ppc64 proposal for push/pop operations on assembly level

push-pop for ppc64.md

The handwritten assembly is located where commented and with less indentation levels(lines 40-80).

The idea is to have a reliable push where the stack pointer $r1 remains unchanged.

On this proposal, the $r10 is used as the pointer to the top of the stack and it'll be initialized by having the same address as the stack pointer $r1:

mr 10, 1

Push will be accomplished with 1 instructions:

stdu @push_element_register,  -16(10)

Pop will be accomplished with 2 instructions:

ld  @pop_element_register,  0(10)
addi 10, 10, 16

When branching out of the function (aka when the stack pointer changes), a fake call stack is built in order to save the pushed elements from being overwritten by the new call stack:

mflr 0
std  0, 16(1)     # save LR
stdu 1, -32(10)   # create 32 bytes additionally for the stack and backchain
mr   1, 10
bl @target

When that branch returns, the fake call stack restores the stack pointer $r1 and the pointer to the top of the stack $r10:

addi 10, 1, 32    # restore top of the stack address
ld   1,  0(1)     # backchain
ld   0, 16(1)     # restore LR
mtlr 0

source.s

	.file	"simple.c"
	.abiversion 2
	.section	".toc","aw"
	.section	".text"
	.align 2
	.globl ret1
	.type	ret1, @function
ret1:
	std 31,-8(1)
	stdu 1,-80(1)
	mr 31,1
	li 9,1
	stw 9,32(31)
	lwz 9,32(31)
	extsw 9,9
	std 9,48(31)
	lfd 0,48(31)
	stfd 0,48(31)
	ld 9,48(31)
	mr 3,9
	addi 1,31,80
	ld 31,-8(1)
	blr
	.long 0
	.byte 0,0,0,0,128,1,0,1
	.size	ret1,.-ret1
	.align 2
	.globl main
	.type	main, @function
main:
0:	addis 2,12,.TOC.-0b@ha
	addi 2,2,.TOC.-0b@l
	.localentry	main,.-main
	mflr 0
	std 0,16(1)
	std 31,-8(1)
	stdu 1,-64(1)
	mr 31,1

  # initialization for pushing
  mr 10, 1          # r10 is the top of the stack address
  li 9,0            # element to be pushed

# pushing elements
  addi 9,   9,  1
  stdu 9,  -16(10)     # pushing

  addi 9,   9,  1
  stdu 9,  -16(10)     # pushing

  addi 9,   9,  1
  stdu 9,  -16(10)     # pushing

  # build a callstack due to the 'call ret' below
  # save r10 so the size is known and update the stack pointer
  mflr 0
  std  0, 16(1)     # save LR
  stdu 10,-16(10)   # save top of the stack address
  stdu 1, -32(10)   # create 32 bytes additionally for the stack and backchain
  mr 1, 10

	bl ret1

  # pop the callstack, revert operations
  ld  10, 32(1)     # restore top of the stack address
  ld  1,   0(1)     # backchain
  ld   0, 16(1)
  mtlr 0

  # popping elements, r10 restored
  ld  9,  0(10)    # a pop
  addi 10, 10, 16

  ld  9,  0(10)    # a pop
  addi 10, 10, 16

  ld  9,  0(10)    # a pop
  addi 10, 10, 16

	mr 9,3
	mr 3,9

	addi 1,31,64
	ld 0,16(1)
	mtlr 0
	ld 31,-8(1)
	blr
	.long 0
	.byte 0,0,0,1,128,1,0,1
	.size	main,.-main
	.ident	"GCC: (Debian 4.9.3-3) 4.9.3"
	.section	.note.GNU-stack,"",@progbits

vret and vcall will be changed in order to be consistent with push/pop? i.e., will r10 be saved in stack before a (v)call?will r10 be copied to r1 before a call (in order to not overwrite the stack)?

This push/pop operations are not specifically for HHVM (as you point out about vret and vcall). On the proposal it says "When branching out of the function"... The source code related to that are 54-68.

As you see on line #58, yes, it's saved.