[llvm-commits] [llvm] r76094 - in /llvm/trunk: lib/Target/X86/X86InstrInfo.cpp test/CodeGen/X86/2009-07-16-LoadFoldingBug.ll
Evan Cheng
evan.cheng at apple.com
Thu Jul 16 11:44:06 PDT 2009
Author: evancheng
Date: Thu Jul 16 13:44:05 2009
New Revision: 76094
URL: http://llvm.org/viewvc/llvm-project?rev=76094&view=rev
Log:
With recent MC changes, RIP base register is explicitly modeled. Make sure we add it when x86 V_SET0 / V_SETALLONES (by transforming it into a constpool load) into the use instruction.
Added:
llvm/trunk/test/CodeGen/X86/2009-07-16-LoadFoldingBug.ll
Modified:
llvm/trunk/lib/Target/X86/X86InstrInfo.cpp
Modified: llvm/trunk/lib/Target/X86/X86InstrInfo.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/X86/X86InstrInfo.cpp?rev=76094&r1=76093&r2=76094&view=diff
==============================================================================
--- llvm/trunk/lib/Target/X86/X86InstrInfo.cpp (original)
+++ llvm/trunk/lib/Target/X86/X86InstrInfo.cpp Thu Jul 16 13:44:05 2009
@@ -2269,13 +2269,15 @@
// x86-32 PIC requires a PIC base register for constant pools.
unsigned PICBase = 0;
- if (TM.getRelocationModel() == Reloc::PIC_ &&
- !TM.getSubtarget<X86Subtarget>().is64Bit())
- // FIXME: PICBase = TM.getInstrInfo()->getGlobalBaseReg(&MF);
- // This doesn't work for several reasons.
- // 1. GlobalBaseReg may have been spilled.
- // 2. It may not be live at MI.
- return false;
+ if (TM.getRelocationModel() == Reloc::PIC_)
+ if (TM.getSubtarget<X86Subtarget>().is64Bit())
+ PICBase = X86::RIP;
+ else
+ // FIXME: PICBase = TM.getInstrInfo()->getGlobalBaseReg(&MF);
+ // This doesn't work for several reasons.
+ // 1. GlobalBaseReg may have been spilled.
+ // 2. It may not be live at MI.
+ return false;
// Create a v4i32 constant-pool entry.
MachineConstantPool &MCP = *MF.getConstantPool();
Added: llvm/trunk/test/CodeGen/X86/2009-07-16-LoadFoldingBug.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/X86/2009-07-16-LoadFoldingBug.ll?rev=76094&view=auto
==============================================================================
--- llvm/trunk/test/CodeGen/X86/2009-07-16-LoadFoldingBug.ll (added)
+++ llvm/trunk/test/CodeGen/X86/2009-07-16-LoadFoldingBug.ll Thu Jul 16 13:44:05 2009
@@ -0,0 +1,102 @@
+; RUN: llvm-as < %s | llc -mtriple=x86_64-apple-darwin10 | FileCheck %s
+
+; CHECK: _foo:
+; CHECK: pavgw LCPI1_4(%rip)
+
+; rdar://7057804
+
+define void @foo(i16* %out8x8, i16* %in8x8, i32 %lastrow) optsize ssp {
+entry:
+ %0 = call <8 x i16> @llvm.x86.sse2.pmulh.w(<8 x i16> <i16 6518, i16 6518, i16 6518, i16 6518, i16 6518, i16 6518, i16 6518, i16 6518>, <8 x i16> undef) nounwind readnone ; <<8 x i16>> [#uses=2]
+ %1 = call <8 x i16> @llvm.x86.sse2.pcmpeq.w(<8 x i16> %0, <8 x i16> <i16 16384, i16 16384, i16 16384, i16 16384, i16 16384, i16 16384, i16 16384, i16 16384>) nounwind readnone ; <<8 x i16>> [#uses=1]
+ %2 = call <8 x i16> @llvm.x86.sse2.psrli.w(<8 x i16> zeroinitializer, i32 14) nounwind readnone ; <<8 x i16>> [#uses=1]
