Re: [PATCH] match.pd: extend the MIN/MAX narrowing through a cast to vectors

On 2026-08-09 00:18, Andrew Pinski wrote:
> On Tue, Aug 4, 2026 at 3:00 AM <[email protected]> wrote:
>>
>> From: Kyrylo Tkachov <[email protected]>
>>
>> The rule folding (type) minmax ((wide_type) a, (wide_type) b) to
>> minmax (a, b) is restricted to scalars.  Extension is monotone, so it
>> commutes with the comparison and the outer truncation is exact, and the
>> argument is lanewise, so it holds for vectors unchanged.
>>
>> The vector case additionally needs the narrow operation to be available,
>> otherwise vector lowering would scalarise what used to be a single wide
>> instruction.
>>
>>    typedef int  v2si __attribute__((vector_size (8)));
>>    typedef long v2di __attribute__((vector_size (16)));
>>    v2si h (v2si a, v2si b)
>>    {
>>      v2di x = __builtin_convertvector (a, v2di);
>>      v2di y = __builtin_convertvector (b, v2di);
>>      return __builtin_convertvector (x < y ? x : y, v2si);
>>    }
>>
>> aarch64 -O3 before:
>>
>>          sshll   v0.2d, v0.2s, 0
>>          sshll   v1.2d, v1.2s, 0
>>          cmgt    v2.2d, v1.2d, v0.2d
>>          bsl     v2.16b, v0.16b, v1.16b
>>          xtn     v0.2s, v2.2d
>>
>> after:
>>
>>          smin    v0.2s, v0.2s, v1.2s
>>
>> Require the wide vector MIN or MAX to have one use.  A shared wide
>> result would otherwise gain a second narrow operation.
>>
>> Bootstrapped and tested on aarch64-none-linux-gnu.
>> Ok for trunk?
>> Thanks,
>> Kyrill
>>
>> gcc/ChangeLog:
>>
>>          * match.pd ((type) minmax ((wide_type) a, (wide_type) b)): Accept
>>          vector types, using element_precision and requiring the narrow
>>          operation to be supported.
>>
>> gcc/testsuite/ChangeLog:
>>
>>          * g++.dg/tree-ssa/vec-narrow-1.C: New test.
>>          * g++.dg/tree-ssa/vec-narrow-minmax-2.C: New test.
>>
>> Signed-off-by: Kyrylo Tkachov <[email protected]>
>> ---
>>   gcc/match.pd                                  | 20 +++++++++++++------
>>   gcc/testsuite/g++.dg/tree-ssa/vec-narrow-1.C  | 11 ++++++++++
>>   .../g++.dg/tree-ssa/vec-narrow-minmax-2.C     | 18 +++++++++++++++++
>>   3 files changed, 43 insertions(+), 6 deletions(-)
>>   create mode 100644 gcc/testsuite/g++.dg/tree-ssa/vec-narrow-1.C
>>   create mode 100644 gcc/testsuite/g++.dg/tree-ssa/vec-narrow-minmax-2.C
>>
>> diff --git a/gcc/match.pd b/gcc/match.pd
>> index e0ec2626842..adbdda91b51 100644
>> --- a/gcc/match.pd
>> +++ b/gcc/match.pd
>> @@ -4649,17 +4649,25 @@ DEFINE_INT_AND_FLOAT_ROUND_FN (RINT)
>>
>>   /* (type) minmax ((wide_type) a, (wide_type) b) -> minmax (a, b)
>>      when type matches the type of a and b, and wide_type is a wider
>> -   type with the same signedness as type. */
>> +   type with the same signedness as type.  Extension is monotone, so it
>> +   commutes with the comparison, and the truncation is then exact.  The
>> +   same holds lanewise for a vector, where the narrow operation has to be
>> +   available so that lowering does not scalarise it.  The wide operation
>> +   must become dead before a narrow vector operation is introduced.  */
>>   (for minmax (min max)
>> +     MINMAX (MIN_EXPR MAX_EXPR)
>>    (simplify
>> -  (convert (minmax:c (convert@2 @0) (convert@3 @1)))
>> -  (if (INTEGRAL_TYPE_P (type)
>> -       && INTEGRAL_TYPE_P (TREE_TYPE (@2))
>> +  (convert (minmax:c@4 (convert@2 @0) (convert@3 @1)))
>> +  (if (ANY_INTEGRAL_TYPE_P (type)
>> +       && ANY_INTEGRAL_TYPE_P (TREE_TYPE (@2))
>>          && types_match (type, TREE_TYPE (@0))
>>          && types_match (type, TREE_TYPE (@1))
>>          && types_match (TREE_TYPE (@2), TREE_TYPE (@3))
>> -       && TYPE_PRECISION (TREE_TYPE (@2)) > TYPE_PRECISION (type)
>> -       && TYPE_UNSIGNED (TREE_TYPE (@2)) == TYPE_UNSIGNED (type))
>> +       && element_precision (TREE_TYPE (@2)) > element_precision (type)
>> +       && TYPE_UNSIGNED (TREE_TYPE (@2)) == TYPE_UNSIGNED (type)
>> +       && (!VECTOR_TYPE_P (type)
>> +          || (single_use (@4)
>> +              && target_supports_op_p (type, MINMAX, optab_vector))))
>>      (minmax @0 @1))))
>>
>>   /* max (a, a + CST) -> a + CST where CST is positive.  */
>> diff --git a/gcc/testsuite/g++.dg/tree-ssa/vec-narrow-1.C b/gcc/testsuite/g++.dg/tree-ssa/vec-narrow-1.C
>> new file mode 100644
>> index 00000000000..de738e33c30
>> --- /dev/null
>> +++ b/gcc/testsuite/g++.dg/tree-ssa/vec-narrow-1.C
>> @@ -0,0 +1,11 @@
>> +// { dg-do compile }
>> +// { dg-options "-O2 -fdump-tree-optimized" }
>> +// Extension is monotone, so it commutes with the comparison and the outer
>> +// truncation is exact.  The argument is lanewise, so a widened vector
>> +// MIN/MAX feeding a truncating conversion narrows.
>> +typedef int  v2si __attribute__((vector_size (8)));
>> +typedef long long v2di __attribute__((vector_size (16)));
>> +v2si f (v2si a, v2si b)
>> +{ v2di x = __builtin_convertvector (a, v2di), y = __builtin_convertvector (b, v2di);
>> +  return __builtin_convertvector (x < y ? x : y, v2si); }
>> +// { dg-final { scan-tree-dump-not "vector\\(2\\) long" "optimized" } }
> 
> 
> This testcase fails on x86_64.

