-
Notifications
You must be signed in to change notification settings - Fork 10.8k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
[GlobalISel] Combine (X == 0) & (Y == 0) -> (X | Y) == 0 #71949
[GlobalISel] Combine (X == 0) & (Y == 0) -> (X | Y) == 0 #71949
Conversation
@llvm/pr-subscribers-llvm-globalisel @llvm/pr-subscribers-backend-aarch64 Author: Dávid Ferenc Szabó (dfszabo) ChangesAlso combine (X != 0) | (Y != 0) -> (X | Y) != 0 Patch is 30.18 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/71949.diff 3 Files Affected:
diff --git a/llvm/include/llvm/Target/GlobalISel/Combine.td b/llvm/include/llvm/Target/GlobalISel/Combine.td
index 7e0691e1ee95048..551fa812cee21ed 100644
--- a/llvm/include/llvm/Target/GlobalISel/Combine.td
+++ b/llvm/include/llvm/Target/GlobalISel/Combine.td
@@ -903,6 +903,37 @@ def redundant_binop_in_equality : GICombineRule<
[{ return Helper.matchRedundantBinOpInEquality(*${root}, ${info}); }]),
(apply [{ Helper.applyBuildFn(*${root}, ${info}); }])>;
+// Transform: (X == 0 & Y == 0) -> (X | Y) == 0
+def double_icmp_zero_and_combine: GICombineRule<
+ (defs root:$root),
+ (match (G_ICMP $d1, $p, $s1, 0),
+ (G_ICMP $d2, $p, $s2, 0),
+ (G_AND $root, $d1, $d2),
+ [{ return ${p}.getPredicate() == CmpInst::ICMP_EQ &&
+ !MRI.getType(${s1}.getReg()).isPointer() &&
+ (MRI.getType(${s1}.getReg()) ==
+ MRI.getType(${s2}.getReg())); }]),
+ (apply (G_OR i32:$ordst, $s1, $s2),
+ (G_ICMP $root, $p, $ordst, 0))
+>;
+
+// Transform: (X != 0 | Y != 0) -> (X | Y) != 0
+def double_icmp_zero_or_combine: GICombineRule<
+ (defs root:$root),
+ (match (G_ICMP $d1, $p, $s1, 0),
+ (G_ICMP $d2, $p, $s2, 0),
+ (G_OR $root, $d1, $d2),
+ [{ return ${p}.getPredicate() == CmpInst::ICMP_NE &&
+ !MRI.getType(${s1}.getReg()).isPointer() &&
+ (MRI.getType(${s1}.getReg()) ==
+ MRI.getType(${s2}.getReg())); }]),
+ (apply (G_OR i32:$ordst, $s1, $s2),
+ (G_ICMP $root, $p, $ordst, 0))
+>;
+
+def double_icmp_zero_and_or_combine : GICombineGroup<[double_icmp_zero_and_combine,
+ double_icmp_zero_or_combine]>;
+
def and_or_disjoint_mask : GICombineRule<
(defs root:$root, build_fn_matchinfo:$info),
(match (wip_match_opcode G_AND):$root,
@@ -1265,7 +1296,7 @@ def all_combines : GICombineGroup<[trivial_combines, insert_vec_elt_combines,
intdiv_combines, mulh_combines, redundant_neg_operands,
and_or_disjoint_mask, fma_combines, fold_binop_into_select,
sub_add_reg, select_to_minmax, redundant_binop_in_equality,
- fsub_to_fneg, commute_constant_to_rhs]>;
+ double_icmp_zero_and_or_combine, fsub_to_fneg, commute_constant_to_rhs]>;
// A combine group used to for prelegalizer combiners at -O0. The combines in
// this group have been selected based on experiments to balance code size and
diff --git a/llvm/test/CodeGen/AArch64/GlobalISel/combine-2-icmps-of-0-and-or.mir b/llvm/test/CodeGen/AArch64/GlobalISel/combine-2-icmps-of-0-and-or.mir
new file mode 100644
index 000000000000000..a285ccf0c88fea4
--- /dev/null
+++ b/llvm/test/CodeGen/AArch64/GlobalISel/combine-2-icmps-of-0-and-or.mir
@@ -0,0 +1,764 @@
+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
+# RUN: llc -mtriple aarch64 -run-pass=aarch64-prelegalizer-combiner -verify-machineinstrs %s -o - | FileCheck %s
+# REQUIRES: asserts
+
+
+---
+name: valid_and_eq_0_eq_0_s32
+tracksRegLiveness: true
+legalized: true
+body: |
+ bb.0:
+ liveins: $w0, $w1
+
+ ; CHECK-LABEL: name: valid_and_eq_0_eq_0_s32
+ ; CHECK: liveins: $w0, $w1
+ ; CHECK-NEXT: {{ $}}
+ ; CHECK-NEXT: %x:_(s32) = COPY $w0
+ ; CHECK-NEXT: %y:_(s32) = COPY $w1
+ ; CHECK-NEXT: %zero:_(s32) = G_CONSTANT i32 0
+ ; CHECK-NEXT: [[OR:%[0-9]+]]:_(s32) = G_OR %x, %y
+ ; CHECK-NEXT: %and:_(s1) = G_ICMP intpred(eq), [[OR]](s32), %zero
+ ; CHECK-NEXT: %zext:_(s32) = G_ZEXT %and(s1)
+ ; CHECK-NEXT: $w0 = COPY %zext(s32)
+ ; CHECK-NEXT: RET_ReallyLR implicit $w0
+ %x:_(s32) = COPY $w0
+ %y:_(s32) = COPY $w1
+ %zero:_(s32) = G_CONSTANT i32 0
+ %cmp1:_(s1) = G_ICMP intpred(eq), %x:_(s32), %zero:_
+ %cmp2:_(s1) = G_ICMP intpred(eq), %y:_(s32), %zero:_
+ %and:_(s1) = G_AND %cmp1, %cmp2
+ %zext:_(s32) = G_ZEXT %and:_(s1)
+ $w0 = COPY %zext
+ RET_ReallyLR implicit $w0
+
+...
