Hi! The following patch fixes ICE, when we try to split a double-word TLS load or store. The problem is that x86_64_immediate_operand disallows CONST_INT offsets with -mcmodel=large. I guess it is intentional for SYMBOL_REFs, LABEL_REFs etc., but for the TLS LE UNSPECs it should be fine, there are no 64-bit relocations for these, we don't really support > 2GB thread local segments.
Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk? 2019-02-12 Jakub Jelinek <ja...@redhat.com> PR target/89290 * config/i386/predicates.md (x86_64_immediate_operand): Allow TLS UNSPECs offsetted by signed 32-bit CONST_INT even with -mcmodel=large. * gcc.target/i386/pr89290.c: New test. --- gcc/config/i386/predicates.md.jj 2019-01-01 12:37:32.267727037 +0100 +++ gcc/config/i386/predicates.md 2019-02-12 17:07:15.937097266 +0100 @@ -182,7 +182,7 @@ (define_predicate "x86_64_immediate_oper rtx op1 = XEXP (XEXP (op, 0), 0); rtx op2 = XEXP (XEXP (op, 0), 1); - if (ix86_cmodel == CM_LARGE) + if (ix86_cmodel == CM_LARGE && GET_CODE (op1) != UNSPEC) return false; if (!CONST_INT_P (op2)) return false; --- gcc/testsuite/gcc.target/i386/pr89290.c.jj 2019-02-12 17:32:49.291750588 +0100 +++ gcc/testsuite/gcc.target/i386/pr89290.c 2019-02-12 17:34:00.847568163 +0100 @@ -0,0 +1,19 @@ +/* PR target/89290 */ +/* { dg-do compile { target { tls && lp64 } } } */ +/* { dg-options "-O0 -mcmodel=large" } */ + +struct S { long int a, b; } e; +__thread struct S s; +__thread struct S t[2]; + +void +foo (void) +{ + s = e; +} + +void +bar (void) +{ + t[1] = e; +} Jakub