On Thu, Oct 23, 2014 at 9:30 AM, Jakub Jelinek <ja...@redhat.com> wrote:
> Hi!
>
> As I wrote in the PR, vt_stack_adjustments can often compute wrong offsets,
> because it never considers pops with autoinc addressing, which can lead
> either to wrong debug info, or turning off -fvar-tracking altogether for
> a function on which that issue resulted in stack depth inconsistencies on
> the edges.
>
> Here are some stats from --enable-checking=yes,rtl cc1plus bootstrapped
> with/without the patch (without the patch got then rebuilt stage3 with the
> patch, i.e. just var-tracking.o and cc1plus-checksum.o got recompiled, so
> I'm comparing identical code, different debug info):
>
> x86_64-linux cc1plus built without the patch:
> cov% samples cumul
> 0.0 506230/38% 506230/38%
> 0..10 10327/0% 516557/39%
> 11..20 12390/0% 528947/39%
> 21..30 31265/2% 560212/42%
> 31..40 18775/1% 578987/43%
> 41..50 20631/1% 599618/45%
> 51..60 24921/1% 624539/47%
> 61..70 40959/3% 665498/50%
> 71..80 23771/1% 689269/52%
> 81..90 41771/3% 731040/55%
> 91..99 81667/6% 812707/61%
> 100 510564/38% 1323271/100%
>
> x86_64-linux cc1plus built with the patch:
> cov% samples cumul
> 0.0 382214/28% 382214/28%
> 0..10 13100/0% 395314/29%
> 11..20 14568/1% 409882/30%
> 21..30 33708/2% 443590/33%
> 31..40 21927/1% 465517/35%
> 41..50 23924/1% 489441/36%
> 51..60 28736/2% 518177/39%
> 61..70 45847/3% 564024/42%
> 71..80 29284/2% 593308/44%
> 81..90 52085/3% 645393/48%
> 91..99 99971/7% 745364/56%
> 100 577907/43% 1323271/100%
>
> i686-linux cc1plus built without the patch:
> cov% samples cumul
> 0.0 631348/48% 631348/48%
> 0..10 7764/0% 639112/48%
> 11..20 9690/0% 648802/49%
> 21..30 25036/1% 673838/51%
> 31..40 16113/1% 689951/52%
> 41..50 19753/1% 709704/54%
> 51..60 14563/1% 724267/55%
> 61..70 34093/2% 758360/58%
> 71..80 17450/1% 775810/59%
> 81..90 31339/2% 807149/61%
> 91..99 60368/4% 867517/66%
> 100 437548/33% 1305065/100%
>
> i686-linux cc1plus built with the patch:
> cov% samples cumul
> 0.0 377352/28% 377352/28%
> 0..10 16077/1% 393429/30%
> 11..20 15390/1% 408819/31%
> 21..30 31790/2% 440609/33%
> 31..40 23889/1% 464498/35%
> 41..50 29267/2% 493765/37%
> 51..60 22902/1% 516667/39%
> 61..70 45629/3% 562296/43%
> 71..80 29511/2% 591807/45%
> 81..90 50536/3% 642343/49%
> 91..99 93584/7% 735927/56%
> 100 569138/43% 1305065/100%
>
> .debug_info/.debug_loc sizes in bytes:
> x86_64-linux cc1plus without patch .debug_info 75411710, .debug_loc 75421077
> x86_64-linux cc1plus with patch .debug_info 78498790, .debug_loc 90530117
> i686-linux cc1plus without patch .debug_info 59921183, .debug_loc 37823166
> i686-linux cc1plus with patch .debug_info 63009554, .debug_loc 59535100
>
> I've also performed instrumented bootstraps/regtests (x86_64-linux and
> i686-linux), where I've logged in how many functions the result of
> vt_stack_adjustments differed between the bad old way and new way.
> In both the bootstraps/regtests, it affected 16892 32-bit and 6646 64-bit
> functions, in all cases it was old way giving up and new way succeeding.
