diff options
| author | Roman Lebedev <lebedev.ri@gmail.com> | 2018-06-15 09:56:39 +0000 |
|---|---|---|
| committer | Roman Lebedev <lebedev.ri@gmail.com> | 2018-06-15 09:56:39 +0000 |
| commit | 9c17dad8f2528ebb9505ab2e549cc25d307d0f2b (patch) | |
| tree | c459ec7d41c6609a88179da60e424fa565c78d84 /llvm/lib | |
| parent | aa8587d1fcfba6b54a826d2590ddf477261b6800 (diff) | |
| download | bcm5719-llvm-9c17dad8f2528ebb9505ab2e549cc25d307d0f2b.tar.gz bcm5719-llvm-9c17dad8f2528ebb9505ab2e549cc25d307d0f2b.zip | |
[AMDGPU] Recognize x & ((1 << y) - 1) pattern.
Summary:
As a followup for D48007.
Since we already handle `x << (bitwidth - y) >> (bitwidth - y)` pattern,
which does not have ub for both the edge cases (`y == 0`, `y == bitwidth`),
i think also handling a pattern that is ub for `y == bitwidth` should be fine.
Reviewers: nhaehnle, bogner, tstellar, arsenm
Reviewed By: arsenm
Subscribers: arsenm, kzhuravl, wdng, yaxunl, dstuttard, tpr, t-tye, llvm-commits
Tags: #amdgpu
Differential Revision: https://reviews.llvm.org/D48010
llvm-svn: 334816
Diffstat (limited to 'llvm/lib')
| -rw-r--r-- | llvm/lib/Target/AMDGPU/AMDGPUInstructions.td | 7 |
1 files changed, 7 insertions, 0 deletions
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUInstructions.td b/llvm/lib/Target/AMDGPU/AMDGPUInstructions.td index 22a7a9c44a3..65d72fcb46f 100644 --- a/llvm/lib/Target/AMDGPU/AMDGPUInstructions.td +++ b/llvm/lib/Target/AMDGPU/AMDGPUInstructions.td @@ -126,6 +126,7 @@ def or_oneuse : HasOneUseBinOp<or>; def xor_oneuse : HasOneUseBinOp<xor>; } // Properties = [SDNPCommutative, SDNPAssociative] +def add_oneuse : HasOneUseBinOp<add>; def sub_oneuse : HasOneUseBinOp<sub>; def srl_oneuse : HasOneUseBinOp<srl>; @@ -682,6 +683,12 @@ multiclass BFEPattern <Instruction UBFE, Instruction SBFE, Instruction MOV> { (UBFE $src, $rshift, (MOV (i32 (IMMPopCount $mask)))) >; + // x & ((1 << y) - 1) + def : AMDGPUPat < + (and i32:$src, (add_oneuse (shl_oneuse 1, i32:$width), -1)), + (UBFE $src, (i32 0), $width) + >; + // x & (-1 >> (bitwidth - y)) def : AMDGPUPat < (and i32:$src, (srl_oneuse -1, (sub 32, i32:$width))), |