This test case fails for arm-none-eabi target thumb/arch=armv7ve+simd/tune=cortex-a7/float-abi=hard/fpu=auto (even after Jakubs tweaks in r17-3353-ge9d5aaf572e8d1):

Testing tree-ssa/vec-narrow-1.C,  -std=gnu++20
doing compile
Executing on host: /build/r17-3152-gbc8c49094cdfbb/bin/arm-none-eabi-g++  /build/gcc_src/gcc/testsuite/g++.dg/tree-ssa/vec-narrow-1.C  -mthumb -march=armv7ve+simd -mtune=cortex-a7 -mfloat-abi=hard -mfpu=auto   -fdiagnostics-plain-output   -fmessage-length=0  -std=gnu++20 -O2 -fdump-tree-optimized  -S     -o vec-narrow-1.s    (timeout = 800)
spawn -ignore SIGHUP /build/r17-3152-gbc8c49094cdfbb/bin/arm-none-eabi-g++ /build/gcc_src/gcc/testsuite/g++.dg/tree-ssa/vec-narrow-1.C -mthumb -march=armv7ve+simd -mtune=cortex-a7 -mfloat-abi=hard -mfpu=auto -fdiagnostics-plain-output -fmessage-length=0 -std=gnu++20 -O2 -fdump-tree-optimized -S -o vec-narrow-1.s
pid is 185 -185
pid is -1
output is  status 0
PASS: g++.dg/tree-ssa/vec-narrow-1.C  -std=gnu++20 (test for excess errors)
FAIL: g++.dg/tree-ssa/vec-narrow-1.C  -std=gnu++20  scan-tree-dump-not optimized "vector\\(2\\) long"

Content of vec-narrow-1.C.272t.optimized is:

;; Function f (_Z1fDv2_iS_, funcdef_no=0, decl_uid=8035, cgraph_uid=1, symbol_order=0)

v2si f (v2si a, v2si b)
{
   vector(2) long long int _1;
   int _8;
   long long int _9;
   int _10;
   long long int _11;
   int _12;
   long long int _13;
   int _14;
   long long int _15;
   long long int _16;
   long long int _17;
   vector(4) int _18;
   vector(2) int _19;

   <bb 2> [local count: 1073741824]:
   _8 = BIT_FIELD_REF <a_2(D), 32, 0>;
   _9 = (long long int) _8;
   _10 = BIT_FIELD_REF <a_2(D), 32, 32>;
   _11 = (long long int) _10;
   _12 = BIT_FIELD_REF <b_4(D), 32, 0>;
   _13 = (long long int) _12;
   _14 = BIT_FIELD_REF <b_4(D), 32, 32>;
   _15 = (long long int) _14;
   _16 = MIN_EXPR <_9, _13>;
   _17 = MIN_EXPR <_11, _15>;
   _1 = {_16, _17};
   _18 = VEC_PACK_TRUNC_EXPR <_1, { 0, 0 }>;
   _19 = BIT_FIELD_REF <_18, 64, 0>;
   return _19;

