[llvm] 1d1515a - [X86] Add an isel pattern for (i8 (trunc (i16 (bitconvert (v16i1 X))))) to avoid an extra EXTRACT_SUBREG

[Craig Topper via llvm-commits]

Author: Craig Topper
Date: 2020-08-25T18:20:43-07:00
New Revision: 1d1515a9e28c7f32af548584ebed66cd8caae84f

URL: https://github.com/llvm/llvm-project/commit/1d1515a9e28c7f32af548584ebed66cd8caae84f
DIFF: https://github.com/llvm/llvm-project/commit/1d1515a9e28c7f32af548584ebed66cd8caae84f.diff

LOG: [X86] Add an isel pattern for (i8 (trunc (i16 (bitconvert (v16i1 X))))) to avoid an extra EXTRACT_SUBREG

Since we can only copy to GR32 we had to EXTRACT from GR32, but
we would first go to GR16 and then the truncate would extra again
to GR8. This adds a special case to go directly from GR32 to GR8.
This would eventually get cleaned up, but though maybe we should
avoid doing it in the first place. Our k-register handling is weird
and we could probably stand to have some more special ISD nodes
for the conversions so the i32 type would be explicit.

Added: 
    

Modified: 
    llvm/lib/Target/X86/X86InstrAVX512.td

Removed: 
    


################################################################################
diff  --git a/llvm/lib/Target/X86/X86InstrAVX512.td b/llvm/lib/Target/X86/X86InstrAVX512.td
index 0514a3a3611c..e1eb66484124 100644
--- a/llvm/lib/Target/X86/X86InstrAVX512.td
+++ b/llvm/lib/Target/X86/X86InstrAVX512.td
@@ -2884,6 +2884,8 @@ def : Pat<(v16i1 (bitconvert (i16 GR16:$src))),
           (COPY_TO_REGCLASS (i32 (INSERT_SUBREG (IMPLICIT_DEF), GR16:$src, sub_16bit)), VK16)>;
 def : Pat<(i16 (bitconvert (v16i1 VK16:$src))),
           (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS VK16:$src, GR32)), sub_16bit)>;
+def : Pat<(i8 (trunc (i16 (bitconvert (v16i1 VK16:$src))))),
+          (EXTRACT_SUBREG (i32 (COPY_TO_REGCLASS VK16:$src, GR32)), sub_8bit)>;
 
 def : Pat<(v8i1 (bitconvert (i8 GR8:$src))),
           (COPY_TO_REGCLASS (i32 (INSERT_SUBREG (IMPLICIT_DEF), GR8:$src, sub_8bit)), VK8)>;