>
> Not adding a testcase, as the one in the PR failed to produce proper debug
> info only in 4.8 (then got latent), there already are some guality
> improvements with the
> patch:
> -FAIL: gcc.dg/guality/pr54693-2.c -O3 -fomit-frame-pointer
> -funroll-all-loops -finline-functions line 21 i == v + 1
> -FAIL: gcc.dg/guality/pr54693-2.c -O3 -fomit-frame-pointer -funroll-loops
> line 21 i == v + 1
> on x86_64 and:
> -FAIL: gcc.dg/guality/pr54519-1.c -O2 line 20 y == 25
> -FAIL: gcc.dg/guality/pr54519-1.c -O2 line 20 z == 6
> -FAIL: gcc.dg/guality/pr54519-1.c -O2 line 23 y == 117
> -FAIL: gcc.dg/guality/pr54519-1.c -O2 line 23 z == 8
> -FAIL: gcc.dg/guality/pr54519-1.c -O3 -fomit-frame-pointer line 20 x == 36
> -FAIL: gcc.dg/guality/pr54519-1.c -O3 -fomit-frame-pointer line 20 y == 25
> -FAIL: gcc.dg/guality/pr54519-1.c -O3 -fomit-frame-pointer line 20 z == 6
> -FAIL: gcc.dg/guality/pr54519-1.c -O3 -g line 20 x == 36
> -FAIL: gcc.dg/guality/pr54519-1.c -O3 -g line 20 y == 25
> -FAIL: gcc.dg/guality/pr54519-1.c -O3 -g line 20 z == 6
> -FAIL: gcc.dg/guality/pr54519-3.c -O2 line 20 x == 36
> -FAIL: gcc.dg/guality/pr54519-3.c -O2 line 20 y == 25
> -FAIL: gcc.dg/guality/pr54519-3.c -O2 line 20 z == 6
> -FAIL: gcc.dg/guality/pr54519-3.c -O2 line 23 x == 98
> -FAIL: gcc.dg/guality/pr54519-3.c -O2 line 23 y == 117
> -FAIL: gcc.dg/guality/pr54519-3.c -O2 line 23 z == 8
> -FAIL: gcc.dg/guality/pr54519-3.c -O2 -flto line 20 x == 36
> -FAIL: gcc.dg/guality/pr54519-3.c -O2 -flto line 23 x == 98
> -FAIL: gcc.dg/guality/pr54519-3.c -O2 -flto -flto-partition=none line 20 x
> == 36
> -FAIL: gcc.dg/guality/pr54519-3.c -O2 -flto -flto-partition=none line 23 x
> == 98
> -FAIL: gcc.dg/guality/pr54519-3.c -O3 -fomit-frame-pointer line 20 x == 36
> -FAIL: gcc.dg/guality/pr54519-3.c -O3 -fomit-frame-pointer line 20 y == 25
> -FAIL: gcc.dg/guality/pr54519-3.c -O3 -fomit-frame-pointer line 20 z == 6
> -FAIL: gcc.dg/guality/pr54519-3.c -O3 -fomit-frame-pointer line 23 x == 98
> -FAIL: gcc.dg/guality/pr54519-3.c -O3 -fomit-frame-pointer line 23 y == 117
> -FAIL: gcc.dg/guality/pr54519-3.c -O3 -fomit-frame-pointer line 23 z == 8
> -FAIL: gcc.dg/guality/pr54519-3.c -O3 -g line 20 x == 36
> -FAIL: gcc.dg/guality/pr54519-3.c -O3 -g line 20 y == 25
> -FAIL: gcc.dg/guality/pr54519-3.c -O3 -g line 20 z == 6
> -FAIL: gcc.dg/guality/pr54519-3.c -O3 -g line 23 x == 98
> -FAIL: gcc.dg/guality/pr54519-3.c -O3 -g line 23 y == 117
> -FAIL: gcc.dg/guality/pr54519-3.c -O3 -g line 23 z == 8
> -FAIL: gcc.dg/guality/pr54693-2.c -O3 -fomit-frame-pointer
> -funroll-all-loops -finline-functions line 21 i == v + 1
> -FAIL: gcc.dg/guality/pr54693-2.c -O3 -fomit-frame-pointer -funroll-loops
> line 21 i == v + 1
> on i686.
>
> Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
>
> Not sure about branches, to some extent this must be a serious debug info
> quality regression, since the problem got significantly worse with omitting
> frame pointer by default on i686, and/or shrink wrapping and simple_return
> changes where there are more pops in the middle of functions. So with a short
> patch we could significantly improve debug info. The risk is compile time
> regressions, if we gave up on var-tracking, we threw away all the hard tracked
> debug stmts/debug insns, but don't need to costly propagate the values through
> the dataflow.
I think if the patch does not hit 4.9.2 (which I guess it won't) then both
4.8 and 4.9 are too mature to get this kind of patch.
Richard.
> 2014-10-23 Jakub Jelinek <ja...@redhat.com>
>
> PR debug/63623
> * var-tracking.c (stack_adjust_offset_pre_post_cb): New function.
> (stack_adjust_offset_pre_post): Use it through for_each_inc_dec,
> instead of only handling autoinc in dest if it is a MEM.
> (vt_stack_adjustments): Fix up formatting.
>
> --- gcc/var-tracking.c.jj 2014-09-01 09:43:58.000000000 +0200
> +++ gcc/var-tracking.c 2014-10-22 22:08:42.985717561 +0200
> @@ -700,6 +700,39 @@ static void vt_add_function_parameters (
> static bool vt_initialize (void);
> static void vt_finalize (void);
>
> +/* Callback for stack_adjust_offset_pre_post, called via for_each_inc_dec.
