Nikita Popov 90ba33099c
[InstCombine] Canonicalize constant GEPs to i8 source element type (#68882)
This patch canonicalizes getelementptr instructions with constant
indices to use the `i8` source element type. This makes it easier for
optimizations to recognize that two GEPs are identical, because they
don't need to see past many different ways to express the same offset.

This is a first step towards
https://discourse.llvm.org/t/rfc-replacing-getelementptr-with-ptradd/68699.
This is limited to constant GEPs only for now, as they have a clear
canonical form, while we're not yet sure how exactly to deal with
variable indices.

The test llvm/test/Transforms/PhaseOrdering/switch_with_geps.ll gives
two representative examples of the kind of optimization improvement we
expect from this change. In the first test SimplifyCFG can now realize
that all switch branches are actually the same. In the second test it
can convert it into simple arithmetic. These are representative of
common optimization failures we see in Rust.

Fixes https://github.com/llvm/llvm-project/issues/69841.
2024-01-24 15:25:29 +01:00

148 lines
5.6 KiB
LLVM

; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
; Verify that strncmp calls with members of constant structs are folded.
;
; RUN: opt < %s -passes=instcombine -S | FileCheck %s
; struct A { char a[3], b[4], c[4]; };
%struct.A = type { [3 x i8], [4 x i8], [4 x i8] }
@a = constant %struct.A { [3 x i8] c"123", [4 x i8] c"1231", [4 x i8] c"2345" }
declare i32 @strncmp(ptr, ptr, i64)
define void @fold_strncmp_Aa_b(ptr %pcmp) {
; CHECK-LABEL: @fold_strncmp_Aa_b(
; CHECK-NEXT: store i32 0, ptr [[PCMP:%.*]], align 4
; CHECK-NEXT: [[PCMP1:%.*]] = getelementptr i8, ptr [[PCMP]], i64 4
; CHECK-NEXT: store i32 0, ptr [[PCMP1]], align 4
; CHECK-NEXT: [[PCMP2:%.*]] = getelementptr i8, ptr [[PCMP]], i64 8
; CHECK-NEXT: store i32 0, ptr [[PCMP2]], align 4
; CHECK-NEXT: [[PCMP3:%.*]] = getelementptr i8, ptr [[PCMP]], i64 12
; CHECK-NEXT: store i32 0, ptr [[PCMP3]], align 4
; CHECK-NEXT: [[PCMP4:%.*]] = getelementptr i8, ptr [[PCMP]], i64 16
; CHECK-NEXT: store i32 0, ptr [[PCMP4]], align 4
; CHECK-NEXT: [[PCMP5:%.*]] = getelementptr i8, ptr [[PCMP]], i64 20
; CHECK-NEXT: store i32 0, ptr [[PCMP5]], align 4
; CHECK-NEXT: [[PCMP6:%.*]] = getelementptr i8, ptr [[PCMP]], i64 24
; CHECK-NEXT: store i32 0, ptr [[PCMP6]], align 4
; CHECK-NEXT: [[PCMP7:%.*]] = getelementptr i8, ptr [[PCMP]], i64 28
; CHECK-NEXT: store i32 -1, ptr [[PCMP7]], align 4
; CHECK-NEXT: [[PCMP8:%.*]] = getelementptr i8, ptr [[PCMP]], i64 32
; CHECK-NEXT: store i32 -1, ptr [[PCMP8]], align 4
; CHECK-NEXT: [[PCMP9:%.*]] = getelementptr i8, ptr [[PCMP]], i64 36
; CHECK-NEXT: store i32 -1, ptr [[PCMP9]], align 4
; CHECK-NEXT: ret void
;
; p1 = a.a
; p2 = a.b
%p2 = getelementptr %struct.A, ptr @a, i32 0, i32 1, i32 0
; Fold strncmp(a.a = "123", a.b = "1231", 0) to 0.
%cmp0 = call i32 @strncmp(ptr @a, ptr %p2, i64 0)
store i32 %cmp0, ptr %pcmp
; Fold strncmp(a.a = "123", a.b = "1231", 1) to 0.
%cmp1 = call i32 @strncmp(ptr @a, ptr %p2, i64 1)
%pcmp1 = getelementptr i32, ptr %pcmp, i64 1
store i32 %cmp1, ptr %pcmp1
; Fold strncmp(a.a = "123", a.b = "1231", 2) to 0.
%cmp2 = call i32 @strncmp(ptr @a, ptr %p2, i64 2)
%pcmp2 = getelementptr i32, ptr %pcmp, i64 2
store i32 %cmp2, ptr %pcmp2
