On Sat, Feb 10, 2024 at 10:21 PM Toon Moene <t...@moene.org> wrote: > > I managed to try this patch on aarch64-linux-gnu: > > This is the test run without your patch: > > https://gcc.gnu.org/pipermail/gcc-testresults/2024-February/807637.html > > And this is the "result" with your patch: > > https://gcc.gnu.org/pipermail/gcc-testresults/2024-February/807645.html
Interesting .. this looks like a miscompile of stage2. Note that the Linaro CI was happy: https://patchwork.sourceware.org/project/gcc/patch/20240209102649.09c543858...@sourceware.org/ Richard. > For me, as for you, it works for x86_64-linux-gnu: > > https://gcc.gnu.org/pipermail/gcc-testresults/2024-February/807609.html > > I hope this helps. > > Kind regards, > Toon Moene. > > On 2/9/24 11:26, Richard Biener wrote: > > The following allows a base term to be derived from an existing > > MEM_EXPR, notably the points-to set of a MEM_REF base. For the > > testcase in the PR this helps RTL DSE elide stores to a stack > > temporary. This covers pointers to NONLOCAL which can be mapped > > to arg_base_value, helping to disambiguate against other special > > bases (ADDRESS) as well as PARM_DECL accesses. > > > > Bootstrapped and tested on x86_64-unknown-linux-gnu. > > > > This is an attempt to recover some of the losses from dumbing down > > find_base_{term,value}. I did give my ideas how to properly do > > this during stage1 a start, I will post a short incomplete RFC series > > later today. > > > > OK for trunk? > > > > I've included all languages in testing and also tested with -m32 but > > details of RTL alias analysis might escape me ... > > > > Thanks, > > Richard. > > > > PR rtl-optimization/113597 > > * alias.cc (find_base_term): Add argument for the whole mem > > and derive a base term from its MEM_EXPR. > > (true_dependence_1): Pass down the MEMs to find_base_term. > > (write_dependence_p): Likewise. > > (may_alias_p): Likewise. > > --- > > gcc/alias.cc | 43 ++++++++++++++++++++++++++++++++++++------- > > 1 file changed, 36 insertions(+), 7 deletions(-) > > > > diff --git a/gcc/alias.cc b/gcc/alias.cc > > index 6fad4b29d31..e33c56b0e80 100644 > > --- a/gcc/alias.cc > > +++ b/gcc/alias.cc > > @@ -40,6 +40,9 @@ along with GCC; see the file COPYING3. If not see > > #include "rtl-iter.h" > > #include "cgraph.h" > > #include "ipa-utils.h" > > +#include "stringpool.h" > > +#include "value-range.h" > > +#include "tree-ssanames.h" > > > > /* The aliasing API provided here solves related but different problems: > > > > @@ -190,6 +193,10 @@ static struct { > > arguments, since we do not know at this level whether accesses > > based on different arguments can alias. The ADDRESS has id 0. > > > > + This is solely useful to disambiguate against other ADDRESS > > + bases as we know incoming pointers cannot point to local > > + stack, frame or argument space. > > + > > 2. stack_pointer_rtx, frame_pointer_rtx, hard_frame_pointer_rtx > > (if distinct from frame_pointer_rtx) and arg_pointer_rtx. > > Each of these rtxes has a separate ADDRESS associated with it, > > @@ -2113,12 +2120,34 @@ find_base_term (rtx x, vec<std::pair<cselib_val *, > > to avoid visiting them multiple times. We unwind that changes here. > > */ > > > > static rtx > > -find_base_term (rtx x) > > +find_base_term (rtx x, const_rtx mem = NULL_RTX) > > { > > auto_vec<std::pair<cselib_val *, struct elt_loc_list *>, 32> > > visited_vals; > > rtx res = find_base_term (x, visited_vals); > > for (unsigned i = 0; i < visited_vals.length (); ++i) > > visited_vals[i].first->locs = visited_vals[i].second; > > + if (!res && mem && MEM_EXPR (mem)) > > + { > > + tree base = get_base_address (MEM_EXPR (mem)); > > + if (TREE_CODE (base) == PARM_DECL > > + && DECL_RTL_SET_P (base)) > > + /* We need to look at how we expanded a PARM_DECL. It might be in > > + the argument space (UNIQUE_BASE_VALUE_ARGP) or it might > > + be spilled (UNIQUE_BASE_VALUE_FP/UNIQUE_BASE_VALUE_HFP). */ > > + res = find_base_term (DECL_RTL (base)); > > + else if (TREE_CODE (base) == MEM_REF > > + && TREE_CODE (TREE_OPERAND (base, 0)) == SSA_NAME > > + && SSA_NAME_PTR_INFO (TREE_OPERAND (base, 0))) > > + { > > + auto pt = &SSA_NAME_PTR_INFO (TREE_OPERAND (base, 0))->pt; > > + if (pt->nonlocal > > + && !pt->anything > > + && !pt->escaped > > + && !pt->ipa_escaped > > + && bitmap_empty_p (pt->vars)) > > + res = arg_base_value; > > + } > > + } > > return res; > > } > > > > @@ -3035,13 +3064,13 @@ true_dependence_1 (const_rtx mem, machine_mode > > mem_mode, rtx mem_addr, > > if (MEM_ADDR_SPACE (mem) != MEM_ADDR_SPACE (x)) > > return true; > > > > - base = find_base_term (x_addr); > > + base = find_base_term (x_addr, x); > > if (base && (GET_CODE (base) == LABEL_REF > > || (GET_CODE (base) == SYMBOL_REF > > && CONSTANT_POOL_ADDRESS_P (base)))) > > return false; > > > > - rtx mem_base = find_base_term (true_mem_addr); > > + rtx mem_base = find_base_term (true_mem_addr, mem); > > if (! base_alias_check (x_addr, base, true_mem_addr, mem_base, > > GET_MODE (x), mem_mode)) > > return false; > > @@ -3142,7 +3171,7 @@ write_dependence_p (const_rtx mem, > > if (MEM_ADDR_SPACE (mem) != MEM_ADDR_SPACE (x)) > > return true; > > > > - base = find_base_term (true_mem_addr); > > + base = find_base_term (true_mem_addr, mem); > > if (! writep > > && base > > && (GET_CODE (base) == LABEL_REF > > @@ -3150,7 +3179,7 @@ write_dependence_p (const_rtx mem, > > && CONSTANT_POOL_ADDRESS_P (base)))) > > return false; > > > > - rtx x_base = find_base_term (true_x_addr); > > + rtx x_base = find_base_term (true_x_addr, x); > > if (! base_alias_check (true_x_addr, x_base, true_mem_addr, base, > > GET_MODE (x), GET_MODE (mem))) > > return false; > > @@ -3265,8 +3294,8 @@ may_alias_p (const_rtx mem, const_rtx x) > > if (MEM_ADDR_SPACE (mem) != MEM_ADDR_SPACE (x)) > > return true; > > > > - rtx x_base = find_base_term (x_addr); > > - rtx mem_base = find_base_term (mem_addr); > > + rtx x_base = find_base_term (x_addr, x); > > + rtx mem_base = find_base_term (mem_addr, mem); > > if (! base_alias_check (x_addr, x_base, mem_addr, mem_base, > > GET_MODE (x), GET_MODE (mem_addr))) > > return false; > > -- > Toon Moene - e-mail: t...@moene.org - phone: +31 346 214290 > Saturnushof 14, 3738 XG Maartensdijk, The Netherlands >
