On 21/02/13 14:05, Andi Kleen wrote:
An auto generated program with a 6.4mio line asm statement gave with 4.7 and 4.8:xxx.c:6400017:1: internal compiler error: in account_size_time, at ipa-inline-analysis.c:601 The problem is that the inliner counts the number of lines in the asm statement and multiplies that with a weight. With the weight this overflows 32bit signed int, and triggers an assert for negative time. Fix this by limiting the number of lines to 1000 for asm cost estimation. The RTL backend also does similar multiplications for jump shortening. I haven't tried to address this, but presumably it's less likely to result in a failure. Passes test suite on x86_64-linux. Ok for 4.7 and 4.8? 2013-02-17 Andi Kleen <a...@linux.intel.com> * tree-inline.c (estimate_num_insns): Limit asm cost to 1000. diff --git a/gcc/tree-inline.c b/gcc/tree-inline.c index 2a1b692..7f8f2f2 100644 --- a/gcc/tree-inline.c +++ b/gcc/tree-inline.c @@ -3595,7 +3595,14 @@ estimate_num_insns (gimple stmt, eni_weights *weights) return 0; case GIMPLE_ASM: - return asm_str_count (gimple_asm_string (stmt)); + { + int count = asm_str_count (gimple_asm_string (stmt)); + /* 1000 means infinity. This avoids overflows later + with very long asm statements. */ + if (count > 1000) + count = 1000; + return count; + } case GIMPLE_RESX: /* This is either going to be an external function call with one
That doesn't sound enough, unless there is already code out there that respects this count. 1000 at 4 bytes per instruction is only 4k. More that small enough for the rest of the compiler to think that it could jump around such blocks cheaply.
I think a limit of 1M or more might be more appropriate. R.
