[llvm] 346ee54 - [GlobalISel][X86] Add G_CTTZ_ZERO_UNDEF/G_CTTZ legalization handling

Simon Pilgrim via llvm-commits llvm-commits at lists.llvm.org
Mon Jun 5 03:18:42 PDT 2023


Author: Simon Pilgrim
Date: 2023-06-05T11:18:27+01:00
New Revision: 346ee549e5e25a68681b678020df8c38e7b68737

URL: https://github.com/llvm/llvm-project/commit/346ee549e5e25a68681b678020df8c38e7b68737
DIFF: https://github.com/llvm/llvm-project/commit/346ee549e5e25a68681b678020df8c38e7b68737.diff

LOG: [GlobalISel][X86] Add G_CTTZ_ZERO_UNDEF/G_CTTZ legalization handling

G_CTTZ_ZERO_UNDEF is always legal using the BSF instruction, G_CTTZ requires the BMI1 TZCNT instruction

Added: 
    llvm/test/CodeGen/X86/GlobalISel/legalize-trailing-zeros-undef.mir
    llvm/test/CodeGen/X86/GlobalISel/legalize-trailing-zeros.mir

Modified: 
    llvm/lib/Target/X86/X86LegalizerInfo.cpp

Removed: 
    


################################################################################
diff  --git a/llvm/lib/Target/X86/X86LegalizerInfo.cpp b/llvm/lib/Target/X86/X86LegalizerInfo.cpp
index 626ef556bf172..2d092f25355eb 100644
--- a/llvm/lib/Target/X86/X86LegalizerInfo.cpp
+++ b/llvm/lib/Target/X86/X86LegalizerInfo.cpp
@@ -237,6 +237,16 @@ X86LegalizerInfo::X86LegalizerInfo(const X86Subtarget &STI,
       .widenScalarToNextPow2(1, /*Min=*/16)
       .clampScalar(1, s16, sMaxScalar);
 
+  // count trailing zeros
+  getActionDefinitionsBuilder({G_CTTZ_ZERO_UNDEF, G_CTTZ})
+      .legalIf([=](const LegalityQuery &Query) -> bool {
+        return (Query.Opcode == G_CTTZ_ZERO_UNDEF || Subtarget.hasBMI()) &&
+               (typePairInSet(0, 1, {{s16, s16}, {s32, s32}})(Query) ||
+                (Is64Bit && typePairInSet(0, 1, {{s64, s64}})(Query)));
+      })
+      .widenScalarToNextPow2(1, /*Min=*/16)
+      .clampScalar(1, s16, sMaxScalar);
+
   // pointer handling
   const std::initializer_list<LLT> PtrTypes32 = {s1, s8, s16, s32};
   const std::initializer_list<LLT> PtrTypes64 = {s1, s8, s16, s32, s64};