+---
+name: invalid_and_eq_1_eq_0_s32
+tracksRegLiveness: true
+legalized: true
+body: |
+ bb.0:
+ liveins: $w0, $w1
+
+ ; CHECK-LABEL: name: invalid_and_eq_1_eq_0_s32
+ ; CHECK: liveins: $w0, $w1
+ ; CHECK-NEXT: {{ $}}
+ ; CHECK-NEXT: %x:_(s32) = COPY $w0
+ ; CHECK-NEXT: %y:_(s32) = COPY $w1
+ ; CHECK-NEXT: %zero:_(s32) = G_CONSTANT i32 0
+ ; CHECK-NEXT: %one:_(s32) = G_CONSTANT i32 1
+ ; CHECK-NEXT: %cmp1:_(s1) = G_ICMP intpred(eq), %x(s32), %one
+ ; CHECK-NEXT: %cmp2:_(s1) = G_ICMP intpred(eq), %y(s32), %zero
+ ; CHECK-NEXT: %and:_(s1) = G_AND %cmp1, %cmp2
+ ; CHECK-NEXT: %zext:_(s32) = G_ZEXT %and(s1)
+ ; CHECK-NEXT: $w0 = COPY %zext(s32)
+ ; CHECK-NEXT: RET_ReallyLR implicit $w0
+ %x:_(s32) = COPY $w0
+ %y:_(s32) = COPY $w1
+ %zero:_(s32) = G_CONSTANT i32 0
+ %one:_(s32) = G_CONSTANT i32 1
+ %cmp1:_(s1) = G_ICMP intpred(eq), %x:_(s32), %one:_
+ %cmp2:_(s1) = G_ICMP intpred(eq), %y:_(s32), %zero:_
+ %and:_(s1) = G_AND %cmp1, %cmp2
+ %zext:_(s32) = G_ZEXT %and:_(s1)
+ $w0 = COPY %zext
+ RET_ReallyLR implicit $w0
+
+...
+---
+name: invalid_and_eq_0_eq_1_s32
+tracksRegLiveness: true
+legalized: true
+body: |
+ bb.0:
+ liveins: $w0, $w1
+
+ ; CHECK-LABEL: name: invalid_and_eq_0_eq_1_s32
+ ; CHECK: liveins: $w0, $w1
+ ; CHECK-NEXT: {{ $}}
+ ; CHECK-NEXT: %x:_(s32) = COPY $w0
+ ; CHECK-NEXT: %y:_(s32) = COPY $w1
+ ; CHECK-NEXT: %zero:_(s32) = G_CONSTANT i32 0
+ ; CHECK-NEXT: %one:_(s32) = G_CONSTANT i32 1
+ ; CHECK-NEXT: %cmp1:_(s1) = G_ICMP intpred(eq), %x(s32), %zero
+ ; CHECK-NEXT: %cmp2:_(s1) = G_ICMP intpred(eq), %y(s32), %one
+ ; CHECK-NEXT: %and:_(s1) = G_AND %cmp1, %cmp2
+ ; CHECK-NEXT: %zext:_(s32) = G_ZEXT %and(s1)
+ ; CHECK-NEXT: $w0 = COPY %zext(s32)
+ ; CHECK-NEXT: RET_ReallyLR implicit $w0
+ %x:_(s32) = COPY $w0
+ %y:_(s32) = COPY $w1
+ %zero:_(s32) = G_CONSTANT i32 0
+ %one:_(s32) = G_CONSTANT i32 1
+ %cmp1:_(s1) = G_ICMP intpred(eq), %x:_(s32), %zero:_
+ %cmp2:_(s1) = G_ICMP intpred(eq), %y:_(s32), %one:_
+ %and:_(s1) = G_AND %cmp1, %cmp2
+ %zext:_(s32) = G_ZEXT %and:_(s1)
+ $w0 = COPY %zext
+ RET_ReallyLR implicit $w0
+
+...
+---
+name: invalid_and_ne_0_eq_0_s32
+tracksRegLiveness: true
+legalized: true
+body: |
+ bb.0:
+ liveins: $w0, $w1
+
+ ; CHECK-LABEL: name: invalid_and_ne_0_eq_0_s32
+ ; CHECK: liveins: $w0, $w1
+ ; CHECK-NEXT: {{ $}}
+ ; CHECK-NEXT: %x:_(s32) = COPY $w0
+ ; CHECK-NEXT: %y:_(s32) = COPY $w1
+ ; CHECK-NEXT: %zero:_(s32) = G_CONSTANT i32 0
+ ; CHECK-NEXT: %cmp1:_(s1) = G_ICMP intpred(ne), %x(s32), %zero
+ ; CHECK-NEXT: %cmp2:_(s1) = G_ICMP intpred(eq), %y(s32), %zero
+ ; CHECK-NEXT: %and:_(s1) = G_AND %cmp1, %cmp2
+ ; CHECK-NEXT: %zext:_(s32) = G_ZEXT %and(s1)
+ ; CHECK-NEXT: $w0 = COPY %zext(s32)
+ ; CHECK-NEXT: RET_ReallyLR implicit $w0
+ %x:_(s32) = COPY $w0
+ %y:_(s32) = COPY $w1
+ %zero:_(s32) = G_CONSTANT i32 0
+ %cmp1:_(s1) = G_ICMP intpred(ne), %x:_(s32), %zero:_
+ %cmp2:_(s1) = G_ICMP intpred(eq), %y:_(s32), %zero:_
+ %and:_(s1) = G_AND %cmp1, %cmp2
+ %zext:_(s32) = G_ZEXT %and:_(s1)
+ $w0 = COPY %zext
+ RET_ReallyLR implicit $w0
+
+...
