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..
---
I'll wait until next week before pushing this.
gcc/config/arm/arm.cc | 4 +++-
gcc/config/arm/arm.h | 6 +++++-
gcc/testsuite/gcc.target/arm/armv8_2-fp16-arith-1.c | 3 +--
3 files changed, 9 insertions(+), 4 deletions(-)
diff --git a/gcc/config/arm/arm.cc b/gcc/config/arm/arm.cc
index 670f487bcce..fccddb0e7bc 100644
--- a/gcc/config/arm/arm.cc
+++ b/gcc/config/arm/arm.cc
@@ -16218,7 +16218,9 @@ arm_select_cc_mode (enum rtx_code op, rtx x, rtx y)
case LE:
case GT:
case GE:
- return CCFPEmode;
+ return (flag_finite_math_only
+ ? CCFPmode
+ : CCFPEmode);
default:
gcc_unreachable ();
diff --git a/gcc/config/arm/arm.h b/gcc/config/arm/arm.h
index 8472b756127..08d3f0dae3d 100644
--- a/gcc/config/arm/arm.h
+++ b/gcc/config/arm/arm.h
@@ -2257,7 +2257,11 @@ extern int making_const_table;
#define SELECT_CC_MODE(OP, X, Y) arm_select_cc_mode (OP, X, Y)
-#define REVERSIBLE_CC_MODE(MODE) 1
+/* Floating-point modes cannot be reversed unless we don't care about
+ NaNs. */
+#define REVERSIBLE_CC_MODE(MODE) \
+ (flag_finite_math_only \
+ || !((MODE) == CCFPmode || (MODE) == CCFPEmode))
#define REVERSE_CONDITION(CODE,MODE) \
(((MODE) == CCFPmode || (MODE) == CCFPEmode) \
diff --git a/gcc/testsuite/gcc.target/arm/armv8_2-fp16-arith-1.c
b/gcc/testsuite/gcc.target/arm/armv8_2-fp16-arith-1.c
index 52b87376dc7..f3fea524809 100644
--- a/gcc/testsuite/gcc.target/arm/armv8_2-fp16-arith-1.c
+++ b/gcc/testsuite/gcc.target/arm/armv8_2-fp16-arith-1.c
@@ -106,8 +106,7 @@ TEST_CMP (greaterthanqual, >=, int16x8_t, float16x8_t)
/* { dg-final { scan-assembler-times {vdiv\.f16\ts[0-9]+, s[0-9]+, s[0-9]+} 13
} } */
/* For float16_t. */
-/* { dg-final { scan-assembler-times {vcmp\.f32\ts[0-9]+, s[0-9]+} 2 } } */
-/* { dg-final { scan-assembler-times {vcmpe\.f32\ts[0-9]+, s[0-9]+} 4 } } */
+/* { dg-final { scan-assembler-times {vcmp\.f32\ts[0-9]+, s[0-9]+} 6 } } */
/* For float16x4_t. */
/* { dg-final { scan-assembler-times {vceq\.f16\td[0-9]+, d[0-9]+} 2 } } */
--
2.34.1