diff  --git a/llvm/test/CodeGen/X86/GlobalISel/legalize-trailing-zeros-undef.mir b/llvm/test/CodeGen/X86/GlobalISel/legalize-trailing-zeros-undef.mir
new file mode 100644
index 0000000000000..b8e972c5e6492
--- /dev/null
+++ b/llvm/test/CodeGen/X86/GlobalISel/legalize-trailing-zeros-undef.mir
@@ -0,0 +1,140 @@
+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py UTC_ARGS: --version 2
+# RUN: llc -mtriple=x86_64-linux-gnu -run-pass=legalizer %s -o - | FileCheck %s --check-prefixes=CHECK,X64
+# RUN: llc -mtriple=i386-linux-gnu -run-pass=legalizer -global-isel-abort=2 -pass-remarks-missed='gisel*'  %s 2>%t -o - | FileCheck %s --check-prefixes=CHECK,X86
+# RUN: FileCheck -check-prefix=ERR32  %s < %t
+
+# ERR32: remark: <unknown>:0:0: unable to legalize instruction: %4:_(s64) = G_SELECT %8:_(s1), %11:_, %12:_ (in function: test_cttz35)
+# ERR32: remark: <unknown>:0:0: unable to legalize instruction: %1:_(s64) = G_SELECT %6:_(s1), %9:_, %10:_ (in function: test_cttz64)
+
+# test count leading zeros for s16, s32, and s64
+
+---
+name:            test_cttz35
+alignment:       16
+legalized:       false
+regBankSelected: false
+registers:
+  - { id: 0, class: _, preferred-register: '' }
+  - { id: 1, class: _, preferred-register: '' }
+body:             |
+  bb.1:
+    ; X64-LABEL: name: test_cttz35
+    ; X64: [[DEF:%[0-9]+]]:_(s35) = IMPLICIT_DEF
+    ; X64-NEXT: [[ANYEXT:%[0-9]+]]:_(s64) = G_ANYEXT [[DEF]](s35)
+    ; X64-NEXT: [[CTTZ_ZERO_UNDEF:%[0-9]+]]:_(s64) = G_CTTZ_ZERO_UNDEF [[ANYEXT]](s64)
+    ; X64-NEXT: [[TRUNC:%[0-9]+]]:_(s35) = G_TRUNC [[CTTZ_ZERO_UNDEF]](s64)
+    ; X64-NEXT: [[COPY:%[0-9]+]]:_(s35) = COPY [[TRUNC]](s35)
+    ; X64-NEXT: RET 0, implicit [[COPY]](s35)
+    ; X86-LABEL: name: test_cttz35
+    ; X86: [[DEF:%[0-9]+]]:_(s35) = IMPLICIT_DEF
+    ; X86-NEXT: [[ANYEXT:%[0-9]+]]:_(s64) = G_ANYEXT [[DEF]](s35)
+    ; X86-NEXT: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[ANYEXT]](s64)
+    ; X86-NEXT: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 0
+    ; X86-NEXT: [[ICMP:%[0-9]+]]:_(s1) = G_ICMP intpred(eq), [[UV]](s32), [[C]]
+    ; X86-NEXT: [[CTTZ_ZERO_UNDEF:%[0-9]+]]:_(s64) = G_CTTZ_ZERO_UNDEF [[UV1]](s32)
+    ; X86-NEXT: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 32
+    ; X86-NEXT: [[ADD:%[0-9]+]]:_(s64) = G_ADD [[CTTZ_ZERO_UNDEF]], [[C1]]
+    ; X86-NEXT: [[CTTZ_ZERO_UNDEF1:%[0-9]+]]:_(s64) = G_CTTZ_ZERO_UNDEF [[UV]](s32)
+    ; X86-NEXT: [[SELECT:%[0-9]+]]:_(s64) = G_SELECT [[ICMP]](s1), [[ADD]], [[CTTZ_ZERO_UNDEF1]]
+    ; X86-NEXT: [[TRUNC:%[0-9]+]]:_(s35) = G_TRUNC [[SELECT]](s64)
+    ; X86-NEXT: [[COPY:%[0-9]+]]:_(s35) = COPY [[TRUNC]](s35)
+    ; X86-NEXT: RET 0, implicit [[COPY]](s35)
+    %0:_(s35) = IMPLICIT_DEF
+    %1:_(s35) = G_CTTZ_ZERO_UNDEF %0
+    %2:_(s35) = COPY %1(s35)
+    RET 0, implicit %2
+...
+---
+name:            test_cttz8
+alignment:       16
+legalized:       false
+regBankSelected: false
+registers:
+  - { id: 0, class: _, preferred-register: '' }
+  - { id: 1, class: _, preferred-register: '' }
+body:             |
+  bb.1:
+    ; CHECK-LABEL: name: test_cttz8
+    ; CHECK: [[DEF:%[0-9]+]]:_(s8) = IMPLICIT_DEF
+    ; CHECK-NEXT: [[ANYEXT:%[0-9]+]]:_(s16) = G_ANYEXT [[DEF]](s8)