+---
+name: invalid_and_eq_0_ne_0_s32
+tracksRegLiveness: true
+legalized: true
+body: |
+ bb.0:
+ liveins: $w0, $w1
+
+ ; CHECK-LABEL: name: invalid_and_eq_0_ne_0_s32
+ ; CHECK: liveins: $w0, $w1
+ ; CHECK-NEXT: {{ $}}
+ ; CHECK-NEXT: %x:_(s32) = COPY $w0
+ ; CHECK-NEXT: %y:_(s32) = COPY $w1
+ ; CHECK-NEXT: %zero:_(s32) = G_CONSTANT i32 0
+ ; CHECK-NEXT: %cmp1:_(s1) = G_ICMP intpred(eq), %x(s32), %zero
+ ; CHECK-NEXT: %cmp2:_(s1) = G_ICMP intpred(ne), %y(s32), %zero
+ ; CHECK-NEXT: %and:_(s1) = G_AND %cmp1, %cmp2
+ ; CHECK-NEXT: %zext:_(s32) = G_ZEXT %and(s1)
+ ; CHECK-NEXT: $w0 = COPY %zext(s32)
+ ; CHECK-NEXT: RET_ReallyLR implicit $w0
+ %x:_(s32) = COPY $w0
+ %y:_(s32) = COPY $w1
+ %zero:_(s32) = G_CONSTANT i32 0
+ %cmp1:_(s1) = G_ICMP intpred(eq), %x:_(s32), %zero:_
+ %cmp2:_(s1) = G_ICMP intpred(ne), %y:_(s32), %zero:_
+ %and:_(s1) = G_AND %cmp1, %cmp2
+ %zext:_(s32) = G_ZEXT %and:_(s1)
+ $w0 = COPY %zext
+ RET_ReallyLR implicit $w0
+
+...
+---
+name: invalid_and_ne_0_ne_0_s32
+tracksRegLiveness: true
+legalized: true
+body: |
+ bb.0:
+ liveins: $w0, $w1
+
+ ; CHECK-LABEL: name: invalid_and_ne_0_ne_0_s32
+ ; CHECK: liveins: $w0, $w1
+ ; CHECK-NEXT: {{ $}}
+ ; CHECK-NEXT: %x:_(s32) = COPY $w0
+ ; CHECK-NEXT: %y:_(s32) = COPY $w1
+ ; CHECK-NEXT: %zero:_(s32) = G_CONSTANT i32 0
+ ; CHECK-NEXT: %cmp1:_(s1) = G_ICMP intpred(ne), %x(s32), %zero
+ ; CHECK-NEXT: %cmp2:_(s1) = G_ICMP intpred(ne), %y(s32), %zero
+ ; CHECK-NEXT: %and:_(s1) = G_AND %cmp1, %cmp2
+ ; CHECK-NEXT: %zext:_(s32) = G_ZEXT %and(s1)
+ ; CHECK-NEXT: $w0 = COPY %zext(s32)
+ ; CHECK-NEXT: RET_ReallyLR implicit $w0
+ %x:_(s32) = COPY $w0
+ %y:_(s32) = COPY $w1
+ %zero:_(s32) = G_CONSTANT i32 0
+ %cmp1:_(s1) = G_ICMP intpred(ne), %x:_(s32), %zero:_
+ %cmp2:_(s1) = G_ICMP intpred(ne), %y:_(s32), %zero:_
+ %and:_(s1) = G_AND %cmp1, %cmp2
+ %zext:_(s32) = G_ZEXT %and:_(s1)
+ $w0 = COPY %zext
+ RET_ReallyLR implicit $w0
+
+...
+---
+name: valid_or_ne_0_ne_0_s32
+tracksRegLiveness: true
+legalized: true
+body: |
+ bb.0:
+ liveins: $w0, $w1
+
+ ; CHECK-LABEL: name: valid_or_ne_0_ne_0_s32
+ ; CHECK: liveins: $w0, $w1
+ ; CHECK-NEXT: {{ $}}
+ ; CHECK-NEXT: %x:_(s32) = COPY $w0
+ ; CHECK-NEXT: %y:_(s32) = COPY $w1
+ ; CHECK-NEXT: %zero:_(s32) = G_CONSTANT i32 0
+ ; CHECK-NEXT: [[OR:%[0-9]+]]:_(s32) = G_OR %x, %y
+ ; CHECK-NEXT: %or:_(s1) = G_ICMP intpred(ne), [[OR]](s32), %zero
+ ; CHECK-NEXT: %zext:_(s32) = G_ZEXT %or(s1)
+ ; CHECK-NEXT: $w0 = COPY %zext(s32)
+ ; CHECK-NEXT: RET_ReallyLR implicit $w0
+ %x:_(s32) = COPY $w0
+ %y:_(s32) = COPY $w1
+ %zero:_(s32) = G_CONSTANT i32 0
+ %cmp1:_(s1) = G_ICMP intpred(ne), %x:_(s32), %zero:_
+ %cmp2:_(s1) = G_ICMP intpred(ne), %y:_(s32), %zero:_
+ %or:_(s1) = G_OR %cmp1, %cmp2
+ %zext:_(s32) = G_ZEXT %or:_(s1)
+ $w0 = COPY %zext
+ RET_ReallyLR implicit $w0
+
+...
