| Message ID | 1291274381-26572-1-git-send-email-sebpop@gmail.com |
|---|---|
| State | New |
| Headers | show |
On Thu, 2 Dec 2010, Sebastian Pop wrote: > Hi, > > This fixes the compilation of the testcase of PR45297 on 32-bit > machines: the ICE occurs because we rewrite out-of-SSA a phi node for > which scev manages to analyze: in this code > > <bb 3>: > q_5 = &MEM[(void *)q_1 + 4294967288B]; > > <bb 4>: > # q_1 = PHI <q_3(2), q_5(3)> > if (q_1 != p_2(D)) > goto <bb 3>; > else > goto <bb 5>; > > scev analysis is able to determine that q1 = {q_3, +, -8}_1 and in the > same time, scev fails to analyze q5. And that triggers a rewrite of > q_5 out of SSA making the scev analysis of q1 impossible in the code > generator. > > I am testing this patch on amd64-linux, ok for trunk after test? Ok with ... > Thanks, > Sebastian > > 2010-12-01 Sebastian Pop <sebastian.pop@amd.com> > > PR middle-end/45297 > * tree-scalar-evolution.c (interpret_rhs_expr): Handle ADDR_EXPR > with MEM_REFs as POINTER_PLUS_EXPR. > --- > gcc/ChangeLog | 6 ++++++ > gcc/tree-scalar-evolution.c | 21 ++++++++++++++++++--- > 2 files changed, 24 insertions(+), 3 deletions(-) > > diff --git a/gcc/ChangeLog b/gcc/ChangeLog > index c9e7072..d38bc01 100644 > --- a/gcc/ChangeLog > +++ b/gcc/ChangeLog > @@ -1,5 +1,11 @@ > 2010-12-01 Sebastian Pop <sebastian.pop@amd.com> > > + PR middle-end/45297 > + * tree-scalar-evolution.c (interpret_rhs_expr): Handle ADDR_EXPR > + with MEM_REFs as POINTER_PLUS_EXPR. > + > +2010-12-01 Sebastian Pop <sebastian.pop@amd.com> > + > * graphite-sese-to-poly.c (analyze_drs_in_stmts): Fix set but > unused warning. > (rewrite_cross_bb_scalar_deps_out_of_ssa): Same. > diff --git a/gcc/tree-scalar-evolution.c b/gcc/tree-scalar-evolution.c > index 8a5797e..b65e295 100644 > --- a/gcc/tree-scalar-evolution.c > +++ b/gcc/tree-scalar-evolution.c > @@ -1698,7 +1698,7 @@ static tree > interpret_rhs_expr (struct loop *loop, gimple at_stmt, > tree type, tree rhs1, enum tree_code code, tree rhs2) > { > - tree res, chrec1, chrec2; > + tree res, chrec1, chrec2, expr; > > if (get_gimple_rhs_class (code) == GIMPLE_SINGLE_RHS) > { > @@ -1715,12 +1715,27 @@ interpret_rhs_expr (struct loop *loop, gimple at_stmt, > return chrec_convert (type, analyze_scalar_evolution (loop, rhs1), > at_stmt); > } > - > - return chrec_dont_know; > } > > switch (code) > { > + case ADDR_EXPR: > + /* Handle &MEM[ptr + CST] which is equivalent to POINTER_PLUS_EXPR. */ > + if (TREE_CODE (TREE_OPERAND (rhs1, 0)) != MEM_REF) > + { > + res = chrec_dont_know; > + break; > + } > + > + expr = TREE_OPERAND (rhs1, 0); > + rhs1 = TREE_OPERAND (expr, 0); > + rhs2 = TREE_OPERAND (expr, 1); > + type = TREE_TYPE (rhs1); > + STRIP_USELESS_TYPE_CONVERSION (rhs1); > + STRIP_USELESS_TYPE_CONVERSION (rhs2); The STRIP_USELESS_TYPE_CONVERSION removed and type retained (not set to that of rhs1). You can then also easily get rid of the new expr temporary. Thus, rhs1 = TREE_OPERAND (TREE_OPERAND (rhs1, 0), 0); rhs2 = TREE_OPERAND (TREE_OPERAND (rhs1, 0), 1); that should be all that is necessary. Thanks, Richard.
