-
Notifications
You must be signed in to change notification settings - Fork 13.5k
release/20.x: [Hexagon] Handle Call Operand vxi1 in Hexagon Backend (#128027) #129311
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Conversation
@iajbar What do you think about merging this PR to the release branch? |
@llvm/pr-subscribers-backend-hexagon Author: None (llvmbot) ChangesBackport 37559c8 Requested by: @androm3da Full diff: https://github.com/llvm/llvm-project/pull/129311.diff 7 Files Affected:
diff --git a/llvm/lib/Target/Hexagon/HexagonCallingConv.td b/llvm/lib/Target/Hexagon/HexagonCallingConv.td
index cc41b569e4904..2378bbc928d49 100644
--- a/llvm/lib/Target/Hexagon/HexagonCallingConv.td
+++ b/llvm/lib/Target/Hexagon/HexagonCallingConv.td
@@ -65,6 +65,8 @@ def CC_Hexagon: CallingConv<[
CCIfType<[i32],
CCIfSplit<
CCCustom<"CC_SkipOdd">>>,
+ CCIfType<[v4i1], CCPromoteToType<v4i16>>,
+ CCIfType<[v8i1], CCPromoteToType<v8i8>>,
CCIfType<[i32,v2i16,v4i8],
CCAssignToReg<[R0,R1,R2,R3,R4,R5]>>,
@@ -111,6 +113,14 @@ class CCIfHvx128<CCAction A>
def CC_Hexagon_HVX: CallingConv<[
// HVX 64-byte mode
+
+ CCIfHvx64<
+ CCIfType<[v16i1], CCPromoteToType<v16i32>>>,
+ CCIfHvx64<
+ CCIfType<[v32i1], CCPromoteToType<v32i16>>>,
+ CCIfHvx64<
+ CCIfType<[v64i1], CCPromoteToType<v64i8>>>,
+
CCIfHvx64<
CCIfType<[v16i32,v32i16,v64i8],
CCAssignToReg<[V0,V1,V2,V3,V4,V5,V6,V7,V8,V9,V10,V11,V12,V13,V14,V15]>>>,
@@ -125,6 +135,14 @@ def CC_Hexagon_HVX: CallingConv<[
CCAssignToStack<128,64>>>,
// HVX 128-byte mode
+
+ CCIfHvx128<
+ CCIfType<[v32i1], CCPromoteToType<v32i32>>>,
+ CCIfHvx128<
+ CCIfType<[v64i1], CCPromoteToType<v64i16>>>,
+ CCIfHvx128<
+ CCIfType<[v128i1], CCPromoteToType<v128i8>>>,
+
CCIfHvx128<
CCIfType<[v32i32,v64i16,v128i8,v32f32,v64f16],
CCAssignToReg<[V0,V1,V2,V3,V4,V5,V6,V7,V8,V9,V10,V11,V12,V13,V14,V15]>>>,
diff --git a/llvm/test/CodeGen/Hexagon/calloperand-v128i1.ll b/llvm/test/CodeGen/Hexagon/calloperand-v128i1.ll
new file mode 100644
index 0000000000000..ddac8c1cd8279
--- /dev/null
+++ b/llvm/test/CodeGen/Hexagon/calloperand-v128i1.ll
@@ -0,0 +1,39 @@
+;RUN: llc -mtriple=hexagon -mattr=+hvxv79,+hvx-length128b < %s -o - | FileCheck %s
+
+; CHECK-LABEL: compare_vectors
+; CHECK: [[REG1:(q[0-9]+)]] = vcmp.eq(v{{[0-9]+}}.b,v{{[0-9]+}}.b)
+; CHECK: [[REG2:(r[0-9]+)]] = #-1
+; CHECK: v0 = vand([[REG1]],[[REG2]])
+
+define void @compare_vectors(<128 x i8> %a, <128 x i8> %b) {
+entry:
+ %result = icmp eq <128 x i8> %a, %b
+ call i32 @f.1(<128 x i1> %result)
+ ret void
+}
+
+; CHECK-LABEL: f.1:
+; CHECK: [[REG3:(q[0-9]+)]] = vand(v0,r{{[0-9]+}})
+; CHECK: [[REG4:(v[0-9]+)]] = vand([[REG3]],r{{[0-9]+}})
+; CHECK: r{{[0-9]+}} = vextract([[REG4]],r{{[0-9]+}})
