On x86, it's safe to treat i32 load anyext as a normal i32 load. Ditto for i8 anyext load to i16.

llvm-svn: 51019

llvm/lib/Target/X86/README-SSE.txt

@@ -757,31 +757,6 @@ or iseling it.
 
 //===---------------------------------------------------------------------===//
 
-Take the following code:
-
-#include <xmmintrin.h>
-__m128i doload64(short x) {return _mm_set_epi16(x,x,x,x,x,x,x,x);}
-
-LLVM currently generates the following on x86:
-doload64:
-        movzwl  4(%esp), %eax
-        movd    %eax, %xmm0
-        punpcklwd       %xmm0, %xmm0
-        pshufd  $0, %xmm0, %xmm0
-        ret
-
-gcc's generated code:
-doload64:
-        movd    4(%esp), %xmm0
-        punpcklwd       %xmm0, %xmm0
-        pshufd  $0, %xmm0, %xmm0
-        ret
-
-LLVM should be able to generate the same thing as gcc.  This looks like it is
-just a matter of matching (scalar_to_vector (load x)) to movd.
-
-//===---------------------------------------------------------------------===//
-
 LLVM currently generates stack realignment code, when it is not necessary
 needed. The problem is that we need to know about stack alignment too early,
 before RA runs.

llvm/lib/Target/X86/X86InstrInfo.td

@@ -229,9 +229,35 @@ def i32immSExt8  : PatLeaf<(i32 imm), [{
 }]>;
 
 // Helper fragments for loads.
+// It's always safe to treat a anyext i16 load as a i32 load. Ditto for
+// i8 to i16.
+def loadi16 : PatFrag<(ops node:$ptr), (i16 (ld node:$ptr)), [{
+  if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
+    if (LD->getAddressingMode() != ISD::UNINDEXED)
+      return false;
+    ISD::LoadExtType ExtType = LD->getExtensionType();
+    if (ExtType == ISD::NON_EXTLOAD)
+      return true;
+    if (ExtType == ISD::EXTLOAD)
+      return LD->getAlignment() >= 16;
+  }
+  return false;
+}]>;
+
+def loadi32 : PatFrag<(ops node:$ptr), (i32 (ld node:$ptr)), [{
+  if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
+    if (LD->getAddressingMode() != ISD::UNINDEXED)
+      return false;
+    ISD::LoadExtType ExtType = LD->getExtensionType();
+    if (ExtType == ISD::NON_EXTLOAD)
+      return true;
+    if (ExtType == ISD::EXTLOAD)
+      return LD->getAlignment() >= 16;
+  }
+  return false;
+}]>;
+
 def loadi8  : PatFrag<(ops node:$ptr), (i8  (load node:$ptr))>;
-def loadi16 : PatFrag<(ops node:$ptr), (i16 (load node:$ptr))>;
-def loadi32 : PatFrag<(ops node:$ptr), (i32 (load node:$ptr))>;
 def loadi64 : PatFrag<(ops node:$ptr), (i64 (load node:$ptr))>;
 
 def loadf32 : PatFrag<(ops node:$ptr), (f32 (load node:$ptr))>;

llvm/test/CodeGen/X86/vec_set-H.ll

+; RUN: llvm-as < %s | llc -march=x86 -mattr=+sse2 | not grep movz
+
+define <2 x i64> @doload64(i16 signext  %x) nounwind  {
+entry:
+	%tmp36 = insertelement <8 x i16> undef, i16 %x, i32 0		; <<8 x i16>> [#uses=1]
+	%tmp37 = insertelement <8 x i16> %tmp36, i16 %x, i32 1		; <<8 x i16>> [#uses=1]
+	%tmp38 = insertelement <8 x i16> %tmp37, i16 %x, i32 2		; <<8 x i16>> [#uses=1]
+	%tmp39 = insertelement <8 x i16> %tmp38, i16 %x, i32 3		; <<8 x i16>> [#uses=1]
+	%tmp40 = insertelement <8 x i16> %tmp39, i16 %x, i32 4		; <<8 x i16>> [#uses=1]
+	%tmp41 = insertelement <8 x i16> %tmp40, i16 %x, i32 5		; <<8 x i16>> [#uses=1]
+	%tmp42 = insertelement <8 x i16> %tmp41, i16 %x, i32 6		; <<8 x i16>> [#uses=1]
+	%tmp43 = insertelement <8 x i16> %tmp42, i16 %x, i32 7		; <<8 x i16>> [#uses=1]
+	%tmp46 = bitcast <8 x i16> %tmp43 to <2 x i64>		; <<2 x i64>> [#uses=1]
+	ret <2 x i64> %tmp46
+}