}


> 
>> diff --git a/gcc/testsuite/g++.dg/tree-ssa/vec-narrow-minmax-2.C b/gcc/testsuite/g++.dg/tree-ssa/vec-narrow-minmax-2.C
>> new file mode 100644
>> index 00000000000..d7fb7f06f16
>> --- /dev/null
>> +++ b/gcc/testsuite/g++.dg/tree-ssa/vec-narrow-minmax-2.C
>> @@ -0,0 +1,18 @@
>> +// { dg-do compile }
>> +// { dg-options "-O2 -fdump-tree-optimized" }
>> +
>> +typedef int v2si __attribute__((vector_size (8)));
>> +typedef long long v2di __attribute__((vector_size (16)));
>> +
>> +v2si
>> +f (v2si a, v2si b, v2di *p)
>> +{
>> +  v2di x = __builtin_convertvector (a, v2di);
>> +  v2di y = __builtin_convertvector (b, v2di);
>> +  v2di z = x < y ? x : y;
>> +  *p = z;
>> +  return __builtin_convertvector (z, v2si);
>> +}
>> +
>> +// A shared wide MIN must not gain a second narrow MIN.
>> +// { dg-final { scan-tree-dump-times "MIN_EXPR" 1 "optimized" } }

And this testcase fails for all arm-none-eabi targets I test (even after Jakubs tweaks in r17-3353-ge9d5aaf572e8d1):

Testing tree-ssa/vec-narrow-minmax-2.C,  -std=gnu++20
doing compile
Executing on host: /build/r17-3152-gbc8c49094cdfbb/bin/arm-none-eabi-g++  /build/gcc_src/gcc/testsuite/g++.dg/tree-ssa/vec-narrow-minmax-2.C  -mthumb -march=armv7ve -mtune=cortex-a7 -mfloat-abi=soft -mfpu=auto   -fdiagnostics-plain-output   -fmessage-length=0  -std=gnu++20 -O2 -fdump-tree-optimized -Wno-psabi  -S     -o vec-narrow-minmax-2.s    (timeout = 800)
spawn -ignore SIGHUP /build/r17-3152-gbc8c49094cdfbb/bin/arm-none-eabi-g++ /build/gcc_src/gcc/testsuite/g++.dg/tree-ssa/vec-narrow-minmax-2.C -mthumb -march=armv7ve -mtune=cortex-a7 -mfloat-abi=soft -mfpu=auto -fdiagnostics-plain-output -fmessage-length=0 -std=gnu++20 -O2 -fdump-tree-optimized -Wno-psabi -S -o vec-narrow-minmax-2.s
pid is 185 -185
pid is -1
output is  status 0
PASS: g++.dg/tree-ssa/vec-narrow-minmax-2.C  -std=gnu++20 (test for excess errors)
g++.dg/tree-ssa/vec-narrow-minmax-2.C  -std=gnu++20 : pattern found 4 times
FAIL: g++.dg/tree-ssa/vec-narrow-minmax-2.C  -std=gnu++20  scan-tree-dump-times optimized "MIN_EXPR" 1


Content of vec-narrow-minmax-2.C.272t.optimized is:

;; Function f (_Z1fDv2_iS_PDv2_x, funcdef_no=0, decl_uid=4948, cgraph_uid=1, symbol_order=0)

v2si f (v2si a, v2si b, v2di * p)
{
   v2di z;
   v2si _9;
   int _10;
   long long int _11;
   int _12;
   long long int _13;
   int _14;
   long long int _15;
   int _16;
   long long int _17;
   long long int _18;
   long long int _19;
   int _20;
   int _21;

   <bb 2> [local count: 1073741824]:
   _10 = BIT_FIELD_REF <a_1(D), 32, 0>;
   _11 = (long long int) _10;
   _12 = BIT_FIELD_REF <a_1(D), 32, 32>;
   _13 = (long long int) _12;
   _14 = BIT_FIELD_REF <b_3(D), 32, 0>;
   _15 = (long long int) _14;
   _16 = BIT_FIELD_REF <b_3(D), 32, 32>;
   _17 = (long long int) _16;
   _18 = MIN_EXPR <_11, _15>;
   _19 = MIN_EXPR <_13, _17>;
   z_5 = {_18, _19};
   *p_7(D) = z_5;
   _20 = MIN_EXPR <_10, _14>;
   _21 = MIN_EXPR <_12, _16>;
   _9 = {_20, _21};
   return _9;

}


Kind regards,
Torbjörn

>> --
>> 2.50.1 (Apple Git-155)
>>