+
+define i32 @f.1(<128 x i1> %vec) {
+ %element = extractelement <128 x i1> %vec, i32 6
+ %is_true = icmp eq i1 %element, true
+ br i1 %is_true, label %if_true, label %if_false
+
+if_true:
+ call void @action_if_true()
+ br label %end
+
+if_false:
+ call void @action_if_false()
+ br label %end
+
+end:
+ %result = phi i32 [1, %if_true], [0, %if_false]
+ ret i32 %result
+}
+
+declare void @action_if_true()
+declare void @action_if_false()
diff --git a/llvm/test/CodeGen/Hexagon/calloperand-v16i1.ll b/llvm/test/CodeGen/Hexagon/calloperand-v16i1.ll
new file mode 100644
index 0000000000000..bbb2697246df1
--- /dev/null
+++ b/llvm/test/CodeGen/Hexagon/calloperand-v16i1.ll
@@ -0,0 +1,40 @@
+;RUN: llc -mtriple=hexagon -mattr=+hvxv79,+hvx-length64b < %s -o - | FileCheck %s --check-prefix=CHECK
+;RUN: llc -mtriple=hexagon -mattr=+hvxv79,+hvx-length128b < %s -o - | FileCheck %s --check-prefix=CHECK
+
+; CHECK-LABEL: compare_vectors
+; CHECK: [[REG1:(q[0-9]+)]] = vcmp.eq(v{{[0-9]+}}.w,v{{[0-9]+}}.w)
+; CHECK: [[REG2:(r[0-9]+)]] = #-1
+; CHECK: v0 = vand([[REG1]],[[REG2]])
+
+define void @compare_vectors(<16 x i32> %a, <16 x i32> %b) {
+entry:
+ %result = icmp eq <16 x i32> %a, %b
+ call i32 @f.1(<16 x i1> %result)
+ ret void
+}
+
+; CHECK-LABEL: f.1:
+; CHECK: [[REG3:(q[0-9]+)]] = vand(v0,r{{[0-9]+}})
+; CHECK: [[REG4:(v[0-9]+)]] = vand([[REG3]],r{{[0-9]+}})
+; CHECK: r{{[0-9]+}} = vextract([[REG4]],r{{[0-9]+}})
+
+define i32 @f.1(<16 x i1> %vec) {
+ %element = extractelement <16 x i1> %vec, i32 6
+ %is_true = icmp eq i1 %element, true
+ br i1 %is_true, label %if_true, label %if_false
+
+if_true:
+ call void @action_if_true()
+ br label %end
+
+if_false:
+ call void @action_if_false()
+ br label %end
+
+end:
+ %result = phi i32 [1, %if_true], [0, %if_false]
+ ret i32 %result
+}
+
+declare void @action_if_true()
+declare void @action_if_false()
diff --git a/llvm/test/CodeGen/Hexagon/calloperand-v32i1.ll b/llvm/test/CodeGen/Hexagon/calloperand-v32i1.ll
new file mode 100644
index 0000000000000..a73478728d910
--- /dev/null
+++ b/llvm/test/CodeGen/Hexagon/calloperand-v32i1.ll
@@ -0,0 +1,50 @@
+; RUN: llc -mtriple=hexagon -mattr=+hvxv79,+hvx-length64b < %s -o - | FileCheck %s --check-prefix=CHECK-64
+; RUN: llc -mtriple=hexagon -mattr=+hvxv79,+hvx-length128b < %s -o - | FileCheck %s --check-prefix=CHECK-128
+
+; CHECK-LABEL: compare_vectors
+; CHECK-64: [[REG1:(q[0-9]+)]] = vcmp.eq(v{{[0-9]+}}.h,v{{[0-9]+}}.h)
+; CHECK-64: [[REG2:(r[0-9]+)]] = #-1
+; CHECK-64: v0 = vand([[REG1]],[[REG2]])
+; CHECK-128: r{{[0-9]+}}:{{[0-9]+}} = combine(##.LCPI0_0,#-1)
+; CHECK-128: [[REG1:(q[0-9]+)]] = vcmp.eq(v0.h,v1.h)
+; CHECK-128: [[REG2:(v[0-9]+)]] = vand([[REG1]],r{{[0-9]+}})