; Fold strncmp(a.a = "123", a.b = "1231", 3) to 0.
%cmp3 = call i32 @strncmp(ptr @a, ptr %p2, i64 3)
%pcmp3 = getelementptr i32, ptr %pcmp, i64 3
store i32 %cmp3, ptr %pcmp3
; Fold strncmp(a.a = "123", a.b = "1231", 4) to 0.
; In this and the subsequent tests, reading past the end of a.a is
; strictly undefined in C/C++ (because it forms a pointer to a distinct
; subobject) but handling such cases as if they were well-defined is
; simpler than trying to exclude them.
%cmp4 = call i32 @strncmp(ptr @a, ptr %p2, i64 4)
%pcmp4 = getelementptr i32, ptr %pcmp, i64 4
store i32 %cmp4, ptr %pcmp4
; Fold strncmp("123", "1231" "2", 5) to 0.
%cmp5 = call i32 @strncmp(ptr @a, ptr %p2, i64 5)
%pcmp5 = getelementptr i32, ptr %pcmp, i64 5
store i32 %cmp5, ptr %pcmp5
; Fold strncmp("123", "1231" "23", 6) to 0.
%cmp6 = call i32 @strncmp(ptr @a, ptr %p2, i64 6)
%pcmp6 = getelementptr i32, ptr %pcmp, i64 6
store i32 %cmp6, ptr %pcmp6
; Fold strncmp("123", "1231" "2345", 7) to 1.
%cmp7 = call i32 @strncmp(ptr @a, ptr %p2, i64 7)
%pcmp7 = getelementptr i32, ptr %pcmp, i64 7
store i32 %cmp7, ptr %pcmp7
; Fold strncmp("123", "1231" "2345", 8) to 1.
%cmp8 = call i32 @strncmp(ptr @a, ptr %p2, i64 8)
%pcmp8 = getelementptr i32, ptr %pcmp, i64 8
store i32 %cmp8, ptr %pcmp8
; Fold strncmp("123", "1231" "2345", 9) to 1.
%cmp9 = call i32 @strncmp(ptr @a, ptr %p2, i64 9)
%pcmp9 = getelementptr i32, ptr %pcmp, i64 9
store i32 %cmp9, ptr %pcmp9
ret void
}
define void @fold_strncmp_Ab_a(ptr %pcmp) {
; CHECK-LABEL: @fold_strncmp_Ab_a(
; CHECK-NEXT: store i32 0, ptr [[PCMP:%.*]], align 4
; CHECK-NEXT: [[PCMP1:%.*]] = getelementptr i8, ptr [[PCMP]], i64 4
; CHECK-NEXT: store i32 0, ptr [[PCMP1]], align 4
; CHECK-NEXT: [[PCMP2:%.*]] = getelementptr i8, ptr [[PCMP]], i64 8
; CHECK-NEXT: store i32 0, ptr [[PCMP2]], align 4
; CHECK-NEXT: [[PCMP3:%.*]] = getelementptr i8, ptr [[PCMP]], i64 12
; CHECK-NEXT: store i32 0, ptr [[PCMP3]], align 4
; CHECK-NEXT: [[PCMP4:%.*]] = getelementptr i8, ptr [[PCMP]], i64 16
; CHECK-NEXT: store i32 1, ptr [[PCMP4]], align 4
; CHECK-NEXT: [[PCMP5:%.*]] = getelementptr i8, ptr [[PCMP]], i64 20
; CHECK-NEXT: store i32 1, ptr [[PCMP5]], align 4
; CHECK-NEXT: ret void
;
; p1 = &a.b[3]
%p1 = getelementptr %struct.A, ptr @a, i32 0, i32 1, i32 3
; p2 = &a.a
; Fold strncmp(a.a = "123", &a.b[3] = "1" "2345", 0) to 0.
%cmp0 = call i32 @strncmp(ptr %p1, ptr @a, i64 0)
store i32 %cmp0, ptr %pcmp
; Fold strncmp(a.a = "123", &a.b[3] = "1" "2345", 1) to 0.
%cmp1 = call i32 @strncmp(ptr %p1, ptr @a, i64 1)
%pcmp1 = getelementptr i32, ptr %pcmp, i64 1
store i32 %cmp1, ptr %pcmp1
; Fold strncmp(a.a = "123", &a.b[3] = "1" "2345", 2) to 0.
%cmp2 = call i32 @strncmp(ptr %p1, ptr @a, i64 2)
%pcmp2 = getelementptr i32, ptr %pcmp, i64 2
store i32 %cmp2, ptr %pcmp2
; Fold strncmp(a.a = "123", &a.b[3] = "1" "2345", 3) to 0.
%cmp3 = call i32 @strncmp(ptr %p1, ptr @a, i64 3)
%pcmp3 = getelementptr i32, ptr %pcmp, i64 3
store i32 %cmp3, ptr %pcmp3
; Fold strncmp(a.a = "123", &a.b[3] = "1" "2345", 4) to 1.
%cmp4 = call i32 @strncmp(ptr %p1, ptr @a, i64 4)
%pcmp4 = getelementptr i32, ptr %pcmp, i64 4
store i32 %cmp4, ptr %pcmp4
; Fold strncmp(a.a = "123", &a.b[3] = "1" "2345", 5) to 1.
%cmp5 = call i32 @strncmp(ptr %p1, ptr @a, i64 5)
%pcmp5 = getelementptr i32, ptr %pcmp, i64 5
store i32 %cmp5, ptr %pcmp5
ret void
}