+ %3 = call <8 x i16> @llvm.x86.sse2.pavg.w(<8 x i16> %2, <8 x i16> zeroinitializer) nounwind readnone ; <<8 x i16>> [#uses=1]
+ %tmp.i.i10 = add <8 x i16> %0, %3 ; <<8 x i16>> [#uses=1]
+ %4 = call <8 x i16> @llvm.x86.sse2.psubs.w(<8 x i16> zeroinitializer, <8 x i16> %1) nounwind readnone ; <<8 x i16>> [#uses=1]
+ %5 = call <8 x i16> @llvm.x86.sse2.padds.w(<8 x i16> %tmp.i.i10, <8 x i16> %4) nounwind readnone ; <<8 x i16>> [#uses=3]
+ %6 = call <8 x i16> @llvm.x86.sse2.psubs.w(<8 x i16> %5, <8 x i16> undef) nounwind readnone ; <<8 x i16>> [#uses=1]
+ %7 = call <8 x i16> @llvm.x86.sse2.pmulh.w(<8 x i16> <i16 6518, i16 6518, i16 6518, i16 6518, i16 6518, i16 6518, i16 6518, i16 6518>, <8 x i16> undef) nounwind readnone ; <<8 x i16>> [#uses=2]
+ %8 = call <8 x i16> @llvm.x86.sse2.pcmpeq.w(<8 x i16> %7, <8 x i16> <i16 16384, i16 16384, i16 16384, i16 16384, i16 16384, i16 16384, i16 16384, i16 16384>) nounwind readnone ; <<8 x i16>> [#uses=1]
+ %9 = call <8 x i16> @llvm.x86.sse2.psrli.w(<8 x i16> zeroinitializer, i32 14) nounwind readnone ; <<8 x i16>> [#uses=1]
+ %10 = call <8 x i16> @llvm.x86.sse2.pavg.w(<8 x i16> %9, <8 x i16> zeroinitializer) nounwind readnone ; <<8 x i16>> [#uses=1]
+ %tmp.i.i8 = add <8 x i16> %7, %10 ; <<8 x i16>> [#uses=1]
+ %11 = call <8 x i16> @llvm.x86.sse2.psubs.w(<8 x i16> undef, <8 x i16> %8) nounwind readnone ; <<8 x i16>> [#uses=1]
+ %12 = call <8 x i16> @llvm.x86.sse2.padds.w(<8 x i16> %tmp.i.i8, <8 x i16> %11) nounwind readnone ; <<8 x i16>> [#uses=1]
+ %13 = call <8 x i16> @llvm.x86.sse2.padds.w(<8 x i16> undef, <8 x i16> undef) nounwind readnone ; <<8 x i16>> [#uses=1]
+ %14 = call <8 x i16> @llvm.x86.sse2.psubs.w(<8 x i16> %5, <8 x i16> undef) nounwind readnone ; <<8 x i16>> [#uses=1]
+ %15 = call <8 x i16> @llvm.x86.sse2.padds.w(<8 x i16> %5, <8 x i16> undef) nounwind readnone ; <<8 x i16>> [#uses=1]
+ %16 = call <8 x i16> @llvm.x86.sse2.padds.w(<8 x i16> %6, <8 x i16> undef) nounwind readnone ; <<8 x i16>> [#uses=1]
+ %17 = call <8 x i16> @llvm.x86.sse2.padds.w(<8 x i16> %12, <8 x i16> undef) nounwind readnone ; <<8 x i16>> [#uses=1]
+ %18 = call <8 x i16> @llvm.x86.sse2.psubs.w(<8 x i16> %13, <8 x i16> %15) nounwind readnone ; <<8 x i16>> [#uses=1]
+ %19 = call <8 x i16> @llvm.x86.sse2.psubs.w(<8 x i16> undef, <8 x i16> %14) nounwind readnone ; <<8 x i16>> [#uses=2]
+ %20 = call <8 x i16> @llvm.x86.sse2.psubs.w(<8 x i16> undef, <8 x i16> undef) nounwind readnone ; <<8 x i16>> [#uses=4]
+ %21 = call <8 x i16> @llvm.x86.sse2.psubs.w(<8 x i16> undef, <8 x i16> %17) nounwind readnone ; <<8 x i16>> [#uses=1]
+ %22 = bitcast <8 x i16> %21 to <2 x i64> ; <<2 x i64>> [#uses=1]
+ %23 = call <8 x i16> @llvm.x86.sse2.pmulh.w(<8 x i16> <i16 23170, i16 23170, i16 23170, i16 23170, i16 23170, i16 23170, i16 23170, i16 23170>, <8 x i16> undef) nounwind readnone ; <<8 x i16>> [#uses=2]
