maxjhandsome
/
llvm-project
forked from OSchip/llvm-project
1
0
Fork 0
Code Issues Pull Requests Packages Releases Wiki Activity

Fix SSE/AVX integer comparison patterns to understand that all integer vector loads are promoted to i64 vector loads so patterns need a bitconvert. Also slightly simplify the AVX2 variable shift patterns by using the predefined bitconvert pattern fragments. 

llvm-svn: 144896
This commit is contained in:
Craig Topper 2011-11-17 07:49:38 +00:00
parent 15b2498e88
commit f41e1d0246
1 changed files with 42 additions and 24 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

66
llvm/lib/Target/X86/X86InstrSSE.td
View File

@ -3891,28 +3891,34 @@ let Predicates = [HasAVX] in {
def : Pat<(v16i8 (X86pcmpeqb VR128:$src1, VR128:$src2)),
(VPCMPEQBrr VR128:$src1, VR128:$src2)>;
def : Pat<(v16i8 (X86pcmpeqb VR128:$src1, (memop addr:$src2))),
def : Pat<(v16i8 (X86pcmpeqb VR128:$src1,
(bc_v16i8 (memopv2i64 addr:$src2)))),
(VPCMPEQBrm VR128:$src1, addr:$src2)>;
def : Pat<(v8i16 (X86pcmpeqw VR128:$src1, VR128:$src2)),
(VPCMPEQWrr VR128:$src1, VR128:$src2)>;
def : Pat<(v8i16 (X86pcmpeqw VR128:$src1, (memop addr:$src2))),
def : Pat<(v8i16 (X86pcmpeqw VR128:$src1,
(bc_v8i16 (memopv2i64 addr:$src2)))),
(VPCMPEQWrm VR128:$src1, addr:$src2)>;
def : Pat<(v4i32 (X86pcmpeqd VR128:$src1, VR128:$src2)),
(VPCMPEQDrr VR128:$src1, VR128:$src2)>;
def : Pat<(v4i32 (X86pcmpeqd VR128:$src1, (memop addr:$src2))),
def : Pat<(v4i32 (X86pcmpeqd VR128:$src1,
(bc_v4i32 (memopv2i64 addr:$src2)))),
(VPCMPEQDrm VR128:$src1, addr:$src2)>;
def : Pat<(v16i8 (X86pcmpgtb VR128:$src1, VR128:$src2)),
(VPCMPGTBrr VR128:$src1, VR128:$src2)>;
def : Pat<(v16i8 (X86pcmpgtb VR128:$src1, (memop addr:$src2))),
def : Pat<(v16i8 (X86pcmpgtb VR128:$src1,
(bc_v16i8 (memopv2i64 addr:$src2)))),
(VPCMPGTBrm VR128:$src1, addr:$src2)>;
def : Pat<(v8i16 (X86pcmpgtw VR128:$src1, VR128:$src2)),
(VPCMPGTWrr VR128:$src1, VR128:$src2)>;
def : Pat<(v8i16 (X86pcmpgtw VR128:$src1, (memop addr:$src2))),
def : Pat<(v8i16 (X86pcmpgtw VR128:$src1,
(bc_v8i16 (memopv2i64 addr:$src2)))),
(VPCMPGTWrm VR128:$src1, addr:$src2)>;
def : Pat<(v4i32 (X86pcmpgtd VR128:$src1, VR128:$src2)),
(VPCMPGTDrr VR128:$src1, VR128:$src2)>;
def : Pat<(v4i32 (X86pcmpgtd VR128:$src1, (memop addr:$src2))),
def : Pat<(v4i32 (X86pcmpgtd VR128:$src1,
(bc_v4i32 (memopv2i64 addr:$src2)))),
(VPCMPGTDrm VR128:$src1, addr:$src2)>;
}
@ -3932,28 +3938,34 @@ let Predicates = [HasAVX2] in {
def : Pat<(v32i8 (X86pcmpeqb VR256:$src1, VR256:$src2)),
