This is an archive of the discontinued LLVM Phabricator instance.

Differential D72511

GlobalISel: Handle more cases of G_SEXT narrowing
ClosedPublic

Authored by arsenm on Jan 10 2020, 7:57 AM.

Details

Reviewers
aemerson
paquette
aditya_nandakumar
dsanders
Summary

This now develops the same problem G_ZEXT/G_ANYEXT have where the
requested type is assumed to be the source type. This will be fixed
separately by creating intermediate merges.

Diff Detail

Revision Contents

PathSize
llvm/
lib/
CodeGen/
GlobalISel/
39 lines
test/
CodeGen/
AMDGPU/
GlobalISel/
67 lines

Diff 237337

llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp

Show First 20 Lines • Show All 651 Lines • ▼ Show 20 Lines case TargetOpcode::G_CONSTANT: {
} }
insertParts(MI.getOperand(0).getReg(), insertParts(MI.getOperand(0).getReg(),
Ty, NarrowTy, PartRegs, LeftoverTy, LeftoverRegs); Ty, NarrowTy, PartRegs, LeftoverTy, LeftoverRegs);
MI.eraseFromParent(); MI.eraseFromParent();
return Legalized; return Legalized;
} }
case TargetOpcode::G_SEXT: { case TargetOpcode::G_SEXT:
if (TypeIdx != 0)
return UnableToLegalize;
Register SrcReg = MI.getOperand(1).getReg();
LLT SrcTy = MRI.getType(SrcReg);
// FIXME: support the general case where the requested NarrowTy may not be
// the same as the source type. E.g. s128 = sext(s32)
if ((SrcTy.getSizeInBits() != SizeOp0 / 2) ||
SrcTy.getSizeInBits() != NarrowTy.getSizeInBits()) {
LLVM_DEBUG(dbgs() << "Can't narrow sext to type " << NarrowTy << "\n");
return UnableToLegalize;
}
// Shift the sign bit of the low register through the high register.
auto ShiftAmt =
MIRBuilder.buildConstant(LLT::scalar(64), NarrowTy.getSizeInBits() - 1);
auto Shift = MIRBuilder.buildAShr(NarrowTy, SrcReg, ShiftAmt);
MIRBuilder.buildMerge(MI.getOperand(0).getReg(), {SrcReg, Shift.getReg(0)});
MI.eraseFromParent();
return Legalized;
}
case TargetOpcode::G_ZEXT: case TargetOpcode::G_ZEXT:
case TargetOpcode::G_ANYEXT: { case TargetOpcode::G_ANYEXT: {
if (TypeIdx != 0) if (TypeIdx != 0)
return UnableToLegalize; return UnableToLegalize;
LLT SrcTy = MRI.getType(MI.getOperand(1).getReg()); Register SrcReg = MI.getOperand(1).getReg();
LLT SrcTy = MRI.getType(SrcReg);
uint64_t SizeOp1 = SrcTy.getSizeInBits(); uint64_t SizeOp1 = SrcTy.getSizeInBits();
if (SizeOp0 % SizeOp1 != 0) if (SizeOp0 % SizeOp1 != 0)
return UnableToLegalize; return UnableToLegalize;
Register PadReg; Register PadReg;
if (MI.getOpcode() == TargetOpcode::G_ZEXT) if (MI.getOpcode() == TargetOpcode::G_ZEXT)
PadReg = MIRBuilder.buildConstant(SrcTy, 0).getReg(0); PadReg = MIRBuilder.buildConstant(SrcTy, 0).getReg(0);
else else if (MI.getOpcode() == TargetOpcode::G_ANYEXT)
PadReg = MIRBuilder.buildUndef(SrcTy).getReg(0); PadReg = MIRBuilder.buildUndef(SrcTy).getReg(0);
else {
// Shift the sign bit of the low register through the high register.
auto ShiftAmt =
MIRBuilder.buildConstant(LLT::scalar(64), SrcTy.getSizeInBits() - 1);
PadReg = MIRBuilder.buildAShr(SrcTy, SrcReg, ShiftAmt).getReg(0);
}
// Generate a merge where the bottom bits are taken from the source, and // Generate a merge where the bottom bits are taken from the source, and
// zero/impdef everything else. // zero/impdef/sign bit everything else.
unsigned NumParts = SizeOp0 / SizeOp1; unsigned NumParts = SizeOp0 / SizeOp1;
SmallVector<Register, 4> Srcs = {MI.getOperand(1).getReg()}; SmallVector<Register, 4> Srcs = {SrcReg};
for (unsigned Part = 1; Part < NumParts; ++Part) for (unsigned Part = 1; Part < NumParts; ++Part)
Srcs.push_back(PadReg); Srcs.push_back(PadReg);
MIRBuilder.buildMerge(MI.getOperand(0).getReg(), Srcs); MIRBuilder.buildMerge(MI.getOperand(0).getReg(), Srcs);
MI.eraseFromParent(); MI.eraseFromParent();
return Legalized; return Legalized;
} }
case TargetOpcode::G_TRUNC: { case TargetOpcode::G_TRUNC: {
if (TypeIdx != 1) if (TypeIdx != 1)
▲ Show 20 Lines • Show All 3,788 Lines • Show Last 20 Lines

llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-sext.mir

# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py # NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
# RUN: llc -O0 -mtriple=amdgcn-mesa-mesa3d -mcpu=fiji -run-pass=legalizer -global-isel-abort=0 %s -o - | FileCheck %s # RUN: llc -O0 -mtriple=amdgcn-mesa-mesa3d -mcpu=fiji -run-pass=legalizer %s -o - | FileCheck %s
--- ---
name: test_sext_s32_to_s64 name: test_sext_s32_to_s64
body: | body: |
bb.0: bb.0:
liveins: $vgpr0 liveins: $vgpr0
; CHECK-LABEL: name: test_sext_s32_to_s64 ; CHECK-LABEL: name: test_sext_s32_to_s64
▲ Show 20 Lines • Show All 320 Lines • ▼ Show 20 Lines
--- ---
name: test_sext_s32_to_s128 name: test_sext_s32_to_s128
body: | body: |
bb.0: bb.0:
liveins: $vgpr0 liveins: $vgpr0
; CHECK-LABEL: name: test_sext_s32_to_s128 ; CHECK-LABEL: name: test_sext_s32_to_s128
; CHECK: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0 ; CHECK: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0
; CHECK: [[SEXT:%[0-9]+]]:_(s128) = G_SEXT [[COPY]](s32) ; CHECK: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 31
; CHECK: S_ENDPGM 0, implicit [[SEXT]](s128) ; CHECK: [[ASHR:%[0-9]+]]:_(s32) = G_ASHR [[COPY]], [[C]](s32)
; CHECK: [[MV:%[0-9]+]]:_(s128) = G_MERGE_VALUES [[COPY]](s32), [[ASHR]](s32), [[ASHR]](s32), [[ASHR]](s32)
; CHECK: S_ENDPGM 0, implicit [[MV]](s128)
%0:_(s32) = COPY $vgpr0 %0:_(s32) = COPY $vgpr0
%1:_(s128) = G_SEXT %0 %1:_(s128) = G_SEXT %0
S_ENDPGM 0, implicit %1 S_ENDPGM 0, implicit %1
... ...
--- ---
name: test_sext_s32_to_s256 name: test_sext_s32_to_s256
body: | body: |
bb.0: bb.0:
liveins: $vgpr0 liveins: $vgpr0
; CHECK-LABEL: name: test_sext_s32_to_s256 ; CHECK-LABEL: name: test_sext_s32_to_s256
; CHECK: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0 ; CHECK: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0
; CHECK: [[SEXT:%[0-9]+]]:_(s256) = G_SEXT [[COPY]](s32) ; CHECK: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 31
; CHECK: S_ENDPGM 0, implicit [[SEXT]](s256) ; CHECK: [[ASHR:%[0-9]+]]:_(s32) = G_ASHR [[COPY]], [[C]](s32)
; CHECK: [[MV:%[0-9]+]]:_(s256) = G_MERGE_VALUES [[COPY]](s32), [[ASHR]](s32), [[ASHR]](s32), [[ASHR]](s32), [[ASHR]](s32), [[ASHR]](s32), [[ASHR]](s32), [[ASHR]](s32)
; CHECK: S_ENDPGM 0, implicit [[MV]](s256)
%0:_(s32) = COPY $vgpr0 %0:_(s32) = COPY $vgpr0
%1:_(s256) = G_SEXT %0 %1:_(s256) = G_SEXT %0
S_ENDPGM 0, implicit %1 S_ENDPGM 0, implicit %1
... ...
--- ---
name: test_sext_s32_to_s512 name: test_sext_s32_to_s512
body: | body: |
bb.0: bb.0:
liveins: $vgpr0 liveins: $vgpr0
; CHECK-LABEL: name: test_sext_s32_to_s512 ; CHECK-LABEL: name: test_sext_s32_to_s512
; CHECK: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0 ; CHECK: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0
; CHECK: [[SEXT:%[0-9]+]]:_(s512) = G_SEXT [[COPY]](s32) ; CHECK: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 31
; CHECK: S_ENDPGM 0, implicit [[SEXT]](s512) ; CHECK: [[ASHR:%[0-9]+]]:_(s32) = G_ASHR [[COPY]], [[C]](s32)
; CHECK: [[MV:%[0-9]+]]:_(s512) = G_MERGE_VALUES [[COPY]](s32), [[ASHR]](s32), [[ASHR]](s32), [[ASHR]](s32), [[ASHR]](s32), [[ASHR]](s32), [[ASHR]](s32), [[ASHR]](s32), [[ASHR]](s32), [[ASHR]](s32), [[ASHR]](s32), [[ASHR]](s32), [[ASHR]](s32), [[ASHR]](s32), [[ASHR]](s32), [[ASHR]](s32)
; CHECK: S_ENDPGM 0, implicit [[MV]](s512)
%0:_(s32) = COPY $vgpr0 %0:_(s32) = COPY $vgpr0
%1:_(s512) = G_SEXT %0 %1:_(s512) = G_SEXT %0
S_ENDPGM 0, implicit %1 S_ENDPGM 0, implicit %1
... ...
--- ---
name: test_sext_s32_to_s1024 name: test_sext_s32_to_s1024
body: | body: |
bb.0: bb.0:
liveins: $vgpr0 liveins: $vgpr0
; CHECK-LABEL: name: test_sext_s32_to_s1024 ; CHECK-LABEL: name: test_sext_s32_to_s1024
; CHECK: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0 ; CHECK: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0
; CHECK: [[SEXT:%[0-9]+]]:_(s1024) = G_SEXT [[COPY]](s32) ; CHECK: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 31
; CHECK: S_ENDPGM 0, implicit [[SEXT]](s1024) ; CHECK: [[ASHR:%[0-9]+]]:_(s32) = G_ASHR [[COPY]], [[C]](s32)
; CHECK: [[MV:%[0-9]+]]:_(s1024) = G_MERGE_VALUES [[COPY]](s32), [[ASHR]](s32), [[ASHR]](s32), [[ASHR]](s32), [[ASHR]](s32), [[ASHR]](s32), [[ASHR]](s32), [[ASHR]](s32), [[ASHR]](s32), [[ASHR]](s32), [[ASHR]](s32), [[ASHR]](s32), [[ASHR]](s32), [[ASHR]](s32), [[ASHR]](s32), [[ASHR]](s32), [[ASHR]](s32), [[ASHR]](s32), [[ASHR]](s32), [[ASHR]](s32), [[ASHR]](s32), [[ASHR]](s32), [[ASHR]](s32), [[ASHR]](s32), [[ASHR]](s32), [[ASHR]](s32), [[ASHR]](s32), [[ASHR]](s32), [[ASHR]](s32), [[ASHR]](s32), [[ASHR]](s32), [[ASHR]](s32)
; CHECK: S_ENDPGM 0, implicit [[MV]](s1024)
%0:_(s32) = COPY $vgpr0 %0:_(s32) = COPY $vgpr0
%1:_(s1024) = G_SEXT %0 %1:_(s1024) = G_SEXT %0
S_ENDPGM 0, implicit %1 S_ENDPGM 0, implicit %1
... ...
--- ---
name: test_sext_s64_to_s128 name: test_sext_s64_to_s128
body: | body: |
Show All 14 Lines
--- ---
name: test_sext_s64_to_s256 name: test_sext_s64_to_s256
body: | body: |
bb.0: bb.0:
liveins: $vgpr0_vgpr1 liveins: $vgpr0_vgpr1
; CHECK-LABEL: name: test_sext_s64_to_s256 ; CHECK-LABEL: name: test_sext_s64_to_s256
; CHECK: [[COPY:%[0-9]+]]:_(s64) = COPY $vgpr0_vgpr1 ; CHECK: [[COPY:%[0-9]+]]:_(s64) = COPY $vgpr0_vgpr1
; CHECK: [[SEXT:%[0-9]+]]:_(s256) = G_SEXT [[COPY]](s64) ; CHECK: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 63
; CHECK: S_ENDPGM 0, implicit [[SEXT]](s256) ; CHECK: [[ASHR:%[0-9]+]]:_(s64) = G_ASHR [[COPY]], [[C]](s32)
; CHECK: [[MV:%[0-9]+]]:_(s256) = G_MERGE_VALUES [[COPY]](s64), [[ASHR]](s64), [[ASHR]](s64), [[ASHR]](s64)
; CHECK: S_ENDPGM 0, implicit [[MV]](s256)
%0:_(s64) = COPY $vgpr0_vgpr1 %0:_(s64) = COPY $vgpr0_vgpr1
%1:_(s256) = G_SEXT %0 %1:_(s256) = G_SEXT %0
S_ENDPGM 0, implicit %1 S_ENDPGM 0, implicit %1
... ...
--- ---
name: test_sext_s64_to_s512 name: test_sext_s64_to_s512
body: | body: |
bb.0: bb.0:
liveins: $vgpr0_vgpr1 liveins: $vgpr0_vgpr1
; CHECK-LABEL: name: test_sext_s64_to_s512 ; CHECK-LABEL: name: test_sext_s64_to_s512
; CHECK: [[COPY:%[0-9]+]]:_(s64) = COPY $vgpr0_vgpr1 ; CHECK: [[COPY:%[0-9]+]]:_(s64) = COPY $vgpr0_vgpr1
; CHECK: [[SEXT:%[0-9]+]]:_(s512) = G_SEXT [[COPY]](s64) ; CHECK: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 63
; CHECK: S_ENDPGM 0, implicit [[SEXT]](s512) ; CHECK: [[ASHR:%[0-9]+]]:_(s64) = G_ASHR [[COPY]], [[C]](s32)
; CHECK: [[MV:%[0-9]+]]:_(s512) = G_MERGE_VALUES [[COPY]](s64), [[ASHR]](s64), [[ASHR]](s64), [[ASHR]](s64), [[ASHR]](s64), [[ASHR]](s64), [[ASHR]](s64), [[ASHR]](s64)
; CHECK: S_ENDPGM 0, implicit [[MV]](s512)
%0:_(s64) = COPY $vgpr0_vgpr1 %0:_(s64) = COPY $vgpr0_vgpr1
%1:_(s512) = G_SEXT %0 %1:_(s512) = G_SEXT %0
S_ENDPGM 0, implicit %1 S_ENDPGM 0, implicit %1
... ...
--- ---
name: test_sext_s64_to_s1024 name: test_sext_s64_to_s1024
body: | body: |
bb.0: bb.0:
liveins: $vgpr0_vgpr1 liveins: $vgpr0_vgpr1
; CHECK-LABEL: name: test_sext_s64_to_s1024 ; CHECK-LABEL: name: test_sext_s64_to_s1024
; CHECK: [[COPY:%[0-9]+]]:_(s64) = COPY $vgpr0_vgpr1 ; CHECK: [[COPY:%[0-9]+]]:_(s64) = COPY $vgpr0_vgpr1
; CHECK: [[SEXT:%[0-9]+]]:_(s1024) = G_SEXT [[COPY]](s64) ; CHECK: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 63
; CHECK: S_ENDPGM 0, implicit [[SEXT]](s1024) ; CHECK: [[ASHR:%[0-9]+]]:_(s64) = G_ASHR [[COPY]], [[C]](s32)
; CHECK: [[MV:%[0-9]+]]:_(s1024) = G_MERGE_VALUES [[COPY]](s64), [[ASHR]](s64), [[ASHR]](s64), [[ASHR]](s64), [[ASHR]](s64), [[ASHR]](s64), [[ASHR]](s64), [[ASHR]](s64), [[ASHR]](s64), [[ASHR]](s64), [[ASHR]](s64), [[ASHR]](s64), [[ASHR]](s64), [[ASHR]](s64), [[ASHR]](s64), [[ASHR]](s64)
; CHECK: S_ENDPGM 0, implicit [[MV]](s1024)
%0:_(s64) = COPY $vgpr0_vgpr1 %0:_(s64) = COPY $vgpr0_vgpr1
%1:_(s1024) = G_SEXT %0 %1:_(s1024) = G_SEXT %0
S_ENDPGM 0, implicit %1 S_ENDPGM 0, implicit %1
... ...
# --- # ---
# name: test_sext_s96_to_s128 # name: test_sext_s96_to_s128
# body: | # body: |
# bb.0: # bb.0:
# liveins: $vgpr0_vgpr1_vgpr2 # liveins: $vgpr0_vgpr1_vgpr2
# %0:_(s96) = COPY $vgpr0_vgpr1_vgpr2 # %0:_(s96) = COPY $vgpr0_vgpr1_vgpr2
# %1:_(s128) = G_SEXT %0 # %1:_(s128) = G_SEXT %0
# S_ENDPGM 0, implicit %1 # S_ENDPGM 0, implicit %1
# ... # ...
--- ---
name: test_sext_s128_to_s256 name: test_sext_s128_to_s256
body: | body: |
bb.0: bb.0:
liveins: $vgpr0_vgpr1_vgpr2_vgpr3 liveins: $vgpr0_vgpr1_vgpr2_vgpr3
; CHECK-LABEL: name: test_sext_s128_to_s256 ; CHECK-LABEL: name: test_sext_s128_to_s256
; CHECK: [[COPY:%[0-9]+]]:_(s128) = COPY $vgpr0_vgpr1_vgpr2_vgpr3 ; CHECK: [[COPY:%[0-9]+]]:_(s128) = COPY $vgpr0_vgpr1_vgpr2_vgpr3
; CHECK: [[SEXT:%[0-9]+]]:_(s256) = G_SEXT [[COPY]](s128) ; CHECK: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 127
; CHECK: S_ENDPGM 0, implicit [[SEXT]](s256) ; CHECK: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 64
; CHECK: [[UV:%[0-9]+]]:_(s64), [[UV1:%[0-9]+]]:_(s64) = G_UNMERGE_VALUES [[COPY]](s128)
; CHECK: [[SUB:%[0-9]+]]:_(s32) = G_SUB [[C]], [[C1]]
; CHECK: [[SUB1:%[0-9]+]]:_(s32) = G_SUB [[C1]], [[C]]
; CHECK: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 0
; CHECK: [[ICMP:%[0-9]+]]:_(s1) = G_ICMP intpred(ult), [[C]](s32), [[C1]]
; CHECK: [[ICMP1:%[0-9]+]]:_(s1) = G_ICMP intpred(eq), [[C]](s32), [[C2]]
; CHECK: [[ASHR:%[0-9]+]]:_(s64) = G_ASHR [[UV1]], [[C]](s32)
; CHECK: [[LSHR:%[0-9]+]]:_(s64) = G_LSHR [[UV]], [[C]](s32)
; CHECK: [[SHL:%[0-9]+]]:_(s64) = G_SHL [[UV1]], [[SUB1]](s32)
; CHECK: [[OR:%[0-9]+]]:_(s64) = G_OR [[LSHR]], [[SHL]]
; CHECK: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 63
; CHECK: [[ASHR1:%[0-9]+]]:_(s64) = G_ASHR [[UV1]], [[C3]](s32)
; CHECK: [[ASHR2:%[0-9]+]]:_(s64) = G_ASHR [[UV1]], [[SUB]](s32)
; CHECK: [[SELECT:%[0-9]+]]:_(s64) = G_SELECT [[ICMP]](s1), [[OR]], [[ASHR2]]
; CHECK: [[SELECT1:%[0-9]+]]:_(s64) = G_SELECT [[ICMP1]](s1), [[UV]], [[SELECT]]
; CHECK: [[SELECT2:%[0-9]+]]:_(s64) = G_SELECT [[ICMP]](s1), [[ASHR]], [[ASHR1]]
; CHECK: [[MV:%[0-9]+]]:_(s128) = G_MERGE_VALUES [[SELECT1]](s64), [[SELECT2]](s64)
; CHECK: [[MV1:%[0-9]+]]:_(s256) = G_MERGE_VALUES [[COPY]](s128), [[MV]](s128)
; CHECK: S_ENDPGM 0, implicit [[MV1]](s256)
%0:_(s128) = COPY $vgpr0_vgpr1_vgpr2_vgpr3 %0:_(s128) = COPY $vgpr0_vgpr1_vgpr2_vgpr3
%1:_(s256) = G_SEXT %0 %1:_(s256) = G_SEXT %0
S_ENDPGM 0, implicit %1 S_ENDPGM 0, implicit %1
... ...