> -----Original Message----- > From: Richard Biener <[email protected]> > Sent: Friday, October 15, 2021 12:31 PM > To: Tamar Christina <[email protected]> > Cc: [email protected]; Jakub Jelinek <[email protected]>; nd > <[email protected]> > Subject: Re: [PATCH] middle-end: fix de-optimizations with bitclear patterns > on signed values > > On Fri, 15 Oct 2021, Tamar Christina wrote: > > > Hi All, > > > > During testing after rebasing to commit I noticed a failing testcase > > with the bitmask compare patch. > > > > Consider the following C++ testcase: > > > > #include <compare> > > > > #define A __attribute__((noipa)) > > A bool f5 (double i, double j) { auto c = i <=> j; return c >= 0; } > > > > This turns into a comparison against chars, on systems where chars are > > signed the pattern inserts an unsigned convert such that it's able to > > do the transformation. > > > > i.e.: > > > > # RANGE [-1, 2] > > # c$_M_value_22 = PHI <-1(3), 0(2), 2(5), 1(4)> > > # RANGE ~[3, 254] > > _11 = (unsigned char) c$_M_value_22; > > _19 = _11 <= 1; > > # .MEM_24 = VDEF <.MEM_6(D)> > > D.10434 ={v} {CLOBBER}; > > # .MEM_14 = VDEF <.MEM_24> > > D.10407 ={v} {CLOBBER}; > > # VUSE <.MEM_14> > > return _19; > > > > instead of: > > > > # RANGE [-1, 2] > > # c$_M_value_5 = PHI <-1(3), 0(2), 2(5), 1(4)> > > # RANGE [-2, 2] > > _3 = c$_M_value_5 & -2; > > _19 = _3 == 0; > > # .MEM_24 = VDEF <.MEM_6(D)> > > D.10440 ={v} {CLOBBER}; > > # .MEM_14 = VDEF <.MEM_24> > > D.10413 ={v} {CLOBBER}; > > # VUSE <.MEM_14> > > return _19; > > > > This causes much worse codegen under -ffast-math due to phiops no > > longer recognizing the pattern. It turns out that phiopts > > spaceship_replacement is looking for the exact form that was just changed. > > > > Trying to get it to recognize the new form is not trivial as the > > transformation doesn't look to work when the thing it's pointing to is > > itself > a phi-node. > > What do you mean? Where it handles the BIT_AND it could also handle the > conversion, no? The later handling would probably more explicitely need to > distinguish between the BIT_AND and the conversion forms.
Looks like I misunderstood the code, it was looking at the uses not the defs of
the value.
--- inline copy of patch ---
The comments seems to suggest this code only checks for (res & ~1) == 0 but the
implementation seems to suggest it's broader.
As such I added a case to check to see if the value comparison we found is a
type cast. and strips away the type cast and continues.
In match.pd the typecasts are only added for signed comparisons to == 0 and != 0
which are then rewritten into comparisons with 1.
As such I only check for 1 and LE and GT, which is what match.pd would have
rewritten it to.
This fixes the regression but this is not code I 100% understand, since I don't
really know the semantics of the spaceship operator so would appreciate an extra
look.
Bootstrapped Regtested on aarch64-none-linux-gnu,
x86_64-pc-linux-gnu and no regressions.
Ok for master?
Thanks,
Tamar
gcc/ChangeLog:
* tree-ssa-phiopt.c (spaceship_replacement): Handle new canonical
codegen.
diff --git a/gcc/tree-ssa-phiopt.c b/gcc/tree-ssa-phiopt.c
index
0e339c46afa29fa97f90d9bc4394370cd9b4b396..65b25be3399b75d5e9cab0f78aa2340418571a33
100644
--- a/gcc/tree-ssa-phiopt.c
+++ b/gcc/tree-ssa-phiopt.c
@@ -2037,6 +2037,7 @@ spaceship_replacement (basic_block cond_bb, basic_block
middle_bb,
tree lhs, rhs;
gimple *orig_use_stmt = use_stmt;
tree orig_use_lhs = NULL_TREE;
+ bool is_canon = false;
int prec = TYPE_PRECISION (TREE_TYPE (phires));
if (is_gimple_assign (use_stmt)
&& gimple_assign_rhs_code (use_stmt) == BIT_AND_EXPR
@@ -2063,6 +2064,26 @@ spaceship_replacement (basic_block cond_bb, basic_block
middle_bb,
}
else if (is_gimple_assign (use_stmt))
{
+ /* Deal with if match.pd has rewritten the (res & ~1) == 0
+ into res <= 1 and has left a type-cast for signed types. */
+ if (gimple_assign_cast_p (use_stmt))
+ {
+ orig_use_lhs = gimple_assign_lhs (use_stmt);
+ if (SSA_NAME_OCCURS_IN_ABNORMAL_PHI (orig_use_lhs))
+ return false;
+ if (EDGE_COUNT (phi_bb->preds) != 4)
+ return false;
+ if (!TYPE_UNSIGNED (TREE_TYPE (orig_use_lhs)))
+ return false;
+ if (!single_imm_use (orig_use_lhs, &use_p, &use_stmt))
+ return false;
+ tree_code cmp;
+ if (is_gimple_assign (use_stmt)
+ && (cmp = gimple_assign_rhs_code (use_stmt))
+ && (cmp == LE_EXPR || cmp == GT_EXPR)
+ && wi::eq_p (wi::to_wide (gimple_assign_rhs2 (use_stmt)), 1))
+ is_canon = true;
+ }
if (gimple_assign_rhs_class (use_stmt) == GIMPLE_BINARY_RHS)
{
cmp = gimple_assign_rhs_code (use_stmt);
@@ -2099,7 +2120,9 @@ spaceship_replacement (basic_block cond_bb, basic_block
middle_bb,
|| !tree_fits_shwi_p (rhs)
|| !IN_RANGE (tree_to_shwi (rhs), -1, 1))
return false;
- if (orig_use_lhs)
+ /* If we're already in the canonical form we need to keep the original
+ comparison. */
+ if (orig_use_lhs && !is_canon)
{
if ((cmp != EQ_EXPR && cmp != NE_EXPR) || !integer_zerop (rhs))
return false;
@@ -2310,6 +2333,7 @@ spaceship_replacement (basic_block cond_bb, basic_block
middle_bb,
one_cmp = GT_EXPR;
enum tree_code res_cmp;
+
switch (cmp)
{
case EQ_EXPR:
@@ -2345,6 +2369,8 @@ spaceship_replacement (basic_block cond_bb, basic_block
middle_bb,
res_cmp = one_cmp == LT_EXPR ? GE_EXPR : LE_EXPR;
else if (integer_minus_onep (rhs))
res_cmp = one_cmp == LT_EXPR ? GT_EXPR : LT_EXPR;
+ else if (integer_onep (rhs) && is_canon)
+ res_cmp = GE_EXPR;
else
return false;
break;
@@ -2353,6 +2379,8 @@ spaceship_replacement (basic_block cond_bb, basic_block
middle_bb,
res_cmp = one_cmp == LT_EXPR ? LE_EXPR : GE_EXPR;
else if (integer_zerop (rhs))
res_cmp = one_cmp;
+ else if (integer_onep (rhs) && is_canon)
+ res_cmp = LE_EXPR;
else
return false;
break;
rb14938.patch
Description: rb14938.patch
