From: Icenowy Zheng Date: Tue, 30 Jun 2026 16:11:42 +0000 (+0800) Subject: RISC-V: Change initial value for fmin/fmax autovec reduce [PR126049] X-Git-Url: http://git.ipfire.org/cgi-bin/gitweb.cgi?a=commitdiff_plain;h=ad45faeb14a4acb551ecbb61bc1b0e1f6ce500b0;p=thirdparty%2Fgcc.git RISC-V: Change initial value for fmin/fmax autovec reduce [PR126049] Currently the auto-vectorization of C fmin()/fmax() uses infinity/-infinity as the initial value for reduction, which introduces bogus infinity values when iterating over an array with only NaNs. As all C fmin()/fmax(), RV F/D fmin/fmax, RVV vfmin/vfmax and RVV vfredmin/vfredmax are implementing the IEEE754 minimumNumber or maximumNumber behavior, an initial value of NaN is more suitable than infinity when reducing the vector (if the input vector is all NaN, the result will still be NaN and if the input vector contains non-NaN elements they will cover the initial NaN). Change the initial value during reduction from corresponding inf to a quiet NaN. PR target/126049 gcc/ChangeLog: * config/riscv/autovec.md: Change fmin/fmax reduction initial value from inf/-inf to NaN for proper semantics of corresponding C funtion. gcc/testsuite/ChangeLog: * gcc.target/riscv/rvv/autovec/pr126049.c: New. --- diff --git a/gcc/config/riscv/autovec.md b/gcc/config/riscv/autovec.md index 31584f4918c..f6ec19d0025 100644 --- a/gcc/config/riscv/autovec.md +++ b/gcc/config/riscv/autovec.md @@ -2436,7 +2436,7 @@ "TARGET_VECTOR && !HONOR_SNANS (mode)" { REAL_VALUE_TYPE rv; - real_inf (&rv, true); + real_nan (&rv, "", 1, VOIDmode); rtx f = const_double_from_real_value (rv, mode); riscv_vector::expand_reduction (UNSPEC_REDUC_MAX, UNSPEC_REDUC_MAX_VL0_SAFE, @@ -2451,7 +2451,7 @@ "TARGET_VECTOR && !HONOR_SNANS (mode)" { REAL_VALUE_TYPE rv; - real_inf (&rv, false); + real_nan (&rv, "", 1, VOIDmode); rtx f = const_double_from_real_value (rv, mode); riscv_vector::expand_reduction (UNSPEC_REDUC_MIN, UNSPEC_REDUC_MIN_VL0_SAFE, diff --git a/gcc/testsuite/gcc.target/riscv/rvv/autovec/pr126049.c b/gcc/testsuite/gcc.target/riscv/rvv/autovec/pr126049.c new file mode 100644 index 00000000000..1fab4466525 --- /dev/null +++ b/gcc/testsuite/gcc.target/riscv/rvv/autovec/pr126049.c @@ -0,0 +1,26 @@ +/* { dg-do run } */ +/* { dg-require-effective-target riscv_v_ok } */ +/* { dg-options "-march=rv64gcv -mabi=lp64d -O2 -ftree-vectorize" } */ + +#define NAN (__builtin_nan("")) + +__attribute__ ((noipa)) float +maxx (float *arr, unsigned int sz) +{ + float val = NAN; + unsigned int i; + + for (i = 0; i < sz; i++) + val = __builtin_fmaxf (val, arr[i]); + + return val; +} + +int +main () +{ + float arr[] = { NAN, NAN, NAN, NAN }; + float res = maxx (arr, 4); + if (!__builtin_isnan (res)) + __builtin_abort (); +}