-
Notifications
You must be signed in to change notification settings - Fork 10.9k
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
[AMDGPU] - Add constant folding to s_wqm intrinsic #72382
Conversation
Fold any constant input to the s_wqm intrinsic.
@llvm/pr-subscribers-backend-amdgpu @llvm/pr-subscribers-llvm-analysis Author: Jessica Del (OutOfCache) ChangesFold any constant input to the Full diff: https://github.com/llvm/llvm-project/pull/72382.diff 2 Files Affected:
diff --git a/llvm/lib/Analysis/ConstantFolding.cpp b/llvm/lib/Analysis/ConstantFolding.cpp
index 966a65ac26b8017..f3f0d079747e13e 100644
--- a/llvm/lib/Analysis/ConstantFolding.cpp
+++ b/llvm/lib/Analysis/ConstantFolding.cpp
@@ -1533,6 +1533,7 @@ bool llvm::canConstantFoldCallTo(const CallBase *Call, const Function *F) {
case Intrinsic::amdgcn_perm:
case Intrinsic::amdgcn_wave_reduce_umin:
case Intrinsic::amdgcn_wave_reduce_umax:
+ case Intrinsic::amdgcn_s_wqm:
case Intrinsic::arm_mve_vctp8:
case Intrinsic::arm_mve_vctp16:
case Intrinsic::arm_mve_vctp32:
@@ -2422,6 +2423,21 @@ static Constant *ConstantFoldScalarCall1(StringRef Name,
return ConstantFP::get(Ty->getContext(), Val);
}
+
+ case Intrinsic::amdgcn_s_wqm: {
+ uint64_t Val = Op->getZExtValue();
+ uint64_t WQM = 0;
+ uint64_t Quad = 0xF;
+ for (unsigned i = 0; i < Op->getBitWidth() / 4;
+ ++i, Val >>= 4, Quad <<= 4) {
+ if (!(Val & 0xF))
+ continue;
+
+ WQM |= Quad;
+ }
+ return ConstantInt::get(Ty, WQM);
+ }
+
default:
return nullptr;
}
diff --git a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.wqm.ll b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.wqm.ll
index 6676dac19ba797f..e44043ffacc07d1 100644
--- a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.wqm.ll
+++ b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.wqm.ll
@@ -9,10 +9,9 @@ define i32 @test_s_wqm_constant_i32() {
; GFX11-LABEL: test_s_wqm_constant_i32:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_wqm_b32 s0, 0x85fe3a92
-; GFX11-NEXT: v_mov_b32_e32 v0, s0
+; GFX11-NEXT: v_mov_b32_e32 v0, 0xff00ff0f
; GFX11-NEXT: s_setpc_b64 s[30:31]
- %br = call i32 @llvm.amdgcn.s.wqm.i32(i32 u0x85FE3A92)
+ %br = call i32 @llvm.amdgcn.s.wqm.i32(i32 u0x85003A02)
ret i32 %br
}
@@ -48,12 +47,10 @@ define i64 @test_s_wqm_constant_i64() {
; GFX11-LABEL: test_s_wqm_constant_i64:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: s_mov_b32 s0, 0x85fe3a92
-; GFX11-NEXT: s_mov_b32 s1, 0x3a9285fe
-; GFX11-NEXT: s_wqm_b64 s[0:1], s[0:1]
-; GFX11-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-NEXT: v_mov_b32_e32 v0, 0xff00ffff
+; GFX11-NEXT: v_mov_b32_e32 v1, 0xffff0fff
; GFX11-NEXT: s_setpc_b64 s[30:31]
- %br = call i64 @llvm.amdgcn.s.wqm.i64(i64 u0x3A9285FE85FE3A92)
+ %br = call i64 @llvm.amdgcn.s.wqm.i64(i64 u0x12480FDBAC00753E)
ret i64 %br
}
|
; GFX11-NEXT: s_setpc_b64 s[30:31] | ||
%br = call i32 @llvm.amdgcn.s.wqm.i32(i32 u0x85FE3A92) | ||
%br = call i32 @llvm.amdgcn.s.wqm.i32(i32 u0x85003A02) |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Added zeroes to the constant so it would not just become 0xffffffff
. Same thing below.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
LGTM with nit.
✅ With the latest revision this PR passed the C/C++ code formatter. |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
LGTM w/ a squash.
BTW, I'm not a fan of merge commits in PRs. I find it makes it more confusing to review.
Understandable. I just tried using the website UI instead of rebasing for the minor merge conflict with the other PRs. Will avoid them in the future. |
Fold any constant input to the
s_wqm
intrinsic.