]> git.ipfire.org Git - thirdparty/gcc.git/commit
arm: Only reverse FP inequalities when -ffinite-math-only [PR110796...]
authorRichard Earnshaw <rearnsha@arm.com>
Fri, 28 Mar 2025 12:59:03 +0000 (12:59 +0000)
committerRichard Earnshaw <rearnsha@arm.com>
Wed, 7 May 2025 16:49:02 +0000 (17:49 +0100)
commit0a339746e7646bacf2c8aa5512268d23660f26f9
tree67ea79f4d5dcaec8b7f722ab359fa2380527e608
parente0c783f31b39a56cbfcc27d84633f9254b5b508d
arm: Only reverse FP inequalities when -ffinite-math-only [PR110796...]

On Arm we have been failing to fully implement support for IEEE NaNs
in inequality comparisons because we have allowed reversing of
inequalities in a way that allows SELECT_CC_MODE to produce different
answers.  For example, the reverse of GT is UNLE, but if we pass these
two RTL codes to SELECT_CC_MODE, the former will return CCFPEmode,
while the latter CCFPmode.

It would be possible to allow fully reversible FPmodes, but to do so
would involve adding yet more RTL codes, something like NOT_GT and
NOT_UNLE, for the cases we cannot currently reverse.  NOT_GT would
then have the same condition code generation as UNLT, but the same
mode selection as GT.

In the mean time, we need to restrict REVERSIBLE_CC_MODE to
non-floating modes unless we are compiling with -ffinite-math-only.  In
that case we can continue to reverse the comparisons, but now we want
to always select CCFPmode as there's no need to consider the exception
raising cases.

PR target/110796
PR target/118446

gcc/ChangeLog:

* config/arm/arm.h (REVERSIBLE_CC_MODE): FP modes are only
reversible if flag_finite_math_only.
* config/arm/arm.cc (arm_select_cc_mode): Return CCFPmode for all
FP comparisons if flag_finite_math_only.

gcc/testsuite/ChangeLog:

* gcc.target/arm/armv8_2-fp16-arith-1.c: Adjust due to no-longer
emitting VCMPE when -ffast-math..
gcc/config/arm/arm.cc
gcc/config/arm/arm.h
gcc/testsuite/gcc.target/arm/armv8_2-fp16-arith-1.c