[LLVMdev] X86TargetLowering::LowerToBT

[Chris Sears]

Thanks Fiona and Stehpen, let me go over this. I've been reading the Intel
optimization manual and not Agner.

The bug actually has 2 parts:

(1) in LowerToBT() there is some logic that determines whether to use a BT
or a TEST but it gets the dividing line between the two wrong:

11994,11995c11994,11997
<     // Use BT if the immediate can't be encoded in a TEST instruction.
<     if (!isUInt<32>(AndRHSVal) && isPowerOf2_64(AndRHSVal)) {
---
>     // 16 bit mode:          32 bit mode:             64 bit mode:
>     //    TEST reg,imm  24b     TEST reg,imm   40b       TEST reg,imm
80b
>     //    BT   reg,imm  32b     BT   reg,imm   32b       BT   reg,imm
40b
>     if (AndRHSVal >= 256 && isPowerOf2_64(AndRHSVal)) {

(2) For an expression like (var & (1 << 37)) Clang emits LSHR/AND.
This actually confuses the hell out of LLVM generating fairly similar
assembly even with O3. It would be better for Clang to emit AND reg,
#(1<<37) and let LLVM recognize + optimize that.


You can argue whether BT should be emitted and I'll work through the
Intel/Agner docs. But LLVM is generating bad code from bad IR. I'll have a
patch for Clang later today and you can look these over.

BTW, macrofusion of TEST only applies for conditional jumps and not to CMOV.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20150122/d13ba24e/attachment.html>