+; CHECK-128: [[REG3:(v[0-9]+)]] = vmem(r{{[0-9]+}}+#0)
+; CHECK-128: [[REG4:(v[0-9]+)]] = vdelta([[REG2]],[[REG3]])
+; CHECK-128: [[REG5:(q[0-9]+)]] = vand([[REG4]],r{{[0-9]+}})
+; CHECK-128: v0 = vand([[REG5]],r{{[0-9]+}})
+
+define void @compare_vectors(<32 x i16> %a, <32 x i16> %b) {
+entry:
+ %result = icmp eq <32 x i16> %a, %b
+ call i32 @f.1(<32 x i1> %result)
+ ret void
+}
+
+; CHECK-LABEL: f.1:
+; CHECK-64: [[REG3:(q[0-9]+)]] = vand(v0,r{{[0-9]+}})
+; CHECK-64: [[REG4:(v[0-9]+)]] = vand([[REG3]],r{{[0-9]+}})
+; CHECK-64: r{{[0-9]+}} = vextract([[REG4]],r{{[0-9]+}})
+; CHECK-128: [[REG6:(q[0-9]+)]] = vand(v0,r{{[0-9]+}})
+; CHECK-128: [[REG7:(v[0-9]+)]] = vand([[REG6]],r{{[0-9]+}})
+; CHECK-128: r{{[0-9]+}} = vextract([[REG7]],r{{[0-9]+}})
+
+define i32 @f.1(<32 x i1> %vec) {
+ %element = extractelement <32 x i1> %vec, i32 6
+ %is_true = icmp eq i1 %element, true
+ br i1 %is_true, label %if_true, label %if_false
+
+if_true:
+ call void @action_if_true()
+ br label %end
+
+if_false:
+ call void @action_if_false()
+ br label %end
+
+end:
+ %result = phi i32 [1, %if_true], [0, %if_false]
+ ret i32 %result
+}
+
+declare void @action_if_true()
+declare void @action_if_false()
diff --git a/llvm/test/CodeGen/Hexagon/calloperand-v4i1.ll b/llvm/test/CodeGen/Hexagon/calloperand-v4i1.ll
new file mode 100644
index 0000000000000..af23e6c788ea5
--- /dev/null
+++ b/llvm/test/CodeGen/Hexagon/calloperand-v4i1.ll
@@ -0,0 +1,39 @@
+;RUN: llc -mtriple=hexagon < %s -o - | FileCheck %s --check-prefix=CHECK
+;RUN: llc -mtriple=hexagon -mattr=+hvxv79,+hvx-length64b < %s -o - | FileCheck %s --check-prefix=CHECK
+;RUN: llc -mtriple=hexagon -mattr=+hvxv79,+hvx-length128b < %s -o - | FileCheck %s --check-prefix=CHECK
+
+; CHECK-LABEL: compare_vectors
+; CHECK: [[REG0:(p[0-9]+)]] = vcmph.eq([[REG1:(r[0-9]+):[0-9]]],[[REG2:(r[0-9]+):[0-9]]])
+; CHECK: [[REG1:(r[0-9]+):[0-9]]] = CONST64(#281479271743489)
+; CHECK: [[REG2:(r[0-9]+):[0-9]]] = mask([[REG0]])
+; CHECK: r{{[0-9]+}}:{{[0-9]+}} = and([[REG2]],[[REG1]])
+
+define void @compare_vectors(<4 x i16> %a, <4 x i16> %b) {
+entry:
+ %result = icmp eq <4 x i16> %a, %b
+ call i32 @f.1(<4 x i1> %result)
+ ret void
+}
+; CHECK-LABEL: f.1:
+; CHECK: [[REG3:(r[0-9]+)]] = and([[REG3]],##65537)
+; CHECK: [[REG4:(r[0-9]+)]] = and([[REG4]],##65537)
+define i32 @f.1(<4 x i1> %vec) {
+ %element = extractelement <4 x i1> %vec, i32 2
+ %is_true = icmp eq i1 %element, true
+ br i1 %is_true, label %if_true, label %if_false
+
+if_true:
+ call void @action_if_true()
+ br label %end
+
+if_false:
+ call void @action_if_false()
+ br label %end
+
+end:
+ %result = phi i32 [1, %if_true], [0, %if_false]
+ ret i32 %result
+}
+
+declare void @action_if_true()
+declare void @action_if_false()
diff --git a/llvm/test/CodeGen/Hexagon/calloperand-v64i1.ll b/llvm/test/CodeGen/Hexagon/calloperand-v64i1.ll
new file mode 100644
index 0000000000000..7cc562085a7e6
--- /dev/null
+++ b/llvm/test/CodeGen/Hexagon/calloperand-v64i1.ll
@@ -0,0 +1,50 @@
+; RUN: llc -mtriple=hexagon -mattr=+hvxv79,+hvx-length64b < %s -o - | FileCheck %s --check-prefix=CHECK-64
+; RUN: llc -mtriple=hexagon -mattr=+hvxv79,+hvx-length128b < %s -o - | FileCheck %s --check-prefix=CHECK-128
+
+; CHECK-LABEL: compare_vectors
+; CHECK-64: [[REG1:(q[0-9]+)]] = vcmp.eq(v{{[0-9]+}}.b,v{{[0-9]+}}.b)
+; CHECK-64: [[REG2:(r[0-9]+)]] = #-1
+; CHECK-64: v0 = vand([[REG1]],[[REG2]])
+; CHECK-128: r{{[0-9]+}}:{{[0-9]+}} = combine(##.LCPI0_0,#-1)
+; CHECK-128: [[REG1:(q[0-9]+)]] = vcmp.eq(v0.b,v1.b)
+; CHECK-128: [[REG2:(v[0-9]+)]] = vand([[REG1]],r{{[0-9]+}})
+; CHECK-128: [[REG3:(v[0-9]+)]] = vmem(r{{[0-9]+}}+#0)
+; CHECK-128: [[REG4:(v[0-9]+)]] = vdelta([[REG2]],[[REG3]])
+; CHECK-128: [[REG5:(q[0-9]+)]] = vand([[REG4]],r{{[0-9]+}})
+; CHECK-128: v0 = vand([[REG5]],r{{[0-9]+}})
+
+define void @compare_vectors(<64 x i8> %a, <64 x i8> %b) {
+entry:
+ %result = icmp eq <64 x i8> %a, %b
+ call i32 @f.1(<64 x i1> %result)
+ ret void
+}
+
+; CHECK-LABEL: f.1:
+; CHECK-64: [[REG3:(q[0-9]+)]] = vand(v0,r{{[0-9]+}})
+; CHECK-64: [[REG4:(v[0-9]+)]] = vand([[REG3]],r{{[0-9]+}})
+; CHECK-64: r{{[0-9]+}} = vextract([[REG4]],r{{[0-9]+}})
+; CHECK-128: [[REG6:(q[0-9]+)]] = vand(v0,r{{[0-9]+}})
+; CHECK-128: [[REG7:(v[0-9]+)]] = vand([[REG6]],r{{[0-9]+}})
+; CHECK-128: r{{[0-9]+}} = vextract([[REG7]],r{{[0-9]+}})
+
+define i32 @f.1(<64 x i1> %vec) {
+ %element = extractelement <64 x i1> %vec, i32 6
+ %is_true = icmp eq i1 %element, true
+ br i1 %is_true, label %if_true, label %if_false
+
+if_true:
+ call void @action_if_true()
+ br label %end
+
+if_false:
+ call void @action_if_false()
+ br label %end
+
+end:
+ %result = phi i32 [1, %if_true], [0, %if_false]
+ ret i32 %result
+}
+
+declare void @action_if_true()
+declare void @action_if_false()
diff --git a/llvm/test/CodeGen/Hexagon/calloperand-v8i1.ll b/llvm/test/CodeGen/Hexagon/calloperand-v8i1.ll
new file mode 100644
index 0000000000000..2ec163acd6ae7
--- /dev/null
+++ b/llvm/test/CodeGen/Hexagon/calloperand-v8i1.ll
@@ -0,0 +1,39 @@
+;RUN: llc -mtriple=hexagon < %s -o - | FileCheck %s --check-prefix=CHECK
+;RUN: llc -mtriple=hexagon -mattr=+hvxv79,+hvx-length64b < %s -o - | FileCheck %s --check-prefix=CHECK
+;RUN: llc -mtriple=hexagon -mattr=+hvxv79,+hvx-length128b < %s -o - | FileCheck %s --check-prefix=CHECK
+
+; CHECK-LABEL: compare_vectors
+; CHECK: [[REG0:(p[0-9]+)]] = vcmpb.eq([[REG1:(r[0-9]+):[0-9]]],[[REG2:(r[0-9]+):[0-9]]])
+; CHECK: [[REG1:(r[0-9]+):[0-9]]] = CONST64(#72340172838076673)
+; CHECK: [[REG2:(r[0-9]+):[0-9]]] = mask([[REG0]])
+; CHECK: r{{[0-9]+}}:{{[0-9]+}} = and([[REG2]],[[REG1]])
+
+define void @compare_vectors(<8 x i8> %a, <8 x i8> %b) {
+entry:
+ %result = icmp eq <8 x i8> %a, %b
+ call i32 @f.1(<8 x i1> %result)
+ ret void
+}
+; CHECK-LABEL: f.1:
+; CHECK: [[REG3:(r[0-9]+)]] = and([[REG3]],##16843009)
+; CHECK: [[REG4:(r[0-9]+)]] = and([[REG4]],##16843009)
+define i32 @f.1(<8 x i1> %vec) {
+ %element = extractelement <8 x i1> %vec, i32 6
+ %is_true = icmp eq i1 %element, true
+ br i1 %is_true, label %if_true, label %if_false
+
+if_true:
+ call void @action_if_true()
+ br label %end
+
+if_false:
+ call void @action_if_false()
+ br label %end
+
+end:
+ %result = phi i32 [1, %if_true], [0, %if_false]
+ ret i32 %result
+}
+
+declare void @action_if_true()
+declare void @action_if_false()
|
Does this need to be in 20.1.0 or could it wait until 20.1.1 ? |
It can wait. |
Hmm -- or, rather -- I should ask @alexrp whether it can wait. |
Strictly speaking this represents an ABI change from the perspective of LLVM backend users (not Clang users), so per the LLVM release policy, I don't see how it could be included in a post-20.1.0 release? If possible, it'd certainly be nice to get it into LLVM 20 from Zig's perspective. IIRC, it's the most prominent blocker for running Zig's language behavior tests on Hexagon. |
@alexrp Usually when we talk about ABI stability we are talking about the ABI of libLLVM.so and not the calling conventions for the individual targets. We have a little more flexibility to change the targets than we do the shared library. What was the old behavior before this change if there was a function with a Vxi1 argument? |
It would assert (and I suppose emit the wrong code in a build w/o asserts). This has never worked IIRC so it's safe to put in either release number AFAICT. |
But this is compelling, so it'd be nice to see if this change could make it. On the upside, it's very well confined to the hexagon backend. |
This commit updates the Hexagon backend to handle vxi1 call operands. It ensures compatibility for vector types of sizes 4, 8, 16, 32, 64, and 128 x i1 when HVX is enabled. ~Fixes llvm#59009 and llvm#118879~ (cherry picked from commit 37559c8)
@androm3da (or anyone else). If you would like to add a note about this fix in the release notes (completely optional). Please reply to this comment with a one or two sentence description of the fix. When you are done, please add the release:note label to this PR. |
Backport 37559c8
Requested by: @androm3da