[llvm] r253795 - [MachineInstrBuilder] Support for adding a ConstantPoolIndex MO with an additional offset.
Simon Pilgrim via llvm-commits
llvm-commits at lists.llvm.org
Sat Nov 21 13:42:27 PST 2015
Author: rksimon
Date: Sat Nov 21 15:42:26 2015
New Revision: 253795
URL: http://llvm.org/viewvc/llvm-project?rev=253795&view=rev
Log:
[MachineInstrBuilder] Support for adding a ConstantPoolIndex MO with an additional offset.
MachineInstrBuilder::addDisp can already add an immediate or global address MO with an adjusted offset, this patch adds support for constant pool indices as well.
All remaining MO types still assert - there are a number of other types that could support adjusted offsets but I have no test cases at this time.
Required to fix a regression in D13988 found by Mikael Holmén during stress testing (test case attached).
Differential Revision: http://reviews.llvm.org/D14867
Added:
llvm/trunk/test/CodeGen/X86/insertps-from-constantpool.ll
Modified:
llvm/trunk/include/llvm/CodeGen/MachineInstrBuilder.h
Modified: llvm/trunk/include/llvm/CodeGen/MachineInstrBuilder.h
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/include/llvm/CodeGen/MachineInstrBuilder.h?rev=253795&r1=253794&r2=253795&view=diff
==============================================================================
--- llvm/trunk/include/llvm/CodeGen/MachineInstrBuilder.h (original)
+++ llvm/trunk/include/llvm/CodeGen/MachineInstrBuilder.h Sat Nov 21 15:42:26 2015
@@ -200,22 +200,24 @@ public:
// Add a displacement from an existing MachineOperand with an added offset.
const MachineInstrBuilder &addDisp(const MachineOperand &Disp, int64_t off,
unsigned char TargetFlags = 0) const {
+ // If caller specifies new TargetFlags then use it, otherwise the
+ // default behavior is to copy the target flags from the existing
+ // MachineOperand. This means if the caller wants to clear the
+ // target flags it needs to do so explicitly.
+ if (0 == TargetFlags)
+ TargetFlags = Disp.getTargetFlags();
+
switch (Disp.getType()) {
default:
llvm_unreachable("Unhandled operand type in addDisp()");
case MachineOperand::MO_Immediate:
return addImm(Disp.getImm() + off);
- case MachineOperand::MO_GlobalAddress: {
- // If caller specifies new TargetFlags then use it, otherwise the
- // default behavior is to copy the target flags from the existing
- // MachineOperand. This means if the caller wants to clear the
- // target flags it needs to do so explicitly.
- if (TargetFlags)
- return addGlobalAddress(Disp.getGlobal(), Disp.getOffset() + off,
- TargetFlags);
+ case MachineOperand::MO_ConstantPoolIndex:
+ return addConstantPoolIndex(Disp.getIndex(), Disp.getOffset() + off,
+ TargetFlags);
+ case MachineOperand::MO_GlobalAddress:
return addGlobalAddress(Disp.getGlobal(), Disp.getOffset() + off,
- Disp.getTargetFlags());
- }
+ TargetFlags);
}
}
Added: llvm/trunk/test/CodeGen/X86/insertps-from-constantpool.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/X86/insertps-from-constantpool.ll?rev=253795&view=auto
==============================================================================
--- llvm/trunk/test/CodeGen/X86/insertps-from-constantpool.ll (added)
+++ llvm/trunk/test/CodeGen/X86/insertps-from-constantpool.ll Sat Nov 21 15:42:26 2015
@@ -0,0 +1,20 @@
+; RUN: llc -mtriple=i686-unknown-unknown -mattr=+sse4.1 < %s | FileCheck %s --check-prefix=X32
+; RUN: llc -mtriple=x86_64-unknown-unknown -mattr=+sse4.1 < %s | FileCheck %s --check-prefix=X64
+
+; Test for case where insertps folds the load of an insertion element from a constant pool.
+
+define <4 x float> @fold_from_constantpool(<4 x float> %a) {
+; X32-LABEL: fold_from_constantpool:
+; X32: # BB#0:
+; X32-NEXT: insertps {{.*#+}} xmm0 = mem[0],xmm0[1,2,3]
+; X32-NEXT: retl
+;
+; X64-LABEL: fold_from_constantpool:
+; X64: # BB#0:
+; X64-NEXT: insertps {{.*#+}} xmm0 = mem[0],xmm0[1,2,3]
+; X64-NEXT: retq
+ %1 = call <4 x float> @llvm.x86.sse41.insertps(<4 x float> %a, <4 x float> <float 0.0, float 1.0, float 0.0, float 0.0>, i8 64)
+ ret <4 x float> %1
+}
+
+declare <4 x float> @llvm.x86.sse41.insertps(<4 x float>, <4 x float>, i8) nounwind readnone
More information about the llvm-commits
mailing list