The following avoids to increase hard register lifetime by not sinking loads from those.
Bootstrapped on x86_64-unknown-linux-gnu, testing in progress. I wonder if it would make sense to not treat hardregs as memory but instead rewrite them into SSA form but with SSA_NAME_OCCURS_IN_ABNORMAL_PHI set (which really tells that all SSA names of the underlying decl must be able to coalesce to the same register). Richard. 2014-06-24 Richard Biener <rguent...@suse.de> PR tree-optimization/61572 * tree-ssa-sink.c (statement_sink_location): Do not sink loads from hard registers. * gcc.target/i386/pr61572.c: New testcase. Index: gcc/tree-ssa-sink.c =================================================================== --- gcc/tree-ssa-sink.c (revision 211928) +++ gcc/tree-ssa-sink.c (working copy) @@ -374,6 +374,12 @@ statement_sink_location (gimple stmt, ba nearest to commondom. */ if (gimple_vuse (stmt)) { + /* Do not sink loads from hard registers. */ + if (gimple_assign_single_p (stmt) + && TREE_CODE (gimple_assign_rhs1 (stmt)) == VAR_DECL + && DECL_HARD_REGISTER (gimple_assign_rhs1 (stmt))) + return false; + imm_use_iterator imm_iter; use_operand_p use_p; basic_block found = NULL; Index: gcc/testsuite/gcc.target/i386/pr61572.c =================================================================== --- gcc/testsuite/gcc.target/i386/pr61572.c (revision 0) +++ gcc/testsuite/gcc.target/i386/pr61572.c (working copy) @@ -0,0 +1,46 @@ +/* { dg-do compile } */ +/* { dg-options "-O2" } */ + +struct autofs_sb_info +{ + int exp_timeout; +}; +void *f; +int g; +static int fn1 (struct autofs_sb_info *p1) +{ + int a, b; + a = ( + { + register __typeof__(0) c +#if defined __x86_64__ + asm("rdx") +#endif + ; + b = c; + int d; + __typeof__(0) e; + e = p1->exp_timeout / 1000; + switch (0) + default: + asm("" : "=a"(d) : "0"(e), ""(0)); + d; + }); + if (a) + return 1; + if (b) + p1->exp_timeout = 0; + return 0; +} + +int fn2 () +{ + struct autofs_sb_info *h = f; + switch (g) + { + case 0 ?: +0 : return fn1 (h); + default: + return 0; + } +}