arm: Use one less instruction for prolog in slightly complex functions - lightning.git - Portable just-in-time compiler library

Portable just-in-time compiler library

diff options

author	pcpa <paulo.cesar.pereira.de.andrade@gmail.com>	2023年02月03日 11:25:45 -0300
committer	pcpa <paulo.cesar.pereira.de.andrade@gmail.com>	2023年02月03日 11:25:45 -0300
commit	8c3d3b123661a4ccceb252c4494c68126bfdada0 (patch)
tree	e47441ea12e0e0f9ffbfc88ee3160adaa23ca5e2
parent	28020887ac68fbd7437b2c481efe6c39cd0b01ce (diff)
download	lightning-8c3d3b123661a4ccceb252c4494c68126bfdada0.tar.gz

arm: Use one less instruction for prolog in slightly complex functions

Slightly complex functions would be functions that need to save r0-r3 in prolog because of calls C functions. The optimization is only for hard float (regardless or abi); could be done for software float in most cases but keep the code, already too complex, a bit simpler. The implementation still assumes there is no division by hardware. Almost certainly modern arm chips (armv7r) have it in hardware, otherwise, could use vfp/neon instructions and do division with double values. This could/should be changed. Currently only qemu is used to test different abis and setups.

Diffstat

-rw-r--r--

lib/jit_arm-cpu.c

1 files changed, 12 insertions, 3 deletions

diff --git a/lib/jit_arm-cpu.c b/lib/jit_arm-cpu.c
index 6beae05..06e38ff 100644
--- a/lib/jit_arm-cpu.c
+++ b/lib/jit_arm-cpu.c

@@ -3911,6 +3911,9 @@ _prolog(jit_state_t *_jit, jit_node_t *node)

}

if (_jitc->function->need_frame)

mask |= (1 << _FP_REGNO) | (1 << _LR_REGNO);

+ if (!jit_swf_p() && _jitc->function->save_reg_args &&

+ !(_jitc->function->self.call & jit_call_varargs))

+ mask |= 0xf;

if (jit_thumb_p()) {

/* switch to thumb mode (better approach would be to

@@ -3921,13 +3924,15 @@ _prolog(jit_state_t *_jit, jit_node_t *node)

BX(_R12_REGNO);

if (!_jitc->thumb)

_jitc->thumb = _jit->pc.w;

- if (_jitc->function->save_reg_args)

+ if (jit_swf_p() || (_jitc->function->save_reg_args &&

+ (_jitc->function->self.call & jit_call_varargs)))

T2_PUSH(0xf);

if (mask)

T2_PUSH(mask);

}

else {

- if (_jitc->function->save_reg_args)

+ if (jit_swf_p() || (_jitc->function->save_reg_args &&

+ (_jitc->function->self.call & jit_call_varargs)))

PUSH(0xf);

if (mask)

PUSH(mask);

@@ -3970,13 +3975,17 @@ _epilog(jit_state_t *_jit, jit_node_t *node)

mask |= (1 << _FP_REGNO) | (1 << _LR_REGNO);

movr(_SP_REGNO, _FP_REGNO);

}

+ if (!jit_swf_p() && _jitc->function->save_reg_args &&

+ !(_jitc->function->self.call & jit_call_varargs))

+ addi(_SP_REGNO, _SP_REGNO, 16);

if (mask) {

if (jit_thumb_p())

T2_POP(mask);

else

POP(mask);

}

- if (_jitc->function->save_reg_args)

+ if (jit_swf_p() || (_jitc->function->save_reg_args &&

+ (_jitc->function->self.call & jit_call_varargs)))

addi(_SP_REGNO, _SP_REGNO, 16);

if (jit_thumb_p())

T1_BX(_LR_REGNO);

generated by cgit v1.2.3 (git 2.39.1) at 2025年09月15日 18:37:31 +0000