[PPC] Do not emit extswsli in 32BIT mode when using -mcpu=pwr9

[ZarkoCA]

Summary

It looks like in some circumstances when compiling with -mcpu=pwr9 we create an EXTSWSLI node when which causes llc to fail. No such error occurs in pwr8 or lower.

This occurs in 32BIT AIX and BE Linux. the cause seems to be that the default return in combineSHL is to create an EXTSWSLI node. Adding a check for whether we are in PPC64 before that fixes the issue.

[Xiangling_L]

llvm/lib/Target/PowerPC/PPCISelLowering.cpp

16066

It seems you can move Subtarget.isPPC64() here instead. Since the following code returns SDValue() for 32bit no matter what.

llvm/test/CodeGen/PowerPC/ppc-32bit-shift.ll

10

The failing reason has nothing to do with endian but only related to 32bit or 64bit mode. So I am suggesting you can test -mtriple=powerpc and -mtriple=powerpc64 only

17

We can simplify the testcase like the following:
define void @a(i32 %add, i64* %sum_a) {
entry:

%conv = sext i32 %add to i64
%shl = shl nsw i64 %conv, 8
store i64 %shl, i64* %sum_a, align 8
ret void

}

[ZarkoCA]

llvm/lib/Target/PowerPC/PPCISelLowering.cpp

16066

It is better to add it there.

llvm/test/CodeGen/PowerPC/ppc-32bit-shift.ll

10

Good point, we only care about 32/64bit.

17

That's much better, thanks.

[ZarkoCA]

Moved !Subtarget.isPPC64() check earlier alongside other checks.

Reworked test case.

[sfertile]

Does PPCMIPeephole::combineSEXTAndSHL have a similar problem that needs to be addressed?

[ZarkoCA]

In D87046#2285772, @sfertile wrote:

Does PPCMIPeephole::combineSEXTAndSHL have a similar problem that needs to be addressed?

I thought so, but it looks like that function is only called when PPCMIPeephole::simplifyCode sees an RLDICR Opcode which is a 64Bit only instruction. So it looks like we avoid this problem there.

[Xiangling_L]

llvm/test/CodeGen/PowerPC/ppc-32bit-shift.ll

18

It seems, you missed matching the third part of srawi instruction. Any reason of doing so?

19

I am suggesting to use eg. [[REG3:[0-9]+]] to match register, use {{[0-9]+}} to match numbers. Or since 8 is hardcoded in the IR, we don't need to worry about this value will change in the assembly .

For example,

rotlwi [[REG3:[0-9]+]], [[REG2]], 8

[sfertile]

In D87046#2288155, @ZarkoCA wrote:

In D87046#2285772, @sfertile wrote:

Does PPCMIPeephole::combineSEXTAndSHL have a similar problem that needs to be addressed?

I thought so, but it looks like that function is only called when PPCMIPeephole::simplifyCode sees an RLDICR Opcode which is a 64Bit only instruction. So it looks like we avoid this problem there.

Thanks for looking into it.

llvm/test/CodeGen/PowerPC/ppc-32bit-shift.ll

16

Is the . character treated as a regex? Neither 32-bit or 64-bit assembly will produce .a in the assembly, both produce a:, but the test works, so I am clearly missing something.

19

+1, and same for the ME/MB fields of the rlwimi instruction.

23

Real minor nit, but there is a trailing space at the end of this line.

[ZarkoCA]

Updated test case.

[ZarkoCA]

llvm/test/CodeGen/PowerPC/ppc-32bit-shift.ll

16

I changed it to foo and removed the period for clarity. It's strange that it wasn't caught.

18

I wanted to only check the register usage but leaving out the SH value makes it more confusing so I added it.

19

Thanks for catching that. I left the actual values for the shifts and the mask end/begin, and used variables for the registers in the 32BIT case. Is that better?

[nemanjai]

This LGTM. My comment isn't really meant to require you to change the test, just to indicate that there are a lot of moving parts that are very difficult to account for completely. So if update_llc_test_checks.py works, you're better off just using that.

llvm/test/CodeGen/PowerPC/ppc-32bit-shift.ll

2

I assume you are not using the script to produce the checks here because it doesn't work with 32-bit PPC codegen or something along those lines. If it does work, it would be preferable to use it.

For example, you have the srawi necessarily preceding the rotlwi but they are independent. So you go through a fair bit of trouble to keep the register allocation flexible, but the instruction order could change. And technically, there is no requirement that the target of the slwi is the same as the source.
Plus the stores are missing so you are not testing that the sign bits are ending up in the right place.

[ZarkoCA]

Generated test case using update_llc_test_checks.py, added stores and used DAG instead of NEXT when appropriate.

[Xiangling_L]

llvm/test/CodeGen/PowerPC/ppc-32bit-shift.ll

22

It seems, [[REG3]] does not need to be the same register as the one used in rotlwi?

23

I am wondering why we don't use regex to match register 4 here?

28

Same question as above.