+    ; CHECK-NEXT: [[CTTZ_ZERO_UNDEF:%[0-9]+]]:_(s16) = G_CTTZ_ZERO_UNDEF [[ANYEXT]](s16)
+    ; CHECK-NEXT: [[TRUNC:%[0-9]+]]:_(s8) = G_TRUNC [[CTTZ_ZERO_UNDEF]](s16)
+    ; CHECK-NEXT: [[COPY:%[0-9]+]]:_(s8) = COPY [[TRUNC]](s8)
+    ; CHECK-NEXT: RET 0, implicit [[COPY]](s8)
+    %0:_(s8) = IMPLICIT_DEF
+    %1:_(s8) = G_CTTZ_ZERO_UNDEF %0
+    %2:_(s8) = COPY %1(s8)
+    RET 0, implicit %2
+...
+---
+name:            test_cttz64
+alignment:       16
+legalized:       false
+regBankSelected: false
+registers:
+  - { id: 0, class: _, preferred-register: '' }
+  - { id: 1, class: _, preferred-register: '' }
+body:             |
+  bb.1:
+    ; X64-LABEL: name: test_cttz64
+    ; X64: [[DEF:%[0-9]+]]:_(s64) = IMPLICIT_DEF
+    ; X64-NEXT: [[CTTZ_ZERO_UNDEF:%[0-9]+]]:_(s64) = G_CTTZ_ZERO_UNDEF [[DEF]](s64)
+    ; X64-NEXT: [[COPY:%[0-9]+]]:_(s64) = COPY [[CTTZ_ZERO_UNDEF]](s64)
+    ; X64-NEXT: RET 0, implicit [[COPY]](s64)
+    ; X86-LABEL: name: test_cttz64
+    ; X86: [[DEF:%[0-9]+]]:_(s64) = IMPLICIT_DEF
+    ; X86-NEXT: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[DEF]](s64)
+    ; X86-NEXT: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 0
+    ; X86-NEXT: [[ICMP:%[0-9]+]]:_(s1) = G_ICMP intpred(eq), [[UV]](s32), [[C]]
+    ; X86-NEXT: [[CTTZ_ZERO_UNDEF:%[0-9]+]]:_(s64) = G_CTTZ_ZERO_UNDEF [[UV1]](s32)
+    ; X86-NEXT: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 32
+    ; X86-NEXT: [[ADD:%[0-9]+]]:_(s64) = G_ADD [[CTTZ_ZERO_UNDEF]], [[C1]]
+    ; X86-NEXT: [[CTTZ_ZERO_UNDEF1:%[0-9]+]]:_(s64) = G_CTTZ_ZERO_UNDEF [[UV]](s32)
+    ; X86-NEXT: [[SELECT:%[0-9]+]]:_(s64) = G_SELECT [[ICMP]](s1), [[ADD]], [[CTTZ_ZERO_UNDEF1]]
+    ; X86-NEXT: [[COPY:%[0-9]+]]:_(s64) = COPY [[SELECT]](s64)
+    ; X86-NEXT: RET 0, implicit [[COPY]](s64)
+    %0:_(s64) = IMPLICIT_DEF
+    %1:_(s64) = G_CTTZ_ZERO_UNDEF %0
+    %2:_(s64) = COPY %1(s64)
+    RET 0, implicit %2
+...
+---
+name:            test_cttz32
+alignment:       16
+legalized:       false
+regBankSelected: false
+registers:
+  - { id: 0, class: _, preferred-register: '' }
+  - { id: 1, class: _, preferred-register: '' }
+body:             |
+  bb.1:
+    ; CHECK-LABEL: name: test_cttz32
+    ; CHECK: [[DEF:%[0-9]+]]:_(s32) = IMPLICIT_DEF
+    ; CHECK-NEXT: [[CTTZ_ZERO_UNDEF:%[0-9]+]]:_(s32) = G_CTTZ_ZERO_UNDEF [[DEF]](s32)
+    ; CHECK-NEXT: [[COPY:%[0-9]+]]:_(s32) = COPY [[CTTZ_ZERO_UNDEF]](s32)
+    ; CHECK-NEXT: RET 0, implicit [[COPY]](s32)
+    %0:_(s32) = IMPLICIT_DEF
+    %1:_(s32) = G_CTTZ_ZERO_UNDEF %0
+    %2:_(s32) = COPY %1(s32)
+    RET 0, implicit %2
+...
+---
+name:            test_cttz16
+alignment:       16
+legalized:       false
+regBankSelected: false
+registers:
+  - { id: 0, class: _, preferred-register: '' }
+  - { id: 1, class: _, preferred-register: '' }
+body:             |
+  bb.1:
+    ; CHECK-LABEL: name: test_cttz16
+    ; CHECK: [[DEF:%[0-9]+]]:_(s16) = IMPLICIT_DEF
+    ; CHECK-NEXT: [[CTTZ_ZERO_UNDEF:%[0-9]+]]:_(s16) = G_CTTZ_ZERO_UNDEF [[DEF]](s16)
+    ; CHECK-NEXT: [[COPY:%[0-9]+]]:_(s16) = COPY [[CTTZ_ZERO_UNDEF]](s16)
+    ; CHECK-NEXT: RET 0, implicit [[COPY]](s16)
+    %0:_(s16) = IMPLICIT_DEF
+    %1:_(s16) = G_CTTZ_ZERO_UNDEF %0
+    %2:_(s16) = COPY %1(s16)
+    RET 0, implicit %2
+...

diff  --git a/llvm/test/CodeGen/X86/GlobalISel/legalize-trailing-zeros.mir b/llvm/test/CodeGen/X86/GlobalISel/legalize-trailing-zeros.mir
new file mode 100644
index 0000000000000..7ee2b3644928c
--- /dev/null
+++ b/llvm/test/CodeGen/X86/GlobalISel/legalize-trailing-zeros.mir
@@ -0,0 +1,146 @@
+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py UTC_ARGS: --version 2
+# RUN: llc -mtriple=x86_64-linux-gnu -mattr=+bmi -run-pass=legalizer %s -o - | FileCheck %s --check-prefixes=CHECK,X64
+# RUN: llc -mtriple=i386-linux-gnu -mattr=+bmi -run-pass=legalizer -global-isel-abort=2 -pass-remarks-missed='gisel*'  %s 2>%t -o - | FileCheck %s --check-prefixes=CHECK,X86
+# RUN: FileCheck -check-prefix=ERR32  %s < %t
+
+# ERR32: remark: <unknown>:0:0: unable to legalize instruction: %6:_(s64) = G_SELECT %10:_(s1), %13:_, %14:_ (in function: test_cttz35)
+# ERR32: remark: <unknown>:0:0: unable to legalize instruction: %1:_(s64) = G_SELECT %6:_(s1), %9:_, %10:_ (in function: test_cttz64)
+
+# test count leading zeros for s16, s32, and s64
+
+---
+name:            test_cttz35
+alignment:       16
+legalized:       false
+regBankSelected: false
+registers:
+  - { id: 0, class: _, preferred-register: '' }
+  - { id: 1, class: _, preferred-register: '' }
+body:             |
+  bb.1:
+    ; X64-LABEL: name: test_cttz35
+    ; X64: [[DEF:%[0-9]+]]:_(s35) = IMPLICIT_DEF
+    ; X64-NEXT: [[ANYEXT:%[0-9]+]]:_(s64) = G_ANYEXT [[DEF]](s35)
+    ; X64-NEXT: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 34359738368
+    ; X64-NEXT: [[OR:%[0-9]+]]:_(s64) = G_OR [[ANYEXT]], [[C]]
+    ; X64-NEXT: [[CTTZ_ZERO_UNDEF:%[0-9]+]]:_(s64) = G_CTTZ_ZERO_UNDEF [[OR]](s64)
+    ; X64-NEXT: [[TRUNC:%[0-9]+]]:_(s35) = G_TRUNC [[CTTZ_ZERO_UNDEF]](s64)
+    ; X64-NEXT: [[COPY:%[0-9]+]]:_(s35) = COPY [[TRUNC]](s35)
+    ; X64-NEXT: RET 0, implicit [[COPY]](s35)
+    ; X86-LABEL: name: test_cttz35
+    ; X86: [[DEF:%[0-9]+]]:_(s35) = IMPLICIT_DEF
+    ; X86-NEXT: [[ANYEXT:%[0-9]+]]:_(s64) = G_ANYEXT [[DEF]](s35)
+    ; X86-NEXT: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 34359738368
+    ; X86-NEXT: [[OR:%[0-9]+]]:_(s64) = G_OR [[ANYEXT]], [[C]]
+    ; X86-NEXT: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[OR]](s64)
