[clang] added regcall struct by reg support (PR #95257)
mahesh-attarde wrote: Thanks aaron for clearing assumptions. Will work on it with separate PR. https://github.com/llvm/llvm-project/pull/95257 ___ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] added regcall struct by reg support (PR #95257)
https://github.com/mahesh-attarde closed https://github.com/llvm/llvm-project/pull/95257 ___ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] added regcall struct by reg support (PR #95257)
AaronBallman wrote: > I don't have any context beyond what's written on > https://reviews.llvm.org/D25204 . And at this point, I'm not sure what > compiler we're trying to be compatible with. (I added @erichkeane in case he > remembers.) The original feature was written for compatibility with ICC and the fixes are also for compatibility with ICC. > Do you know why currently clang does not do that ? is there different version > of regcall spec it follows? I suspect this was an oversight, but @erichkeane may recall more specific details (note, he's out on leave until Sept, so I don't expect to hear back from him any time soon). AFAIK, the spec you linked is the correct one for us to follow. > Also if we add new regcall struct implementation, would that break binaries > across clang version? It would be an ABI break because we'd start passing structures differently. However, we have ABI tags; we can implement the new functionality under an ABI tag so users can get the old ABI still if they need to (search for `AbiTagAttr` to see examples of how we do this elsewhere and how we test for it). https://github.com/llvm/llvm-project/pull/95257 ___ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] added regcall struct by reg support (PR #95257)
mahesh-attarde wrote: @erichkeane Regcall spec has mentioned allocation strategy that is chunk based. Do you know why currently clang does not do that ? is there different version of regcall spec it follows? Also if we add new regcall struct implementation, would that break binaries across clang version? https://github.com/llvm/llvm-project/pull/95257 ___ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] added regcall struct by reg support (PR #95257)
efriedma-quic wrote: I don't have any context beyond what's written on https://reviews.llvm.org/D25204 . And at this point, I'm not sure what compiler we're trying to be compatible with. (I added @erichkeane in case he remembers.) https://github.com/llvm/llvm-project/pull/95257 ___ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] added regcall struct by reg support (PR #95257)
efriedma-quic wrote: Right, the specification requires splitting the whole structure into chunks; if we add a special-case for 8-byte structs, we'll just have to throw it away when we implement the right algorithm. Also, I'm not sure what the isBuiltinType() check is supposed to handle. It seems to exclude values you'd want to handle, like pointers. https://github.com/llvm/llvm-project/pull/95257 ___ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] added regcall struct by reg support (PR #95257)
mahesh-attarde wrote: > What you're implementing in this change doesn't seem like it brings us much > closer to what the document says. I mean, it handles the specific structs in > your testcase, but the algorithm you're using doesn't generalize. Can you suggest case which can be useful in generalizing? https://github.com/llvm/llvm-project/pull/95257 ___ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] added regcall struct by reg support (PR #95257)
efriedma-quic wrote: What you're implementing in this change doesn't seem like it brings us much closer to what the document says. I mean, it handles the specific structs in your testcase, but the algorithm you're using doesn't generalize. https://github.com/llvm/llvm-project/pull/95257 ___ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] added regcall struct by reg support (PR #95257)
mahesh-attarde wrote: There are not failures on `buildkite/github-pull-requests/linux-linux-x64.` something else is off. https://github.com/llvm/llvm-project/pull/95257 ___ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] added regcall struct by reg support (PR #95257)
mahesh-attarde wrote: Regcall ABI is described on https://cdrdv2-public.intel.com/679047/Intel-ABI-Vector-Function-v0.9.8.pdf Page No. 18 >  https://github.com/llvm/llvm-project/pull/95257 ___ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] added regcall struct by reg support (PR #95257)
efriedma-quic wrote: Do you have a reference for the rules you're implementing? https://github.com/llvm/llvm-project/pull/95257 ___ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] added regcall struct by reg support (PR #95257)
llvmbot wrote:
@llvm/pr-subscribers-clang
Author: None (mahesh-attarde)
Changes
In Context of __regcall,
Whenever struct can fit in a register we must use single register to pass
whole struct object.