+---
+name: invalid_or_ne_1_ne_0_s32
+tracksRegLiveness: true
+legalized: true
+body: |
+ bb.0:
+ liveins: $w0, $w1
+
+ ; CHECK-LABEL: name: invalid_or_ne_1_ne_0_s32
+ ; CHECK: liveins: $w0, $w1
+ ; CHECK-NEXT: {{ $}}
+ ; CHECK-NEXT: %x:_(s32) = COPY $w0
+ ; CHECK-NEXT: %y:_(s32) = COPY $w1
+ ; CHECK-NEXT: %zero:_(s32) = G_CONSTANT i32 0
+ ; CHECK-NEXT: %one:_(s32) = G_CONSTANT i32 1
+ ; CHECK-NEXT: %cmp1:_(s1) = G_ICMP intpred(ne), %x(s32), %one
+ ; CHECK-NEXT: %cmp2:_(s1) = G_ICMP intpred(ne), %y(s32), %zero
+ ; CHECK-NEXT: %or:_(s1) = G_OR %cmp1, %cmp2
+ ; CHECK-NEXT: %zext:_(s32) = G_ZEXT %or(s1)
+ ; CHECK-NEXT: $w0 = COPY %zext(s32)
+ ; CHECK-NEXT: RET_ReallyLR implicit $w0
+ %x:_(s32) = COPY $w0
+ %y:_(s32) = COPY $w1
+ %zero:_(s32) = G_CONSTANT i32 0
+ %one:_(s32) = G_CONSTANT i32 1
+ %cmp1:_(s1) = G_ICMP intpred(ne), %x:_(s32), %one:_
+ %cmp2:_(s1) = G_ICMP intpred(ne), %y:_(s32), %zero:_
+ %or:_(s1) = G_OR %cmp1, %cmp2
+ %zext:_(s32) = G_ZEXT %or:_(s1)
+ $w0 = COPY %zext
+ RET_ReallyLR implicit $w0
+
+...
+---
+name: invalid_or_ne_0_ne_1_s32
+tracksRegLiveness: true
+legalized: true
+body: |
+ bb.0:
+ liveins: $w0, $w1
+
+ ; CHECK-LABEL: name: invalid_or_ne_0_ne_1_s32
+ ; CHECK: liveins: $w0, $w1
+ ; CHECK-NEXT: {{ $}}
+ ; CHECK-NEXT: %x:_(s32) = COPY $w0
+ ; CHECK-NEXT: %y:_(s32) = COPY $w1
+ ; CHECK-NEXT: %zero:_(s32) = G_CONSTANT i32 0
+ ; CHECK-NEXT: %one:_(s32) = G_CONSTANT i32 1
+ ; CHECK-NEXT: %cmp1:_(s1) = G_ICMP intpred(ne), %x(s32), %zero
+ ; CHECK-NEXT: %cmp2:_(s1) = G_ICMP intpred(ne), %y(s32), %one
+ ; CHECK-NEXT: %or:_(s1) = G_OR %cmp1, %cmp2
+ ; CHECK-NEXT: %zext:_(s32) = G_ZEXT %or(s1)
+ ; CHECK-NEXT: $w0 = COPY %zext(s32)
+ ; CHECK-NEXT: RET_ReallyLR implicit $w0
+ %x:_(s32) = COPY $w0
+ %y:_(s32) = COPY $w1
+ %zero:_(s32) = G_CONSTANT i32 0
+ %one:_(s32) = G_CONSTANT i32 1
+ %cmp1:_(s1) = G_ICMP intpred(ne), %x:_(s32), %zero:_
+ %cmp2:_(s1) = G_ICMP intpred(ne), %y:_(s32), %one:_
+ %or:_(s1) = G_OR %cmp1, %cmp2
+ %zext:_(s32) = G_ZEXT %or:_(s1)
+ $w0 = COPY %zext
+ RET_ReallyLR implicit $w0
+
+...
+---
+name: invalid_or_eq_0_ne_0_s32
+tracksRegLiveness: true
+legalized: true
+body: |
+ bb.0:
+ liveins: $w0, $w1
+
+ ; CHECK-LABEL: name: invalid_or_eq_0_ne_0_s32
+ ; CHECK: liveins: $w0, $w1
+ ; CHECK-NEXT: {{ $}}
+ ; CHECK-NEXT: %x:_(s32) = COPY $w0
+ ; CHECK-NEXT: %y:_(s32) = COPY $w1
+ ; CHECK-NEXT: %zero:_(s32) = G_CONSTANT i32 0
+ ; CHECK-NEXT: %cmp1:_(s1) = G_ICMP intpred(eq), %x(s32), %zero
+ ; CHECK-NEXT: %cmp2:_(s1) = G_ICMP intpred(ne), %y(s32), %zero
+ ; CHECK-NEXT: %or:_(s1) = G_OR %cmp1, %cmp2
+ ; CHECK-NEXT: %zext:_(s32) = G_ZEXT %or(s1)
+ ; CHECK-NEXT: $w0 = COPY %zext(s32)
+ ; CHECK-NEXT: RET_ReallyLR implicit $w0
+ %x:_(s32) = COPY $w0
+ %y:_(s32) = COPY $w1
+ %zero:_(s32) = G_CONSTANT i32 0
+ %cmp1:_(s1) = G_ICMP intpred(eq), %x:_(s32), %zero:_
+ %cmp2:_(s1) = G_ICMP intpred(ne), %y:_(s32), %zero:_
+ %or:_(s1) = G_OR %cmp1, %cmp2
+ %zext:_(s32) = G_ZEXT %or:_(s1)
+ $w0 = COPY %zext
+ RET_ReallyLR implicit $w0
+
+...