+    ; X86-NEXT: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 0
+    ; X86-NEXT: [[ICMP:%[0-9]+]]:_(s1) = G_ICMP intpred(eq), [[UV]](s32), [[C1]]
+    ; X86-NEXT: [[CTTZ_ZERO_UNDEF:%[0-9]+]]:_(s64) = G_CTTZ_ZERO_UNDEF [[UV1]](s32)
+    ; X86-NEXT: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 32
+    ; X86-NEXT: [[ADD:%[0-9]+]]:_(s64) = G_ADD [[CTTZ_ZERO_UNDEF]], [[C2]]
+    ; X86-NEXT: [[CTTZ_ZERO_UNDEF1:%[0-9]+]]:_(s64) = G_CTTZ_ZERO_UNDEF [[UV]](s32)
+    ; X86-NEXT: [[SELECT:%[0-9]+]]:_(s64) = G_SELECT [[ICMP]](s1), [[ADD]], [[CTTZ_ZERO_UNDEF1]]
+    ; X86-NEXT: [[TRUNC:%[0-9]+]]:_(s35) = G_TRUNC [[SELECT]](s64)
+    ; X86-NEXT: [[COPY:%[0-9]+]]:_(s35) = COPY [[TRUNC]](s35)
+    ; X86-NEXT: RET 0, implicit [[COPY]](s35)
+    %0:_(s35) = IMPLICIT_DEF
+    %1:_(s35) = G_CTTZ %0
+    %2:_(s35) = COPY %1(s35)
+    RET 0, implicit %2
+...
+---
+name:            test_cttz8
+alignment:       16
+legalized:       false
+regBankSelected: false
+registers:
+  - { id: 0, class: _, preferred-register: '' }
+  - { id: 1, class: _, preferred-register: '' }
+body:             |
+  bb.1:
+    ; CHECK-LABEL: name: test_cttz8
+    ; CHECK: [[DEF:%[0-9]+]]:_(s8) = IMPLICIT_DEF
+    ; CHECK-NEXT: [[ANYEXT:%[0-9]+]]:_(s16) = G_ANYEXT [[DEF]](s8)
+    ; CHECK-NEXT: [[C:%[0-9]+]]:_(s16) = G_CONSTANT i16 256
+    ; CHECK-NEXT: [[OR:%[0-9]+]]:_(s16) = G_OR [[ANYEXT]], [[C]]
+    ; CHECK-NEXT: [[CTTZ_ZERO_UNDEF:%[0-9]+]]:_(s16) = G_CTTZ_ZERO_UNDEF [[OR]](s16)
+    ; CHECK-NEXT: [[TRUNC:%[0-9]+]]:_(s8) = G_TRUNC [[CTTZ_ZERO_UNDEF]](s16)
+    ; CHECK-NEXT: [[COPY:%[0-9]+]]:_(s8) = COPY [[TRUNC]](s8)
+    ; CHECK-NEXT: RET 0, implicit [[COPY]](s8)
+    %0:_(s8) = IMPLICIT_DEF
+    %1:_(s8) = G_CTTZ %0
+    %2:_(s8) = COPY %1(s8)
+    RET 0, implicit %2
+...
+---
+name:            test_cttz64
+alignment:       16
+legalized:       false
+regBankSelected: false
+registers:
+  - { id: 0, class: _, preferred-register: '' }
+  - { id: 1, class: _, preferred-register: '' }
+body:             |
+  bb.1:
+    ; X64-LABEL: name: test_cttz64
+    ; X64: [[DEF:%[0-9]+]]:_(s64) = IMPLICIT_DEF
+    ; X64-NEXT: [[CTTZ:%[0-9]+]]:_(s64) = G_CTTZ [[DEF]](s64)
+    ; X64-NEXT: [[COPY:%[0-9]+]]:_(s64) = COPY [[CTTZ]](s64)
+    ; X64-NEXT: RET 0, implicit [[COPY]](s64)
