Try to take advantage of the nan check behavior of fcmp.
Details
Diff Detail
Event Timeline
llvm/include/llvm/CodeGen/CodeGenCommonISel.h | ||
---|---|---|
224 | compare instructions which can test for nan? | |
llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp | ||
8200–8201 | Maybe drop the const - I expect the OrigTestMask rename will have caught any changes in the function | |
llvm/test/CodeGen/X86/is_fpclass.ll | ||
250 | Do you know if we have any legality/TLI callbacks that we could use to fallback to UseFP = false to avoid these regressions? |
Rebase
llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp | ||
---|---|---|
8200–8201 | And the const will keep it that way | |
llvm/test/CodeGen/X86/is_fpclass.ll | ||
250 | Is this a regression? I was thinking it might be beneficial to leave all the FP in the f registers rather than copying to integer. I don't know what I would check for here, the excess instructions seems to just be from the weird stackiness of x87. |
compare instructions which can test for nan?