---
See test/tls.ssa for usage. Essentially:

threadlocal data $i = { w 42 }

%ret =l loadl threadlocal $i

Seeking sign-off on the approach and syntax. Still to-do:
- rv64, arm64 support
- Documentation updates

Points of note for the review:
- Adding -lpthread to every test
- use of strcmp in gasemitlnk
- Any edge cases I've forgotten about?

 all.h         |  2 ++
 amd64/emit.c  |  7 ++++++-
 amd64/isel.c  |  1 +
 gas.c         |  6 ++++++
 parse.c       | 24 ++++++++++++++++++++++++
 test/tls.ssa  | 23 +++++++++++++++++++++++
 tools/lexh.c  |  2 +-
 tools/test.sh | 10 +++++-----
 8 files changed, 68 insertions(+), 7 deletions(-)
 create mode 100644 test/tls.ssa

diff --git a/all.h b/all.h
index c8ec93c..3418f63 100644
--- a/all.h
+++ b/all.h
@@ -309,6 +309,7 @@ struct Con {
 		CUndef,
 		CBits,
 		CAddr,
+		CThreadLocal,
 	} type;
 	uint32_t label;
 	union {
@@ -334,6 +335,7 @@ struct Lnk {
 	char align;
 	char *sec;
 	char *secf;
+	char istls;
 };
 
 struct Fn {
diff --git a/amd64/emit.c b/amd64/emit.c
index b8e9e8e..9870e0b 100644
--- a/amd64/emit.c
+++ b/amd64/emit.c
@@ -171,6 +171,10 @@ emitcon(Con *con, FILE *f)
 		if (con->bits.i)
 			fprintf(f, "%+"PRId64, con->bits.i);
 		break;
+	case CThreadLocal:
+		l = str(con->label);
+		fprintf(f, "%%fs:%s@tpoff", l);
+		break;
 	case CBits:
 		fprintf(f, "%"PRId64, con->bits.i);
 		break;
@@ -306,7 +310,8 @@ Next:
 			fputc(')', f);
 			break;
 		case RCon:
-			fputc('$', f);
+			if (fn->con[ref.val].type != CThreadLocal)
+				fputc('$', f);
 			emitcon(&fn->con[ref.val], f);
 			break;
 		default:
diff --git a/amd64/isel.c b/amd64/isel.c
index 5a64429..592f7fc 100644
--- a/amd64/isel.c
+++ b/amd64/isel.c
@@ -36,6 +36,7 @@ noimm(Ref r, Fn *fn)
 		return 0;
 	switch (fn->con[r.val].type) {
 	case CAddr:
+	case CThreadLocal:
 		/* we only support the 'small'
 		 * code model of the ABI, this
 		 * means that we can always
diff --git a/gas.c b/gas.c
index 865edba..92ffd7f 100644
--- a/gas.c
+++ b/gas.c
@@ -29,6 +29,12 @@ gasemitlnk(char *n, Lnk *l, char *s, FILE *f)
 		fprintf(f, ".section %s", l->sec);
 		if (l->secf)
 			fprintf(f, ", %s", l->secf);
+	} else if (l->istls) {
+		if (strcmp(s, ".bss") == 0) {
+			fprintf(f, ".section .tbss,\"awT\"");
+		} else {
+			fprintf(f, ".section .tdata,\"awT\"");
+		}
 	} else {
 		fputs(s, f);
 	}
diff --git a/parse.c b/parse.c
index 1912c8b..5e09a21 100644
--- a/parse.c
+++ b/parse.c
@@ -42,6 +42,7 @@ enum {
 	Ttype,
 	Tdata,
 	Tsection,
+	Tthreadlocal,
 	Talign,
 	Tl,
 	Tw,
@@ -92,6 +93,7 @@ static char *kwmap[Ntok] = {
 	[Ttype] = "type",
 	[Tdata] = "data",
 	[Tsection] = "section",
+	[Tthreadlocal] = "threadlocal",
 	[Talign] = "align",
 	[Tl] = "l",
 	[Tw] = "w",
@@ -403,6 +405,13 @@ parseref()
 	case Tglo:
 		c.type = CAddr;
 		c.label = intern(tokval.str);
+		goto Look;
+	case Tthreadlocal:
+		c.type = CThreadLocal;
+		if (next() != Tglo)
+			err("function name expected");
+		c.label = intern(tokval.str);
+		goto Look;
 	Look:
 		return newcon(&c, curf);
 	default:
@@ -798,6 +807,9 @@ parsefn(Lnk *lnk)
 	int i;
 	PState ps;
 