diff --git a/gcc/ChangeLog b/gcc/ChangeLog index c9e7072..d38bc01 100644 --- a/gcc/ChangeLog +++ b/gcc/ChangeLog @@ -1,5 +1,11 @@ 2010-12-01 Sebastian Pop <sebastian.pop@amd.com> + PR middle-end/45297 + * tree-scalar-evolution.c (interpret_rhs_expr): Handle ADDR_EXPR + with MEM_REFs as POINTER_PLUS_EXPR. + +2010-12-01 Sebastian Pop <sebastian.pop@amd.com> + * graphite-sese-to-poly.c (analyze_drs_in_stmts): Fix set but unused warning. (rewrite_cross_bb_scalar_deps_out_of_ssa): Same. diff --git a/gcc/tree-scalar-evolution.c b/gcc/tree-scalar-evolution.c index 8a5797e..b65e295 100644 --- a/gcc/tree-scalar-evolution.c +++ b/gcc/tree-scalar-evolution.c @@ -1698,7 +1698,7 @@ static tree interpret_rhs_expr (struct loop *loop, gimple at_stmt, tree type, tree rhs1, enum tree_code code, tree rhs2) { - tree res, chrec1, chrec2; + tree res, chrec1, chrec2, expr; if (get_gimple_rhs_class (code) == GIMPLE_SINGLE_RHS) { @@ -1715,12 +1715,27 @@ interpret_rhs_expr (struct loop *loop, gimple at_stmt, return chrec_convert (type, analyze_scalar_evolution (loop, rhs1), at_stmt); } - - return chrec_dont_know; } switch (code) { + case ADDR_EXPR: + /* Handle &MEM[ptr + CST] which is equivalent to POINTER_PLUS_EXPR. */ + if (TREE_CODE (TREE_OPERAND (rhs1, 0)) != MEM_REF) + { + res = chrec_dont_know; + break; + } + + expr = TREE_OPERAND (rhs1, 0); + rhs1 = TREE_OPERAND (expr, 0); + rhs2 = TREE_OPERAND (expr, 1); + type = TREE_TYPE (rhs1); + STRIP_USELESS_TYPE_CONVERSION (rhs1); + STRIP_USELESS_TYPE_CONVERSION (rhs2); + + /* Fall through. */ + case POINTER_PLUS_EXPR: chrec1 = analyze_scalar_evolution (loop, rhs1); chrec2 = analyze_scalar_evolution (loop, rhs2);
Hi, This fixes the compilation of the testcase of PR45297 on 32-bit machines: the ICE occurs because we rewrite out-of-SSA a phi node for which scev manages to analyze: in this code <bb 3>: q_5 = &MEM[(void *)q_1 + 4294967288B]; <bb 4>: # q_1 = PHI <q_3(2), q_5(3)> if (q_1 != p_2(D)) goto <bb 3>; else goto <bb 5>; scev analysis is able to determine that q1 = {q_3, +, -8}_1 and in the same time, scev fails to analyze q5. And that triggers a rewrite of q_5 out of SSA making the scev analysis of q1 impossible in the code generator. I am testing this patch on amd64-linux, ok for trunk after test? Thanks, Sebastian 2010-12-01 Sebastian Pop <sebastian.pop@amd.com> PR middle-end/45297 * tree-scalar-evolution.c (interpret_rhs_expr): Handle ADDR_EXPR with MEM_REFs as POINTER_PLUS_EXPR. --- gcc/ChangeLog | 6 ++++++ gcc/tree-scalar-evolution.c | 21 ++++++++++++++++++--- 2 files changed, 24 insertions(+), 3 deletions(-)