[llvm-commits] CVS: llvm/lib/Target/X86/X86ISelLowering.cpp X86InstrSSE.td
Evan Cheng
evan.cheng at apple.com
Tue Mar 28 19:05:01 PST 2006
Changes in directory llvm/lib/Target/X86:
X86ISelLowering.cpp updated: 1.142 -> 1.143
X86InstrSSE.td updated: 1.41 -> 1.42
---
Log message:
- More shuffle related bug fixes.
- Whenever possible use ops of the right packed types for vector shuffles /
splats.
---
Diffs of the changes: (+30 -47)
X86ISelLowering.cpp | 35 +++++------------------------------
X86InstrSSE.td | 42 +++++++++++++++++++++++++-----------------
2 files changed, 30 insertions(+), 47 deletions(-)
Index: llvm/lib/Target/X86/X86ISelLowering.cpp
diff -u llvm/lib/Target/X86/X86ISelLowering.cpp:1.142 llvm/lib/Target/X86/X86ISelLowering.cpp:1.143
--- llvm/lib/Target/X86/X86ISelLowering.cpp:1.142 Tue Mar 28 19:30:51 2006
+++ llvm/lib/Target/X86/X86ISelLowering.cpp Tue Mar 28 21:04:49 2006
@@ -2403,43 +2403,18 @@
return SDOperand();
// PSHUFD's 2nd vector must be undef.
- if (MVT::isInteger(VT) && X86::isPSHUFDMask(PermMask.Val))
- if (V2.getOpcode() == ISD::UNDEF)
- return SDOperand();
- else
+ if (MVT::isInteger(VT) && X86::isPSHUFDMask(PermMask.Val)) {
+ if (V2.getOpcode() != ISD::UNDEF)
return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1,
- DAG.getNode(ISD::UNDEF, V1.getValueType()),
- PermMask);
+ DAG.getNode(ISD::UNDEF, V1.getValueType()),PermMask);
+ return SDOperand();
+ }
if (NumElems == 2 ||
X86::isSplatMask(PermMask.Val) ||
X86::isSHUFPMask(PermMask.Val)) {
return NormalizeVectorShuffle(V1, V2, PermMask, VT, DAG);
}
-#if 0
- else if (X86::isSplatMask(PermMask.Val)) {
- // Handle splat cases.
- if (V2.getOpcode() == ISD::UNDEF)
- // Leave the VECTOR_SHUFFLE alone. It matches SHUFP*.
- return SDOperand();
- else
- // Make it match SHUFP* or UNPCKLPD. Second vector is undef since it's
- // not needed.
- return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1,
- DAG.getNode(ISD::UNDEF, V1.getValueType()),
- PermMask);
- } else if (X86::isPSHUFDMask(PermMask.Val)) {
- if (V2.getOpcode() == ISD::UNDEF)
- // Leave the VECTOR_SHUFFLE alone. It matches PSHUFD.
- return SDOperand();
- else
- // Make it match PSHUFD. Second vector is undef since it's not needed.
- return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1,
- DAG.getNode(ISD::UNDEF, V1.getValueType()),
- PermMask);
- } else if (X86::isSHUFPMask(PermMask.Val))
- return NormalizeVectorShuffle(V1, V2, PermMask, VT, DAG);
-#endif
assert(0 && "Unexpected VECTOR_SHUFFLE to lower");
abort();
Index: llvm/lib/Target/X86/X86InstrSSE.td
diff -u llvm/lib/Target/X86/X86InstrSSE.td:1.41 llvm/lib/Target/X86/X86InstrSSE.td:1.42
--- llvm/lib/Target/X86/X86InstrSSE.td:1.41 Tue Mar 28 19:30:51 2006
+++ llvm/lib/Target/X86/X86InstrSSE.td Tue Mar 28 21:04:49 2006
@@ -55,11 +55,7 @@
return getI8Imm(X86::getShuffleSHUFImmediate(N));
}]>;
-def SHUFP_splat_mask : PatLeaf<(build_vector), [{
- return X86::isSplatMask(N);
-}], SHUFFLE_get_shuf_imm>;
-
-def MOVLHPS_splat_mask : PatLeaf<(build_vector), [{
+def v2f64_v2i64_splat_mask : PatLeaf<(build_vector), [{
return X86::isSplatMask(N);
}]>;
@@ -87,6 +83,12 @@
return X86::isSHUFPMask(N);
}], SHUFFLE_get_shuf_imm>;
+// Only use SHUFP for v4i32 if no other options are available.
