segmented stacks
SVN=125267
diff --git a/src/cmd/6l/l.h b/src/cmd/6l/l.h
index 20bac85..8966ed2 100644
--- a/src/cmd/6l/l.h
+++ b/src/cmd/6l/l.h
@@ -346,8 +346,7 @@
EXTERN Prog undefp;
EXTERN ulong stroffset;
EXTERN vlong textstksiz;
-EXTERN vlong textinarg;
-EXTERN vlong textoutarg;
+EXTERN vlong textarg;
#define UP (&undefp)
diff --git a/src/cmd/6l/list.c b/src/cmd/6l/list.c
index 9832156..789e7ce 100644
--- a/src/cmd/6l/list.c
+++ b/src/cmd/6l/list.c
@@ -106,11 +106,11 @@
goto brk;
}
parsetextconst(a->offset);
- if(textinarg == 0 && textoutarg == 0) {
+ if(textarg == 0) {
sprint(str, "$%lld", textstksiz);
goto brk;
}
- sprint(str, "$%lld-%lld-%lld", textstksiz, textinarg, textoutarg);
+ sprint(str, "$%lld-%lld", textstksiz, textarg);
goto brk;
}
@@ -422,18 +422,15 @@
if(textstksiz & 0x80000000LL)
textstksiz = -(-textstksiz & 0xffffffffLL);
-
- // the following throws away one bit
- // of precision, but maintains compat
- textinarg = (arg >> 32) & 0xffffLL;
- if(textinarg & 0x8000LL)
- textinarg = -(-textinarg & 0xffffLL);
- if(textinarg <= 0)
- textinarg = 100;
-
- textoutarg = (arg >> 48) & 0xffffLL;
- if(textoutarg & 0x8000LL)
- textoutarg = -(-textoutarg & 0xffffLL);
- if(textoutarg <= 0)
- textoutarg = 0;
+ textarg = (arg >> 32) & 0xffffffffLL;
+ if(textarg & 0x80000000LL)
+ textarg = 0;
+ if(textarg <= 0)
+ textarg = 100;
+ if(textarg > textstksiz) {
+ textarg = textstksiz;
+ if(textarg <= 0)
+ textarg = 0;
+ }
+ textarg = (textarg+7) & ~7LL;
}
diff --git a/src/cmd/6l/pass.c b/src/cmd/6l/pass.c
index 468318a..d593908 100644
--- a/src/cmd/6l/pass.c
+++ b/src/cmd/6l/pass.c
@@ -668,7 +668,7 @@
q = P;
if(pmorestack != P)
if(!(p->from.scale & NOSPLIT)) {
- if(autoffset <= 50) {
+ if(autoffset <= 75) {
// small stack
p = appendp(p);
p->as = ACMPQ;
@@ -678,14 +678,9 @@
} else {
// large stack
p = appendp(p);
- p->as = AMOVQ;
- p->from.type = D_SP;
- p->to.type = D_AX;
-
- p = appendp(p);
- p->as = ASUBQ;
- p->from.type = D_CONST;
- p->from.offset = autoffset-50;
+ p->as = ALEAQ;
+ p->from.type = D_INDIR+D_SP;
+ p->from.offset = -(autoffset-75);
p->to.type = D_AX;
p = appendp(p);
@@ -693,6 +688,7 @@
p->from.type = D_AX;
p->to.type = D_INDIR+D_R15;
}
+
// common
p = appendp(p);
p->as = AJHI;
@@ -703,9 +699,14 @@
p = appendp(p);
p->as = AMOVQ;
p->from.type = D_CONST;
- p->from.offset = curtext->to.offset;
+ p->from.offset = 0;
p->to.type = D_AX;
+ /* 160 comes from 3 calls (3*8) 4 safes (4*8) and 104 guard */
+ if(autoffset+160 > 4096)
+ p->from.offset = (autoffset+160) & ~7LL;