[ZarkoCA]

llvm/test/CodeGen/PowerPC/ppc-32bit-shift.ll

2

Thanks for the review, I added the stores and tried to address your points.

[ZarkoCA]

Fixed register usage.

[Xiangling_L]

llvm/test/CodeGen/PowerPC/ppc-32bit-shift.ll

27

Sorry, I should've mentioned this earlier, we want to follow the same rules here for 64bit as above 32bit cases.
For example:

; 64BIT:               extswsli [[REG1:[0-9]+]], {{REG2:[0-9]+}}, 8

[ZarkoCA]

llvm/test/CodeGen/PowerPC/ppc-32bit-shift.ll

27

Does REG2 need to be matched anywhere?

[ZarkoCA]

Used REG variables in 64Bit part of test case.

[ZarkoCA]

Code changes didn't get included in last diff.

[Xiangling_L]

LGTM. Thanks

llvm/lib/Target/PowerPC/PPCISelLowering.cpp

SDValue PPCTargetLowering::combineSHL(SDNode *N, DAGCombinerInfo &DCI) const {
   if (auto Value = stripModuloOnShift(*this, N, DCI.DAG))
     return Value;
 
   SDValue N0 = N->getOperand(0);
   ConstantSDNode *CN1 = dyn_cast<ConstantSDNode>(N->getOperand(1));
   if (!Subtarget.isISA3_0() ||
       N0.getOpcode() != ISD::SIGN_EXTEND ||
       N0.getOperand(0).getValueType() != MVT::i32 ||
       CN1 == nullptr || N->getValueType(0) != MVT::i64)
     return SDValue();
 
   // We can't save an operation here if the value is already extended, and
   // the existing shift is easier to combine.
   SDValue ExtsSrc = N0.getOperand(0);
   if (ExtsSrc.getOpcode() == ISD::TRUNCATE &&
       ExtsSrc.getOperand(0).getOpcode() == ISD::AssertSext)
     return SDValue();
 
   SDLoc DL(N0);
   SDValue ShiftBy = SDValue(CN1, 0);
   // We want the shift amount to be i32 on the extswli, but the shift could
   // have an i64.
   if (ShiftBy.getValueType() == MVT::i64)
     ShiftBy = DCI.DAG.getConstant(CN1->getZExtValue(), DL, MVT::i32);
 
+  if (Subtarget.isPPC64())
     return DCI.DAG.getNode(PPCISD::EXTSWSLI, DL, MVT::i64, N0->getOperand(0),
                            ShiftBy);
+  return SDValue();
 }
 
 SDValue PPCTargetLowering::combineSRA(SDNode *N, DAGCombinerInfo &DCI) const {
   if (auto Value = stripModuloOnShift(*this, N, DCI.DAG))
     return Value;
 
   return SDValue();
 }

llvm/test/CodeGen/PowerPC/ppc-32bit-shift.ll

+; RUN: llc -verify-machineinstrs -mtriple=powerpc-ibm-aix-xcoff \
+; RUN:     -mattr=-altivec -mcpu=pwr9 < %s | FileCheck %s --check-prefix=32BIT
+; RUN: llc -verify-machineinstrs -mtriple=powerpc-unknown-linux-gnu \
+; RUN:     -mattr=-altivec -mcpu=pwr9 < %s | FileCheck %s --check-prefix=32BIT
+
+; RUN: llc -verify-machineinstrs -mtriple=powerpc64-ibm-aix-xcoff \
+; RUN:     -mattr=-altivec -mcpu=pwr9 < %s | FileCheck %s --check-prefix=64BIT
+; RUN: llc -verify-machineinstrs -mtriple=powerpc64-unknown-linux-gnu \
+; RUN:     -mattr=-altivec -mcpu=pwr9 < %s | FileCheck %s --check-prefix=64BIT
+; RUN: llc -verify-machineinstrs -mtriple=powerpc64le-unknown-linux-gnu \
+; RUN:     -mattr=-altivec -mcpu=pwr9 < %s | FileCheck %s --check-prefix=64BIT
+
+define dso_local void @a(i32 %inta, i32 %intb, i64 %longint, i64* %sum_a) {
+entry:  
+
+  %add = add nsw i32 %inta, %intb
+  %conv = sext i32 %add to i64
+  %shl = shl nsw i64 %conv, 8
+  %add15 = add nsw i64 %longint, %shl
+  store i64 %add15, i64* %sum_a, align 8
+  ret void
+}
+
+; CHECK-LABEL:     .a
+
+; 32BIT:           srawi [[REG1:[0-9]+]], [[REG2:[0-9]+]]
+; 32BIT-NEXT:      rotlwi [[REG3:[0-9]+]], [[REG2]], [[REG3]]
+; 32BIT-NEXT:      slwi [[REG2]], [[REG2]], [[REG3]]
+; 32BIT-NEXT:      rlwimi [[REG3]], {{[0-9]+}}, [[REG3]], {{[0-9]+}}, {{[0-9]+}}
+
+; 64BIT:           extswsli {{[0-9]+}}, {{[0-9]+}}, {{[0-9]+}}