+---
+name: invalid_or_ne_0_eq_0_s32
+tracksRegLiveness: true
+legalized: true
+body: |
+ bb.0:
+ liveins: $w0, $w1
+
+ ; CHECK-LABEL: name: invalid_or_ne_0_eq_0_s32
+ ; CHECK: liveins: $w0, $w1
+ ; CHECK-NEXT: {{ $}}
+ ; CHECK-NEXT: %x:_(s32) = COPY $w0
+ ; CHECK-NEXT: %y:_(s32) = COPY $w1
+ ; CHECK-NEXT: %zero:_(s32) = G_CONSTANT i32 0
+ ; CHECK-NEXT: %cmp1:_(s1) = G_ICMP intpred(ne), %x(s32), %zero
+ ; CHECK-NEXT: %cmp2:_(s1) = G_ICMP intpred(eq), %y(s32), %zero
+ ; CHECK-NEXT: %or:_(s1) = G_OR %cmp1, %cmp2
+ ; CHECK-NEXT: %zext:_(s32) = G_ZEXT %or(s1)
+ ; CHECK-NEXT: $w0 = COPY %zext(s32)
+ ; CHECK-NEXT: RET_ReallyLR implicit $w0
+ %x:_(s32) = COPY $w0
+ %y:_(s32) = COPY $w1
+ %zero:_(s32) = G_CONSTANT i32 0
+ %cmp1:_(s1) = G_ICMP intpred(ne), %x:_(s32), %zero:_
+ %cmp2:_(s1) = G_ICMP intpred(eq), %y:_(s32), %zero:_
+ %or:_(s1) = G_OR %cmp1, %cmp2
+ %zext:_(s32) = G_ZEXT %or:_(s1)
+ $w0 = COPY %zext
+ RET_ReallyLR implicit $w0
+
+...
+
+---
+name: valid_and_eq_0_eq_0_s64
+tracksRegLiveness: true
+legalized: true
+body: |
+ bb.0:
+ liveins: $x0, $x1
+
+ ; CHECK-LABEL: name: valid_and_eq_0_eq_0_s64
+ ; CHECK: liveins: $x0, $x1
+ ; CHECK-NEXT: {{ $}}
+ ; CHECK-NEXT: %x:_(s64) = COPY $x0
+ ; CHECK-NEXT: %y:_(s64) = COPY $x1
+ ; CHECK-NEXT: %zero:_(s64) = G_CONSTANT i64 0
+ ; CHECK-NEXT: [[OR:%[0-9]+]]:_(s64) = G_OR %x, %y
+ ; CHECK-NEXT: %and:_(s1) = G_ICMP intpred(eq), [[OR]](s64), %zero
+ ; CHECK-NEXT: %zext:_(s64) = G_ZEXT %and(s1)
+ ; CHECK-NEXT: $x0 = COPY %zext(s64)
+ ; CHECK-NEXT: RET_ReallyLR implicit $x0
+ %x:_(s64) = COPY $x0
+ %y:_(s64) = COPY $x1
+ %zero:_(s64) = G_CONSTANT i64 0
+ %cmp1:_(s1) = G_ICMP intpred(eq), %x:_(s64), %zero:_
+ %cmp2:_(s1) = G_ICMP intpred(eq), %y:_(s64), %zero:_
+ %and:_(s1) = G_AND %cmp1, %cmp2
+ %zext:_(s64) = G_ZEXT %and:_(s1)
+ $x0 = COPY %zext
+ RET_ReallyLR implicit $x0
+
+...
+---
+name: invalid_and_eq_1_eq_0_s64
+tracksRegLiveness: true
+legalized: true
+body: |
+ bb.0:
+ liveins: $x0, $x1
+
+ ; CHECK-LABEL: name: invalid_and_eq_1_eq_0_s64
+ ; CHECK: liveins: $x0, $x1
+ ; CHECK-NEXT: {{ $}}
+ ; CHECK-NEXT: %x:_(s64) = COPY $x0
+ ; CHECK-NEXT: %y:_(s64) = COPY $x1
+ ; CHECK-NEXT: %zero:_(s64) = G_CONSTANT i64 0
+ ; CHECK-NEXT: %one:_(s64) = G_CONSTANT i64 1
+ ; CHECK-NEXT: %cmp1:_(s1) = G_ICMP intpred(eq), %x(s64), %one
+ ; CHECK-NEXT: %cmp2:_(s1) = G_ICMP intpred(eq), %y(s64), %zero
+ ; CHECK-NEXT: %and:_(s1) = G_AND %cmp1, %cmp2
+ ; CHECK-NEXT: %zext:_(s64) = G_ZEXT %and(s1)
+ ; CHECK-NEXT: $x0 = COPY %zext(s64)
+ ; CHECK-NEXT: RET_ReallyLR implicit $x0
+ %x:_(s64) = COPY $x0
+ %y:_(s64) = COPY $x1
+ %zero:_(s64) = G_CONSTANT i64 0
+ %one:_(s64) = G_CONSTANT i64 1
+ %cmp1:_(s1) = G_ICMP intpred(eq), %x:_(s64), %one:_
+ %cmp2:_(s1) = G_ICMP intpred(eq), %y:_(s64), %zero:_
+ %and:_(s1) = G_AND %cmp1, %cmp2
+ %zext:_(s64) = G_ZEXT %and:_(s1)
+ $x0 = COPY %zext
+ RET_ReallyLR implicit $x0
+
+...
