[llvm] r224746 - X86: Don't over-align combined loads.

[Jim Grosbach]

Author: grosbach
Date: Mon Dec 22 18:35:23 2014
New Revision: 224746

URL: http://llvm.org/viewvc/llvm-project?rev=224746&view=rev
Log:
X86: Don't over-align combined loads.

When combining consecutive loads+inserts into a single vector load,
we should keep the alignment of the base load. Doing otherwise can, and does,
lead to using overly aligned instructions. In the included test case, for
example, using a 32-byte vmovaps on a 16-byte aligned value. Oops.

rdar://19190968

Added:
    llvm/trunk/test/CodeGen/X86/vec-loadsingles-alignment.ll
Modified:
    llvm/trunk/lib/Target/X86/X86ISelLowering.cpp

Modified: llvm/trunk/lib/Target/X86/X86ISelLowering.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/X86/X86ISelLowering.cpp?rev=224746&r1=224745&r2=224746&view=diff
==============================================================================
--- llvm/trunk/lib/Target/X86/X86ISelLowering.cpp (original)
+++ llvm/trunk/lib/Target/X86/X86ISelLowering.cpp Mon Dec 22 18:35:23 2014
@@ -6042,15 +6042,10 @@ static SDValue EltsFromConsecutiveLoads(
 
     SDValue NewLd = SDValue();
 
-    if (DAG.InferPtrAlignment(LDBase->getBasePtr()) >= 16)
-      NewLd = DAG.getLoad(VT, DL, LDBase->getChain(), LDBase->getBasePtr(),
-                          LDBase->getPointerInfo(),
-                          LDBase->isVolatile(), LDBase->isNonTemporal(),
-                          LDBase->isInvariant(), 0);
     NewLd = DAG.getLoad(VT, DL, LDBase->getChain(), LDBase->getBasePtr(),
-                        LDBase->getPointerInfo(),
-                        LDBase->isVolatile(), LDBase->isNonTemporal(),
-                        LDBase->isInvariant(), LDBase->getAlignment());
+                        LDBase->getPointerInfo(), LDBase->isVolatile(),
+                        LDBase->isNonTemporal(), LDBase->isInvariant(),
+                        LDBase->getAlignment());
 
     if (LDBase->hasAnyUseOfValue(1)) {
       SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,

Added: llvm/trunk/test/CodeGen/X86/vec-loadsingles-alignment.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/X86/vec-loadsingles-alignment.ll?rev=224746&view=auto
==============================================================================
--- llvm/trunk/test/CodeGen/X86/vec-loadsingles-alignment.ll (added)
+++ llvm/trunk/test/CodeGen/X86/vec-loadsingles-alignment.ll Mon Dec 22 18:35:23 2014
@@ -0,0 +1,35 @@
+; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx2 | FileCheck %s
+
+ at e = global [8 x i32] [i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8], align 16
+ at d = global [8 x i32] [i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1], align 16
+
+; The global 'e' has 16 byte alignment, so make sure we don't generate an
+; aligned 32-byte load instruction when we combine the load+insert sequence.
+
+define i32 @subb() nounwind ssp {
+; CHECK-LABEL: subb:
+; CHECK:  vmovups e(%rip), %ymm
+entry:
+  %0 = load i32* getelementptr inbounds ([8 x i32]* @e, i64 0, i64 7), align 4
+  %1 = load i32* getelementptr inbounds ([8 x i32]* @e, i64 0, i64 6), align 8
+  %2 = load i32* getelementptr inbounds ([8 x i32]* @e, i64 0, i64 5), align 4
+  %3 = load i32* getelementptr inbounds ([8 x i32]* @e, i64 0, i64 4), align 16
+  %4 = load i32* getelementptr inbounds ([8 x i32]* @e, i64 0, i64 3), align 4
+  %5 = load i32* getelementptr inbounds ([8 x i32]* @e, i64 0, i64 2), align 8
+  %6 = load i32* getelementptr inbounds ([8 x i32]* @e, i64 0, i64 1), align 4
+  %7 = load i32* getelementptr inbounds ([8 x i32]* @e, i64 0, i64 0), align 16
+  %vecinit.i = insertelement <8 x i32> undef, i32 %7, i32 0
+  %vecinit1.i = insertelement <8 x i32> %vecinit.i, i32 %6, i32 1
+  %vecinit2.i = insertelement <8 x i32> %vecinit1.i, i32 %5, i32 2
+  %vecinit3.i = insertelement <8 x i32> %vecinit2.i, i32 %4, i32 3
+  %vecinit4.i = insertelement <8 x i32> %vecinit3.i, i32 %3, i32 4
+  %vecinit5.i = insertelement <8 x i32> %vecinit4.i, i32 %2, i32 5
+  %vecinit6.i = insertelement <8 x i32> %vecinit5.i, i32 %1, i32 6
+  %vecinit7.i = insertelement <8 x i32> %vecinit6.i, i32 %0, i32 7
+  %8 = bitcast <8 x i32> %vecinit7.i to <32 x i8>
+  tail call void @llvm.x86.avx.storeu.dq.256(i8* bitcast ([8 x i32]* @d to i8*), <32 x i8> %8)
+  ret i32 0
+}
+
+declare void @llvm.x86.avx.storeu.dq.256(i8*, <32 x i8>) nounwind
+