Re: [PATCH 2/2] Enable elimination of zext/sext

[Richard Biener <richard.guenther@gmail.com>]

On Mon, Jul 7, 2014 at 8:55 AM, Kugan <kugan.vivekanandarajah@linaro.org> wrote:
>> For -fwrapv I don't see why you'd get into trouble ever, the VRP computation
>> should be well aware of the -fwrapv semantics and the value ranges should
>> reflect that.
>>
>> For -fno-strict-overflow, I have no idea since it is very weirdly defined.
>>
>> In any case, for your example above, the loop is always well defined,
>> because for char/short a++ is performed as:
>> a = (short) ((int) a + 1)
>> So, if the patch turns it into infinite loop, with -Os -fno-strict-overflow
>> or -Os, it is simply a problem with the patch.  VR [1, 32768] looks correct,
>> a++ is performed only if a is >= 0, therefore before addition [0, 32767].
>> But from VR [1, 32768] you can't optimize away the sign extension, make sure
>> you don't have there off-by-one?

I have fixed the above bug yesterday.

>> It would be nice if the patch contained some testcases, it is easy
>> to construct testcases where you have arbitrary VRs on some SSA_NAMEs,
>> you just need something to stick the VR on, so you can do something like:
>> type foo (type a)
>> {
>>   if (a < VR_min + 1 || a > VR_max + 1) return; // If VR_min is type minimum or VR_max type maximum this needs to be adjusted of course.
>>   a = a + 1;
>>   // now you can try some cast that your optimization would try to optimize
>>   return a;
>> }
>> Or void bar (type a) { a = (a & mask) + bias; (or similarly) }
>> Make sure to cover the boundary cases, where VR minimum or maximum still
>> allow optimizing away zero and/or sign extensions, and another case where
>> they are +- 1 and already don't allow it.
>
>
> Hi Jakub,
>
> For -fwrapv, it is due to how PROMOTE_MODE is defined in arm back-end.
> In the test-case, a function (which has signed char return type) returns
> -1 in one of the paths. ARM PROMOTE_MODE changes that to 255 and relies
> on zero/sign extension generated by RTL again for the correct value. I
> saw some other targets also defining similar think. I am therefore
> skipping removing zero/sign extension if the ssa variable can be set to
> negative integer constants.

Hm?  I think you should rather check that you are removing a
sign-/zero-extension - PROMOTE_MODE tells you if it will sign- or
zero-extend.  Definitely

+  /* In some architectures, negative integer constants are truncated and
+     sign changed with target defined PROMOTE_MODE macro. This will impact
+     the value range seen here and produce wrong code if zero/sign extensions
+     are eliminated. Therefore, return false if this SSA can have negative
+     integers.  */
+  if (is_gimple_assign (stmt)
+      && (TREE_CODE_CLASS (gimple_assign_rhs_code (stmt)) == tcc_unary))
+    {
+      tree rhs1 = gimple_assign_rhs1 (stmt);
+      if (TREE_CODE (rhs1) == INTEGER_CST
+         && !TYPE_UNSIGNED (TREE_TYPE (ssa))
+         && tree_int_cst_compare (rhs1, integer_zero_node) == -1)
+       return false;

looks completely bogus ... (an unary op with a constant operand?)

instead you want to do sth like

  mode = TYPE_MODE (TREE_TYPE (ssa));
  rhs_uns = TYPE_UNSIGNED (TREE_TYPE (ssa));
  PROMOTE_MODE (mode, rhs_uns, TREE_TYPE (ssa));

instead of initializing rhs_uns from ssas type.  That is, if
PROMOTE_MODE tells you to promote _not_ according to ssas sign then
honor that.

> As for the -fno-strict-overflow case, if the variables overflows, in VRP
> dumps, I see +INF(OVF), but the value range stored in ssa has TYPE_MAX.
> We therefore should limit the comparison to (TYPE_MIN < VR_MIN && VR_MAX
> < TYPE_MAX) instead of (TYPE_MIN <= VR_MIN && VR_MAX <= TYPE_MAX) when
> checking to be sure that this is not the overflowing case. Attached
> patch changes this.

I don't think that's necessary - the overflow cases happen only when
that overflow has undefined behavior, thus any valid program will have
values <= MAX.

Richard.

> I have bootstrapped on x86_64-unknown-linux-gnu and regression tested
> for x86_64-unknown-linux-gnu, arm-none-linux-gnueabi (using qemu),
> aarch64_be-none-elf (Foundation model), aarch64-none-elf
> --with-abi=ilp32 (Foundation model) and s390x-ibm-linux (64bit, using
> qemu) with no new regression.
>
> Is this OK?
>
> Thanks,
> Kugan
>
> gcc/
> 2014-07-07  Kugan Vivekanandarajah  <kuganv@linaro.org>
>
>         * calls.c (precompute_arguments): Check is_promoted_for_type
>         and set the promoted mode.
>         (is_promoted_for_type): New function.
>         (expand_expr_real_1): Check is_promoted_for_type
>         and set the promoted mode.
>         * expr.h (is_promoted_for_type): New function definition.
>         * cfgexpand.c (expand_gimple_stmt_1): Call emit_move_insn if
>         SUBREG is promoted with SRP_SIGNED_AND_UNSIGNED.
>
>
> gcc/testsuite
>
> 2014-07-07  Kugan Vivekanandarajah  <kuganv@linaro.org>
>
>         * gcc.dg/zero_sign_ext_test.c: New test.