+ p->from.offset |= textarg<<32;
+
p = appendp(p);
p->as = ACALL;
p->to.type = D_BRANCH;
diff --git a/src/runtime/rt0_amd64_darwin.s b/src/runtime/rt0_amd64_darwin.s
index 2219489..8f2aed6 100644
--- a/src/runtime/rt0_amd64_darwin.s
+++ b/src/runtime/rt0_amd64_darwin.s
@@ -5,30 +5,40 @@
TEXT _rt0_amd64_darwin(SB),7,$-8
-// copy arguments forward on an even stack
-
+ // copy arguments forward on an even stack
MOVQ 0(SP), AX // argc
LEAQ 8(SP), BX // argv
+ SUBQ $(4*8+7), SP // 2args 2auto
ANDQ $~7, SP
- SUBQ $32, SP
MOVQ AX, 16(SP)
MOVQ BX, 24(SP)
-// allocate the per-user block
+ // allocate the per-user block
LEAQ peruser<>(SB), R15 // dedicated u. register
- MOVQ SP, AX
- SUBQ $4096, AX
- MOVQ AX, 0(R15)
+
+ LEAQ (-4096+104+4*8)(SP), AX
+ MOVQ AX, 0(R15) // 0(R15) is stack limit (w 104b guard)
+
+ MOVL $1024, AX
+ MOVL AX, 0(SP)
+ CALL mal(SB)
+
+ LEAQ 104(AX), BX
+ MOVQ BX, 16(R15) // 16(R15) is limit of istack (w 104b guard)
+
+ ADDQ 0(SP), AX
+ LEAQ (-4*8)(AX), BX
+ MOVQ BX, 24(R15) // 24(R15) is base of istack (w auto*4)
CALL check(SB)
-// process the arguments
+ // process the arguments
- MOVL 16(SP), AX
+ MOVL 16(SP), AX // copy argc
MOVL AX, 0(SP)
- MOVQ 24(SP), AX
+ MOVQ 24(SP), AX // copy argv
MOVQ AX, 8(SP)
CALL args(SB)
@@ -38,15 +48,131 @@
MOVQ AX, 0(SP) // exit status
CALL sys·exit(SB)
- CALL notok(SB)
-
- ADDQ $32, SP
+ CALL notok(SB) // fault
RET
+//
+// the calling sequence for a routine that
+// needs N bytes stack, A args.
+//
+// N1 = (N+160 > 4096)? N+160: 0
+// A1 = A
+//
+// if N <= 75
+// CMPQ SP, 0(R15)
+// JHI 3(PC)
+// MOVQ $(N1<<0) | (A1<<32)), AX
+// CALL _morestack
+//
+// if N > 75
+// LEAQ (-N-75)(SP), AX
+// CMPQ AX, 0(R15)
+// JHI 3(PC)
+// MOVQ $(N1<<0) | (A1<<32)), AX
+// CALL _morestack
+//
+
TEXT _morestack(SB), 7, $0
- MOVQ SP, AX
- SUBQ $1024, AX
+ // save stuff on interrupt stack
+
+ MOVQ 24(R15), BX // istack
+ MOVQ SP, 8(BX) // old SP
+ MOVQ AX, 16(BX) // magic number
+ MOVQ 0(R15), AX // old limit
+ MOVQ AX, 24(BX)
+
+ // switch and set up new limit
+
+ MOVQ BX, SP
+ MOVQ 16(R15), AX // istack limit
MOVQ AX, 0(R15)
+
+ // allocate a new stack max of request and 4k
+
+ MOVL 16(SP), AX // magic number
+ CMPL AX, $4096
+ JHI 2(PC)
+ MOVL $4096, AX
+ MOVL AX, 0(SP)
+ CALL mal(SB)
+
+ // switch to new stack
+
+ MOVQ SP, BX // istack
+ ADDQ $104, AX // new stack limit
+ MOVQ AX, 0(R15)
+ ADDQ 0(SP), AX
+ LEAQ (-104-4*8)(AX), SP // new SP
+ MOVQ 8(R15), AX
+ MOVQ AX, 0(SP) // old base
+ MOVQ SP, 8(R15) // new base
