Add support for breaking 256-bit v16i16 and v32i8 VSETCC into two 128-bit ones, avoiding sclarization. Add vex form of pcmpeqq and pcmpgtq. Fixes more cases for PR10712.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@138321 91177308-0d34-0410-b5e6-96231b3b80d8
diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index e261fca..dd78aa9 100644
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -989,6 +989,8 @@
setOperationAction(ISD::SRA, MVT::v8i32, Custom);
setOperationAction(ISD::SRA, MVT::v16i16, Custom);
+ setOperationAction(ISD::VSETCC, MVT::v32i8, Custom);
+ setOperationAction(ISD::VSETCC, MVT::v16i16, Custom);
setOperationAction(ISD::VSETCC, MVT::v8i32, Custom);
setOperationAction(ISD::VSETCC, MVT::v4i64, Custom);
diff --git a/lib/Target/X86/X86InstrSSE.td b/lib/Target/X86/X86InstrSSE.td
index d189ed3..d1f7484 100644
--- a/lib/Target/X86/X86InstrSSE.td
+++ b/lib/Target/X86/X86InstrSSE.td
@@ -4889,6 +4889,11 @@
0>, VEX_4V;
defm VPMULDQ : SS41I_binop_rm_int<0x28, "vpmuldq", int_x86_sse41_pmuldq,
0>, VEX_4V;
+
+ def : Pat<(v2i64 (X86pcmpeqq VR128:$src1, VR128:$src2)),
+ (VPCMPEQQrr VR128:$src1, VR128:$src2)>;
+ def : Pat<(v2i64 (X86pcmpeqq VR128:$src1, (memop addr:$src2))),
+ (VPCMPEQQrm VR128:$src1, addr:$src2)>;
}
let Constraints = "$src1 = $dst" in {
@@ -5099,9 +5104,16 @@
(bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
}
-let Predicates = [HasAVX] in
+let Predicates = [HasAVX] in {
defm VPCMPGTQ : SS42I_binop_rm_int<0x37, "vpcmpgtq", int_x86_sse42_pcmpgtq,
0>, VEX_4V;
+
+ def : Pat<(v2i64 (X86pcmpgtq VR128:$src1, VR128:$src2)),
+ (VPCMPGTQrr VR128:$src1, VR128:$src2)>;
+ def : Pat<(v2i64 (X86pcmpgtq VR128:$src1, (memop addr:$src2))),
+ (VPCMPGTQrm VR128:$src1, addr:$src2)>;
+}
+
let Constraints = "$src1 = $dst" in
defm PCMPGTQ : SS42I_binop_rm_int<0x37, "pcmpgtq", int_x86_sse42_pcmpgtq>;
@@ -5660,6 +5672,11 @@
def : Pat<(v2f64 (extract_subvector (v4f64 VR256:$src), (i32 0))),
(v2f64 (EXTRACT_SUBREG (v4f64 VR256:$src), sub_xmm))>;
+def : Pat<(v8i16 (extract_subvector (v16i16 VR256:$src), (i32 0))),
+ (v8i16 (EXTRACT_SUBREG (v16i16 VR256:$src), sub_xmm))>;
+def : Pat<(v16i8 (extract_subvector (v32i8 VR256:$src), (i32 0))),
+ (v16i8 (EXTRACT_SUBREG (v32i8 VR256:$src), sub_xmm))>;
+
//===----------------------------------------------------------------------===//
// VMASKMOV - Conditional SIMD Packed Loads and Stores