6502/llvm-6502
1
0
Fork 0
mirror of https://github.com/c64scene-ar/llvm-6502.git synced 2024-12-24 06:30:19 +00:00
Code Issues Projects Releases Wiki Activity
1a5f7f54f4
llvm-6502/test/Transforms/SeparateConstOffsetFromGEP/NVPTX/split-gep.ll
Jingyue Wu 1d56cda023 Partially revert r210444 due to performance regression 
Summary:
Converting outermost zext(a) to sext(a) causes worse code when the
computation of zext(a) could be reused. For example, after converting

... = array[zext(a)]
... = array[zext(a) + 1]

to

... = array[sext(a)]
... = array[zext(a) + 1],

the program computes sext(a), which is actually unnecessary. I added one
test in split-gep-and-gvn.ll to illustrate this scenario.

Also, with r211281 and r211084, we annotate more "nuw" tags to
computation involving CUDA intrinsics such as threadIdx.x. These
annotations help with splitting GEP a lot, rendering the benefit we get
from this reverted optimization only marginal.

Test Plan: make check-all

Reviewers: eliben, meheff

Reviewed By: meheff

Subscribers: jholewinski, llvm-commits

Differential Revision: http://reviews.llvm.org/D4542

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@213209 91177308-0d34-0410-b5e6-96231b3b80d8
2014-07-16 23:25:00 +00:00