+---
+name: invalid_and_eq_0_eq_1_s64
+tracksRegLiveness: true
+legalized: true
+body: |
+ bb.0:
+ liveins: $x0, $x1
+
+ ; CHECK-LABEL: name: invalid_and_eq_0_eq_1_s64
+ ; CHECK: liveins: $x0, $x1
+ ; CHECK-NEXT: {{ $}}
+ ; CHECK-NEXT: %x:_(s64) = COPY $x0
+ ; CHECK-NEXT: %y:_(s64) = COPY $x1
+ ; CHECK-NEXT: %zero:_(s64) = G_CONSTANT i64 0
+ ; CHECK-NEXT: %one:_(s64) = G_CONSTANT i64 1
+ ; CHECK-NEXT: %cmp1:_(s1) = G_ICMP intpred(eq), %x(s64), %zero
+ ; CHECK-NEXT: %cmp2:_(s1) = G_ICMP intpred(eq), %y(s64), %one
+ ; CHECK-NEXT: %and:_(s1) = G_AND %cmp1, %cmp2
+ ; CHECK-NEXT: %zext:_(s64) = G_ZEXT %and(s1)
+ ; CHECK-NEXT: $x0 = COPY %zext(s64)
+ ; CHECK-NEXT: RET_ReallyLR implicit $x0
+ %x:_(s64) = COPY $x0
+ %y:_(s64) = COPY $x1
+ %zero:_(s64) = G_CONSTANT i64 0
+ %one:_(s64) = G_CONSTANT i64 1
+ %cmp1:_(s1) = G_ICMP intpred(eq), %x:_(s64), %zero:_
+ %cmp2:_(s1) = G_ICMP intpred(eq), %y:_(s64), %one:_
+ %and:_(s1) = G_AND %cmp1, %cmp2
+ %zext:_(s64) = G_ZEXT %and:_(s1)
+ $x0 = COPY %zext
+ RET_ReallyLR implicit $x0
+
+...
+---
+name: invalid_and_ne_0_eq_0_s64
+tracksRegLiveness: true
+legalized: true
+body: |
+ bb.0:
+ liveins: $x0, $x1
+
+ ; CHECK-LABEL: name: invalid_and_ne_0_eq_0_s64
+ ; CHECK: liveins: $x0, $x1
+ ; CHECK-NEXT: {{ $}}
+ ; CHECK-NEXT: %x:_(s64) = COPY $x0
+ ; CHECK-NEXT: %y:_(s64) = COPY $x1
+ ; CHECK-NEXT: %zero:_(s64) = G_CONSTANT i64 0
+ ; CHECK-NEXT: %cmp1:_(s1) = G_ICMP intpred(ne), %x(s64), %zero
+ ; CHECK-NEXT: %cmp2:_(s1) = G_ICMP intpred(eq), %y(s64), %zero
+ ; CHECK-NEXT: %and:_(s1) = G_AND %cmp1, %cmp2
+ ; CHECK-NEXT: %zext:_(s64) = G_ZEXT %and(s1)
+ ; CHECK-NEXT: $x0 = COPY %zext(s64)
+ ; CHECK-NEXT: RET_ReallyLR implicit $x0
+ %x:_(s64) = COPY $x0
+ %y:_(s64) = COPY $x1
+ %zero:_(s64) = G_CONSTANT i64 0
+ %cmp1:_(s1) = G_ICMP intpred(ne), %x:_(s64), %zero:_
+ %cmp2:_(s1) = G_ICMP intpred(eq), %y:_(s64), %zero:_
+ %and:_(s1) = G_AND %cmp1, %cmp2
+ %zext:_(s64) = G_ZEXT %and:_(s1)
+ $x0 = COPY %zext
+ RET_ReallyLR implicit $x0
+
+...
+---
+name: invalid_and_eq_0_ne_0_s64
+tracksRegLiveness: true
+legalized: true
+body: |
+ bb.0:
+ liveins: $x0, $x1
+
+ ; CHECK-LABEL: name: invalid_and_eq_0_ne_0_s64
+ ; CHECK: liveins: $x0, $x1
+ ; CHECK-NEXT: {{ $}}
+ ; CHECK-NEXT: %x:_(s64) = COPY $x0
+ ; CHECK-NEXT: %y:_(s64) = COPY $x1
+ ; CHECK-NEXT: %zero:_(s64) = G_CONSTANT i64 0
+ ; CHECK-NEXT: %cmp1:_(s1) = G_ICMP intpred(eq), %x(s64), %zero
+ ; CHECK-NEXT: %cmp2:_(s1) = G_ICMP intpred(ne), %y(s64), %zero
+ ; CHECK-NEXT: %and:_(s1) = G_AND %cmp1, %cmp2
+ ; CHECK-NEXT: %zext:_(s64) = G_ZEXT %and(s1)
+ ; CHECK-NEXT: $x0 = COPY %zext(s64)
+ ; CHECK-NEXT: RET_ReallyLR implicit $x0
+ %x:_(s64) = COPY $x0
+ %y:_(s64) = COPY $x1
+ %zero:_(s64) = G_CONSTANT i64 0
+ %cmp1:_(s1) = G_ICMP intpred(eq), %x:_(s64), %zero:_
+ %cmp2:_(s1) = G_ICMP intpred(ne), %y:_(s64), %zero:_
+ %and:_(s1) = G_AND %cmp1, %cmp2
+ %zext:_(s64) = G_ZEXT %and:_(s1)
+ $x0 = COPY %zext
+ RET_ReallyLR implicit $x0
+
+...
