[llvm] [AMDGPU] Support double type in atomic optimizer. (PR #84307)

[Jay Foad via llvm-commits]

================
@@ -955,6 +955,543 @@ define amdgpu_ps void @global_atomic_fadd_div_address_div_value_system_scope_str
   ret void
 }
 
+define amdgpu_ps void @global_atomic_fadd_double_uni_address_uni_value_agent_scope_unsafe(ptr addrspace(1) inreg %ptr, double inreg %val) #0 {
+; IR-ITERATIVE-LABEL: @global_atomic_fadd_double_uni_address_uni_value_agent_scope_unsafe(
+; IR-ITERATIVE-NEXT:    [[TMP1:%.*]] = call i1 @llvm.amdgcn.ps.live()
+; IR-ITERATIVE-NEXT:    br i1 [[TMP1]], label [[TMP2:%.*]], label [[TMP17:%.*]]
+; IR-ITERATIVE:       2:
+; IR-ITERATIVE-NEXT:    [[TMP3:%.*]] = call i64 @llvm.amdgcn.ballot.i64(i1 true)
+; IR-ITERATIVE-NEXT:    [[TMP4:%.*]] = trunc i64 [[TMP3]] to i32
+; IR-ITERATIVE-NEXT:    [[TMP5:%.*]] = lshr i64 [[TMP3]], 32
+; IR-ITERATIVE-NEXT:    [[TMP6:%.*]] = trunc i64 [[TMP5]] to i32
+; IR-ITERATIVE-NEXT:    [[TMP7:%.*]] = call i32 @llvm.amdgcn.mbcnt.lo(i32 [[TMP4]], i32 0)
+; IR-ITERATIVE-NEXT:    [[TMP8:%.*]] = call i32 @llvm.amdgcn.mbcnt.hi(i32 [[TMP6]], i32 [[TMP7]])
+; IR-ITERATIVE-NEXT:    [[TMP9:%.*]] = call i64 @llvm.ctpop.i64(i64 [[TMP3]])
+; IR-ITERATIVE-NEXT:    [[TMP10:%.*]] = trunc i64 [[TMP9]] to i32
+; IR-ITERATIVE-NEXT:    [[TMP11:%.*]] = uitofp i32 [[TMP10]] to double
+; IR-ITERATIVE-NEXT:    [[TMP12:%.*]] = fmul double [[VAL:%.*]], [[TMP11]]
+; IR-ITERATIVE-NEXT:    [[TMP13:%.*]] = icmp eq i32 [[TMP8]], 0
+; IR-ITERATIVE-NEXT:    br i1 [[TMP13]], label [[TMP14:%.*]], label [[TMP16:%.*]]
+; IR-ITERATIVE:       14:
+; IR-ITERATIVE-NEXT:    [[TMP15:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR:%.*]], double [[TMP12]] syncscope("agent") monotonic, align 4
+; IR-ITERATIVE-NEXT:    br label [[TMP16]]
+; IR-ITERATIVE:       16:
+; IR-ITERATIVE-NEXT:    br label [[TMP17]]
+; IR-ITERATIVE:       17:
+; IR-ITERATIVE-NEXT:    ret void
+;
+; IR-DPP-LABEL: @global_atomic_fadd_double_uni_address_uni_value_agent_scope_unsafe(
----------------
jayfoad wrote:

The idea is that if you have both iterative and DPP RUN lines in the same test then you can use `-check-prefixes=GFX10,GFX10-DPP` and `-check-prefixes=GFX10,GFX10-ITERATIVE` respectively, and then the update script is clever enough to the share checks using the common prefix `GFX10` for functions that generate identical code for both runs.

https://github.com/llvm/llvm-project/pull/84307