237 lines
9.3 KiB
LLVM
Raw Blame History

; RUN: opt < %s -separate-const-offset-from-gep -dce -S | FileCheck %s
; Several unit tests for -separate-const-offset-from-gep. The transformation
; heavily relies on TargetTransformInfo, so we put these tests under
; target-specific folders.
target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
; target triple is necessary; otherwise TargetTransformInfo rejects any
; addressing mode.
target triple = "nvptx64-unknown-unknown"
%struct.S = type { float, double }
@struct_array = global [1024 x %struct.S] zeroinitializer, align 16
@float_2d_array = global [32 x [32 x float]] zeroinitializer, align 4
; We should not extract any struct field indices, because fields in a struct
; may have different types.
define double* @struct(i32 %i) {
entry:
%add = add nsw i32 %i, 5
%idxprom = sext i32 %add to i64
%p = getelementptr inbounds [1024 x %struct.S]* @struct_array, i64 0, i64 %idxprom, i32 1
ret double* %p
}
; CHECK-LABEL: @struct(
; CHECK: getelementptr [1024 x %struct.S]* @struct_array, i64 0, i64 %{{[a-zA-Z0-9]+}}, i32 1
; We should be able to trace into sext(a + b) if a + b is non-negative
; (e.g., used as an index of an inbounds GEP) and one of a and b is
; non-negative.
define float* @sext_add(i32 %i, i32 %j) {
entry:
%0 = add i32 %i, 1
%1 = sext i32 %0 to i64 ; inbound sext(i + 1) = sext(i) + 1
%2 = add i32 %j, -2
; However, inbound sext(j + -2) != sext(j) + -2, e.g., j = INT_MIN
%3 = sext i32 %2 to i64
%p = getelementptr inbounds [32 x [32 x float]]* @float_2d_array, i64 0, i64 %1, i64 %3
ret float* %p
}
; CHECK-LABEL: @sext_add(
; CHECK-NOT: = add
; CHECK: add i32 %j, -2
; CHECK: sext
; CHECK: getelementptr [32 x [32 x float]]* @float_2d_array, i64 0, i64 %{{[a-zA-Z0-9]+}}, i64 %{{[a-zA-Z0-9]+}}
; CHECK: getelementptr float* %{{[a-zA-Z0-9]+}}, i64 32
; We should be able to trace into sext/zext if it can be distributed to both
; operands, e.g., sext (add nsw a, b) == add nsw (sext a), (sext b)
;
; This test verifies we can transform
; gep base, a + sext(b +nsw 1), c + zext(d +nuw 1)
; to
; gep base, a + sext(b), c + zext(d); gep ..., 1 * 32 + 1
define float* @ext_add_no_overflow(i64 %a, i32 %b, i64 %c, i32 %d) {
%b1 = add nsw i32 %b, 1
%b2 = sext i32 %b1 to i64
%i = add i64 %a, %b2 ; i = a + sext(b +nsw 1)
%d1 = add nuw i32 %d, 1
%d2 = zext i32 %d1 to i64
%j = add i64 %c, %d2 ; j = c + zext(d +nuw 1)
%p = getelementptr inbounds [32 x [32 x float]]* @float_2d_array, i64 0, i64 %i, i64 %j
ret float* %p
}
; CHECK-LABEL: @ext_add_no_overflow(
; CHECK: [[BASE_PTR:%[a-zA-Z0-9]+]] = getelementptr [32 x [32 x float]]* @float_2d_array, i64 0, i64 %{{[a-zA-Z0-9]+}}, i64 %{{[a-zA-Z0-9]+}}
; CHECK: getelementptr float* [[BASE_PTR]], i64 33
; Verifies we handle nested sext/zext correctly.
define void @sext_zext(i32 %a, i32 %b, float** %out1, float** %out2) {
entry:
%0 = add nsw nuw i32 %a, 1
%1 = sext i32 %0 to i48
%2 = zext i48 %1 to i64 ; zext(sext(a +nsw nuw 1)) = zext(sext(a)) + 1
%3 = add nsw i32 %b, 2
%4 = sext i32 %3 to i48
%5 = zext i48 %4 to i64 ; zext(sext(b +nsw 2)) != zext(sext(b)) + 2
%p1 = getelementptr [32 x [32 x float]]* @float_2d_array, i64 0, i64 %2, i64 %5
store float* %p1, float** %out1
%6 = add nuw i32 %a, 3
%7 = zext i32 %6 to i48
%8 = sext i48 %7 to i64 ; sext(zext(a +nuw 3)) = zext(a +nuw 3) = zext(a) + 3
%9 = add nsw i32 %b, 4
%10 = zext i32 %9 to i48
%11 = sext i48 %10 to i64 ; sext(zext(b +nsw 4)) != zext(b) + 4
%p2 = getelementptr [32 x [32 x float]]* @float_2d_array, i64 0, i64 %8, i64 %11
store float* %p2, float** %out2
ret void
}
; CHECK-LABEL: @sext_zext(
; CHECK: [[BASE_PTR_1:%[a-zA-Z0-9]+]] = getelementptr [32 x [32 x float]]* @float_2d_array, i64 0, i64 %{{[a-zA-Z0-9]+}}, i64 %{{[a-zA-Z0-9]+}}
; CHECK: getelementptr float* [[BASE_PTR_1]], i64 32
; CHECK: [[BASE_PTR_2:%[a-zA-Z0-9]+]] = getelementptr [32 x [32 x float]]* @float_2d_array, i64 0, i64 %{{[a-zA-Z0-9]+}}, i64 %{{[a-zA-Z0-9]+}}
; CHECK: getelementptr float* [[BASE_PTR_2]], i64 96
; Similar to @ext_add_no_overflow, we should be able to trace into s/zext if
; its operand is an OR and the two operands of the OR have no common bits.
define float* @sext_or(i64 %a, i32 %b) {
entry:
%b1 = shl i32 %b, 2
%b2 = or i32 %b1, 1 ; (b << 2) and 1 have no common bits
%b3 = or i32 %b1, 4 ; (b << 2) and 4 may have common bits
%b2.ext = zext i32 %b2 to i64
%b3.ext = sext i32 %b3 to i64
%i = add i64 %a, %b2.ext
%j = add i64 %a, %b3.ext
%p = getelementptr inbounds [32 x [32 x float]]* @float_2d_array, i64 0, i64 %i, i64 %j
ret float* %p
}
; CHECK-LABEL: @sext_or(
; CHECK: [[BASE_PTR:%[a-zA-Z0-9]+]] = getelementptr [32 x [32 x float]]* @float_2d_array, i64 0, i64 %{{[a-zA-Z0-9]+}}, i64 %{{[a-zA-Z0-9]+}}
