6502/llvm-6502
1
0
Fork 0
mirror of https://github.com/c64scene-ar/llvm-6502.git synced 2025-04-05 17:39:16 +00:00
Code Issues Projects Releases Wiki Activity

For ARM subtargets with useNEONForSinglePrecisionFP, double count uses

Browse Source 
of the floating point types less than 64-bits. It's somewhat of a temporary
hack but forces more accurate modeling of register pressure and results
in fewer spills.


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@123811 91177308-0d34-0410-b5e6-96231b3b80d8
This commit is contained in:
Andrew Trick 2011-01-19 02:35:27 +00:00
parent 7fa75ce11d
commit 32cec0a756
1 changed files with 16 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

16
lib/Target/ARM/ARMISelLowering.cpp
View File

@ -700,6 +700,16 @@ ARMTargetLowering::ARMTargetLowering(TargetMachine &TM)
benefitFromCodePlacementOpt = true;
}
// FIXME: It might make sense to define the representative register class as the
// nearest super-register that has a non-null superset. For example, DPR_VFP2 is
// a super-register of SPR, and DPR is a superset if DPR_VFP2. Consequently,
// SPR's representative would be DPR_VFP2. This should work well if register
// pressure tracking were modified such that a register use would increment the
// pressure of the register class's representative and all of it's super
// classes' representatives transitively. We have not implemented this because
// of the difficulty prior to coalescing of modeling operand register classes
// due to the common occurence of cross class copies and subregister insertions
// and extractions.
std::pair<const TargetRegisterClass*, uint8_t>
ARMTargetLowering::findRepresentativeClass(EVT VT) const{
const TargetRegisterClass *RRC = 0;
@ -713,6 +723,12 @@ ARMTargetLowering::findRepresentativeClass(EVT VT) const{
case MVT::f32: case MVT::f64: case MVT::v8i8: case MVT::v4i16:
case MVT::v2i32: case MVT::v1i64: case MVT::v2f32:
RRC = ARM::DPRRegisterClass;
// When NEON is used for SP, only half of the register file is available
// because operations that define both SP and DP results will be constrained
// to the VFP2 class (D0-D15). We currently model this constraint prior to
// coalescing by double-counting the SP regs. See the FIXME above.
if (Subtarget->useNEONForSinglePrecisionFP())
Cost = 2;
break;
case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
case MVT::v4f32: case MVT::v2f64: