From e73e2a976eccadc249d7e9560dec63c41e123426 Mon Sep 17 00:00:00 2001
From: Russel Arbore <russel.jma@gmail.com>
Date: Sun, 16 Feb 2025 18:16:37 -0600
Subject: [PATCH 1/4] emit aligns
---
hercules_cg/src/cpu.rs | 13 ++++++++-----
juno_scheduler/src/pm.rs | 3 +++
2 files changed, 11 insertions(+), 5 deletions(-)
diff --git a/hercules_cg/src/cpu.rs b/hercules_cg/src/cpu.rs
index 7d87170b..a19218c7 100644
--- a/hercules_cg/src/cpu.rs
+++ b/hercules_cg/src/cpu.rs
@@ -103,8 +103,9 @@ impl<'a> CPUContext<'a> {
} else {
write!(
w,
- "{} noalias nofree nonnull noundef %p{}",
+ "{} noalias nofree nonnull noundef align({}) %p{}",
self.get_type(*ty),
+ get_type_alignment(&self.types, *ty),
idx
)?;
}
@@ -506,10 +507,11 @@ impl<'a> CPUContext<'a> {
// load.
write!(
body,
- " {} = load {}, ptr {}\n",
+ " {} = load {}, ptr {}, align {}\n",
self.get_value(id, false),
self.get_type(self_ty),
- index_ptr_name
+ index_ptr_name,
+ get_type_alignment(self.types, collect_ty),
)?;
} else {
// If this read doesn't reach a primitive type, just return
@@ -538,9 +540,10 @@ impl<'a> CPUContext<'a> {
// perform a single store of the data value.
write!(
body,
- " store {}, ptr {}\n",
+ " store {}, ptr {}, align {}\n",
self.get_value(data, true),
- index_ptr_name
+ index_ptr_name,
+ get_type_alignment(self.types, collect_ty),
)?;
} else {
// If the data item being written is not a primitive type,
diff --git a/juno_scheduler/src/pm.rs b/juno_scheduler/src/pm.rs
index 410e614c..1c00e3d0 100644
--- a/juno_scheduler/src/pm.rs
+++ b/juno_scheduler/src/pm.rs
@@ -795,6 +795,7 @@ impl PassManager {
.arg(&llvm_path)
.arg("-c")
.arg("-O3")
+ .arg("-ffast-math")
.arg("-march=native")
.arg("-o")
.arg(&llvm_object)
@@ -830,6 +831,8 @@ impl PassManager {
let mut nvcc_process = Command::new("nvcc")
.arg("-c")
.arg("-O3")
+ .arg("-ffast-math")
+ .arg("-march=native")
.arg("-diag-suppress")
.arg("177")
.arg("-o")
--
GitLab
From 7764ca63040186f02411516448f8e7f81236fc0c Mon Sep 17 00:00:00 2001
From: Russel Arbore <russel.jma@gmail.com>
Date: Sun, 16 Feb 2025 18:19:50 -0600
Subject: [PATCH 2/4] Use larger alignment for arrays
---
hercules_cg/src/cpu.rs | 6 ++----
hercules_cg/src/lib.rs | 5 +++--
hercules_rt/src/lib.rs | 6 ++++--
3 files changed, 9 insertions(+), 8 deletions(-)
diff --git a/hercules_cg/src/cpu.rs b/hercules_cg/src/cpu.rs
index a19218c7..f90657a4 100644
--- a/hercules_cg/src/cpu.rs
+++ b/hercules_cg/src/cpu.rs
@@ -507,11 +507,10 @@ impl<'a> CPUContext<'a> {
// load.
write!(
body,
- " {} = load {}, ptr {}, align {}\n",
+ " {} = load {}, ptr {}\n",
self.get_value(id, false),
self.get_type(self_ty),
index_ptr_name,
- get_type_alignment(self.types, collect_ty),
)?;
} else {
// If this read doesn't reach a primitive type, just return
@@ -540,10 +539,9 @@ impl<'a> CPUContext<'a> {
// perform a single store of the data value.
write!(
body,
- " store {}, ptr {}, align {}\n",
+ " store {}, ptr {}\n",
self.get_value(data, true),
index_ptr_name,
- get_type_alignment(self.types, collect_ty),
)?;
} else {
// If the data item being written is not a primitive type,
diff --git a/hercules_cg/src/lib.rs b/hercules_cg/src/lib.rs
index 15946f72..af2420d8 100644
--- a/hercules_cg/src/lib.rs
+++ b/hercules_cg/src/lib.rs
@@ -16,7 +16,7 @@ use std::collections::BTreeMap;
use hercules_ir::*;
-pub const LARGEST_ALIGNMENT: usize = 8;
+pub const LARGEST_ALIGNMENT: usize = 32;
/*
* The alignment of a type does not depend on dynamic constants.