(VPCMPEQBYrr VR256:$src1, VR256:$src2)>;
def : Pat<(v32i8 (X86pcmpeqb VR256:$src1, (memop addr:$src2))),
def : Pat<(v32i8 (X86pcmpeqb VR256:$src1,
(bc_v32i8 (memopv4i64 addr:$src2)))),
(VPCMPEQBYrm VR256:$src1, addr:$src2)>;
def : Pat<(v16i16 (X86pcmpeqw VR256:$src1, VR256:$src2)),
(VPCMPEQWYrr VR256:$src1, VR256:$src2)>;
def : Pat<(v16i16 (X86pcmpeqw VR256:$src1, (memop addr:$src2))),
def : Pat<(v16i16 (X86pcmpeqw VR256:$src1,
(bc_v16i16 (memopv4i64 addr:$src2)))),
(VPCMPEQWYrm VR256:$src1, addr:$src2)>;
def : Pat<(v8i32 (X86pcmpeqd VR256:$src1, VR256:$src2)),
(VPCMPEQDYrr VR256:$src1, VR256:$src2)>;
def : Pat<(v8i32 (X86pcmpeqd VR256:$src1, (memop addr:$src2))),
def : Pat<(v8i32 (X86pcmpeqd VR256:$src1,
(bc_v8i32 (memopv4i64 addr:$src2)))),
(VPCMPEQDYrm VR256:$src1, addr:$src2)>;
def : Pat<(v32i8 (X86pcmpgtb VR256:$src1, VR256:$src2)),
(VPCMPGTBYrr VR256:$src1, VR256:$src2)>;
def : Pat<(v32i8 (X86pcmpgtb VR256:$src1, (memop addr:$src2))),
def : Pat<(v32i8 (X86pcmpgtb VR256:$src1,
(bc_v32i8 (memopv4i64 addr:$src2)))),
(VPCMPGTBYrm VR256:$src1, addr:$src2)>;
def : Pat<(v16i16 (X86pcmpgtw VR256:$src1, VR256:$src2)),
(VPCMPGTWYrr VR256:$src1, VR256:$src2)>;
def : Pat<(v16i16 (X86pcmpgtw VR256:$src1, (memop addr:$src2))),
def : Pat<(v16i16 (X86pcmpgtw VR256:$src1,
(bc_v16i16 (memopv4i64 addr:$src2)))),
(VPCMPGTWYrm VR256:$src1, addr:$src2)>;
def : Pat<(v8i32 (X86pcmpgtd VR256:$src1, VR256:$src2)),
(VPCMPGTDYrr VR256:$src1, VR256:$src2)>;
def : Pat<(v8i32 (X86pcmpgtd VR256:$src1, (memop addr:$src2))),
def : Pat<(v8i32 (X86pcmpgtd VR256:$src1,
(bc_v8i32 (memopv4i64 addr:$src2)))),
(VPCMPGTDYrm VR256:$src1, addr:$src2)>;
}
@ -3975,28 +3987,34 @@ let Constraints = "$src1 = $dst" in {
let Predicates = [HasSSE2] in {
def : Pat<(v16i8 (X86pcmpeqb VR128:$src1, VR128:$src2)),
(PCMPEQBrr VR128:$src1, VR128:$src2)>;
def : Pat<(v16i8 (X86pcmpeqb VR128:$src1, (memop addr:$src2))),
def : Pat<(v16i8 (X86pcmpeqb VR128:$src1,
(bc_v16i8 (memopv2i64 addr:$src2)))),
(PCMPEQBrm VR128:$src1, addr:$src2)>;
def : Pat<(v8i16 (X86pcmpeqw VR128:$src1, VR128:$src2)),
(PCMPEQWrr VR128:$src1, VR128:$src2)>;
def : Pat<(v8i16 (X86pcmpeqw VR128:$src1, (memop addr:$src2))),
def : Pat<(v8i16 (X86pcmpeqw VR128:$src1,
(bc_v8i16 (memopv2i64 addr:$src2)))),
(PCMPEQWrm VR128:$src1, addr:$src2)>;
def : Pat<(v4i32 (X86pcmpeqd VR128:$src1, VR128:$src2)),
(PCMPEQDrr VR128:$src1, VR128:$src2)>;
def : Pat<(v4i32 (X86pcmpeqd VR128:$src1, (memop addr:$src2))),