> */
> +
> +static int
> +stack_adjust_offset_pre_post_cb (rtx, rtx op, rtx dest, rtx src, rtx srcoff,
> + void *arg)
> +{
> + if (dest != stack_pointer_rtx)
> + return 0;
> +
> + switch (GET_CODE (op))
> + {
> + case PRE_INC:
> + case PRE_DEC:
> + ((HOST_WIDE_INT *)arg)[0] -= INTVAL (srcoff);
> + return 0;
> + case POST_INC:
> + case POST_DEC:
> + ((HOST_WIDE_INT *)arg)[1] -= INTVAL (srcoff);
> + return 0;
> + case PRE_MODIFY:
> + case POST_MODIFY:
> + /* We handle only adjustments by constant amount. */
> + gcc_assert (GET_CODE (src) == PLUS
> + && CONST_INT_P (XEXP (src, 1))
> + && XEXP (src, 0) == stack_pointer_rtx);
> + ((HOST_WIDE_INT *)arg)[GET_CODE (op) == POST_MODIFY]
> + -= INTVAL (XEXP (src, 1));
> + return 0;
> + default:
> + gcc_unreachable ();
> + }
> +}
> +
> /* Given a SET, calculate the amount of stack adjustment it contains
> PRE- and POST-modifying stack pointer.
> This function is similar to stack_adjust_offset. */
> @@ -725,68 +758,12 @@ stack_adjust_offset_pre_post (rtx patter
> *post += INTVAL (XEXP (src, 1));
> else
> *post -= INTVAL (XEXP (src, 1));
> + return;
> }
> - else if (MEM_P (dest))
> - {
> - /* (set (mem (pre_dec (reg sp))) (foo)) */
> - src = XEXP (dest, 0);
> - code = GET_CODE (src);
> -
> - switch (code)
> - {
> - case PRE_MODIFY:
> - case POST_MODIFY:
> - if (XEXP (src, 0) == stack_pointer_rtx)
> - {
> - rtx val = XEXP (XEXP (src, 1), 1);
> - /* We handle only adjustments by constant amount. */
> - gcc_assert (GET_CODE (XEXP (src, 1)) == PLUS &&
> - CONST_INT_P (val));
> -
> - if (code == PRE_MODIFY)
> - *pre -= INTVAL (val);
> - else
> - *post -= INTVAL (val);
> - break;
> - }
> - return;
> -
> - case PRE_DEC:
> - if (XEXP (src, 0) == stack_pointer_rtx)
> - {
> - *pre += GET_MODE_SIZE (GET_MODE (dest));
> - break;
> - }
> - return;
> -
> - case POST_DEC:
> - if (XEXP (src, 0) == stack_pointer_rtx)
> - {
> - *post += GET_MODE_SIZE (GET_MODE (dest));
> - break;
> - }
> - return;
> -
> - case PRE_INC:
> - if (XEXP (src, 0) == stack_pointer_rtx)
> - {
> - *pre -= GET_MODE_SIZE (GET_MODE (dest));
> - break;
> - }
> - return;
> -
> - case POST_INC:
> - if (XEXP (src, 0) == stack_pointer_rtx)
> - {
> - *post -= GET_MODE_SIZE (GET_MODE (dest));
> - break;
> - }
> - return;
> -
> - default:
> - return;
> - }
> - }
> + HOST_WIDE_INT res[2] = { 0, 0 };
> + for_each_inc_dec (pattern, stack_adjust_offset_pre_post_cb, res);
> + *pre += res[0];
> + *post += res[1];
> }
>
> /* Given an INSN, calculate the amount of stack adjustment it contains
> @@ -836,10 +813,10 @@ vt_stack_adjustments (void)
>
> /* Initialize entry block. */
> VTI (ENTRY_BLOCK_PTR_FOR_FN (cfun))->visited = true;
> - VTI (ENTRY_BLOCK_PTR_FOR_FN (cfun))->in.stack_adjust =
> - INCOMING_FRAME_SP_OFFSET;
> - VTI (ENTRY_BLOCK_PTR_FOR_FN (cfun))->out.stack_adjust =
> - INCOMING_FRAME_SP_OFFSET;
> + VTI (ENTRY_BLOCK_PTR_FOR_FN (cfun))->in.stack_adjust
> + = INCOMING_FRAME_SP_OFFSET;
> + VTI (ENTRY_BLOCK_PTR_FOR_FN (cfun))->out.stack_adjust
> + = INCOMING_FRAME_SP_OFFSET;
>
> /* Allocate stack for back-tracking up CFG. */
> stack = XNEWVEC (edge_iterator, n_basic_blocks_for_fn (cfun) + 1);
>
> Jakub
