 0x0, 0x0, 0x7ffff000, 0x7fff8000 qt_vstart: .word qt_vstart$TXT .text qt_vstart$TXT: finop ; cxnop # entry prefix finop ; cxnop # entry prefix finop ; cxnop finop ; cxnop add8.ntr 11,%i31,%i31 ; movi8 512,%c5 finop ; ssub8.ntr 0,%sp,%c5,%sp # fix stack finop ; ld8 8(%sp),%i2 # load `t' as arg to finop ; cxnop # `startup' finop ; cxnop finop ; ld8 16(%sp),%c10 # `startup' const block finop ; cxnop finop ; cxnop finop ; ld8 0(%c10),%c4 # `startup' text loc. finop ; cxnop finop ; cxnop finop ; jsr %c14,16(%c4) # call `startup' finop ; cxnop finop ; cxnop movi8 1, %i0 ; movi8 0,%c8 # nargs brain dmg # # finop ; sadd 0,%sp,128,%sp # alter stack # finop ; ld8 8(%sp),%i2 # load `t' as arg to finop ; ld8 8(%sp),%i2 # load `t' as arg to finop ; ld8 8(%sp),%i2 # load `t' as arg to finop ; ld8 8(%sp),%i2 # load `t' as arg to finop ; ld8 32(%sp),%c10 # `only' constant block finop ; ld8 8(%sp),%i2 # `u' arg. finop ; ld8 16(%sp),%i3 # `t' arg. finop ; ld8 0(%c10),%c4 # `only' text location finop ; ld8 24(%sp),%i4 # `userf' arg. finop ; cxnop finop ; jsr %c4,16(%c4) # call `only' finop ; cxnop finop ; cxnop # # If the callee ever calls `nargs', the following instruction (pair) # will be executed. However, we don't know when we compile this code # how many args are being passed. So we give our best guess: 0. # movi8 0,%i0 ; movi8 0,%c8 # nargs brain dmg # # If we ever return, it's an error. # finop ; jmp qt_error$TXT finop ; cxnop finop ; cxnop movi8 0,%i0 ; movi8 0,%c8 # nargs brain dmg