Without this patch, it uses separate registers for each field. With Patch, if
Struct size is = 64
we can use GPR.
---
Full diff: https://github.com/llvm/llvm-project/pull/95257.diff
2 Files Affected:
- (modified) clang/lib/CodeGen/Targets/X86.cpp (+20)
- (added) clang/test/CodeGen/regcall3.c (+53)
``diff
diff --git a/clang/lib/CodeGen/Targets/X86.cpp
b/clang/lib/CodeGen/Targets/X86.cpp
index 43dadf5e724ac..506d106ad65b0 100644
--- a/clang/lib/CodeGen/Targets/X86.cpp
+++ b/clang/lib/CodeGen/Targets/X86.cpp
@@ -148,6 +148,7 @@ class X86_32ABIInfo : public ABIInfo {
Class classify(QualType Ty) const;
ABIArgInfo classifyReturnType(QualType RetTy, CCState ) const;
+
ABIArgInfo classifyArgumentType(QualType RetTy, CCState ,
unsigned ArgIndex) const;
@@ -1306,6 +1307,8 @@ class X86_64ABIInfo : public ABIInfo {
unsigned ,
unsigned ) const;
+ bool DoesRegcallStructFitInReg(QualType Ty) const;
+
bool IsIllegalVectorType(QualType Ty) const;
/// The 0.98 ABI revision clarified a lot of ambiguities,
@@ -2830,6 +2833,20 @@ X86_64ABIInfo::classifyArgumentType(QualType Ty,
unsigned freeIntRegs,
return ABIArgInfo::getDirect(ResType);
}
+bool X86_64ABIInfo::DoesRegcallStructFitInReg(QualType Ty) const {
+ auto RT = Ty->castAs();
+ // For Integer class, Max GPR Size is 64
+ if (getContext().getTypeSize(Ty) > 64)
+return false;
+ // Struct At hand must not have other non Builtin types
+ for (const auto *FD : RT->getDecl()->fields()) {
+QualType MTy = FD->getType();
+if (!MTy->isBuiltinType())
+ return false;
+ }
+ return true;
+}
+
ABIArgInfo
X86_64ABIInfo::classifyRegCallStructTypeImpl(QualType Ty, unsigned ,
unsigned ,
@@ -2837,6 +2854,9 @@ X86_64ABIInfo::classifyRegCallStructTypeImpl(QualType Ty,
unsigned ,
auto RT = Ty->getAs();
assert(RT && "classifyRegCallStructType only valid with struct types");
+ if (DoesRegcallStructFitInReg(Ty))
+return classifyArgumentType(Ty, UINT_MAX, NeededInt, NeededSSE, true,
true);
+
if (RT->getDecl()->hasFlexibleArrayMember())
return getIndirectReturnResult(Ty);
diff --git a/clang/test/CodeGen/regcall3.c b/clang/test/CodeGen/regcall3.c
new file mode 100644
index 0..1c83407220861
--- /dev/null
+++ b/clang/test/CodeGen/regcall3.c
@@ -0,0 +1,53 @@
+// RUN: %clang_cc1 -S %s -o - -ffreestanding -triple=x86_64-unknown-linux-gnu
| FileCheck %s --check-prefixes=LINUX64
+
+#include
+struct struct1 { int x; int y; };
+void __regcall v6(int a, float b, struct struct1 c) {}
+
+void v6_caller(){
+struct struct1 c0;
+c0.x = 0xa0a0; c0.y = 0xb0b0;
+int x= 0xf0f0, y = 0x0f0f;
+v6(x,y,c0);
+}
+
+// LINUX64-LABEL: __regcall3__v6
+// LINUX64: movq %rcx, -8(%rsp)
+// LINUX64: movl %eax, -12(%rsp)
+// LINUX64: movss %xmm0, -16(%rsp)
+
+// LINUX64-LABEL: v6_caller
+// LINUX64: movl $41120, 16(%rsp)# imm = 0xA0A0
