[llvm] r222504 - [PPC] Use SeparateConstOffsetFromGEP

Hal Finkel hfinkel at anl.gov
Thu Nov 20 20:35:51 PST 2014


Author: hfinkel
Date: Thu Nov 20 22:35:51 2014
New Revision: 222504

URL: http://llvm.org/viewvc/llvm-project?rev=222504&view=rev
Log:
[PPC] Use SeparateConstOffsetFromGEP

This mirrors r222331, which enabled SeparateConstOffsetFromGEP on AArch64, in
the PowerPC backend. Yields, on a POWER7 machine, a 30% speedup on
SingleSource/Benchmarks/Shootout/nestedloop (this might just be from LICM,
there is a store moved out of the inner loop) and a potential speedup on
MultiSource/Benchmarks/mediabench/mpeg2/mpeg2dec/mpeg2decode. Regardless, it
makes some code look cleaner, and synchronizing the backends in this regard
seems like a generally good thing.

Added:
    llvm/trunk/test/CodeGen/PowerPC/ppc64-gep-opt.ll
Modified:
    llvm/trunk/lib/Target/PowerPC/PPCTargetMachine.cpp

Modified: llvm/trunk/lib/Target/PowerPC/PPCTargetMachine.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/PowerPC/PPCTargetMachine.cpp?rev=222504&r1=222503&r2=222504&view=diff
==============================================================================
--- llvm/trunk/lib/Target/PowerPC/PPCTargetMachine.cpp (original)
+++ llvm/trunk/lib/Target/PowerPC/PPCTargetMachine.cpp Thu Nov 20 22:35:51 2014
@@ -22,6 +22,7 @@
 #include "llvm/Support/FormattedStream.h"
 #include "llvm/Support/TargetRegistry.h"
 #include "llvm/Target/TargetOptions.h"
+#include "llvm/Transforms/Scalar.h"
 using namespace llvm;
 
 static cl::
@@ -32,6 +33,11 @@ static cl::opt<bool>
 VSXFMAMutateEarly("schedule-ppc-vsx-fma-mutation-early",
   cl::Hidden, cl::desc("Schedule VSX FMA instruction mutation early"));
 
+static cl::opt<bool>
+EnableGEPOpt("ppc-gep-opt", cl::Hidden,
+             cl::desc("Enable optimizations on complex GEPs"),
+             cl::init(true));
+
 extern "C" void LLVMInitializePowerPCTarget() {
   // Register the targets
   RegisterTargetMachine<PPC32TargetMachine> A(ThePPC32Target);
@@ -168,6 +174,20 @@ TargetPassConfig *PPCTargetMachine::crea
 
 void PPCPassConfig::addIRPasses() {
   addPass(createAtomicExpandPass(&getPPCTargetMachine()));
+
+  if (TM->getOptLevel() == CodeGenOpt::Aggressive && EnableGEPOpt) {
+    // Call SeparateConstOffsetFromGEP pass to extract constants within indices
+    // and lower a GEP with multiple indices to either arithmetic operations or
+    // multiple GEPs with single index.
+    addPass(createSeparateConstOffsetFromGEPPass(TM, true));
+    // Call EarlyCSE pass to find and remove subexpressions in the lowered
+    // result.
+    addPass(createEarlyCSEPass());
+    // Do loop invariant code motion in case part of the lowered result is
+    // invariant.
+    addPass(createLICMPass());
+  }
+
   TargetPassConfig::addIRPasses();
 }
 

Added: llvm/trunk/test/CodeGen/PowerPC/ppc64-gep-opt.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/PowerPC/ppc64-gep-opt.ll?rev=222504&view=auto
==============================================================================
--- llvm/trunk/test/CodeGen/PowerPC/ppc64-gep-opt.ll (added)
+++ llvm/trunk/test/CodeGen/PowerPC/ppc64-gep-opt.ll Thu Nov 20 22:35:51 2014
@@ -0,0 +1,157 @@
+; RUN: llc -O3 -mcpu=pwr7 < %s | FileCheck %s
+; RUN: llc -O3 -print-after=codegenprepare -mcpu=ppc64 < %s >%t 2>&1 && FileCheck --check-prefix=CHECK-NoAA <%t %s
+; RUN: llc -O3 -print-after=codegenprepare -mcpu=pwr7  < %s >%t 2>&1 && FileCheck --check-prefix=CHECK-UseAA <%t %s
+target datalayout = "E-m:e-i64:64-n32:64"
+target triple = "powerpc64-unknown-linux-gnu"
+
+; Following test cases test enabling SeparateConstOffsetFromGEP pass in the PPC
+; backend. If useAA() returns true, it will lower a GEP with multiple indices
+; into GEPs with a single index, otherwise it will lower it into a
+; "ptrtoint+arithmetics+inttoptr" form.
+
+%struct = type { i32, i32, i32, i32, [20 x i32] }
+
+; Check that when two complex GEPs are used in two basic blocks, LLVM can
+; elimilate the common subexpression for the second use.
+define void @test_GEP_CSE([240 x %struct]* %string, i32* %adj, i32 %lib, i64 %idxprom) {
+  %liberties = getelementptr [240 x %struct]* %string, i64 1, i64 %idxprom, i32 3
+  %1 = load i32* %liberties, align 4
+  %cmp = icmp eq i32 %1, %lib
+  br i1 %cmp, label %if.then, label %if.end
+
+if.then:                                          ; preds = %entry
+  %origin = getelementptr [240 x %struct]* %string, i64 1, i64 %idxprom, i32 2
+  %2 = load i32* %origin, align 4
+  store i32 %2, i32* %adj, align 4
+  br label %if.end
+
+if.end:                                           ; preds = %if.then, %entry
+  ret void
+}
+
+; CHECK-NoAA-LABEL: @test_GEP_CSE(
+; CHECK-NoAA: [[PTR0:%[a-zA-Z0-9]+]] = ptrtoint [240 x %struct]* %string to i64
+; CHECK-NoAA: [[PTR1:%[a-zA-Z0-9]+]] = mul i64 %idxprom, 96
+; CHECK-NoAA: [[PTR2:%[a-zA-Z0-9]+]] = add i64 [[PTR0]], [[PTR1]]
+; CHECK-NoAA: add i64 [[PTR2]], 23052
+; CHECK-NoAA: inttoptr
+; CHECK-NoAA: if.then:
+; CHECK-NoAA-NOT: ptrtoint
+; CHECK-NoAA-NOT: mul
+; CHECK-NoAA: add i64 [[PTR2]], 23048
+; CHECK-NoAA: inttoptr
+
+; CHECK-UseAA-LABEL: @test_GEP_CSE(
+; CHECK-UseAA: [[PTR0:%[a-zA-Z0-9]+]] = bitcast [240 x %struct]* %string to i8*
+; CHECK-UseAA: [[IDX:%[a-zA-Z0-9]+]] = mul i64 %idxprom, 96
+; CHECK-UseAA: [[PTR1:%[a-zA-Z0-9]+]] = getelementptr i8* [[PTR0]], i64 [[IDX]]
+; CHECK-UseAA: getelementptr i8* [[PTR1]], i64 23052
+; CHECK-UseAA: bitcast
+; CHECK-UseAA: if.then:
+; CHECK-UseAA: getelementptr i8* [[PTR1]], i64 23048
+; CHECK-UseAA: bitcast
+
+%class.my = type { i32, [128 x i32], i32, [256 x %struct.pt]}
+%struct.pt = type { %struct.point*, i32, i32 }
+%struct.point = type { i32, i32 }
+
+; Check when a GEP is used across two basic block, LLVM can sink the address
+; calculation and code gen can generate a better addressing mode for the second
+; use.
+define void @test_GEP_across_BB(%class.my* %this, i64 %idx) {
+  %1 = getelementptr %class.my* %this, i64 0, i32 3, i64 %idx, i32 1
+  %2 = load i32* %1, align 4
+  %3 = getelementptr %class.my* %this, i64 0, i32 3, i64 %idx, i32 2