+
+ // copy needed stuff from istack to new stack
+
+ MOVQ 16(BX), AX // magic number
+ MOVQ AX, 16(SP)
+ MOVQ 24(BX), AX // old limit
+ MOVQ AX, 24(SP)
+ MOVQ 8(BX), AX // old SP
+ MOVQ AX, 8(SP)
+
+// are there parameters
+
+ MOVL 20(SP), CX // copy count
+ CMPL CX, $0
+ JEQ easy
+
+// copy in
+
+ LEAQ 16(AX), SI
+ SUBQ CX, SP
+ MOVQ SP, DI
+ SHRL $3, CX
+ CLD
+ REP
+ MOVSQ
+
+ // call the intended
+ CALL 0(AX)
+
+// copy out
+
+ MOVQ SP, SI
+ MOVQ 8(R15), BX // new base
+ MOVQ 8(BX), AX // old SP
+ LEAQ 16(AX), DI
+ MOVL 20(BX), CX // copy count
+ SHRL $3, CX
+ CLD
+ REP
+ MOVSQ
+
+ // restore old SP and limit
+ MOVQ 8(R15), SP // new base
+ MOVQ 24(SP), AX // old limit
+ MOVQ AX, 0(R15)
+ MOVQ 0(SP), AX
+ MOVQ AX, 8(R15) // old base
+ MOVQ 8(SP), AX // old SP
+ MOVQ AX, SP
+
+ // and return to the call behind mine
+ ADDQ $8, SP
+ RET
+
+easy:
+ CALL 0(AX)
+
+ // restore old SP and limit
+ MOVQ 24(SP), AX // old limit
+ MOVQ AX, 0(R15)
+ MOVQ 0(SP), AX
+ MOVQ AX, 8(R15) // old base
+ MOVQ 8(SP), AX // old SP
+ MOVQ AX, SP
+
+ // and return to the call behind mine
+ ADDQ $8, SP
RET
TEXT FLUSH(SB),7,$-8
diff --git a/src/runtime/rt0_amd64_linux.s b/src/runtime/rt0_amd64_linux.s
index 1dd77e4..fdda7e1 100644
--- a/src/runtime/rt0_amd64_linux.s
+++ b/src/runtime/rt0_amd64_linux.s
@@ -5,30 +5,40 @@
TEXT _rt0_amd64_linux(SB),7,$-8
-// copy arguments forward on an even stack
-
+ // copy arguments forward on an even stack
MOVQ 0(SP), AX // argc
LEAQ 8(SP), BX // argv
+ SUBQ $(4*8+7), SP // 2args 2auto
ANDQ $~7, SP
- SUBQ $32, SP
MOVQ AX, 16(SP)
MOVQ BX, 24(SP)
-// allocate the per-user block
+ // allocate the per-user block
LEAQ peruser<>(SB), R15 // dedicated u. register
- MOVQ SP, AX
- SUBQ $4096, AX
- MOVQ AX, 0(R15)
+
+ LEAQ (-4096+104+4*8)(SP), AX
+ MOVQ AX, 0(R15) // 0(R15) is stack limit (w 104b guard)
+
+ MOVL $1024, AX
+ MOVL AX, 0(SP)
+ CALL mal(SB)
+
+ LEAQ 104(AX), BX
+ MOVQ BX, 16(R15) // 16(R15) is limit of istack (w 104b guard)
+
+ ADDQ 0(SP), AX
+ LEAQ (-4*8)(AX), BX
+ MOVQ BX, 24(R15) // 24(R15) is base of istack (w auto*4)
CALL check(SB)
-// process the arguments
+ // process the arguments
- MOVL 16(SP), AX
+ MOVL 16(SP), AX // copy argc
MOVL AX, 0(SP)
- MOVQ 24(SP), AX
+ MOVQ 24(SP), AX // copy argv
MOVQ AX, 8(SP)
CALL args(SB)
@@ -38,15 +48,131 @@
MOVQ AX, 0(SP) // exit status
CALL sys·exit(SB)
- CALL notok(SB)
-
- ADDQ $32, SP
+ CALL notok(SB) // fault
RET
+//
+// the calling sequence for a routine that
+// needs N bytes stack, A args.