@@ -33,7 +33,8 @@ pub fn get_type_alignment(types: &Vec<Type>, ty: TypeID) -> usize {
.map(|id| get_type_alignment(types, *id))
.max()
.unwrap_or(1),
- Type::Array(elem, _) => get_type_alignment(types, elem),
+ // Use a large alignment for arrays to generate better vector code.
+ Type::Array(_, _) => LARGEST_ALIGNMENT,
}
}
diff --git a/hercules_rt/src/lib.rs b/hercules_rt/src/lib.rs
index f8fdf2ef..0c3ffd80 100644
--- a/hercules_rt/src/lib.rs
+++ b/hercules_rt/src/lib.rs
@@ -13,8 +13,10 @@ use std::sync::OnceLock;
* src/rt.rs (the RT backend).
*/
+pub const LARGEST_ALIGNMENT: usize = 32;
+
pub unsafe fn __cpu_alloc(size: usize) -> *mut u8 {
- let ptr = alloc(Layout::from_size_align(size, 16).unwrap());
+ let ptr = alloc(Layout::from_size_align(size, LARGEST_ALIGNMENT).unwrap());
if cfg!(feature = "debug") {
eprintln!("__cpu_alloc: {:?}, {}", ptr, size);
assert!(!ptr.is_null() || size == 0);
@@ -27,7 +29,7 @@ pub unsafe fn __cpu_dealloc(ptr: *mut u8, size: usize) {
eprintln!("__cpu_dealloc: {:?}, {}", ptr, size);
assert!(!ptr.is_null() || size == 0);
}
- dealloc(ptr, Layout::from_size_align(size, 16).unwrap())
+ dealloc(ptr, Layout::from_size_align(size, LARGEST_ALIGNMENT).unwrap())
}
pub unsafe fn __cpu_zero_mem(ptr: *mut u8, size: usize) {
--
GitLab
From 6ba77d30bb843202e19bdc746d80356fc8a652c2 Mon Sep 17 00:00:00 2001
From: Russel Arbore <russel.jma@gmail.com>
Date: Sun, 16 Feb 2025 18:23:53 -0600
Subject: [PATCH 3/4] .
---
juno_scheduler/src/pm.rs | 1 -
1 file changed, 1 deletion(-)
diff --git a/juno_scheduler/src/pm.rs b/juno_scheduler/src/pm.rs
index 1c00e3d0..e1ff113e 100644
--- a/juno_scheduler/src/pm.rs
+++ b/juno_scheduler/src/pm.rs
@@ -831,7 +831,6 @@ impl PassManager {
let mut nvcc_process = Command::new("nvcc")
.arg("-c")
.arg("-O3")
- .arg("-ffast-math")
.arg("-march=native")
.arg("-diag-suppress")
.arg("177")
--
GitLab
From c4f5220dca24514978ab8adfa41f096b12bce54e Mon Sep 17 00:00:00 2001
From: Russel Arbore <russel.jma@gmail.com>
Date: Sun, 16 Feb 2025 18:27:54 -0600
Subject: [PATCH 4/4] .
---
juno_scheduler/src/pm.rs | 1 -
1 file changed, 1 deletion(-)
diff --git a/juno_scheduler/src/pm.rs b/juno_scheduler/src/pm.rs
index e1ff113e..4555e024 100644
--- a/juno_scheduler/src/pm.rs
+++ b/juno_scheduler/src/pm.rs
@@ -831,7 +831,6 @@ impl PassManager {
let mut nvcc_process = Command::new("nvcc")
.arg("-c")
.arg("-O3")
- .arg("-march=native")
.arg("-diag-suppress")
.arg("177")
.arg("-o")
--
GitLab