; CHECK: getelementptr float* [[BASE_PTR]], i64 32
; The subexpression (b + 5) is used in both "i = a + (b + 5)" and "*out = b +
; 5". When extracting the constant offset 5, make sure "*out = b + 5" isn't
; affected.
define float* @expr(i64 %a, i64 %b, i64* %out) {
entry:
%b5 = add i64 %b, 5
%i = add i64 %b5, %a
%p = getelementptr inbounds [32 x [32 x float]]* @float_2d_array, i64 0, i64 %i, i64 0
store i64 %b5, i64* %out
ret float* %p
}
; CHECK-LABEL: @expr(
; CHECK: [[BASE_PTR:%[a-zA-Z0-9]+]] = getelementptr [32 x [32 x float]]* @float_2d_array, i64 0, i64 %{{[a-zA-Z0-9]+}}, i64 0
; CHECK: getelementptr float* [[BASE_PTR]], i64 160
; CHECK: store i64 %b5, i64* %out
; d + sext(a +nsw (b +nsw (c +nsw 8))) => (d + sext(a) + sext(b) + sext(c)) + 8
define float* @sext_expr(i32 %a, i32 %b, i32 %c, i64 %d) {
entry:
%0 = add nsw i32 %c, 8
%1 = add nsw i32 %b, %0
%2 = add nsw i32 %a, %1
%3 = sext i32 %2 to i64
%i = add i64 %d, %3
%p = getelementptr inbounds [32 x [32 x float]]* @float_2d_array, i64 0, i64 0, i64 %i
ret float* %p
}
; CHECK-LABEL: @sext_expr(
; CHECK: sext i32
; CHECK: sext i32
; CHECK: sext i32
; CHECK: getelementptr float* %{{[a-zA-Z0-9]+}}, i64 8
; Verifies we handle "sub" correctly.
define float* @sub(i64 %i, i64 %j) {
%i2 = sub i64 %i, 5 ; i - 5
%j2 = sub i64 5, %j ; 5 - i
%p = getelementptr inbounds [32 x [32 x float]]* @float_2d_array, i64 0, i64 %i2, i64 %j2
ret float* %p
}
; CHECK-LABEL: @sub(
; CHECK: %[[j2:[a-zA-Z0-9]+]] = sub i64 0, %j
; CHECK: [[BASE_PTR:%[a-zA-Z0-9]+]] = getelementptr [32 x [32 x float]]* @float_2d_array, i64 0, i64 %i, i64 %[[j2]]
; CHECK: getelementptr float* [[BASE_PTR]], i64 -155
%struct.Packed = type <{ [3 x i32], [8 x i64] }> ; <> means packed
; Verifies we can emit correct uglygep if the address is not natually aligned.
define i64* @packed_struct(i32 %i, i32 %j) {
entry:
%s = alloca [1024 x %struct.Packed], align 16
%add = add nsw i32 %j, 3
%idxprom = sext i32 %add to i64
%add1 = add nsw i32 %i, 1
%idxprom2 = sext i32 %add1 to i64
%arrayidx3 = getelementptr inbounds [1024 x %struct.Packed]* %s, i64 0, i64 %idxprom2, i32 1, i64 %idxprom
ret i64* %arrayidx3
}
; CHECK-LABEL: @packed_struct(
; CHECK: [[BASE_PTR:%[a-zA-Z0-9]+]] = getelementptr [1024 x %struct.Packed]* %s, i64 0, i64 %{{[a-zA-Z0-9]+}}, i32 1, i64 %{{[a-zA-Z0-9]+}}
; CHECK: [[CASTED_PTR:%[a-zA-Z0-9]+]] = bitcast i64* [[BASE_PTR]] to i8*
; CHECK: %uglygep = getelementptr i8* [[CASTED_PTR]], i64 100
; CHECK: bitcast i8* %uglygep to i64*
; We shouldn't be able to extract the 8 from "zext(a +nuw (b + 8))",
; because "zext(b + 8) != zext(b) + 8"
define float* @zext_expr(i32 %a, i32 %b) {
entry:
%0 = add i32 %b, 8
%1 = add nuw i32 %a, %0
%i = zext i32 %1 to i64
%p = getelementptr [32 x [32 x float]]* @float_2d_array, i64 0, i64 0, i64 %i
ret float* %p
}
; CHECK-LABEL: zext_expr(
; CHECK: getelementptr [32 x [32 x float]]* @float_2d_array, i64 0, i64 0, i64 %i
; Per http://llvm.org/docs/LangRef.html#id181, the indices of a off-bound gep
; should be considered sign-extended to the pointer size. Therefore,
; gep base, (add i32 a, b) != gep (gep base, i32 a), i32 b
; because
; sext(a + b) != sext(a) + sext(b)
;
; This test verifies we do not illegitimately extract the 8 from
; gep base, (i32 a + 8)
define float* @i32_add(i32 %a) {
entry:
%i = add i32 %a, 8
%p = getelementptr [32 x [32 x float]]* @float_2d_array, i64 0, i64 0, i32 %i
ret float* %p
}
; CHECK-LABEL: @i32_add(
; CHECK: getelementptr [32 x [32 x float]]* @float_2d_array, i64 0, i64 0, i64 %{{[a-zA-Z0-9]+}}
; CHECK-NOT: getelementptr
; Verifies that we compute the correct constant offset when the index is
; sign-extended and then zero-extended. The old version of our code failed to
; handle this case because it simply computed the constant offset as the
; sign-extended value of the constant part of the GEP index.
define float* @apint(i1 %a) {
entry:
%0 = add nsw nuw i1 %a, 1
%1 = sext i1 %0 to i4
%2 = zext i4 %1 to i64 ; zext (sext i1 1 to i4) to i64 = 15
%p = getelementptr [32 x [32 x float]]* @float_2d_array, i64 0, i64 0, i64 %2
ret float* %p
}
; CHECK-LABEL: @apint(
; CHECK: [[BASE_PTR:%[a-zA-Z0-9]+]] = getelementptr [32 x [32 x float]]* @float_2d_array, i64 0, i64 0, i64 %{{[a-zA-Z0-9]+}}
; CHECK: getelementptr float* [[BASE_PTR]], i64 15
; Do not trace into binary operators other than ADD, SUB, and OR.
define float* @and(i64 %a) {
entry:
%0 = shl i64 %a, 2
%1 = and i64 %0, 1
%p = getelementptr [32 x [32 x float]]* @float_2d_array, i64 0, i64 0, i64 %1
ret float* %p
}
; CHECK-LABEL: @and(
; CHECK: getelementptr [32 x [32 x float]]* @float_2d_array
; CHECK-NOT: getelementptr
Reference in New Issue View Git Blame Copy Permalink