+// FIXME: add tblgen hook to reduce the complexity of pattern.
+def SHUFP_v4i32_shuffle_mask : PatLeaf<(build_vector), [{
+ return !X86::isUNPCKHMask(N) && !X86::isPSHUFDMask(N) && X86::isSHUFPMask(N);
+}], SHUFFLE_get_shuf_imm>;
+
//===----------------------------------------------------------------------===//
// SSE scalar FP Instructions
//===----------------------------------------------------------------------===//
@@ -1327,6 +1329,8 @@
Requires<[HasSSE2]>;
// bit_convert
+def : Pat<(v2i64 (bitconvert (v4i32 VR128:$src))), (v2i64 VR128:$src)>,
+ Requires<[HasSSE2]>;
def : Pat<(v4i32 (bitconvert (v4f32 VR128:$src))), (v4i32 VR128:$src)>,
Requires<[HasSSE2]>;
def : Pat<(v4f32 (bitconvert (v4i32 VR128:$src))), (v4f32 VR128:$src)>,
@@ -1346,16 +1350,20 @@
def : Pat<(v16i8 (X86zexts2vec R8:$src)),
(MOVZD128rr (V_SET0_PI), (MOVZX32rr8 R8:$src))>, Requires<[HasSSE2]>;
-// Splat v4f32 / v4i32
-def : Pat<(vector_shuffle (v4f32 VR128:$src), (undef), SHUFP_splat_mask:$sm),
- (v4f32 (SHUFPSrr VR128:$src, VR128:$src, SHUFP_splat_mask:$sm))>,
- Requires<[HasSSE1]>;
-def : Pat<(vector_shuffle (v4i32 VR128:$src), (undef), SHUFP_splat_mask:$sm),
- (v4i32 (SHUFPSrr VR128:$src, VR128:$src, SHUFP_splat_mask:$sm))>,
- Requires<[HasSSE2]>;
-
// Splat v2f64 / v2i64
-def : Pat<(vector_shuffle (v2f64 VR128:$src), (undef), MOVLHPS_splat_mask:$sm),
- (v2f64 (MOVLHPSrr VR128:$src, VR128:$src))>, Requires<[HasSSE2]>;
-def : Pat<(vector_shuffle (v2i64 VR128:$src), (undef), MOVLHPS_splat_mask:$sm),
- (v2i64 (MOVLHPSrr VR128:$src, VR128:$src))>, Requires<[HasSSE2]>;
+def : Pat<(vector_shuffle (v2f64 VR128:$src), (v2f64 VR128:$src),
+ v2f64_v2i64_splat_mask:$sm),
+ (v2f64 (UNPCKLPDrr VR128:$src, VR128:$src))>, Requires<[HasSSE2]>;
+def : Pat<(vector_shuffle (v2i64 VR128:$src), (v2i64 VR128:$src),
+ v2f64_v2i64_splat_mask:$sm),
+ (v2i64 (PUNPCKLQDQrr VR128:$src, VR128:$src))>, Requires<[HasSSE2]>;
+
+// Shuffle v4i32 if others do not match
+def : Pat<(vector_shuffle (v4i32 VR128:$src1), (v4i32 VR128:$src2),
+ SHUFP_shuffle_mask:$sm),
+ (v4i32 (SHUFPSrr VR128:$src1, VR128:$src2,
+ SHUFP_v4i32_shuffle_mask:$sm))>, Requires<[HasSSE2]>;
+def : Pat<(vector_shuffle (v4i32 VR128:$src1), (load addr:$src2),
+ SHUFP_shuffle_mask:$sm),
+ (v4i32 (SHUFPSrm VR128:$src1, addr:$src2,
+ SHUFP_v4i32_shuffle_mask:$sm))>, Requires<[HasSSE2]>;
More information about the llvm-commits
mailing list