+ %24 = call <8 x i16> @llvm.x86.sse2.pcmpeq.w(<8 x i16> %23, <8 x i16> <i16 16384, i16 16384, i16 16384, i16 16384, i16 16384, i16 16384, i16 16384, i16 16384>) nounwind readnone ; <<8 x i16>> [#uses=1]
+ %25 = call <8 x i16> @llvm.x86.sse2.psrli.w(<8 x i16> zeroinitializer, i32 14) nounwind readnone ; <<8 x i16>> [#uses=1]
+ %26 = call <8 x i16> @llvm.x86.sse2.pavg.w(<8 x i16> %25, <8 x i16> zeroinitializer) nounwind readnone ; <<8 x i16>> [#uses=1]
+ %tmp.i.i6 = add <8 x i16> %23, %26 ; <<8 x i16>> [#uses=1]
+ %27 = call <8 x i16> @llvm.x86.sse2.psubs.w(<8 x i16> undef, <8 x i16> %24) nounwind readnone ; <<8 x i16>> [#uses=1]
+ %28 = call <8 x i16> @llvm.x86.sse2.padds.w(<8 x i16> %tmp.i.i6, <8 x i16> %27) nounwind readnone ; <<8 x i16>> [#uses=1]
+ %29 = call <8 x i16> @llvm.x86.sse2.pmulh.w(<8 x i16> <i16 -23170, i16 -23170, i16 -23170, i16 -23170, i16 -23170, i16 -23170, i16 -23170, i16 -23170>, <8 x i16> undef) nounwind readnone ; <<8 x i16>> [#uses=2]
+ %30 = call <8 x i16> @llvm.x86.sse2.pcmpeq.w(<8 x i16> %29, <8 x i16> <i16 16384, i16 16384, i16 16384, i16 16384, i16 16384, i16 16384, i16 16384, i16 16384>) nounwind readnone ; <<8 x i16>> [#uses=1]
+ %31 = call <8 x i16> @llvm.x86.sse2.psrli.w(<8 x i16> zeroinitializer, i32 14) nounwind readnone ; <<8 x i16>> [#uses=1]
+ %32 = call <8 x i16> @llvm.x86.sse2.pavg.w(<8 x i16> %31, <8 x i16> zeroinitializer) nounwind readnone ; <<8 x i16>> [#uses=1]
+ %tmp.i.i4 = add <8 x i16> %29, %32 ; <<8 x i16>> [#uses=1]
+ %33 = call <8 x i16> @llvm.x86.sse2.psubs.w(<8 x i16> undef, <8 x i16> %30) nounwind readnone ; <<8 x i16>> [#uses=1]
+ %34 = call <8 x i16> @llvm.x86.sse2.padds.w(<8 x i16> %tmp.i.i4, <8 x i16> %33) nounwind readnone ; <<8 x i16>> [#uses=1]
+ %35 = call <8 x i16> @llvm.x86.sse2.pmulh.w(<8 x i16> <i16 23170, i16 23170, i16 23170, i16 23170, i16 23170, i16 23170, i16 23170, i16 23170>, <8 x i16> %20) nounwind readnone ; <<8 x i16>> [#uses=2]
+ %tmp.i2.i1 = mul <8 x i16> %20, <i16 23170, i16 23170, i16 23170, i16 23170, i16 23170, i16 23170, i16 23170, i16 23170> ; <<8 x i16>> [#uses=1]
+ %36 = call <8 x i16> @llvm.x86.sse2.pcmpeq.w(<8 x i16> %35, <8 x i16> <i16 16384, i16 16384, i16 16384, i16 16384, i16 16384, i16 16384, i16 16384, i16 16384>) nounwind readnone ; <<8 x i16>> [#uses=1]
+ %37 = call <8 x i16> @llvm.x86.sse2.psrli.w(<8 x i16> %tmp.i2.i1, i32 14) nounwind readnone ; <<8 x i16>> [#uses=1]
+ %38 = call <8 x i16> @llvm.x86.sse2.pavg.w(<8 x i16> %37, <8 x i16> zeroinitializer) nounwind readnone ; <<8 x i16>> [#uses=1]
+ %tmp.i.i2 = add <8 x i16> %35, %38 ; <<8 x i16>> [#uses=1]
+ %39 = call <8 x i16> @llvm.x86.sse2.psubs.w(<8 x i16> %19, <8 x i16> %36) nounwind readnone ; <<8 x i16>> [#uses=1]
+ %40 = call <8 x i16> @llvm.x86.sse2.padds.w(<8 x i16> %tmp.i.i2, <8 x i16> %39) nounwind readnone ; <<8 x i16>> [#uses=1]
+ %41 = call <8 x i16> @llvm.x86.sse2.pmulh.w(<8 x i16> <i16 -23170, i16 -23170, i16 -23170, i16 -23170, i16 -23170, i16 -23170, i16 -23170, i16 -23170>, <8 x i16> %20) nounwind readnone ; <<8 x i16>> [#uses=2]