+---
+name: invalid_and_ne_0_ne_0_s64
+tracksRegLiveness: true
+legalized: true
+body: |
+ bb.0:
+ liveins: $x0, $x1
+
+ ; CHECK-LABEL: name: invalid_and_ne_0_ne_0_s64
+ ; CHECK: liveins: $x0, $x1
+ ; CHECK-NEXT: {{ $}}
+ ; CHECK-NEXT: %x:_(s64) = COPY $x0
+ ; CHECK-NEXT: %y:_(s64) = COPY $x1
+ ; CHECK-NEXT: %zero:_(s64) = ...
[truncated]
|
Uploading this version for debug purposes. The original one is here: #69017. Notes for now I used i32 as type for ordst, but it should be the type of the G_ICMP's source. Findings:
With this version:
Seems like the issue is that after creating the G_OR the old one will be deleted already, but it is required to create the G_ICMP as well, since the parent info is came from it for example. In BuildMI the *Stete.MIs[0] will be the already deleted G_OR.
At least this is how I understand it so far. |
Was this crash fixed by 8e8c954? |
fc3f891
to
6328134
Compare
Yes! Nice, I rebased the PR and did some minor change to make it work. |
6328134
to
1eadb1f
Compare
[{ return ${p}.getPredicate() == CmpInst::ICMP_EQ && | ||
!MRI.getType(${s1}.getReg()).isPointer() && | ||
(MRI.getType(${s1}.getReg()) == | ||
MRI.getType(${s2}.getReg())); }]), |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
@Pierre-vh would be nice if we could match the literal G_ICMP predicate inline
(MRI.getType(${s1}.getReg()) == | ||
MRI.getType(${s2}.getReg())); }]), |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
You should be able to match this with GITypeOf<"$s1"> instead of using the manual predicate.
@Pierre-vh is there a nicer way to filter out pointers?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I got this error when using it
error: GISpecialType is not supported in 'match' patterns
Maybe its only usable in the apply
part atm.
$x0 = COPY %zext | ||
RET_ReallyLR implicit $x0 | ||
|
||
... |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Can you add some vector tests? In particular, compares involving vectors of pointers? I believe they are broken as-is
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Indeed they were so added getScalarType
as well.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Need some vector and vector of pointer tests
0bea4c8
to
94f15e6
Compare
Also combine (X != 0) | (Y != 0) -> (X | Y) != 0
94f15e6
to
e762ae9
Compare
Very recently got merged this #82733, which will insert G_FREEZE before the G_OR (https://godbolt.org/z/dcr83rfM6) causing now the above patterns not to work on cases like in |
You can try the equivalent transform in alive2 on the IR to see if it's valid |
Nice tool. Seems like these are valid: https://alive2.llvm.org/ce/z/3AHtx3, https://alive2.llvm.org/ce/z/W8uR9j. With these changes the gains coming back and even improving the regression caused by the G_FREEZE introduction. But I think this might need a separate PR. diff --git a/llvm/include/llvm/Target/GlobalISel/Combine.td b/llvm/include/llvm/Target/GlobalISel/Combine.td
index 18db7a819540..66d3c14547c2 100644
--- a/llvm/include/llvm/Target/GlobalISel/Combine.td
+++ b/llvm/include/llvm/Target/GlobalISel/Combine.td
@@ -952,6 +952,21 @@ def redundant_binop_in_equality : GICombineRule<
[{ return Helper.matchRedundantBinOpInEquality(*${root}, ${info}); }]),
(apply [{ Helper.applyBuildFn(*${root}, ${info}); }])>;
+def remove_freeze_of_constants: GICombineRule <
+ (defs root:$dst),
+ (match (G_FREEZE $dst, $src),
+ [{ return !!isConstantOrConstantSplatVector(*MRI.getVRegDef(${src}.getReg()), MRI); }]),
+ (apply (GIReplaceReg $dst, $src))>;
+
+def canonicalize_icmp_freeze: GICombineRule<
+ (defs root:$root),
+ (match (G_ICMP $dst, $p, $src1, $src2),
+ (G_FREEZE $root, $dst)),
+ (apply (G_FREEZE $new_src1, $src1),
+ (G_FREEZE $new_src2, $src2),
+ (G_ICMP $root, $p, $new_src1, $new_src2))
+>;
+
// Transform: (X == 0 & Y == 0) -> (X | Y) == 0
def double_icmp_zero_and_combine: GICombineRule<
(defs root:$root),
@@ -980,6 +995,8 @@ def double_icmp_zero_or_combine: GICombineRule<
