6502/llvm-6502
1
0
Fork 0
mirror of https://github.com/c64scene-ar/llvm-6502.git synced 2025-11-02 07:17:36 +00:00
Code Issues Projects Releases Wiki Activity

Optimize away unnecessary address casts.

Browse Source 
Removes unnecessary casts from non-generic address spaces to the generic address
space for certain code patterns.

Patch by Jingyue Wu.



git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@205571 91177308-0d34-0410-b5e6-96231b3b80d8
This commit is contained in:
Eli Bendersky
2014-04-03 21:18:25 +00:00
parent 89218827c8
commit 25540a7f39
6 changed files with 299 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

91
test/CodeGen/NVPTX/access-non-generic.ll Normal file
View File

@@ -0,0 +1,91 @@
; RUN: llc < %s -march=nvptx -mcpu=sm_20 | FileCheck %s --check-prefix PTX
; RUN: llc < %s -march=nvptx64 -mcpu=sm_20 | FileCheck %s --check-prefix PTX
; RUN: opt < %s -S -nvptx-favor-non-generic -dce | FileCheck %s --check-prefix IR
@array = internal addrspace(3) global [10 x float] zeroinitializer, align 4
@scalar = internal addrspace(3) global float 0.000000e+00, align 4
; Verifies nvptx-favor-non-generic correctly optimizes generic address space
; usage to non-generic address space usage for the patterns we claim to handle:
; 1. load cast
; 2. store cast
; 3. load gep cast
; 4. store gep cast
; gep and cast can be an instruction or a constant expression. This function
; tries all possible combinations.
define float @ld_st_shared_f32(i32 %i, float %v) {
; IR-LABEL: @ld_st_shared_f32
; IR-NOT: addrspacecast
; PTX-LABEL: ld_st_shared_f32(
; load cast
%1 = load float* addrspacecast (float addrspace(3)* @scalar to float*), align 4
; PTX: ld.shared.f32 %f{{[0-9]+}}, [scalar];
; store cast
store float %v, float* addrspacecast (float addrspace(3)* @scalar to float*), align 4
; PTX: st.shared.f32 [scalar], %f{{[0-9]+}};
; use syncthreads to disable optimizations across components
call void @llvm.cuda.syncthreads()
; PTX: bar.sync 0;
; cast; load
%2 = addrspacecast float addrspace(3)* @scalar to float*
%3 = load float* %2, align 4
; PTX: ld.shared.f32 %f{{[0-9]+}}, [scalar];
; cast; store
store float %v, float* %2, align 4
; PTX: st.shared.f32 [scalar], %f{{[0-9]+}};
call void @llvm.cuda.syncthreads()
; PTX: bar.sync 0;
; load gep cast
%4 = load float* getelementptr inbounds ([10 x float]* addrspacecast ([10 x float] addrspace(3)* @array to [10 x float]*), i32 0, i32 5), align 4
; PTX: ld.shared.f32 %f{{[0-9]+}}, [array+20];
; store gep cast
store float %v, float* getelementptr inbounds ([10 x float]* addrspacecast ([10 x float] addrspace(3)* @array to [10 x float]*), i32 0, i32 5), align 4
; PTX: st.shared.f32 [array+20], %f{{[0-9]+}};
call void @llvm.cuda.syncthreads()
; PTX: bar.sync 0;
; gep cast; load
%5 = getelementptr inbounds [10 x float]* addrspacecast ([10 x float] addrspace(3)* @array to [10 x float]*), i32 0, i32 5
%6 = load float* %5, align 4
; PTX: ld.shared.f32 %f{{[0-9]+}}, [array+20];
; gep cast; store
store float %v, float* %5, align 4
; PTX: st.shared.f32 [array+20], %f{{[0-9]+}};
call void @llvm.cuda.syncthreads()
; PTX: bar.sync 0;
; cast; gep; load
%7 = addrspacecast [10 x float] addrspace(3)* @array to [10 x float]*
%8 = getelementptr inbounds [10 x float]* %7, i32 0, i32 %i
%9 = load float* %8, align 4
; PTX: ld.shared.f32 %f{{[0-9]+}}, [%{{(r|rl|rd)[0-9]+}}];
; cast; gep; store
store float %v, float* %8, align 4
; PTX: st.shared.f32 [%{{(r|rl|rd)[0-9]+}}], %f{{[0-9]+}};
call void @llvm.cuda.syncthreads()
; PTX: bar.sync 0;
%sum2 = fadd float %1, %3
%sum3 = fadd float %sum2, %4
%sum4 = fadd float %sum3, %6
%sum5 = fadd float %sum4, %9
ret float %sum5
}
; Verifies nvptx-favor-non-generic keeps addrspacecasts between pointers of
; different element types.
define i32 @ld_int_from_float() {
; IR-LABEL: @ld_int_from_float
; IR: addrspacecast
; PTX-LABEL: ld_int_from_float(
; PTX: cvta.shared.u{{(32|64)}}
%1 = load i32* addrspacecast(float addrspace(3)* @scalar to i32*), align 4
ret i32 %1
}
declare void @llvm.cuda.syncthreads() #3
attributes #3 = { noduplicate nounwind }