+  %4 = load i32* %3, align 4
+  %5 = icmp eq i32 %2, %4
+  br i1 %5, label %if.true, label %exit
+
+if.true:
+  %6 = shl i32 %4, 1
+  store i32 %6, i32* %3, align 4
+  br label %exit
+
+exit:
+  %7 = add nsw i32 %4, 1
+  store i32 %7, i32* %1, align 4
+  ret void
+}
+; CHECK-LABEL: test_GEP_across_BB:
+; CHECK-NOT: lwzu
+; CHECK: blr
+
+; CHECK-NoAA-LABEL: test_GEP_across_BB(
+; CHECK-NoAA: add i64 [[TMP:%[a-zA-Z0-9]+]], 528
+; CHECK-NoAA: add i64 [[TMP]], 532
+; CHECK-NoAA: if.true:
+; CHECK-NoAA: {{%sunk[a-zA-Z0-9]+}} = add i64 [[TMP]], 532
+; CHECK-NoAA: exit:
+; CHECK-NoAA: {{%sunk[a-zA-Z0-9]+}} = add i64 [[TMP]], 528
+
+; CHECK-UseAA-LABEL: test_GEP_across_BB(
+; CHECK-UseAA: [[PTR0:%[a-zA-Z0-9]+]] = getelementptr
+; CHECK-UseAA: getelementptr i8* [[PTR0]], i64 528
+; CHECK-UseAA: getelementptr i8* [[PTR0]], i64 532
+; CHECK-UseAA: if.true:
+; CHECK-UseAA: {{%sunk[a-zA-Z0-9]+}} = getelementptr i8* [[PTR0]], i64 532
+; CHECK-UseAA: exit:
+; CHECK-UseAA: {{%sunk[a-zA-Z0-9]+}} = getelementptr i8* [[PTR0]], i64 528
+
+%struct.S = type { float, double }
+ at struct_array = global [1024 x %struct.S] zeroinitializer, align 16
+
+; The following two test cases check we can extract constant from indices of
+; struct type.
+; The constant offsets are from indices "i64 %idxprom" and "i32 1". As the
+; alloca size of %struct.S is 16, and "i32 1" is the 2rd element whose field
+; offset is 8, the total constant offset is (5 * 16 + 8) = 88.
+define double* @test-struct_1(i32 %i) {
+entry:
+  %add = add nsw i32 %i, 5
+  %idxprom = sext i32 %add to i64
+  %p = getelementptr [1024 x %struct.S]* @struct_array, i64 0, i64 %idxprom, i32 1
+  ret double* %p
+}
+; CHECK-NoAA-LABEL: @test-struct_1(
+; CHECK-NoAA-NOT: getelementptr
+; CHECK-NoAA: add i64 %{{[a-zA-Z0-9]+}}, 88
+
+; CHECK-UseAA-LABEL: @test-struct_1(
+; CHECK-UseAA: getelementptr i8* %{{[a-zA-Z0-9]+}}, i64 88
+
+%struct3 = type { i64, i32 }
+%struct2 = type { %struct3, i32 }
+%struct1 = type { i64, %struct2 }
+%struct0 = type { i32, i32, i64*, [100 x %struct1] }
+
+; The constant offsets are from indices "i32 3", "i64 %arrayidx" and "i32 1".
+; "i32 3" is the 4th element whose field offset is 16. The alloca size of
+; %struct1 is 32. "i32 1" is the 2rd element whose field offset is 8. So the
+; total constant offset is 16 + (-2 * 32) + 8 = -40
+define %struct2* @test-struct_2(%struct0* %ptr, i64 %idx) {
+entry:
+  %arrayidx = add nsw i64 %idx, -2
+  %ptr2 = getelementptr %struct0* %ptr, i64 0, i32 3, i64 %arrayidx, i32 1
+  ret %struct2* %ptr2
+}
+; CHECK-NoAA-LABEL: @test-struct_2(
+; CHECK-NoAA-NOT: = getelementptr
+; CHECK-NoAA: add i64 %{{[a-zA-Z0-9]+}}, -40
+
+; CHECK-UseAA-LABEL: @test-struct_2(
+; CHECK-UseAA: getelementptr i8* %{{[a-zA-Z0-9]+}}, i64 -40
+
+; Test that when a index is added from two constant, SeparateConstOffsetFromGEP
+; pass does not generate incorrect result.
+define void @test_const_add([3 x i32]* %in) {
+  %inc = add nsw i32 2, 1
+  %idxprom = sext i32 %inc to i64
+  %arrayidx = getelementptr [3 x i32]* %in, i64 %idxprom, i64 2
+  store i32 0, i32* %arrayidx, align 4
+  ret void
+}
+; CHECK-LABEL: test_const_add:
+; CHECK: li [[REG:[0-9]+]], 0
+; CHECK: stw [[REG]], 44(3)
+; CHECK: blr
+





More information about the llvm-commits mailing list