+	if (lnk->istls)
+		err("TLS is not supported for function declarations");
+
 	curb = 0;
 	nblk = 0;
 	curi = insb;
@@ -1073,6 +1085,8 @@ parselnk(Lnk *lnk)
 		case Tsection:
 			if (lnk->sec)
 				err("only one section allowed");
+			if (lnk->istls)
+				err("threadlocal and section cannot co-exist");
 			if (next() != Tstr)
 				err("section \"name\" expected");
 			lnk->sec = tokval.str;
@@ -1081,6 +1095,13 @@ parselnk(Lnk *lnk)
 				lnk->secf = tokval.str;
 			}
 			break;
+		case Tthreadlocal:
+			if (lnk->istls)
+				err("only one TLS directive allowed");
+			if (lnk->sec)
+				err("threadlocal and section cannot co-exist");
+			lnk->istls = 1;
+			break;
 		default:
 			if (haslnk)
 			if (t != Tdata)
@@ -1138,6 +1159,9 @@ printcon(Con *c, FILE *f)
 		if (c->bits.i)
 			fprintf(f, "%+"PRIi64, c->bits.i);
 		break;
+	case CThreadLocal:
+		fprintf(f, "threadlocal $%s", str(c->label));
+		break;
 	case CBits:
 		if (c->flt == 1)
 			fprintf(f, "s_%f", c->bits.s);
diff --git a/test/tls.ssa b/test/tls.ssa
new file mode 100644
index 0000000..5975932
--- /dev/null
+++ b/test/tls.ssa
@@ -0,0 +1,23 @@
+export function w $main() {
+@start
+	%thread =l alloc8 8
+	%retval =l alloc8 8
+
+	storew 1337, threadlocal $i
+	call $pthread_create(l %thread, l 0, l $thread, l 0)
+
+	%thread_id =l loadl %thread
+	call $pthread_join(l %thread_id, l %retval)
+
+	%ret =l loadl %retval
+	%res =l cnel %ret, 42
+	ret %res
+}
+
+function w $thread() {
+@start
+	%ret =l loadl threadlocal $i
+	ret %ret
+}
+
+threadlocal data $i = { w 42 }
diff --git a/tools/lexh.c b/tools/lexh.c
index 8d0af21..5be376e 100644
--- a/tools/lexh.c
+++ b/tools/lexh.c
@@ -26,7 +26,7 @@ char *tok[] = {
 	"vaarg", "vastart", "...", "env",
 
 	"call", "phi", "jmp", "jnz", "ret", "export",
-	"function", "type", "data", "section", "align",
+	"function", "type", "data", "section", "threadlocal", "align",
 	"l", "w", "h", "b", "d", "s", "z", "loadw", "loadl",
 	"loads", "loadd", "alloc1", "alloc2",
 
diff --git a/tools/test.sh b/tools/test.sh
index 4653b83..37105df 100755
--- a/tools/test.sh
+++ b/tools/test.sh
@@ -22,7 +22,7 @@ init() {
 	arm64)
 		for p in aarch64-linux-musl aarch64-linux-gnu
 		do
-			cc="$p-gcc -no-pie -static"
+			cc="$p-gcc -no-pie -static -lpthread"
 			qemu="qemu-aarch64"
 			if
 				$cc -v >/dev/null 2>&1 &&
@@ -46,7 +46,7 @@ init() {
 	rv64)
 		for p in riscv64-linux-musl riscv64-linux-gnu
 		do
-			cc="$p-gcc -no-pie -static"
+			cc="$p-gcc -no-pie -static -lpthread"
 			qemu="qemu-riscv64"
 			if
 				$cc -v >/dev/null 2>&1 &&
@@ -73,13 +73,13 @@ init() {
 			cc="cc -Wl,-no_pie"
 			;;
 		*OpenBSD*)
-			cc="cc -nopie"
+			cc="cc -nopie -lpthread"
 			;;
 		*FreeBSD*)
-			cc="cc"
+			cc="cc -lpthread"
 			;;
 		*)
-			cc="${CC:-cc} -no-pie"
+			cc="${CC:-cc} -no-pie -lpthread"
 			testcc "$cc" || cc="${CC:-cc}"
 			;;
 		esac

base-commit: c8cd2824eae0137505fe46530c3a8e9788ab9a63
-- 
2.37.1