def : Pat<(v4i32 (X86pcmpeqd VR128:$src1,
(bc_v4i32 (memopv2i64 addr:$src2)))),
(PCMPEQDrm VR128:$src1, addr:$src2)>;
def : Pat<(v16i8 (X86pcmpgtb VR128:$src1, VR128:$src2)),
(PCMPGTBrr VR128:$src1, VR128:$src2)>;
def : Pat<(v16i8 (X86pcmpgtb VR128:$src1, (memop addr:$src2))),
def : Pat<(v16i8 (X86pcmpgtb VR128:$src1,
(bc_v16i8 (memopv2i64 addr:$src2)))),
(PCMPGTBrm VR128:$src1, addr:$src2)>;
def : Pat<(v8i16 (X86pcmpgtw VR128:$src1, VR128:$src2)),
(PCMPGTWrr VR128:$src1, VR128:$src2)>;
def : Pat<(v8i16 (X86pcmpgtw VR128:$src1, (memop addr:$src2))),
def : Pat<(v8i16 (X86pcmpgtw VR128:$src1,
(bc_v8i16 (memopv2i64 addr:$src2)))),
(PCMPGTWrm VR128:$src1, addr:$src2)>;
def : Pat<(v4i32 (X86pcmpgtd VR128:$src1, VR128:$src2)),
(PCMPGTDrr VR128:$src1, VR128:$src2)>;
def : Pat<(v4i32 (X86pcmpgtd VR128:$src1, (memop addr:$src2))),
def : Pat<(v4i32 (X86pcmpgtd VR128:$src1,
(bc_v4i32 (memopv2i64 addr:$src2)))),
(PCMPGTDrm VR128:$src1, addr:$src2)>;
}
@ -7754,29 +7772,29 @@ let Predicates = [HasAVX2] in {
(VPSRAVDYrr VR256:$src1, VR256:$src2)>;
def : Pat<(v4i32 (shl (v4i32 VR128:$src1),
(v4i32 (bitconvert (memopv2i64 addr:$src2))))),
(bc_v4i32 (memopv2i64 addr:$src2)))),
(VPSLLVDrm VR128:$src1, addr:$src2)>;
def : Pat<(v2i64 (shl (v2i64 VR128:$src1), (memopv2i64 addr:$src2))),
(VPSLLVQrm VR128:$src1, addr:$src2)>;
def : Pat<(v4i32 (srl (v4i32 VR128:$src1),
(v4i32 (bitconvert (memopv2i64 addr:$src2))))),
(bc_v4i32 (memopv2i64 addr:$src2)))),
(VPSRLVDrm VR128:$src1, addr:$src2)>;
def : Pat<(v2i64 (srl (v2i64 VR128:$src1), (memopv2i64 addr:$src2))),
(VPSRLVQrm VR128:$src1, addr:$src2)>;
def : Pat<(v4i32 (sra (v4i32 VR128:$src1),
(v4i32 (bitconvert (memopv2i64 addr:$src2))))),
(bc_v4i32 (memopv2i64 addr:$src2)))),
(VPSRAVDrm VR128:$src1, addr:$src2)>;
def : Pat<(v8i32 (shl (v8i32 VR256:$src1),
(v8i32 (bitconvert (memopv4i64 addr:$src2))))),
(bc_v8i32 (memopv4i64 addr:$src2)))),
(VPSLLVDYrm VR256:$src1, addr:$src2)>;
def : Pat<(v4i64 (shl (v4i64 VR256:$src1), (memopv4i64 addr:$src2))),
(VPSLLVQYrm VR256:$src1, addr:$src2)>;
def : Pat<(v8i32 (srl (v8i32 VR256:$src1),
(v8i32 (bitconvert (memopv4i64 addr:$src2))))),
(bc_v8i32 (memopv4i64 addr:$src2)))),
(VPSRLVDYrm VR256:$src1, addr:$src2)>;
def : Pat<(v4i64 (srl (v4i64 VR256:$src1), (memopv4i64 addr:$src2))),
(VPSRLVQYrm VR256:$src1, addr:$src2)>;
def : Pat<(v8i32 (sra (v8i32 VR256:$src1),
(v8i32 (bitconvert (memopv4i64 addr:$src2))))),
(bc_v8i32 (memopv4i64 addr:$src2)))),
(VPSRAVDYrm VR256:$src1, addr:$src2)>;
}