(G_ICMP $root, $p, $ordst, 0))
>;
+def freeze_combines : GICombineGroup<[remove_freeze_of_constants, canonicalize_icmp_freeze]>;
+
def double_icmp_zero_and_or_combine : GICombineGroup<[double_icmp_zero_and_combine,
double_icmp_zero_or_combine]>;
@@ -1374,7 +1391,7 @@ def all_combines : GICombineGroup<[trivial_combines, insert_vec_elt_combines,
and_or_disjoint_mask, fma_combines, fold_binop_into_select,
sub_add_reg, select_to_minmax, redundant_binop_in_equality,
fsub_to_fneg, commute_constant_to_rhs, match_ands, match_ors,
- combine_concat_vector, double_icmp_zero_and_or_combine]>;
+ combine_concat_vector, double_icmp_zero_and_or_combine, freeze_combines]>;
// A combine group used to for prelegalizer combiners at -O0. The combines in
// this group have been selected based on experiments to balance code size and
diff --git a/llvm/test/CodeGen/AArch64/cmp-chains.ll b/llvm/test/CodeGen/AArch64/cmp-chains.ll
index 1d9f39e51859..4eaa25608d16 100644
--- a/llvm/test/CodeGen/AArch64/cmp-chains.ll
+++ b/llvm/test/CodeGen/AArch64/cmp-chains.ll
@@ -109,8 +109,7 @@ define i32 @cmp_or2(i32 %0, i32 %1, i32 %2, i32 %3) {
; GISEL-NEXT: cset w8, lo
; GISEL-NEXT: cmp w2, w3
; GISEL-NEXT: cset w9, ne
-; GISEL-NEXT: orr w8, w8, w9
-; GISEL-NEXT: and w0, w8, #0x1
+; GISEL-NEXT: orr w0, w8, w9
; GISEL-NEXT: ret
%5 = icmp ult i32 %0, %1
%6 = icmp ne i32 %2, %3
@@ -138,8 +137,7 @@ define i32 @cmp_or3(i32 %0, i32 %1, i32 %2, i32 %3, i32 %4, i32 %5) {
; GISEL-NEXT: cmp w4, w5
; GISEL-NEXT: orr w8, w8, w9
; GISEL-NEXT: cset w9, ne
-; GISEL-NEXT: orr w8, w8, w9
-; GISEL-NEXT: and w0, w8, #0x1
+; GISEL-NEXT: orr w0, w8, w9
; GISEL-NEXT: ret
%7 = icmp ult i32 %0, %1
%8 = icmp ugt i32 %2, %3
@@ -173,8 +171,7 @@ define i32 @cmp_or4(i32 %0, i32 %1, i32 %2, i32 %3, i32 %4, i32 %5, i32 %6, i32
; GISEL-NEXT: orr w8, w8, w9
; GISEL-NEXT: cset w11, eq
; GISEL-NEXT: orr w9, w10, w11
-; GISEL-NEXT: orr w8, w8, w9
-; GISEL-NEXT: and w0, w8, #0x1
+; GISEL-NEXT: orr w0, w8, w9
; GISEL-NEXT: ret
%9 = icmp ult i32 %0, %1
%10 = icmp ugt i32 %2, %3
@@ -189,22 +186,12 @@ define i32 @cmp_or4(i32 %0, i32 %1, i32 %2, i32 %3, i32 %4, i32 %5, i32 %6, i32
; (x0 != 0) || (x1 != 0)
define i32 @true_or2(i32 %0, i32 %1) {
-; SDISEL-LABEL: true_or2:
-; SDISEL: // %bb.0:
-; SDISEL-NEXT: orr w8, w0, w1
-; SDISEL-NEXT: cmp w8, #0
-; SDISEL-NEXT: cset w0, ne
-; SDISEL-NEXT: ret
-;
-; GISEL-LABEL: true_or2:
-; GISEL: // %bb.0:
-; GISEL-NEXT: cmp w0, #0
-; GISEL-NEXT: cset w8, ne
-; GISEL-NEXT: cmp w1, #0
-; GISEL-NEXT: cset w9, ne
-; GISEL-NEXT: orr w8, w8, w9
-; GISEL-NEXT: and w0, w8, #0x1
-; GISEL-NEXT: ret
+; CHECK-LABEL: true_or2:
+; CHECK: // %bb.0:
+; CHECK-NEXT: orr w8, w0, w1
+; CHECK-NEXT: cmp w8, #0
+; CHECK-NEXT: cset w0, ne
+; CHECK-NEXT: ret
%3 = icmp ne i32 %0, 0
%4 = icmp ne i32 %1, 0
%5 = select i1 %3, i1 true, i1 %4
@@ -214,26 +201,13 @@ define i32 @true_or2(i32 %0, i32 %1) {
; (x0 != 0) || (x1 != 0) || (x2 != 0)
define i32 @true_or3(i32 %0, i32 %1, i32 %2) {
-; SDISEL-LABEL: true_or3:
-; SDISEL: // %bb.0:
-; SDISEL-NEXT: orr w8, w0, w1
-; SDISEL-NEXT: orr w8, w8, w2
-; SDISEL-NEXT: cmp w8, #0
-; SDISEL-NEXT: cset w0, ne
-; SDISEL-NEXT: ret
-;
-; GISEL-LABEL: true_or3:
-; GISEL: // %bb.0:
-; GISEL-NEXT: cmp w0, #0
-; GISEL-NEXT: cset w8, ne
-; GISEL-NEXT: cmp w1, #0
-; GISEL-NEXT: cset w9, ne
-; GISEL-NEXT: cmp w2, #0
-; GISEL-NEXT: orr w8, w8, w9
-; GISEL-NEXT: cset w9, ne
-; GISEL-NEXT: orr w8, w8, w9
-; GISEL-NEXT: and w0, w8, #0x1
-; GISEL-NEXT: ret
+; CHECK-LABEL: true_or3:
+; CHECK: // %bb.0:
+; CHECK-NEXT: orr w8, w0, w1
+; CHECK-NEXT: orr w8, w8, w2
+; CHECK-NEXT: cmp w8, #0
+; CHECK-NEXT: cset w0, ne
+; CHECK-NEXT: ret
%4 = icmp ne i32 %0, 0
%5 = icmp ne i32 %1, 0
%6 = select i1 %4, i1 true, i1 %5
@@ -242,5 +216,3 @@ define i32 @true_or3(i32 %0, i32 %1, i32 %2) {
%9 = zext i1 %8 to i32
ret i32 %9
}
|
Yes, this should be done separately |
Also combine (X != 0) | (Y != 0) -> (X | Y) != 0