+    ; X86-LABEL: name: test_cttz64
+    ; X86: [[DEF:%[0-9]+]]:_(s64) = IMPLICIT_DEF
+    ; X86-NEXT: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[DEF]](s64)
+    ; X86-NEXT: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 0
+    ; X86-NEXT: [[ICMP:%[0-9]+]]:_(s1) = G_ICMP intpred(eq), [[UV]](s32), [[C]]
+    ; X86-NEXT: [[CTTZ:%[0-9]+]]:_(s64) = G_CTTZ [[UV1]](s32)
+    ; X86-NEXT: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 32
+    ; X86-NEXT: [[ADD:%[0-9]+]]:_(s64) = G_ADD [[CTTZ]], [[C1]]
+    ; X86-NEXT: [[CTTZ_ZERO_UNDEF:%[0-9]+]]:_(s64) = G_CTTZ_ZERO_UNDEF [[UV]](s32)
+    ; X86-NEXT: [[SELECT:%[0-9]+]]:_(s64) = G_SELECT [[ICMP]](s1), [[ADD]], [[CTTZ_ZERO_UNDEF]]
+    ; X86-NEXT: [[COPY:%[0-9]+]]:_(s64) = COPY [[SELECT]](s64)
+    ; X86-NEXT: RET 0, implicit [[COPY]](s64)
+    %0:_(s64) = IMPLICIT_DEF
+    %1:_(s64) = G_CTTZ %0
+    %2:_(s64) = COPY %1(s64)
+    RET 0, implicit %2
+...
+---
+name:            test_cttz32
+alignment:       16
+legalized:       false
+regBankSelected: false
+registers:
+  - { id: 0, class: _, preferred-register: '' }
+  - { id: 1, class: _, preferred-register: '' }
+body:             |
+  bb.1:
+    ; CHECK-LABEL: name: test_cttz32
+    ; CHECK: [[DEF:%[0-9]+]]:_(s32) = IMPLICIT_DEF
+    ; CHECK-NEXT: [[CTTZ:%[0-9]+]]:_(s32) = G_CTTZ [[DEF]](s32)
+    ; CHECK-NEXT: [[COPY:%[0-9]+]]:_(s32) = COPY [[CTTZ]](s32)
+    ; CHECK-NEXT: RET 0, implicit [[COPY]](s32)
+    %0:_(s32) = IMPLICIT_DEF
+    %1:_(s32) = G_CTTZ %0
+    %2:_(s32) = COPY %1(s32)
+    RET 0, implicit %2
+...
+---
+name:            test_cttz16
+alignment:       16
+legalized:       false
+regBankSelected: false
+registers:
+  - { id: 0, class: _, preferred-register: '' }
+  - { id: 1, class: _, preferred-register: '' }
+body:             |
+  bb.1:
+    ; CHECK-LABEL: name: test_cttz16
+    ; CHECK: [[DEF:%[0-9]+]]:_(s16) = IMPLICIT_DEF
+    ; CHECK-NEXT: [[CTTZ:%[0-9]+]]:_(s16) = G_CTTZ [[DEF]](s16)
+    ; CHECK-NEXT: [[COPY:%[0-9]+]]:_(s16) = COPY [[CTTZ]](s16)
+    ; CHECK-NEXT: RET 0, implicit [[COPY]](s16)
+    %0:_(s16) = IMPLICIT_DEF
+    %1:_(s16) = G_CTTZ %0
+    %2:_(s16) = COPY %1(s16)
+    RET 0, implicit %2
+...


        


More information about the llvm-commits mailing list