+// LINUX64: movl $45232, 20(%rsp)# imm = 0xB0B0
+// LINUX64: movl $61680, 12(%rsp)# imm = 0xF0F0
+// LINUX64: movl $3855, 8(%rsp) # imm = 0xF0F
+// LINUX64: movl 12(%rsp), %eax
+// LINUX64: cvtsi2ssl 8(%rsp), %xmm0
+// LINUX64: movq 16(%rsp), %rcx
+// LINUX64: callq .L__regcall3__v6$local
+
+
+struct struct2 { int x; float y; };
+void __regcall v31(int a, float b, struct struct2 c) {}
+
+void v31_caller(){
+struct struct2 c0;
+c0.x = 0xa0a0; c0.y = 0xb0b0;
+int x= 0xf0f0, y = 0x0f0f;
+v31(x,y,c0);
+}
+
+// LINUX64: __regcall3__v31:# @__regcall3__v31
+// LINUX64:movq%rcx, -8(%rsp)
+// LINUX64:movl%eax, -12(%rsp)
+// LINUX64:movss %xmm0, -16(%rsp)
+// LINUX64: v31_caller: # @v31_caller
+// LINUX64:movl$41120, 16(%rsp)# imm = 0xA0A0
+// LINUX64:movss .LCPI3_0(%rip), %xmm0 # xmm0 =
[4.5232E+4,0.0E+0,0.0E+0,0.0E+0]
+// LINUX64:movss %xmm0, 20(%rsp)
+// LINUX64:movl$61680, 12(%rsp)# imm = 0xF0F0
+// LINUX64:movl$3855, 8(%rsp) # imm = 0xF0F
+// LINUX64:movl12(%rsp), %eax
+// LINUX64:cvtsi2ssl 8(%rsp), %xmm0
+// LINUX64:movq16(%rsp), %rcx
+// LINUX64:callq .L__regcall3__v31$local
``
https://github.com/llvm/llvm-project/pull/95257
___
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
[clang] added regcall struct by reg support (PR #95257)
https://github.com/mahesh-attarde created
https://github.com/llvm/llvm-project/pull/95257
In Context of __regcall,
Whenever struct can fit in a register we must use single register to pass
whole struct object.
Without this patch, it uses separate registers for each field. With Patch, if
Struct size is <= 64
we can use GPR.
>From 6d6619f8f7a37906ac45791487a4d63b51a48ad1 Mon Sep 17 00:00:00 2001
From: mahesh-attarde
Date: Wed, 12 Jun 2024 06:15:51 -0700
Subject: [PATCH] added regcall strct by reg support
---
clang/lib/CodeGen/Targets/X86.cpp | 20
clang/test/CodeGen/regcall3.c | 53 +++
2 files changed, 73 insertions(+)
create mode 100644 clang/test/CodeGen/regcall3.c
diff --git a/clang/lib/CodeGen/Targets/X86.cpp
b/clang/lib/CodeGen/Targets/X86.cpp
index 43dadf5e724ac..506d106ad65b0 100644
--- a/clang/lib/CodeGen/Targets/X86.cpp
+++ b/clang/lib/CodeGen/Targets/X86.cpp
@@ -148,6 +148,7 @@ class X86_32ABIInfo : public ABIInfo {
Class classify(QualType Ty) const;
ABIArgInfo classifyReturnType(QualType RetTy, CCState ) const;
+
ABIArgInfo classifyArgumentType(QualType RetTy, CCState ,
unsigned ArgIndex) const;
@@ -1306,6 +1307,8 @@ class X86_64ABIInfo : public ABIInfo {
unsigned ,
unsigned ) const;
+ bool DoesRegcallStructFitInReg(QualType Ty) const;
+
bool IsIllegalVectorType(QualType Ty) const;
/// The 0.98 ABI revision clarified a lot of ambiguities,
@@ -2830,6 +2833,20 @@ X86_64ABIInfo::classifyArgumentType(QualType Ty,
unsigned freeIntRegs,
return ABIArgInfo::getDirect(ResType);
