Current for SAE instructions we only allow _MM_FROUND_CUR_DIRECTION(bit 2) or _MM_FROUND_NO_EXC(bit 3) to be used as the immediate passed to the inrinsics. But these instructions don't perform rounding so _MM_FROUND_CUR_DIRECTION is just sort of a default placeholder when you don't want to suppress exceptions. Using _MM_FROUND_NO_EXC by itself is really bit equivalent to (_MM_FROUND_NO_EXC | _MM_FROUND_TO_NEAREST_INT) since _MM_FROUND_TO_NEAREST_INT is 0. Since we aren't rounding on these instructions we should also accept (_MM_FROUND_CUR_DIRECTION | _MM_FROUND_NO_EXC) as equivalent to (_MM_FROUND_NO_EXC). icc allows this, but gcc does not.
Differential Revision: https://reviews.llvm.org/D67289
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@371430
91177308-0d34-0410-b5e6-
96231b3b80d8
return false;
};
auto isRoundModeSAE = [](SDValue Rnd) {
- if (auto *C = dyn_cast<ConstantSDNode>(Rnd))
- return C->getAPIntValue() == X86::STATIC_ROUNDING::NO_EXC;
+ if (auto *C = dyn_cast<ConstantSDNode>(Rnd)) {
+ unsigned RC = C->getZExtValue();
+ if (RC & X86::STATIC_ROUNDING::NO_EXC) {
+ // Clear the NO_EXC bit and check remaining bits.
+ RC ^= X86::STATIC_ROUNDING::NO_EXC;
+ // As a convenience we allow no other bits or explicitly
+ // current direction.
+ return RC == 0 || RC == X86::STATIC_ROUNDING::CUR_DIRECTION;
+ }
+ }
return false;
};
; CHECK: # %bb.0:
; CHECK-NEXT: vgetexppd {sae}, %zmm0, %zmm0
; CHECK-NEXT: ret{{[l|q]}}
- %res = call <8 x double> @llvm.x86.avx512.mask.getexp.pd.512(<8 x double> %a0, <8 x double> zeroinitializer, i8 -1, i32 8)
+ %res = call <8 x double> @llvm.x86.avx512.mask.getexp.pd.512(<8 x double> %a0, <8 x double> zeroinitializer, i8 -1, i32 12)
ret <8 x double> %res
}
declare <8 x double> @llvm.x86.avx512.mask.getexp.pd.512(<8 x double>, <8 x double>, i8, i32) nounwind readnone