+//
+// N1 = (N+160 > 4096)? N+160: 0
+// A1 = A
+//
+// if N <= 75
+// CMPQ SP, 0(R15)
+// JHI 3(PC)
+// MOVQ $(N1<<0) | (A1<<32)), AX
+// CALL _morestack
+//
+// if N > 75
+// LEAQ (-N-75)(SP), AX
+// CMPQ AX, 0(R15)
+// JHI 3(PC)
+// MOVQ $(N1<<0) | (A1<<32)), AX
+// CALL _morestack
+//
+
TEXT _morestack(SB), 7, $0
- MOVQ SP, AX
- SUBQ $1024, AX
+ // save stuff on interrupt stack
+
+ MOVQ 24(R15), BX // istack
+ MOVQ SP, 8(BX) // old SP
+ MOVQ AX, 16(BX) // magic number
+ MOVQ 0(R15), AX // old limit
+ MOVQ AX, 24(BX)
+
+ // switch and set up new limit
+
+ MOVQ BX, SP
+ MOVQ 16(R15), AX // istack limit
MOVQ AX, 0(R15)
+
+ // allocate a new stack max of request and 4k
+
+ MOVL 16(SP), AX // magic number
+ CMPL AX, $4096
+ JHI 2(PC)
+ MOVL $4096, AX
+ MOVL AX, 0(SP)
+ CALL mal(SB)
+
+ // switch to new stack
+
+ MOVQ SP, BX // istack
+ ADDQ $104, AX // new stack limit
+ MOVQ AX, 0(R15)
+ ADDQ 0(SP), AX
+ LEAQ (-104-4*8)(AX), SP // new SP
+ MOVQ 8(R15), AX
+ MOVQ AX, 0(SP) // old base
+ MOVQ SP, 8(R15) // new base
+
+ // copy needed stuff from istack to new stack
+
+ MOVQ 16(BX), AX // magic number
+ MOVQ AX, 16(SP)
+ MOVQ 24(BX), AX // old limit
+ MOVQ AX, 24(SP)
+ MOVQ 8(BX), AX // old SP
+ MOVQ AX, 8(SP)
+
+// are there parameters
+
+ MOVL 20(SP), CX // copy count
+ CMPL CX, $0
+ JEQ easy
+
+// copy in
+
+ LEAQ 16(AX), SI
+ SUBQ CX, SP
+ MOVQ SP, DI
+ SHRL $3, CX
+ CLD
+ REP
+ MOVSQ
+
+ // call the intended
+ CALL 0(AX)
+
+// copy out
+
+ MOVQ SP, SI
+ MOVQ 8(R15), BX // new base
+ MOVQ 8(BX), AX // old SP
+ LEAQ 16(AX), DI
+ MOVL 20(BX), CX // copy count
+ SHRL $3, CX
+ CLD
+ REP
+ MOVSQ
+
+ // restore old SP and limit
+ MOVQ 8(R15), SP // new base
+ MOVQ 24(SP), AX // old limit
+ MOVQ AX, 0(R15)
+ MOVQ 0(SP), AX
+ MOVQ AX, 8(R15) // old base
+ MOVQ 8(SP), AX // old SP
+ MOVQ AX, SP
+
+ // and return to the call behind mine
+ ADDQ $8, SP
+ RET
+
+easy:
+ CALL 0(AX)
+
+ // restore old SP and limit
+ MOVQ 24(SP), AX // old limit
+ MOVQ AX, 0(R15)
+ MOVQ 0(SP), AX
+ MOVQ AX, 8(R15) // old base
+ MOVQ 8(SP), AX // old SP
+ MOVQ AX, SP
+
+ // and return to the call behind mine
+ ADDQ $8, SP
RET
TEXT FLUSH(SB),7,$-8
@@ -145,7 +271,7 @@
TEXT sys·memclr(SB),1,$-8
MOVQ 8(SP), DI // arg 1 addr
- MOVL 16(SP), CX // arg 2 count
+ MOVL 16(SP), CX // arg 2 count (cannot be zero)
ADDL $7, CX
SHRL $3, CX
MOVQ $0, AX
diff --git a/src/runtime/runtime.c b/src/runtime/runtime.c
index 5ff3f85..52ffba7 100644
--- a/src/runtime/runtime.c
+++ b/src/runtime/runtime.c
@@ -183,7 +183,7 @@
{
byte* v;
- v = sys·mmap(nil, NHUNK, PROT_READ|PROT_WRITE, MAP_ANON|MAP_PRIVATE, 0, 0);
+ v = sys·mmap(nil, n, PROT_READ|PROT_WRITE, MAP_ANON|MAP_PRIVATE, 0, 0);
sys·memclr(v, n);
nmmap += n;
return v;
@@ -194,10 +194,8 @@
{
byte* v;
- // round to keep everything 64-bit alligned
- while(n & 7)
- n++;
-
+ // round to keep everything 64-bit aligned
+ n = (n+7) & ~7;
nmal += n;
// do we have enough in contiguous hunk