}
+bool X86_64ABIInfo::DoesRegcallStructFitInReg(QualType Ty) const {
+ auto RT = Ty->castAs();
+ // For Integer class, Max GPR Size is 64
+ if (getContext().getTypeSize(Ty) > 64)
+return false;
+ // Struct At hand must not have other non Builtin types
+ for (const auto *FD : RT->getDecl()->fields()) {
+QualType MTy = FD->getType();
+if (!MTy->isBuiltinType())
+ return false;
+ }
+ return true;
+}
+
ABIArgInfo
X86_64ABIInfo::classifyRegCallStructTypeImpl(QualType Ty, unsigned ,
unsigned ,
@@ -2837,6 +2854,9 @@ X86_64ABIInfo::classifyRegCallStructTypeImpl(QualType Ty,
unsigned ,
auto RT = Ty->getAs();
assert(RT && "classifyRegCallStructType only valid with struct types");
+ if (DoesRegcallStructFitInReg(Ty))
+return classifyArgumentType(Ty, UINT_MAX, NeededInt, NeededSSE, true,
true);
+
if (RT->getDecl()->hasFlexibleArrayMember())
return getIndirectReturnResult(Ty);
diff --git a/clang/test/CodeGen/regcall3.c b/clang/test/CodeGen/regcall3.c
new file mode 100644
index 0..1c83407220861
--- /dev/null
+++ b/clang/test/CodeGen/regcall3.c
@@ -0,0 +1,53 @@
+// RUN: %clang_cc1 -S %s -o - -ffreestanding -triple=x86_64-unknown-linux-gnu
| FileCheck %s --check-prefixes=LINUX64
+
+#include
+struct struct1 { int x; int y; };
+void __regcall v6(int a, float b, struct struct1 c) {}
+
+void v6_caller(){
+struct struct1 c0;
+c0.x = 0xa0a0; c0.y = 0xb0b0;
+int x= 0xf0f0, y = 0x0f0f;
+v6(x,y,c0);
+}
+
+// LINUX64-LABEL: __regcall3__v6
+// LINUX64: movq %rcx, -8(%rsp)
+// LINUX64: movl %eax, -12(%rsp)
+// LINUX64: movss %xmm0, -16(%rsp)
+
+// LINUX64-LABEL: v6_caller
+// LINUX64: movl $41120, 16(%rsp)# imm = 0xA0A0
+// LINUX64: movl $45232, 20(%rsp)# imm = 0xB0B0
+// LINUX64: movl $61680, 12(%rsp)# imm = 0xF0F0
+// LINUX64: movl $3855, 8(%rsp) # imm = 0xF0F
+// LINUX64: movl 12(%rsp), %eax
+// LINUX64: cvtsi2ssl 8(%rsp), %xmm0
+// LINUX64: movq 16(%rsp), %rcx
+// LINUX64: callq .L__regcall3__v6$local
+
+
+struct struct2 { int x; float y; };
+void __regcall v31(int a, float b, struct struct2 c) {}
+
+void v31_caller(){
+struct struct2 c0;
+c0.x = 0xa0a0; c0.y = 0xb0b0;
+int x= 0xf0f0, y = 0x0f0f;
+v31(x,y,c0);
+}
+
+// LINUX64: __regcall3__v31:# @__regcall3__v31
+// LINUX64:movq%rcx, -8(%rsp)
+// LINUX64:movl%eax, -12(%rsp)
+// LINUX64:movss %xmm0, -16(%rsp)
+// LINUX64: v31_caller: # @v31_caller
+// LINUX64:movl$41120, 16(%rsp)# imm = 0xA0A0
+// LINUX64:movss .LCPI3_0(%rip), %xmm0 # xmm0 =
[4.5232E+4,0.0E+0,0.0E+0,0.0E+0]
+// LINUX64:movss %xmm0, 20(%rsp)
+// LINUX64:movl$61680, 12(%rsp)# imm = 0xF0F0
+// LINUX64:movl$3855, 8(%rsp) # imm = 0xF0F
+// LINUX64:movl12(%rsp), %eax
+// LINUX64:cvtsi2ssl 8(%rsp), %xmm0
+// LINUX64:movq16(%rsp), %rcx
+// LINUX64:callq .L__regcall3__v31$local
___
cfe-commits mailing list