+ %tmp.i2.i = mul <8 x i16> %20, <i16 -23170, i16 -23170, i16 -23170, i16 -23170, i16 -23170, i16 -23170, i16 -23170, i16 -23170> ; <<8 x i16>> [#uses=1]
+ %42 = call <8 x i16> @llvm.x86.sse2.pcmpeq.w(<8 x i16> %41, <8 x i16> <i16 16384, i16 16384, i16 16384, i16 16384, i16 16384, i16 16384, i16 16384, i16 16384>) nounwind readnone ; <<8 x i16>> [#uses=1]
+ %43 = call <8 x i16> @llvm.x86.sse2.psrli.w(<8 x i16> %tmp.i2.i, i32 14) nounwind readnone ; <<8 x i16>> [#uses=1]
+ %44 = call <8 x i16> @llvm.x86.sse2.pavg.w(<8 x i16> %43, <8 x i16> zeroinitializer) nounwind readnone ; <<8 x i16>> [#uses=1]
+ %tmp.i.i = add <8 x i16> %41, %44 ; <<8 x i16>> [#uses=1]
+ %45 = call <8 x i16> @llvm.x86.sse2.psubs.w(<8 x i16> %19, <8 x i16> %42) nounwind readnone ; <<8 x i16>> [#uses=1]
+ %46 = call <8 x i16> @llvm.x86.sse2.padds.w(<8 x i16> %tmp.i.i, <8 x i16> %45) nounwind readnone ; <<8 x i16>> [#uses=1]
+ %47 = call <8 x i16> @llvm.x86.sse2.padds.w(<8 x i16> %18, <8 x i16> %16) nounwind readnone ; <<8 x i16>> [#uses=1]
+ %48 = bitcast <8 x i16> %47 to <2 x i64> ; <<2 x i64>> [#uses=1]
+ %49 = bitcast <8 x i16> %28 to <2 x i64> ; <<2 x i64>> [#uses=1]
+ %50 = getelementptr i16* %out8x8, i64 8 ; <i16*> [#uses=1]
+ %51 = bitcast i16* %50 to <2 x i64>* ; <<2 x i64>*> [#uses=1]
+ store <2 x i64> %49, <2 x i64>* %51, align 16
+ %52 = bitcast <8 x i16> %40 to <2 x i64> ; <<2 x i64>> [#uses=1]
+ %53 = getelementptr i16* %out8x8, i64 16 ; <i16*> [#uses=1]
+ %54 = bitcast i16* %53 to <2 x i64>* ; <<2 x i64>*> [#uses=1]
+ store <2 x i64> %52, <2 x i64>* %54, align 16
+ %55 = getelementptr i16* %out8x8, i64 24 ; <i16*> [#uses=1]
+ %56 = bitcast i16* %55 to <2 x i64>* ; <<2 x i64>*> [#uses=1]
+ store <2 x i64> %48, <2 x i64>* %56, align 16
+ %57 = bitcast <8 x i16> %46 to <2 x i64> ; <<2 x i64>> [#uses=1]
+ %58 = getelementptr i16* %out8x8, i64 40 ; <i16*> [#uses=1]
+ %59 = bitcast i16* %58 to <2 x i64>* ; <<2 x i64>*> [#uses=1]
+ store <2 x i64> %57, <2 x i64>* %59, align 16
+ %60 = bitcast <8 x i16> %34 to <2 x i64> ; <<2 x i64>> [#uses=1]
+ %61 = getelementptr i16* %out8x8, i64 48 ; <i16*> [#uses=1]
+ %62 = bitcast i16* %61 to <2 x i64>* ; <<2 x i64>*> [#uses=1]
+ store <2 x i64> %60, <2 x i64>* %62, align 16
+ %63 = getelementptr i16* %out8x8, i64 56 ; <i16*> [#uses=1]
+ %64 = bitcast i16* %63 to <2 x i64>* ; <<2 x i64>*> [#uses=1]
+ store <2 x i64> %22, <2 x i64>* %64, align 16
+ ret void
+}
+
+declare <8 x i16> @llvm.x86.sse2.psubs.w(<8 x i16>, <8 x i16>) nounwind readnone
+
+declare <8 x i16> @llvm.x86.sse2.pmulh.w(<8 x i16>, <8 x i16>) nounwind readnone
+
+declare <8 x i16> @llvm.x86.sse2.pcmpeq.w(<8 x i16>, <8 x i16>) nounwind readnone
+
+declare <8 x i16> @llvm.x86.sse2.pavg.w(<8 x i16>, <8 x i16>) nounwind readnone
+
+declare <8 x i16> @llvm.x86.sse2.padds.w(<8 x i16>, <8 x i16>) nounwind readnone
+
+declare <8 x i16> @llvm.x86.sse2.psrli.w(<8 x i16>, i32) nounwind readnone
More information about the llvm-commits
mailing list